CN118963281A

CN118963281A - A product data analysis method, device, equipment and storage medium

Info

Publication number: CN118963281A
Application number: CN202411047638.0A
Authority: CN
Inventors: 程子健; 周希波; 文晋晓
Original assignee: BOE Technology Group Co Ltd; Beijing BOE Technology Development Co Ltd
Current assignee: BOE Technology Group Co Ltd; Beijing BOE Technology Development Co Ltd
Priority date: 2024-07-31
Filing date: 2024-07-31
Publication date: 2024-11-15

Abstract

The invention discloses a method, a device, equipment and a storage medium for analyzing product data. At least one pre-processing step and a current processing step are included in the product processing process; the current processing step is performed on the basis of the processing result of the at least one preceding processing step. The method comprises the following steps: for any one of the at least one pre-processing step, determining a pre-defect location in the processing result of the pre-processing step concerned; determining a pre-defect high-incidence area corresponding to the pre-processing step according to the distribution condition of the pre-defect positions; and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

Description

Product data analysis method, device, equipment and storage medium

Technical Field

The present invention relates to the field of computer application technologies, and in particular, to a method, an apparatus, a device, and a storage medium for product data analysis.

Background

In the current processing process of products, in order to conveniently improve the processing precision of the products and the processing yield of the products, the processing process data of the products are usually required to be acquired for analysis, and particularly the defects generated in the processing process are required to be analyzed, so that the probability of occurrence of the defects can be conveniently reduced.

However, the data analysis can only be performed manually at present, and the efficiency is low.

Disclosure of Invention

The invention provides a method, a device, equipment and a storage medium for analyzing product data, which are used for solving the defects in the related technology.

According to a first aspect of embodiments of the present invention, there is provided a method for analyzing product data, including at least one pre-processing step and a current processing step during product processing; the current processing step is performed on the basis of the processing result of the at least one preceding processing step;

The method comprises the following steps:

for any one of the at least one pre-processing step, determining a pre-defect location in the processing result of the pre-processing step concerned;

Determining a pre-defect high-incidence area corresponding to the pre-processing step according to the distribution condition of the pre-defect positions;

and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

Optionally, the determining, according to the distribution of the pre-defect positions, a pre-defect high-incidence area corresponding to the pre-processing step includes:

determining a region where the pre-defect positions are gathered according to a preset abnormality detection algorithm based on the pre-defect positions;

And predicting a pre-defect high-incidence area corresponding to the pre-processing step based on the area gathered by the pre-defect positions.

Optionally, the method further comprises any one of the following:

determining an actual defect high-incidence area corresponding to the current processing step; updating parameters of the preset anomaly detection algorithm to reduce the difference between the predicted defect high-incidence area and the actual defect high-incidence area;

determining the actual defect position in the processing result of the current processing step; and updating parameters of the preset abnormality detection algorithm to increase the number of the actual defect positions contained in the predicted defect high-incidence area.

Optionally, the determining, based on the pre-defect positions, a region where the pre-defect positions are gathered according to a preset anomaly detection algorithm includes at least one of the following:

Clustering is carried out according to a preset clustering algorithm aiming at the pre-defect position to obtain a clustering result; determining a region where the pre-defect positions are gathered based on clusters, wherein the number of the clusters including the pre-defect positions meets the preset number condition, in the clustering result;

Inputting the pre-defect position into a pre-trained self-encoder to obtain a corresponding reconstruction position output by the self-encoder; determining a region where the pre-defect positions are aggregated based on the pre-defect positions having a distance to the corresponding reconstructed position less than a preset distance threshold;

determining the pre-defect positions of outliers according to an isolated forest algorithm aiming at the pre-defect positions; the area in which the pre-defect locations are clustered is determined based on other pre-defect locations than the outlier pre-defect locations.

Optionally, for any one of the at least one pre-processing step, determining a pre-defect position in a processing result of the pre-processing step, and determining a pre-defect high-incidence area corresponding to the pre-processing step according to a distribution condition of the pre-defect position, including:

for different pre-processing steps, respectively executing: determining a pre-defect position in a processing result of the pre-processing step, and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the distribution condition of the pre-defect position;

The predicting the predicted defect high incidence area corresponding to the current processing step based on the pre-defect high incidence area corresponding to the pre-processing step comprises the following steps:

And predicting the predicted defect high-incidence areas corresponding to the current processing step based on the pre-defect high-incidence areas corresponding to the different pre-processing steps respectively.

Optionally, the predicting the predicted defect high incidence area corresponding to the current processing step based on the pre-defect high incidence areas corresponding to the different pre-processing steps respectively includes any one of the following:

Determining a union region aiming at the pre-defect high-incidence regions corresponding to the different pre-processing steps respectively, and predicting a predicted defect high-incidence region corresponding to the current processing step based on the union region;

And determining an intersection area aiming at the pre-defect high-incidence areas corresponding to the different pre-processing steps respectively, and predicting the predicted defect high-incidence area corresponding to the current processing step based on the intersection area.

Optionally, the method further comprises:

determining a historical defect position in a historical processing result of the current processing step; determining a high-incidence area of the historical defect according to the distribution condition of the positions of the historical defect;

and predicting a predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step and the history defect high-incidence area.

Optionally, the determining, for any one of the at least one pre-processing step, a pre-defect location in a processing result of the pre-processing step includes:

predicting, for any one of the at least one pre-processing step, a failure processing result from the processing results of the pre-processing step for which it is intended;

And checking the predicted disqualified machining result to determine the pre-defect position.

Optionally, the method further comprises:

Predicting whether any processing result of any processing step in the product processing process is qualified.

Optionally, a prediction method for predicting whether the processing result is qualified, including:

Inputting actual processing parameters of any processing result into a pre-trained self-encoder to obtain corresponding reconstruction processing parameters output by the self-encoder;

and under the condition that the difference between the actual processing parameter and the corresponding reconstruction processing parameter is larger than a preset difference threshold value, predicting any processing result as a disqualified processing result.

Optionally, the training process of the self-encoder includes:

Determining an initial self-encoder;

determining actual processing parameters of qualified processing results;

The following steps are circularly executed until the preset training stop condition is met:

Inputting the determined actual processing parameters into a current initial self-encoder to obtain corresponding reconstructed processing parameters output by the current initial self-encoder;

updating the current initial self-encoder to reduce the difference between the input actual processing parameters and the corresponding reconstructed processing parameters, and re-determining the updated initial self-encoder as the current initial self-encoder;

after the cycle is completed, the current initial self-encoder is determined as the self-encoder with the training completed.

Optionally, the method further comprises at least one of:

Inputting the actual processing parameters of a plurality of unqualified processing results and the actual processing parameters of a plurality of qualified processing results into a pre-trained self-encoder to obtain hidden layer characteristics; determining differences between hidden features of the qualified machining result and hidden features of the unqualified machining result so as to analyze actual machining parameters associated with the unqualified machining result;

Inputting the actual processing parameters of a plurality of unqualified processing results and the actual processing parameters of a plurality of qualified processing results into a pre-trained self-encoder to obtain hidden layer characteristics; determining hidden layer nodes meeting preset screening conditions; the preset screening conditions comprise: aiming at the qualified processing result and the unqualified processing result, the difference between the output hidden layer characteristics is larger than a preset difference threshold; weight parameters between the actual machining parameters and the determined hidden layer nodes are determined to facilitate analysis of the actual machining parameters associated with the failed machining results.

Optionally, the method further comprises at least one of:

Determining actual qualified detection results of a plurality of processing results on the parameters of the target product; determining the contribution degree of an actual processing parameter to the target product parameter based on the actual qualification detection result and the prediction results obtained by the self-encoder aiming at a plurality of processing results so as to analyze the actual processing parameter related to the processing result with unqualified target product parameter;

Determining actual qualified detection results of a plurality of processing results; and determining the contribution degree of the actual processing parameters based on the actual qualified detection result and the prediction results obtained by the self-encoder for a plurality of processing results so as to analyze the actual processing parameters related to the unqualified processing results.

According to a second aspect of embodiments of the present invention, there is provided a product data analysis apparatus including at least one pre-processing step and a current processing step in a product processing process; the current processing step is performed on the basis of the processing result of the at least one preceding processing step;

The device comprises:

A position unit for: for any one of the at least one pre-processing step, determining a pre-defect location in the processing result of the pre-processing step concerned;

a summarizing unit for: determining a pre-defect high-incidence area corresponding to the pre-processing step according to the distribution condition of the pre-defect positions;

A prediction unit for: and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

Optionally, the summarizing unit is configured to:

Optionally, the summarizing unit is further configured to perform any one of the following:

Optionally, the summarizing unit is configured to perform at least one of:

Optionally, the location unit is configured to:

for different pre-processing steps, respectively executing: determining the position of a pre-defect in the processing result of the pre-processing step;

The summarizing unit is used for:

for different pre-processing steps, respectively executing: according to the distribution condition of the determined pre-defect positions, determining a pre-defect high-incidence area corresponding to the pre-processing step;

The prediction unit is used for:

Optionally, the prediction unit is configured to perform any one of the following:

Optionally, the summarizing unit is further configured to:

The prediction unit is used for:

Optionally, the summarizing unit is configured to:

Optionally, the apparatus further comprises:

and the qualification unit is used for predicting whether any processing result of any processing step in the product processing process is qualified.

Optionally, the training process of the self-encoder includes:

Determining an initial self-encoder;

determining actual processing parameters of qualified processing results;

Optionally, the apparatus further comprises an analysis unit.

Optionally, the analysis unit is configured to perform at least one of:

According to the embodiment, the efficiency of analyzing the product defects is improved by automatically analyzing the product data.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention as claimed.

In addition, not all of the above-described effects need be achieved in any one of the embodiments of the present invention.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention.

FIG. 1 is a flow chart of a method of product data analysis according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a product process flow shown in accordance with an embodiment of the present invention;

FIG. 3 is a schematic view showing a distribution of defects in a processing result according to an embodiment of the present invention;

FIG. 4 is a schematic view showing a distribution of defect high incidence areas in a processing result according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a self-encoder according to an embodiment of the present invention;

FIG. 6 is a flow chart illustrating another method of product data analysis according to an embodiment of the present invention;

FIG. 7 is a schematic diagram illustrating a product data analysis in accordance with an embodiment of the present invention;

FIG. 8 is a schematic diagram illustrating another self-encoder according to an embodiment of the present invention;

FIG. 9 is a schematic diagram of an anomaly detection principle for determining correlation of quality inspection parameters according to an embodiment of the present invention;

Fig. 10 is a schematic structural view of a product data analysis device according to an embodiment of the present invention;

Fig. 11 is a schematic diagram of a hardware structure of a computer device for configuring a method according to an embodiment of the present invention.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the invention. Rather, they are merely examples of apparatus and methods consistent with aspects of the invention as detailed in the accompanying claims.

It should be noted that, the user information (including but not limited to user equipment information, user personal information, etc.) and the data (including but not limited to data for analysis, stored data, presented data, etc.) related to the embodiments of the present invention are information and data authorized by the user or fully authorized by each party, and the collection, use and processing of the related data need to comply with the related laws and regulations and standards of the related country and region, and provide a corresponding operation entry for the user to select authorization or rejection.

In the current processing process of products, in order to conveniently improve the processing precision and the processing yield of the products, the processing process data of the products are usually required to be acquired for analysis, and particularly the defects of the products generated in the processing process are required to be analyzed, so that the probability of occurrence of the defects can be conveniently reduced.

The embodiment of the invention provides a product data analysis method for solving the problems.

In the method, the product data can be automatically analyzed through the machine equipment, so that the efficiency and the accuracy of analyzing the product defects are improved.

Specifically, in the method, the analysis of the processing data can be performed for any one of the processing steps in the processing of the product. The process data may include, in particular, defect locations in the process results.

The defect positions in the processing results are analyzed, the defect high-incidence areas in the processing steps can be conveniently determined, and the processing process related to the defect high-incidence areas is conveniently improved, so that the probability of occurrence of defects is reduced, and the processing precision and the processing yield of products are improved. It will be appreciated that if the frequency of defects actually occurring in a region is high, it can be predicted that the region is relatively susceptible to defects, i.e., high defect occurrence regions.

For ease of understanding, in one specific example, for a display panel product, a number of defect locations that occur in the machining result of any one of the machining steps may be obtained for that machining step. Specifically, statistics can be performed on a plurality of processing results of the processing step, and analysis can be performed by integrating defect positions in each processing result.

The present example is not limited to a specific type of defect of the display panel product, and for example, the defect may be specifically a type of crack, breakage, chipping, scratch, burr, water droplet, foreign matter, and the like. When the defect positions appearing in the machining result are obtained, the defect types can be considered, and only the defect positions of a single defect type can be obtained, and the defect positions of a plurality of defect types can also be obtained.

From the acquired number of defect locations, the area in which the defect locations are aggregated may be analyzed by an algorithm. Specifically, the area aggregated by a plurality of defect positions can be determined through a clustering algorithm or an isolated forest algorithm and the like, and the area is further determined to be a defect high-incidence area. It will be appreciated that if the frequency of defects actually occurring in a region, i.e., a region where defect positions are abnormally concentrated, is high, it can be predicted that the region is relatively easily defective, i.e., a defect high-occurrence region.

Then, aiming at the whole processing process of the display panel product, the defect high-incidence area can be focused, and particularly, the quality inspection can be performed on the defect high-incidence area in the processing process, the process level of the defect high-incidence area can be improved in the processing process, and the like, so that the probability of occurrence of defects in the defect high-incidence area can be reduced, and the processing precision and the processing yield of the product can be improved.

Of course, in the method, the analysis of the processing data is performed for any one processing step, and it is understood that the analysis of the processing data may be performed for a plurality of processing steps or for different processing steps, to determine the defect high incidence area of each processing step.

Therefore, in the method, the data can be automatically analyzed through the machine equipment, the defect high-incidence area corresponding to the processing step can be determined, the data analysis efficiency is improved, and the efficiency and accuracy for determining the defect high-incidence area can also be improved.

In the method, in order to conveniently determine the defect high-incidence area aiming at any processing step, the defect high-incidence area of the front processing step of the processing step can be combined for determination, so that the efficiency of determining the defect high-incidence area aiming at the processing step is improved.

In the processing of a product, there may be a plurality of processing steps, and a processing result may be obtained based on each processing step, and there may be a sequential execution order between the processing steps, specifically, a subsequent processing step may be further executed based on the processing result obtained by the preceding processing step.

It will therefore be appreciated that the location of defects occurring in the processing results of the preceding processing steps will typically also occur at the same location in the processing results of the subsequent processing steps as the processing steps are performed sequentially. Often because of these defects in the processing results of the preceding processing steps, either remaining during subsequent processing or causing new defects to be easily present at or near the same location during subsequent processing.

For easy understanding, for example, it is generally required to sequentially coat films for products, specifically, three times of film coating may be performed, and if bubbles occur at a certain position during the previous film coating process, film coating defects may also occur at the same position during the subsequent film coating process. Of course, defects such as breakage, edge breakage, water drops, foreign matters and the like may also occur during the coating process, and these defects generally remain during the subsequent coating process, so that coating defects also occur at the same position.

For another example, if the pixel in the product needs to be sequentially processed, if the pixel is damaged and a defect occurs in the previous processing step, the pixel maintains the damaged state in the subsequent processing step, so that the defect at the same position occurs.

For another example, when a display panel in a product is damaged, a coating leak is likely to occur at the damaged portion during the subsequent coating process, and thus a new defect occurs at the same position.

As another example, for a crack in a product, the crack may gradually expand with subsequent machining operations, resulting in defects at and near the same location.

As another example, foreign matter effects may occur during product processing, which may cause new defects at or near the same location with subsequent processing operations.

Therefore, the region in which a defect is likely to occur in the processing result of the pre-processing step is usually a region in which a defect is likely to occur in the processing result of the post-processing step, and the defect high-incidence region corresponding to the pre-processing step is usually a defect high-incidence region corresponding to the post-processing step, so that the defect high-incidence region corresponding to the post-processing step can be determined, and the defect high-incidence region corresponding to the pre-processing step can be included.

In the method, aiming at the current processing step, the processing result of the pre-processing step before the current processing step is utilized to analyze the defect position and determine the defect high-incidence area, so that the defect high-incidence area corresponding to the current processing step can be further predicted and determined according to the determined defect high-incidence area corresponding to the pre-processing step.

Specifically, the defect high-incidence area corresponding to the determined pre-processing step may be further determined as the defect high-incidence area corresponding to the current processing step, or the defect high-incidence area corresponding to the current processing step may be determined to include the defect high-incidence area corresponding to the determined pre-processing step.

By the method, the data of the processing result of the front processing step can be utilized to help determine the defect high-incidence area of the current processing step, so that the requirement on the processing result of the current processing step can be reduced, more product data can be conveniently collected for analysis, and the defect analysis efficiency and accuracy are improved.

The following explains in detail a method for analyzing product data provided by the embodiment of the present invention.

Referring to fig. 1, fig. 1 is a schematic flow chart of a method for analyzing product data according to an embodiment of the present invention.

The embodiment of the invention is not limited to the execution main body of the flow of the method. Alternatively, the execution body may be any computing device. Such as terminal equipment, server equipment, internal equipment of a process plant or terminals of management personnel, etc.

At least one pre-processing step may be included in the product manufacturing process as well as the current processing step. The current processing step may be performed on the basis of the processing results of at least one of the preceding processing steps. The product processing may include several processing steps. After each processing step is executed, each processing result can be obtained respectively.

Wherein, in the process of processing any type of product, at least one pre-processing step and the current processing step can be sequentially carried out; the current processing step may be performed on the basis of the processing results of at least one preceding processing step.

The flow of the method is not limited to the type of the product, and can be a display panel, a display screen or a display device.

The process of processing different products of the same type may include the same processing step, that is, sequentially through at least one preceding processing step and the current processing step.

The method may comprise the following steps.

S101: for any one of the at least one pre-processing step, the location of the pre-defect in the processing result for the pre-processing step is determined.

S102: and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the determined distribution condition of the pre-defect positions.

S103: and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

The method flow can automatically analyze the product data, so that the efficiency and accuracy of analyzing the product defects can be improved.

And the defect position in the product processing process can be automatically analyzed to determine the defect high-incidence area corresponding to the current processing step, so that the efficiency and accuracy of determining the defect high-incidence area are improved.

The data of the processing result of the front processing step is utilized to help determine the defect high-incidence area of the current processing step, so that the improvement on the current processing step in advance can be conveniently realized, the requirement on the processing result of the current processing step can be reduced, more product data can be conveniently acquired for analysis, and the efficiency and accuracy of defect analysis are improved.

The process flow is not limited to a specific process of product processing, but may include other processing steps in addition to at least one pre-processing step and the current processing step. The current processing step may be any processing step in the product processing having at least one preceding processing step.

It will be appreciated that the flow of the method is interpreted in a manner that determines the defect high incidence area by data analysis for any one of the processing steps as the current processing step. For a plurality of processing steps, the defect high-incidence area can be determined by adopting the method.

The following is a detailed explanation of the respective steps.

1. And (3) processing the product.

The process flow is not limited to a particular type of product. Alternatively, the product may be a display panel, a toy, a tire, or a vehicle screen, among others.

The process flow of the method is not limited to the specific processing process of the product. The processing may be performed to produce a product, or may be performed in addition to the processing performed to modify the product, and the process is not limited to the processing.

Wherein, in the process of processing any type of product, at least one pre-processing step and the current processing step can be sequentially performed. It will be appreciated that the processing steps and sequences that are performed during processing for different products of this type are identical, so that analysis may be facilitated.

The current processing step may be any processing step having at least one preceding processing step in the processing of the product.

The process flow of the method does not limit the number of processing steps in the product processing process.

It is understood that the pre-processing step herein may be performed prior to the current processing step. In the whole product processing process, the front processing step can be preceded by additional other processing steps, the current processing step can also be followed by additional other processing steps, and the front processing step and the current processing step can also be followed by additional other processing steps. The pre-processing step herein may be performed prior to the current processing step.

In a specific example, the product processing may include processing steps 1-9, and specifically, processing steps 1-9 may be sequentially performed. Wherein, processing step 8 can be considered the current processing step, processing steps 1, 4, 5 and 7 can be considered the preceding processing steps, and processing steps 4-7 can also be considered the preceding processing steps.

The specific pre-processing step may be any processing step performed before the current processing step in the product processing process.

The method is not limited to a specific method of determining the pre-processing step, and may be a method of randomly selecting a processing step performed immediately before the current processing step, or a method of fixedly selecting a plurality of processing steps performed immediately before the current processing step.

Of course, the process flow is not limited to a particular processing step.

In a specific example, for the manufacturing process of the display panel product, there may be a plurality of manufacturing stations, in each of which a plurality of complicated manufacturing operations such as assembling, coating, sealing, and the like may be performed. The operations performed by each process production station may be considered a process step.

Further, after the execution of each processing step is completed, each processing result may be obtained respectively. The processing result may be a product intermediate after the processing step is completed.

For example, after the coating processing step is performed, a processing result after coating can be obtained; after the assembly processing step is completed, the assembled processing result can be obtained.

Wherein the current processing step may be performed on the basis of the processing result of at least one preceding processing step.

Taking the production line for product manufacture as an example, 10 steps, that is, 10 processing steps, may be performed sequentially, and any one processing step may be performed based on the processing result of the previous processing step.

Thus, in specifically analyzing the areas of high occurrence of defects that may occur in the current processing step, the analysis may be performed in conjunction with the pre-processing step, since defects in the processing results of the pre-processing step may generally result in the processing results of the current processing step, as well as consent to the occurrence of defects at or near the same location.

For example, it is generally necessary to sequentially perform coating and painting for display panel products. If a foreign object is present at a certain position during the coating process, a corresponding foreign object defect is also present at the same position during the subsequent painting process, or other types of defects, such as bubble defects, are easily caused to occur at the same position.

Therefore, the defect high-frequency region corresponding to the pre-processing step may be determined by determining the defect high-frequency region corresponding to the current processing step, and the defect high-frequency region corresponding to the pre-processing step may be included.

In the flow of the method, the defect high-incidence area corresponding to the pre-processing step is mainly explained and combined, so that the defect high-incidence area corresponding to the current processing step is predicted and determined.

The data of the processing result of the pre-processing step is utilized to help determine the defect high-incidence area of the current processing step, so that the requirement on the processing result of the current processing step can be reduced, more product data can be conveniently collected for analysis, and the efficiency and accuracy of defect analysis are improved.

For ease of understanding, as shown in fig. 2, fig. 2 is a schematic diagram of a product processing flow shown in accordance with an embodiment of the present invention.

Wherein, the products in the same batch can be processed by 3 processing steps 1-3, and specifically, processing result sets 1-3 can be respectively obtained. And the sampling quality inspection can be respectively carried out on the processing result sets processed by each processing step.

2. S101: for any pre-processing step, the pre-defect position in the processing result of the pre-processing step is determined.

1. With respect to determining the pre-defect location.

In an alternative embodiment, in order to facilitate predicting and determining the defect high incidence area corresponding to the current processing step, analysis may be performed in combination with product data of the preceding processing step, and specifically may include a defect location in the processing result of the preceding processing step.

For convenience of description, the defect position in the processing result of the pre-processing step will be referred to as a pre-defect position.

The process flow is not limited to the specific form of the pre-defect location. Optionally, the pre-defect position may specifically be represented by a coordinate point, may also be represented by a region, or may also be represented by a rectangular frame.

Wherein, the defects of different types can also be characterized by using the pre-defect positions of different forms.

For example, for defects of a single pixel point, a pre-defect position in the form of a coordinate point can be used for characterization; for foreign object defects, the pre-defect location in the form of a rectangular box can be used for characterization.

It should be noted that, since the shapes and dimensions of the different processing results are basically the same, the positions of the pre-defects in the different processing results can be regarded as the positions under the same coordinate system, so that the analysis of the subsequent steps can be facilitated to determine the region where the positions of the pre-defects are gathered.

For example, for a display panel product, defects of pixel points can be represented by pre-defect positions in the form of coordinate points between different processing results after film coating in the processing process, and the defects can be specifically: the pixel defect positions in processing result 1 are (1, 1) and (2, 2), the pixel defect positions in processing result 2 are (1, 2) and (3, 2), and the pixel defect positions in processing result 3 are (1, 3) and (2, 3). Therefore, in the subsequent steps, the pixel defect positions can be regarded as the areas where the pixel defect positions are gathered in the same coordinate system through analysis.

In other words, the pre-defect locations in different processing results may be in the same coordinate system.

In a specific example, the pre-defect positions determined from the actual machining result can be regarded as defect positions distributed in the theoretical machining result, so that the defect distribution situation in the theoretical machining result can be conveniently analyzed, and the pre-defect aggregation area of the theoretical machining result can be conveniently analyzed and determined.

The embodiment does not limit the construction basis and manner of the coordinate system. The coordinate system in which the pre-defect position is located can be constructed based on the theoretical processing result, or can be constructed based on the theoretical final product.

In an example, a theoretical processing result of the display panel product in the processing process may include a pixel point matrix of 100×100, so that a coordinate system may be constructed, coordinate points of each pixel point may be divided, and further, for the pixel point position where a defect exists in the actual processing result of the display panel product, the coordinate point in the corresponding pixel point matrix may be determined, so as to obtain the defect position. Accordingly, the determined positions of the pixel defects between the actual processing results may be in the same coordinate system, that is, the coordinate system constructed based on the theoretical processing results.

The method flow is not limited to the manner in which the location of the defect in the machining result is determined.

Optionally, determining the defects in the machining results and the positions of the defects by means of quality inspection, specifically, sampling a plurality of machining results for quality inspection, or performing quality inspection on all the machining results; the defects in the processing result and the positions of the defects can be determined by means of image detection or image recognition; the defects and the defect positions in the processing result can also be determined by a product inspection mode.

For example, the machining result may be subjected to a power-on test to determine whether there is a defect therein, or may be subjected to an optical photographing or laser scanning to determine whether there is a breakage or burr in the machining result. Accordingly, in the case of detecting a defect, the position of the defect may be further detected.

Of course, the processing result with the defect can be detected and screened first, and then the quality inspection can be further carried out on the processing result with the defect, so as to determine the specific defect and the position of the defect in the processing result.

The determined defect location may be directly determined as the pre-defect location.

Accordingly, the present method flow is not limited to the manner in which the pre-defect location is obtained.

Specifically, the method may be directly acquiring the pre-defect position determined based on the method in real time for the processing result, or may be determining the pre-defect position in the processing result before the pre-processing step from the history data.

The number of the processing results is not limited, and the method may specifically determine the position of the pre-defect in any one of the processing results of the pre-processing step, or determine the position of the pre-defect in each of the processing results of the pre-processing step, or determine the positions of the pre-defect in a plurality of the processing results of the pre-processing step. The present embodiment is not limited to the above-mentioned processing results, and may be, in particular, processing results obtained by performing the above-mentioned pre-processing step in real time, processing results obtained by performing the quality inspection, or historical processing results obtained by performing the above-mentioned pre-processing step.

The method flow is also not limited to the number of pre-defect locations determined. In particular, one or more pre-defect locations may be determined.

In the case of determining the location of the pre-defect, a subsequent step of determining a defect high incidence area corresponding to the pre-processing step may be further performed.

For convenience of description, the defect high-incidence area corresponding to the pre-processing step, that is, the defect high-incidence area in the processing result of the pre-processing step, is referred to as a pre-defect high-incidence area.

For ease of understanding, fig. 3 is a schematic diagram illustrating a distribution of defects in a processing result according to an embodiment of the present invention, as shown in fig. 3.

The processing result may specifically be a display panel. Defects of pixel point damage, crack defects and breakage defects can exist in a single display panel respectively, wherein the defect positions of the pixel point damage can be represented in a coordinate point mode, the crack defect positions can be represented in a region mode, particularly an elliptic region, and the breakage defect positions can be represented in a rectangular frame mode.

2. With respect to determining other process data.

In the process of specifically determining the advanced defect high-incidence area, analysis and determination can be performed by combining other processing data of the advanced processing step.

Thus, optionally, for any pre-processing step, other data for the pre-processing step may also be determined, facilitating subsequent analysis to determine pre-defect high incidence areas.

The present embodiment is not limited to the specific other data determined. Optionally, the other data may specifically include production process data of the pre-processing step, also include data related to a processing result, also include data related to a defect, and also include data related to quality inspection.

The present embodiment is not limited to a specific example of these other data.

Wherein the production process data may include at least one of: product information, equipment parameters, important time node information, product batch information, product model information, production temperature, production precision and the like.

The information about the processing result may include at least one of: the size, state, assembly, etc. of the result of the process.

The defect-related data may include at least one of: defect type, defect level, defect identification, etc. Wherein the present embodiment does not limit the defect type. The classification of the defect type may be performed according to the size of the defect, such as a small defect and a large defect, or may be performed according to the form of the defect, such as a crack, a breakage, a chipping, a scratch, a burr, a water droplet, a foreign matter, and the like. The present embodiment is not limited to the division of defect levels, and may be specifically divided into 1-10 levels according to the serious condition of defects.

The quality check related data may include at least one of: whether a defect exists, a defect type, a defect level, a defect identification, a defect existence area, a defect existence coordinate, and the like.

Optionally, using the determined pre-defect locations, subsequent analysis may be aided in determining pre-defect high incidence areas of the pre-processing step.

Other data determined may also be utilized to assist in subsequent analysis in determining pre-defect high incidence areas of the pre-processing steps.

For example, the pre-defect positions may be classified according to the determined other data, specifically, the pre-defect positions may be classified according to the defect type, the defect grade, or the information of the product lot, so that the pre-defect positions of different classifications may be subjected to subsequent analysis.

Alternatively, the determined pre-defect locations may be classified. For example, the determined pre-defect location may be of the same defect type, so that subsequent analysis may be facilitated to determine the high-incidence region of the defect type.

The determined pre-defect location may be of the same product lot, so that subsequent analysis may be facilitated to determine a defect high incidence area for the product lot.

The foregoing explains that for any pre-processing step, the pre-defect location may be determined, as well as other data, for subsequent analysis of the data.

It will be appreciated that the location of the pre-defect in the processing results of the different pre-processing steps may be determined separately for the different pre-processing steps, or other data may be determined separately. The explanation of the different pre-processing steps can be found separately above.

3. S102: and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the determined distribution condition of the pre-defect positions.

1. With respect to the pre-defect high incidence area.

In the step S101, a plurality of pre-defect positions may be determined in the processing result of the pre-processing step. In particular, a number of pre-defect locations in a number of processing results for a pre-processing step are determined.

Based on the determined distribution of the plurality of pre-defect positions, an area in which the pre-defect positions are gathered may be analytically determined, and further, the area in which the pre-defect positions are gathered may be determined as a pre-defect high incidence area. Therefore, the high incidence area of the pre-defect corresponding to the pre-processing step can be determined based on the determined distribution of the pre-defect positions.

It will be appreciated that if the frequency of defects actually occurring in a region is high, or more actual defects occur, or the actual defects are abnormally concentrated in the region, it is predicted that the region is relatively likely to be defective, that is, a defect-rich region. Therefore, the defect aggregation area can be determined by means of the distribution condition of the positions of the pre-defects, and the high incidence area of the pre-defects can be further determined. Of course, the high incidence area of the pre-defect may be determined directly based on the distribution of the pre-defect positions.

The pre-defect high-incidence area can be predicted based on the determined pre-defect position, and can be used for representing an area which is easy to generate defects in the processing result of the pre-processing step and can also be used for representing an area which is easy to generate defects in the theoretical processing result of the pre-processing step. The analysis and determination of the advanced defect high-incidence area can be used for helping to improve and detect quality, so that the yield of products and processing results is conveniently improved, and the probability of defects of the products and the processing results is reduced.

According to the defects actually appearing in the actual processing result, the distribution situation of the actual defect positions can be analyzed, and the area easily appearing the defects in the processing result can be predicted and determined, so that the area gathered by the front defect positions can be determined as the front defect high-incidence area.

The flow of the method is not limited to the corresponding relation between the pre-processing step and the pre-defect high-incidence area, and specifically can be one-to-one correspondence. The pre-defect high incidence region can be used for representing a region which is easy to generate defects in the processing result of the corresponding pre-processing step.

The method flow is not limited to a specific manner of determining the pre-defect location aggregate area. Alternatively, the determination may be performed in a clustering manner, or may be performed in a statistical manner.

For example, for a product of a display panel, the pre-defect positions are generally distributed on the display panel, so that the pre-defect positions can be regarded as coordinate points in one image, clustering is performed in a k-means clustering mode, a plurality of clustered coordinate points, that is, a plurality of clustered pre-defect positions, can be determined, and a region containing any cluster obtained by clustering can be determined to be a pre-defect position clustered region. Of course, for each cluster obtained by clustering, a region containing the cluster may be determined as a pre-defect-location aggregation region.

For another example, the product of the display panel may be divided into a plurality of fixed areas in advance. Specifically, the method may divide the area into a plurality of areas with the same size, and then the number of the pre-defect positions in different divided areas may be counted, so that the divided area including the number of the pre-defect positions greater than the preset number may be determined as the pre-defect position aggregation area. Of course, the number of the pre-defect positions may be sorted from large to small, and the divided area having the sorting order number of the first few bits may be determined as the pre-defect position aggregation area.

Therefore, alternatively, the divided area in which the number of the contained pre-defect positions satisfies the preset number condition may be determined as the pre-defect high-occurrence area among the number of areas divided in advance.

The present embodiment is not limited to a specific area division method, and may be specifically divided into a plurality of areas with equal areas. The embodiment is not limited to a specific preset number condition, and the number may be specifically greater than a preset number threshold, or may be in the order from large to small, where the sequence number is before the preset sequence number.

The pre-defect aggregation area determined by the method can be further determined as a pre-defect high-incidence area.

Of course, the method is not limited to a specific way of determining the pre-defect high-incidence area. Optionally, the pre-defect high incidence area may be further predictively determined based on the determined pre-defect aggregation area. The determined pre-defect aggregation area may be expanded, and the expansion result may be determined as a pre-defect high-incidence area.

For ease of understanding, fig. 4 is a schematic diagram illustrating a distribution of defect high-incidence areas in a processing result according to an embodiment of the present invention.

The processing result may specifically be a display panel. The display panel in the figure can be a theoretical processing result, and the defect positions in the determined actual processing results can be conveniently displayed in a summarized mode.

Therefore, on the display panel in the figures, there may be several defects, each characterized in a different form. The method is characterized by adopting a coordinate point form and a rectangular frame form respectively.

For each determined defect position, the defect high-incidence area can be determined by the following anomaly detection algorithm. Which contains 2 circular areas, i.e. 2 areas of high incidence of defects, in which the defect locations are clustered.

2. With respect to anomaly detection algorithms.

In an alternative embodiment, an anomaly detection algorithm may be employed to analyze the pre-defect locations, predicting the areas where the pre-defect locations are clustered.

The anomaly detection algorithm herein may be specifically used to detect areas of abnormal concentration of defect locations.

Therefore, optionally, according to the determined distribution situation of the pre-defect positions, the pre-defect high-incidence area corresponding to the pre-processing step is determined, which specifically may be: predicting a region where the pre-defect positions are gathered according to a preset anomaly detection algorithm based on the determined pre-defect positions; and predicting a pre-defect high-incidence area corresponding to the pre-processing step based on the determined pre-defect position gathered area.

The embodiment can be combined with a preset anomaly detection algorithm to help predict and determine the concentrated areas of the pre-defect positions, and improve the efficiency and accuracy of predicting and determining the high-incidence areas of the pre-defects.

The present embodiment is not limited to a specific manner of determining the pre-defect high incidence area. Alternatively, the determined pre-defect aggregation area may be directly determined as a pre-defect high incidence area. The determined pre-defect aggregation area may be expanded, and the expansion result may be determined as a pre-defect high-incidence area. The specific extension may be an area extension.

The flow of the method is not limited to a specific preset abnormality detection algorithm. Alternatively, the anomaly detection may be performed by using a clustering algorithm, or by using a decision tree, or by using an isolated forest algorithm, or by using a machine learning model. The preset abnormality detection algorithm may be pre-trained, and specifically may be training update by using a training sample.

For example, the features in the training samples may include pre-defect locations in the machining results of the pre-machining step, the labels in the training samples may be areas where the pre-defect locations are aggregated as determined by other means, and further the training samples may be used in advance to update parameters of the training preset anomaly detection algorithm for fitting.

In an alternative embodiment, a training sample may be obtained to train a machine learning model, where the features in the training sample may include pre-defect positions in the processing result of the pre-processing step, and specifically may be image-form features, where the pre-defect positions include a plurality of pre-defect positions, and the pre-defect positions may be in a coordinate point form or a region form. The labels in the training samples may be areas of pre-defect location concentration determined by other means, such as statistical means or clustering algorithms, i.e. pre-defect high-incidence areas.

Training the machine learning model by using a plurality of training samples can obtain the machine learning model for predicting and determining the high incidence area of the pre-defect aiming at the determined pre-defect position.

The present embodiment is not limited to the structure and training manner of the machine learning model. The method can be a neural network model, a graph convolution model and the like.

Of course, a clustering algorithm may also be used to determine the area in which the positions of the pre-defects are abnormally gathered, so as to predict and determine the high-incidence area of the pre-defects.

The isolated forest algorithm can also be adopted to determine the abnormal discrete pre-defect positions, so that the abnormal aggregated pre-defect positions can be reversely determined, and the pre-defect high-incidence area can be predicted and determined.

Thus, optionally, predicting the area where the pre-defect locations are determined to be aggregated according to a preset anomaly detection algorithm based on the determined pre-defect locations may include at least one of:

1) Clustering is carried out according to a preset clustering algorithm aiming at the determined pre-defect position, so as to obtain a clustering result; and determining the area where the pre-defect positions are gathered based on clusters, wherein the number of the pre-defect positions is contained in the clustering result and meets the preset number condition.

2) Inputting the determined pre-defect position into a pre-trained self-encoder to obtain a corresponding reconstruction position output by the self-encoder; and determining an area where the pre-defect positions are gathered based on the pre-defect positions having a distance to the corresponding reconstructed position less than a preset distance threshold.

3) Determining the position of the outlier pre-defect according to an isolated forest algorithm aiming at the determined position of the pre-defect; the area in which the pre-defect locations are clustered is determined based on other pre-defect locations than the outlier pre-defect locations.

The embodiment is not limited to a specific preset clustering algorithm, and may specifically be a k-means algorithm or a density-based clustering algorithm.

Clustering (C l uster i ng) is the most common unsupervised learning algorithm, which refers to partitioning a dataset into different classes or clusters according to some specific criteria (e.g., distance) such that the similarity of data objects within the same cluster is as large as possible, while the variability of data objects that are not in the same cluster is also as large as possible. That is, the data of the same class after clustering are gathered together as much as possible, and the data of different classes are separated as much as possible.

Clustering is a data analysis and mining technique for grouping similar objects into the same group, called clusters, while dissimilar objects are grouped into different groups.

Clusters can be obtained by a clustering algorithm, so that clusters can be screened. The clustering result may include the determined one or more clusters. The present embodiment is not limited to a specific screening method, and the obtained area where the cluster is located may be directly determined as an area where the pre-defect positions are gathered, or may be screened according to the number of pre-defect positions in the cluster.

The present embodiment is not limited to a specific preset number of conditions. Specifically, clusters containing the number of the pre-defect positions larger than a preset number threshold can be selected, or clusters with the sequence numbers smaller than a preset sequence number can be selected according to the sequence number of the pre-defect positions from large to small. The area where the pre-defect locations are clustered may then be determined based on the screened clusters.

The present embodiment is not limited to the manner in which the area in which the pre-defect positions are concentrated is determined based on the clusters obtained by screening. Specifically, it may be determined that a region contains any cluster obtained by screening, and the region may not contain other pre-defect positions. The divided area where any cluster obtained by screening is located may be determined as the area where the pre-defect positions are gathered according to the pre-divided area.

In addition, in addition to clustering the pre-defect locations, the determined other processing data may be combined to participate in clustering with the pre-defect locations.

For example, the defect type, the defect grade and the product batch are bound with the corresponding pre-defect positions, and further comprehensive clustering is performed. Therefore, the dimension of the clustering can be increased, and the accuracy of the clustering is improved.

It will be appreciated that for each cluster obtained by screening, the area of the corresponding cluster of pre-defect locations may be determined. Specific ways can be seen from the above.

The present embodiment is not limited to a specific machine learning model structure, and may be specifically a self-encoder, or a specific structure in a self-encoder, or the number of encoding layers and the number of decoding layers in a self-encoder.

The self-encoder here can be used for reconstruction of the incoming pre-defect locations. Alternatively, the self-encoder may be specifically trained by using the pre-defect positions in the pre-defect high-incidence area in the machining result.

Specifically, the pre-defect positions in the pre-defect high-incidence area in the processing result are respectively used as sample characteristics and sample labels for training the self-encoder.

The pre-defect positions in the pre-defect high-incidence area in the processing result are concentrated in distribution, so that the self-encoder for better reconstructing the pre-defect positions in the pre-defect high-incidence area can be trained.

The relatively discrete pre-defect positions in the processing result are not located in the pre-defect high-incidence area, and after the trained self-encoder is input, the self-encoder cannot be well rebuilt, so that great difference exists between input and output.

Thus, optionally, the training process of the self-encoder may include: and respectively taking the pre-defect positions in the pre-defect high-incidence area in the processing result as sample characteristics and sample labels to train the self-encoder.

Specifically, the pre-defect position in the high-incidence area of the pre-defect in the processing result is input into the self-encoder to obtain the reconstruction position output by the self-encoder, and the distance between the input pre-defect position and the reconstruction position corresponding to the output is reduced by updating the parameters of the self-encoder.

The pre-defect high-incidence area in the processing result can be predetermined in other modes, so that the pre-defect high-incidence area can be used as a training sample for training.

Of course, other steps may be included in the training process of the self-encoder, and the present embodiment is not limited thereto.

Accordingly, the present embodiment is not limited to a specific preset distance threshold. For the pre-defect positions with the distance from the corresponding reconstruction position smaller than the preset distance threshold value, the pre-defect positions which are positioned in the pre-defect high-incidence area can be determined, so that the area for determining the pre-defect position aggregation can be further predicted.

The present embodiment does not limit the manner of determining the area where the pre-defect positions are gathered according to the pre-defect positions having a distance from the corresponding reconstructed position smaller than the preset distance threshold. Specifically, it may be determined that a region contains pre-defect positions with a distance from the corresponding reconstructed position less than a preset distance threshold, and the region may not contain other pre-defect positions. The divided area where the pre-defect position is located, which has a distance smaller than the preset distance threshold value from the corresponding reconstruction position, may be determined as the area where the pre-defect positions are gathered according to the pre-divided area.

In addition, the input of the machine learning model can also add other determined processing data, such as defect type, defect grade, product batch and the like, so that the dimension of the input characteristic can be increased, and the prediction effect of the machine learning model can be improved.

The implementation process of the specific isolated forest algorithm is not limited, so long as the pre-defect position of the outlier can be determined. The other pre-defect positions than the outlier pre-defect positions can be regarded as the aggregated pre-defect positions, so that the region where the pre-defect positions are aggregated can be determined based on the other pre-defect positions than the outlier pre-defect positions.

The present embodiment is not limited to a specific manner of determining the area where the pre-defect positions are gathered. Specifically, it may be determined that a piece of area contains "other pre-defect positions than the pre-defect positions of the outliers", and the area may not contain the pre-defect positions of the outliers. The divided area where the "other pre-defect positions than the outlier pre-defect position" is located may be determined as the area where the pre-defect positions are gathered, according to the pre-divided area.

In addition, the decision tree may be used to determine the location of the aggregated pre-defect in combination with other determined processing data, such as defect type, process data, defect level, product lot and process data, etc., for the anomaly detection algorithm of the decision tree.

The embodiment provides 3 preset abnormality detection algorithm embodiments, and the efficiency and accuracy of abnormality detection can be improved through a specific algorithm or model, so that the efficiency and accuracy of determining the area where the pre-defect positions are gathered are improved.

Of course, it is understood that the manner of determining the pre-defect high incidence area is not limited to the above-described embodiment. The above embodiments are for illustrative purposes only.

In addition, the foregoing explains that for any pre-processing step, the pre-defect location may be determined, other data may be determined, and the pre-defect high incidence area may be determined.

It is understood that, for different pre-processing steps, the pre-defect positions in the processing results of the different pre-processing steps may be determined respectively, and the pre-defect high-incidence areas corresponding to the respective pre-processing steps may be determined respectively. The explanation of the different pre-processing steps can be found separately above.

Alternatively, the foregoing embodiments may be adopted for different pre-processing steps, respectively, to determine pre-defect high-incidence areas that are in one-to-one correspondence, respectively.

For example, for each of the 3 pre-processing steps, using the above embodiment, each of the 3 corresponding pre-defect high incidence areas may be determined.

And aiming at a plurality of pre-defect high-incidence areas, the follow-up step can synthesize the determined pre-defect high-incidence areas, and forecast and determine the forecast defect high-incidence areas corresponding to the current processing step.

In addition, the above embodiment may be adopted to determine the high incidence area of the pre-defect by using the distribution of the integrated pre-defect positions by integrating the pre-defect positions in the processing results of the respective pre-processing steps for the respective different pre-processing steps. In particular, it is possible to integrate several pre-defect locations determined for different pre-processing steps.

For example, for each of the 3 pre-processing steps, by using the above embodiment, 1 pre-defect high-incidence area may be determined by integrating the pre-defect positions determined for each of the 3 pre-processing steps.

After the pre-defect high-incidence area is determined, the pre-defect high-incidence area corresponding to the current processing step can be predicted and determined by further utilizing the subsequent steps. The predicted defect high incidence area may include a pre-defect high incidence area.

3. Update regarding anomaly detection algorithms.

Further, for the anomaly detection algorithm, in an alternative embodiment, since the actual defect position may also be obtained by detecting in the processing result of the current processing step, an actual defect high incidence area may also be obtained.

Therefore, the preset anomaly detection algorithm can be updated again in the reverse direction by determining the actual defect position in the machining result of the current machining step, so that the accuracy of prediction is improved.

Optionally, the above method flow may further include any one of the following:

1) Determining an actual defect high-incidence area corresponding to the current processing step; updating parameters of the preset anomaly detection algorithm to reduce the difference between the determined predicted defect high incidence area and the determined actual defect high incidence area may be reducing the difference between the predicted defect high incidence area and the actual defect high incidence area.

2) Determining the actual defect position in the processing result of the current processing step; the parameters of the preset anomaly detection algorithm are updated to increase the number of determined actual defect locations contained in the determined predicted defect high incidence area, which may be the number of actual defect locations contained in the predicted defect high incidence area.

The present embodiment is not limited to the manner of determining the actual defect position, and specifically, the actual defect position in the processing result of the current processing step may be determined by means of quality inspection or image recognition, etc., as described in the above embodiments.

The present embodiment is also not limited to the manner in which the actual defect high incidence area is determined. Specifically, the actual defect position in the processing result of the current processing step is determined, and then the actual defect high-incidence area is determined according to the determined actual defect position. For a specific manner, see the above embodiments, clustering or statistical manners may be adopted.

The present embodiment is not limited to the manner of updating the parameters of the preset anomaly detection algorithm. Specifically, for the clustering algorithm, some parameters in the clustering algorithm, such as the k value in k-means, may be updated and adjusted. For machine learning models, parameters in the machine learning model may be updated and adjusted.

According to the embodiment, the accuracy of determining the predicted defect high-incidence area can be improved by updating the parameters of the preset abnormality detection algorithm.

It can be understood that, when the preset anomaly detection algorithm is trained in advance, the parameters of the preset anomaly detection algorithm may also be updated by adopting the method of the above embodiment.

For ease of understanding, in a specific example, a specific preset anomaly detection algorithm may include:

1) And constructing a multi-classification model of the decision tree aiming at each related parameter, calculating the accuracy of the model based on the test data set of the decision tree, and outputting branch rules of the decision tree for judging abnormal classification.

2) For each relevant parameter, a self-codec model is constructed. The four-layer encoder and the two-layer decoder are respectively built, the activation function is Re l u, the defined loss function can be mean square error (Mean Square Error, MSE), and the optimizer learning rate alpha can be set to be 0.001. Based on the comparison of the reconstructed data and the defect high incidence area in the subsequent flow, calculating the accuracy of the model, and judging whether the model is abnormal or not according to the model.

3) For each relevant parameter, a DBSCAN (ε, minP) density-based clustering algorithm is constructed. Where ε represents a given domain radius and minP is the minimum number of points within the domain radius that are targeted at the core point. A Bayesian optimization algorithm can be adopted, the model accuracy is evaluated based on the detection label, an optimal solution enabling the model accuracy to reach the highest is searched, and then whether the model is abnormal or not is judged according to the model.

4. S103: and predicting and determining the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

1. And determining a predicted defect high incidence area according to any pre-processing step.

The predicted defect high-incidence area may specifically be a predicted defect high-incidence area in a processing result of the current processing step. The processing result of the current processing step may specifically be a theoretical processing result of the current processing step.

The predicted defect high-incidence area can be predicted based on the pre-defect high-incidence area, and can be used for representing an area which is easy to generate defects in the processing result of the current processing step and can also be used for representing an area which is easy to generate defects in the theoretical processing result of the current processing step. The analysis and determination of the predicted defect high-incidence area can be used for helping to improve and detect quality, so that the yield of products and processing results is conveniently improved, and the probability of defects of the products and the processing results is reduced.

The method is not limited to a specific mode of predicting and determining the predicted defect high incidence area corresponding to the current processing step. Wherein the predicted defect high incidence area may include a pre-defect high incidence area.

Alternatively, due to the association between the pre-processing step and the current processing step, the current processing step is performed on the basis of the processing result of the pre-processing step, so that the determined pre-defect high incidence area can be directly determined as the predicted defect high incidence area. For a specific explanation, reference may be made to the above explanation, and defects that may easily occur in the processing results of the preceding processing steps may also easily occur in the processing results of the current processing step.

In addition, the method can be further expanded on the basis of the pre-defect high-incidence area to obtain a predicted defect high-incidence area so as to improve the comprehensiveness of determining the predicted defect high-incidence area.

The present embodiment is not limited to a specific expansion mode. Specifically, the size of the defect can be expanded and fixed, or a region containing the high incidence region of the pre-defect can be determined.

The number of the pre-defect high-incidence areas determined is not limited, and one or more pre-defect high-incidence areas may be determined.

For a plurality of pre-defect high-incidence areas, predicting and determining a predicted defect high-incidence area corresponding to the current processing step based on any pre-defect high-incidence area respectively; the method can also integrate a plurality of pre-defect high incidence areas, and forecast and determine the predicted defect high incidence areas corresponding to the current processing step; and the high incidence areas of the pre-defects can be integrated, and the high incidence areas of the pre-defects corresponding to the current processing step can be predicted and determined.

The present embodiment is not limited to a specific manner of determining the predicted defect high incidence area for the plurality of pre-defect high incidence areas.

The predicted defect high incidence areas may be determined for each of the pre-defect high incidence areas by referring to the above embodiment, and the predicted defect high incidence areas may be combined as the predicted defect high incidence areas corresponding to the final current processing step.

Of course, a plurality of pre-defect high incidence areas may be directly determined as predicted defect high incidence areas; the predicted defect high incidence area can be obtained by expanding the predicted defect high incidence area on the basis of a plurality of pre-defect high incidence areas.

Optionally, the determining the predicted defect high incidence area corresponding to the current processing step based on the pre-defect high incidence area corresponding to the pre-processing step may specifically be: and determining the pre-defect high-incidence area corresponding to the pre-processing step as the predicted defect high-incidence area corresponding to the current processing step.

The embodiment can directly determine the pre-defect high-incidence area as the predicted defect high-incidence area, and improves the efficiency of determining the predicted defect high-incidence area.

Further, the high-incidence region of the pre-defect corresponding to the pre-processing step may be expanded, and the expanded region may be determined as the high-incidence region of the predicted defect corresponding to the current processing step. The predicted defect high incidence area may include a pre-defect high incidence area corresponding to the pre-processing step.

2. For a plurality of different pre-processing steps, a predicted defect high incidence region is determined.

The above embodiments explain the manner of predicting and determining the predicted defect high incidence area corresponding to the current processing step specifically for one or more pre-defect high incidence areas corresponding to any pre-processing step.

The manner in which the predicted defect high incidence area corresponding to the current processing step is predicted and determined specifically for a plurality of different pre-processing steps is further explained below.

The plurality of different pre-processing steps herein may be a plurality of different pre-processing steps performed prior to the current processing step during product processing.

The method flow is not limited to a specific manner of determining the predicted defect high incidence area for the plurality of pre-processing steps.

In an alternative embodiment, the above embodiments are primarily explained with respect to determining the corresponding pre-defect high incidence area for any pre-processing step.

It will be appreciated that, for a plurality of pre-processing steps, the steps S101 and S102 may be performed separately, and the pre-defect high-incidence areas corresponding to the steps may be determined separately, so that a plurality of pre-defect high-incidence areas may be obtained. For the manner in which the corresponding pre-defect high incidence areas are determined for each pre-processing step, reference may be made to the above embodiments.

Accordingly, the predicted defect high incidence area corresponding to the current processing step can be predicted and determined further based on the determined pre-defect high incidence areas corresponding to the plurality of pre-processing steps, respectively.

The predicted defect high incidence area may include a plurality of determined pre-defect high incidence areas, so as to improve the comprehensiveness of the predicted defect high incidence area.

Optionally, for any one of the at least one pre-processing step, determining a pre-defect position in a processing result of the pre-processing step, and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the determined distribution of the pre-defect positions, where the pre-defect high-incidence area may specifically be: for different pre-processing steps, respectively executing: and determining the pre-defect positions in the processing result of the pre-processing step, and determining the pre-defect high-incidence area corresponding to the pre-processing step according to the distribution condition of the determined pre-defect positions. That is, S101 and S102 may be performed for different pre-processing steps, respectively.

Correspondingly, based on the pre-defect high-incidence area corresponding to the pre-processing step, the predicted defect high-incidence area corresponding to the current processing step is predicted, which specifically may be: and predicting and determining the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence areas respectively corresponding to the different pre-processing steps. That is, the predicted defect high incidence area corresponding to the current processing step is predicted based on the pre-defect high incidence areas corresponding to the different pre-processing steps.

The method and the device can integrate the processing results of a plurality of different pre-processing steps, are used for predicting and determining the predicted defect high incidence area corresponding to the current processing step, and improve the accuracy of determining the predicted defect high incidence area by increasing the data quantity of the prediction basis.

Whereas S101 and S102 may be performed separately for different pre-processing steps, a specific explanation may be found in the above embodiments.

Of course, for different pre-processing steps, the pre-defect positions in the processing results of a plurality of pre-processing steps may be integrated, and the pre-defect high-incidence area may be determined according to the integrated distribution of the pre-defect positions.

The embodiment is not limited to a specific manner of predicting and determining the predicted defect high incidence area corresponding to the current processing step based on the pre-defect high incidence areas corresponding to the different pre-processing steps.

Alternatively, the pre-defect high-incidence areas corresponding to different pre-processing steps can be directly determined as the predicted defect high-incidence areas. The method can also be used for expanding the pre-defect high-incidence areas corresponding to different pre-processing steps respectively, and determining the expansion result as a predicted defect high-incidence area. Or predicting the intersection area or the union area between the pre-defect high-incidence areas corresponding to different pre-processing steps respectively to determine the predicted defect high-incidence area.

Since the pre-defect high-incidence areas respectively corresponding to the different pre-processing steps can be regarded as a plurality of pre-defect high-incidence areas, the explanation of the predicted defect high-incidence areas can be determined with reference to the plurality of pre-defect high-incidence areas in the above embodiment.

The present embodiment is not limited to a specific expansion method, and may be an expansion area that expands a sheet of the high-incidence area including the pre-defect, or may be a fixed size.

The present embodiment is not limited to the manner of predicting based on the intersection region or the union region, and specifically may directly determine the intersection region or the union region as the predicted defect high incidence region, or may expand based on the intersection region or the union region and determine the expansion result as the predicted defect high incidence region.

Thus, optionally, predicting the predicted defect high incidence area corresponding to the current processing step based on the pre-defect high incidence areas respectively corresponding to the different pre-processing steps may include any one of the following:

1) And determining a union region aiming at the pre-defect high-incidence regions corresponding to different pre-processing steps respectively, and predicting the predicted defect high-incidence region corresponding to the current processing step based on the union region.

2) And determining an intersection area aiming at the pre-defect high-incidence areas corresponding to different pre-processing steps respectively, and predicting the predicted defect high-incidence area corresponding to the current processing step based on the intersection area.

The present embodiment is not limited to the manner in which prediction is performed based on the intersection region or the union region.

The method can improve the comprehensiveness of determining the predicted defect high-incidence areas by utilizing the union areas among the pre-defect high-incidence areas respectively corresponding to different pre-processing steps.

While the present embodiment is not limited to the manner of specifically determining the intersection region in determining the intersection region. Alternatively, the intersection regions between the pre-defect high-incidence regions corresponding to all the pre-processing steps may be the intersection regions between the pre-defect high-incidence regions corresponding to part of the pre-processing steps, the intersection regions between the pre-defect high-incidence regions corresponding to any two pre-processing steps may be the intersection regions between the pre-defect high-incidence regions corresponding to any three pre-processing steps, and so on.

By determining the intersection region, it is possible to conveniently determine a region in which defects are likely to occur in the processing results of the plurality of preceding processing steps, for determining the predicted defect high incidence region, so that it is possible to improve the accuracy of determining the predicted defect high incidence region.

It can be understood that the determined intersection area may be directly determined as the predicted defect high incidence area corresponding to the current processing step, or the expansion may be performed based on the determined intersection area, and the expansion result may be determined as the predicted defect high incidence area corresponding to the current processing step.

In addition, the divided region having an intersection with the determined intersection region may be determined as the predicted defect high incidence region corresponding to the current processing step in combination with the previously divided region.

Similarly, prediction can be performed with reference to the above embodiment for the union region.

The above embodiment explains that the predicted defect high incidence area corresponding to the current processing step is determined by analyzing the data of the processing result in the preceding processing step.

In the product processing process, one or more pre-processing steps can exist before the current processing step, so that analysis can be performed by combining data of processing results in the pre-processing steps, and the analyzed data size is increased, so that the comprehensiveness and accuracy of determining the predicted defect high-incidence area are improved.

In addition, by analyzing the data of the actual machining result in the pre-machining step, the requirement on the actual machining result of the current machining step can be reduced, the predicted defect high-incidence area can be predicted and determined without the actual machining result of the current machining step, the corresponding predicted defect high-incidence area can be predicted and determined conveniently before the current machining step is actually executed, and the current machining step can be adjusted or improved conveniently, so that the product yield is improved.

In a specific example, for a new type of product, during actual processing, it is often necessary to accumulate more actual production data after a certain number of actual products are produced, so that specific defect data can be analyzed. However, these actual products tend to be prone to more rejects and, with subsequent analysis and adjustment processes, they tend to be destroyed, wasting resources. Therefore, the flow of the method, particularly in the above embodiment, may be adopted, and by means of a plurality of processing steps in the product processing process, actual production data of the pre-processing step is used for accumulation and analysis, so as to quickly determine the predicted defect high incidence area corresponding to the current processing step, and thus, the accumulated data may be convenient, and the requirement on the actual processing result of the current processing step is reduced.

The current processing step can be any processing step with a pre-processing step in the product processing process, so that the method can be executed for different current processing steps for multiple times to determine the predicted defect high incidence area so as to be convenient for adjustment.

For ease of understanding, in one specific example, the product manufacturing process may include process steps 1-10.

The processing steps 1-4 may be performed first, resulting in actual processing results for each processing step, where the processing step 5 may be temporarily not continued.

And taking the processing step 5 as the current processing step, and combining the actual processing results of the processing steps 1-4, and predicting and determining the predicted defect high incidence area corresponding to the processing step 5 based on the method flow.

The processing step 5 may be adjusted or modified, specifically, more strict quality detection may be performed on the predicted defect high incidence area, or process modification may be performed on the predicted defect high incidence area, so as to reduce the probability of occurrence of defects.

With the actual execution of the processing step 5, several actual processing results of the processing step 5 can be obtained. The processing step 6 may not be continued for a while.

And then, taking the processing step 6 as the current processing step, and combining the actual processing results of the processing steps 1-5, and predicting and determining the predicted defect high incidence area corresponding to the processing step 6 based on the method flow. And so on.

3. And combining the processing result data of other processing steps to determine the predicted defect high incidence area.

In an alternative embodiment, in addition to the processing data of the preceding processing step, the predictive determination may also be performed in combination with the processing data of other processing steps.

The present embodiment is not limited to specific other processing steps. Alternatively, the processing data of the current processing step may be specifically processing data of a post-processing step after the current processing step.

The specific post-processing step may be any processing step performed after the current processing step in the product processing process.

The method is not limited to a specific manner of determining the post-processing step, and may be a method of randomly selecting a processing step to be performed after the current processing step, or a method of fixedly selecting a plurality of processing steps to be performed immediately after the current processing step.

The post-processing step may be performed based on the processing results of the current processing step.

The defect location in the processing result of the current processing step may also be reflected in the post-processing step, in particular, see the explanation of the pre-processing step above, so that an analysis can be performed in combination with the processing data of the post-processing step.

Optionally, the method flow may further include: determining a post defect position in a machining result of the post machining step for any post machining step; determining a post defect high incidence area corresponding to the post processing step according to the determined distribution condition of the post defect positions; based on the pre-defect high-incidence area corresponding to the pre-processing step, predicting and determining the predicted defect high-incidence area corresponding to the current processing step can be specifically: and predicting and determining a predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step and the post-defect high-incidence area corresponding to the post-processing step.

Wherein, the relevant explanation of the specific post-processing step can be seen from the relevant explanation of the above embodiment, and the relevant explanation of the pre-processing step can be seen from the specific explanation.

The embodiment can combine the processing result data of the pre-processing step and the processing result data of the post-processing step, comprehensively analyze and determine the pre-defect high-incidence area and the post-defect high-incidence area, determine the predicted defect high-incidence area, improve the data volume serving as the analysis and prediction basis, and improve the accuracy of determining the predicted defect high-incidence area.

In addition, the present embodiment does not limit the number of the determined post-defect high incidence areas, and one or more post-defect high incidence areas may be determined.

It will be appreciated that the above embodiments may also be performed separately for a plurality of different post-processing steps, to determine the corresponding post-defect high incidence areas. Specific explanations can be found above.

The embodiment is not limited to a specific way of predicting the pre-defect high incidence area and the post-defect high incidence area. Reference may be made in particular to the relevant explanations above.

Alternatively, the pre-defect high incidence area and the post-defect high incidence area may be directly determined as the predicted defect high incidence areas; the pre-defect high-incidence area and the post-defect high-incidence area can be expanded, and the expansion result is determined to be the predicted defect high-incidence area; the predicted defect high incidence area may be determined based on an intersection area or a union area between the pre-defect high incidence area and the post-defect high incidence area.

The processing data of the current processing step may be processing data in a processing result obtained by performing the current processing step in the past, for example, an actual defect position in the processing result obtained in the past, or other processing data in the current processing step.

For convenience of description and distinction, the processing result obtained by performing the current processing step in the past is referred to as a history processing result. The actual defect position in the history processing result obtained by executing the current processing step in the past is referred to as a history defect position. Specific explanations can be found in connection with the above embodiments.

Accordingly, based on the distribution situation of the historical defect positions, analysis can be performed to determine the areas with dense or concentrated distribution of the historical defect positions, so that the method can be used for determining the areas which are easy to generate defects in the historical processing result of the current processing step, namely the defect high-incidence areas.

The defect high incidence area determined by analysis using the historical processing result of the current processing step may be referred to as a historical defect high incidence area.

The high-incidence area of the historical defect can be predicted based on the determined position of the historical defect, can be used for representing the area which is easy to generate the defect in the historical processing result of the current processing step, and can also be used for representing the area which is easy to generate the defect in the theoretical historical processing result of the current processing step.

The area in which defects easily occur in the history processing result of the current processing step is usually an area in which defects easily occur in the current processing result of the current processing step, or an area in which defects easily occur in the processing result subsequent to the current processing step.

Therefore, the predicted defect high incidence area can be predicted and determined based on analysis of the history processing result of the current processing step together with the above-described pre-defect high incidence area.

Optionally, the method flow may further include: determining a historical defect position in a historical processing result of the current processing step; according to the determined distribution condition of the historical defect positions, determining a high incidence area of the historical defects; based on the pre-defect high-incidence area corresponding to the pre-processing step, the predicted defect high-incidence area corresponding to the current processing step is predicted, which specifically may be: and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step and the determined historical defect high-incidence area.

The embodiment can combine the processing result data of the pre-processing step and the historical processing result data of the current processing step, comprehensively analyze and determine the pre-defect high-incidence area and the historical defect high-incidence area, determine the predicted defect high-incidence area, increase the data quantity and the data source which are taken as analysis and prediction basis, and improve the accuracy of determining the predicted defect high-incidence area.

The present embodiment does not limit the number of the history processing results, and may specifically be one or more history processing results.

The present embodiment is not limited to a specific manner of determining and acquiring the historic defect positions. See above for details. The determination and acquisition may be performed by means of quality inspection or inquiry.

The present embodiment is also not limited to a specific manner of determining the high incidence area of the history defect. In particular, the method can be determined by adopting an anomaly detection algorithm or statistics and the like according to the embodiment.

The present embodiment does not limit the number of the determined history defect high incidence areas, and may be one or more history defect high incidence areas.

The embodiment is not limited to a specific way of predicting and determining the predicted defect high incidence area based on the pre-defect high incidence area and the history defect high incidence area. Specific modes can be seen from the explanation in the above examples.

Alternatively, the pre-defect high incidence area and the history defect high incidence area may be directly determined as the predicted defect high incidence area; the pre-defect high-incidence area and the history defect high-incidence area can be expanded, and the expansion result is determined to be the predicted defect high-incidence area; the predicted defect high incidence area may be determined based on an intersection area or a union area between the pre-defect high incidence area and the history defect high incidence area.

The above examples illustrate that various sources of data can be used to analyze and determine areas that are prone to defects in the current processing step, i.e., areas of high incidence of predicted defects. Specifically, at least one of the following may be included: 1) Data of the pre-processing step; 2) Historical data of the current processing step; 3) And (5) data of a post-processing step. The data may include actual defect locations, and may include other process data such as defect type, product lot, defect level, process data, and the like. Therefore, the above-described embodiments can improve the comprehensiveness and accuracy of determining the predicted defective high-incidence area by increasing the data source and the data amount as the prediction basis.

5. And determining an operation performed after predicting the defect high incidence area.

The above embodiment explains the overall process of determining the predicted defect high incidence area, and the overall process can determine the area where defects are likely to occur in the current processing step through data analysis of the product processing process.

It will be appreciated that for different processing steps in the product processing process, the current processing step may be determined separately, and the area where defects are likely to occur is determined by the above method flow.

After determining the area prone to defects, the method is not limited to specific subsequent operations.

Optionally, the current processing step can be reminded of the predicted defect high-incidence area determined in the mode by sending a notification or a prompt, so as to facilitate improvement or important inspection.

For ease of understanding, in one specific example, a sample inspection is typically required when quality inspection is performed with respect to the machining results. After the predicted defect high-incidence area is determined, whether defects exist in the predicted defect high-incidence area or not can be checked with respect to the total machining results, so that the quality inspection efficiency and the comprehensiveness can be improved.

In another specific example, process modifications may be targeted to predicting defect high incidence areas. For example, the defect that the water drop type easily appears in the predicted defect high-incidence area is determined in the mode, so that improvement can be performed in the current processing step, the defect of the water drop type is removed or eliminated, and subsequent processing operation is performed, so that the yield of products can be improved.

Therefore, optionally, after determining the predicted defect high incidence area, the method flow may further include: sending prompt information; the hint information may be used to characterize at least one of: 1) Aiming at the determined predicted defect high incidence area, the processing level of the current processing step is improved; 2) Aiming at the determined predicted defect high-incidence area, the defect inspection precision is improved; 3) A strategy for adjusting quality inspection, etc., for the determined predicted defect high incidence area.

The embodiment can conveniently reduce the defect occurrence condition through the prompt information, and improve the product yield.

The embodiment is not limited to the specific form of the prompt information, and can adopt an alarm

In addition, in an alternative embodiment, since the current processing step is further performed, an actual processing result is obtained, so that the actual defect position in the actual processing result is conveniently determined. And based on the determined actual defect position, the actual defect high incidence area can be conveniently determined.

Therefore, the method for determining the predicted defect high incidence area in the method flow can be reversely adjusted according to the actual defect position and/or the actual defect high incidence area.

In the above embodiments, it is explained that the parameters of the preset anomaly detection algorithm may be adjusted according to the actual defect position and/or the actual defect high incidence area.

In addition, optionally, the pre-processing steps aimed at may be adjusted according to the actual defect position and/or the actual defect high-incidence area, and specifically, other pre-processing steps may be replaced, or the number of pre-processing steps aimed at may be updated.

The method for determining the pre-defect positions can also be adjusted, and more pre-defect positions can be determined, or the pre-defect positions meeting the specified conditions can be determined.

The method of determining the high incidence area of the pre-defect may be adjusted, for example, the method of replacing the preset abnormality detection algorithm with statistics may be performed, or the method of replacing one preset abnormality detection algorithm with another preset abnormality detection algorithm may be performed.

The embodiment is not limited to a specific adjustment manner, and the adjustment direction may specifically be at least one of the following: 1) Reducing the difference between the determined predicted defect high incidence area and the determined actual defect high incidence area; 2) The number of determined actual defect locations included in the determined predicted defect high incidence area is increased.

6. And (5) a qualified detection mode of the processing result.

In an alternative embodiment, other data may be used for analysis in addition to the above described method flow for defect location.

The present embodiment is not limited to the analyzed data. Alternatively, the analysis may be performed as to whether the processing result of the processing step is acceptable.

It should be noted that, specifically, the data may be automatically analyzed by a machine device to analyze whether the machining result is qualified, and compared with manually checking whether the machining result is qualified, the analysis efficiency may be improved.

And the data is analyzed through the machine equipment, and the qualified processing result and the unqualified processing result can be further analyzed so as to determine specific differences and factors influencing whether the processing result is qualified or not, thereby being convenient for improvement so as to improve the product yield.

It will be appreciated that the processing results may also be specifically considered as products, so that whether the products are acceptable can be analyzed. The following is an explanation of the processing results for convenience of description. It is understood that the processing results hereinafter may also be considered as products.

For convenience of explanation, first, criteria concerning qualification of the processing result are explained.

The present embodiment does not limit the criterion of whether the processing result is acceptable or not. Specifically, the product parameters of the processing result may be within a specified range or meet specified requirements, so that it may be determined that the processing result is acceptable.

The product parameters here may be in particular the properties of the processing result itself, such as size, brightness, color rendering, color difference, error size, and whether there is a breakage or not, etc.

For example, the processing result or product is a display panel whose size needs to be within a specified range or meet specified requirements. If the size of the display panel exceeds the specified range, it can be determined that the display panel is not acceptable.

For another example, the processing result or the product is a display panel, and the brightness of the display panel needs to be within a specified range. If the brightness of the display panel exceeds the specified range, it can be determined that the display panel is not acceptable.

As another example, the result of the process or the product is a toy, which needs to be free of breakage. If a breakage occurs in the toy, it may be determined that the toy is not acceptable.

For one processing result, a plurality of product parameters are generally required to be within a specified range, and can be determined as a qualified processing result.

Conversely, for one process result, if one or more product parameters are outside of the specified range, then it may be determined as a failed process result.

1. And predicting whether the machining result is qualified.

The present embodiment is not limited to a specific predicted processing result object, and may specifically be a qualified prediction for any processing result.

Optionally, the method flow may further include: predicting whether any processing result of any processing step in the product processing process is qualified.

In this embodiment, qualification prediction can be performed for any processing result produced in the processing process of the product. As the qualification prediction is automatically carried out through the machine equipment, the efficiency and the comprehensiveness of the qualification prediction can be improved, the yield of the processing result can be conveniently mastered, and the strategy adjustment can be conveniently carried out according to the yield of the processing result in real time.

Of course, the present embodiment is not limited to the operation after predicting whether the machining result is acceptable. Policy adjustments or data analysis may be performed specifically, for example, to analyze data closely related to the failed process results.

In addition, after the processing result is predicted to be qualified, the unqualified processing result is usually unqualified due to the defect, so that the unqualified processing result can be analyzed and detected in a targeted manner to determine the defect position.

Optionally, for any one of the at least one pre-processing steps, determining the pre-defect location in the processing result of the pre-processing step for which it is intended may specifically be: predicting and determining a disqualification machining result from machining results of at least one pre-machining step; and checking the predicted unqualified machining result to determine the position of the pre-defect.

According to the method, the device and the system, unqualified machining results can be predicted and screened, the position of the pre-defect is determined through inspection in a targeted mode, and the efficiency of determining the position of the pre-defect is improved.

In order to analyze whether the processing result or the product is acceptable, the relevant data in the processing process can be used as a basis for analysis.

The present embodiment is not limited to the relevant data during processing. The relevant data in the process may specifically be process parameters, such as process data, process accuracy, temperature, humidity, pressure values, gas concentrations, etc. Other process and production parameters include: evaporation rate, flow rate, water resistance, number of washes, waiting time, processing time, etc.

It will be appreciated that these processing parameters may affect whether the processing results are acceptable or not, and that the values of the processing parameters may also exhibit different distributions for acceptable processing results and unacceptable processing results.

For example, if there is a large deviation in temperature and humidity, a large defect tends to occur easily in processing a product, resulting in an unacceptable processing result.

The present embodiment is not limited to a specific prediction method. Alternatively, a direct numerical calculation or a numerical analysis may be used, for example, if the temperature exceeds a specified range or the machining precision is less than a preset precision, the machining result may be determined to be failed directly.

In addition, optionally, when the data is analyzed specifically, since the data amount of the processing parameters is often rich, a machine learning model may also be used for analysis. Specifically, the machine learning model which is trained in advance and used for analyzing the processing parameters can be adopted for analysis, and the machine learning model which is trained in advance and used for predicting the processing results or whether the products are qualified can be adopted for analysis.

The present embodiment is not limited to a specific machine learning model structure. Specifically, a neural network model or a convolution model can be used.

In an alternative embodiment, the self-encoder may be used for training, considering the difficulty of obtaining training samples, since the number of acceptable processing results tends to be high, while the number of unacceptable processing results tends to be low, and the duty ratio difference between the positive and negative samples is high.

The self-encoder may be used to reconstruct the input features, which tend to be the same as the tag when the self-encoder is specifically trained.

Wherein only positive samples, i.e. acceptable machining results, may be used as samples for training the self-encoder. Because the processing parameter distribution of the qualified processing result often has a large difference from the processing parameter distribution of the unqualified processing result, the processing parameter distribution of the qualified processing result can be fitted by using the self-encoder so as to be convenient for reconstruction to obtain similar processing parameters. For processing parameters of unqualified processing results, the self-encoder is generally difficult to reconstruct to obtain similar processing parameters, but rather, processing parameters with larger differences are reconstructed.

Of course, the self-encoder may be trained in other ways, and the present embodiment is not limited.

Thus, optionally, the prediction method for predicting whether the machining result is acceptable may specifically include: inputting the actual processing parameters of any processing result into a pre-trained self-encoder to obtain corresponding reconstruction processing parameters output by the self-encoder; and under the condition that the difference between the actual processing parameter and the corresponding reconstruction processing parameter is larger than a preset difference threshold value, predicting the processing result as a disqualified processing result.

The embodiment can rapidly predict whether the machining result is qualified based on the pre-trained self-encoder, so that the efficiency of predicting whether the machining result is qualified can be improved.

The embodiment is not limited to a specific preset difference threshold, to specific actual processing parameters, or to a specific training mode of the self-encoder.

Alternatively, the training process of the self-encoder may include: determining an initial self-encoder; determining actual processing parameters of qualified processing results; the following steps are circularly executed until the preset training stop condition is met: inputting the determined actual processing parameters into a current initial self-encoder to obtain corresponding reconstructed processing parameters output by the current initial self-encoder; updating the current initial self-encoder to reduce the difference between the input actual processing parameters and the corresponding reconstructed processing parameters, and re-determining the updated initial self-encoder as the current initial self-encoder; after the cycle is completed, the current initial self-encoder is determined as the self-encoder with the training completed.

The embodiment can train the self-encoder by utilizing the actual processing parameters of the qualified processing result, so that the self-encoder can learn the distribution condition of the actual processing parameters of the qualified processing result.

The embodiment is not limited to a preset training stop condition, and specifically may be that the difference between the input actual processing parameter and the corresponding reconstructed processing parameter is smaller than a preset difference value, where the preset difference value may be the preset difference threshold, or may be smaller than the preset difference threshold, so as to distinguish a qualified processing result and an unqualified processing result through the preset difference threshold; or the cycle reaches the preset cycle number, etc.

The present embodiment is also not limited to a loss function during training, as long as it can be used to characterize the difference between the input actual process parameters and the corresponding reconstructed process parameters. In particular, the loss function may be in the form of the sum of absolute values of the difference values, or the like.

The present embodiment is not limited to a specific structure of the self-encoder, and alternatively, the self-encoder may include an input layer, a hidden layer, and a reconstruction layer therein. Wherein the hidden layer may specifically include an encoding layer and/or a decoding layer, and the specific number is not limited. The reconstruction layer may in particular comprise one or more decoding layers. The reconstruction layer may be used in particular for reconstructing input data of the input layer.

Of course, the self-encoder may also be trained in other ways, and the above embodiments are for illustrative purposes.

According to the embodiment, the self-encoder is adopted for training and predicting, so that the sample requirement on unqualified processing results can be reduced, the actual condition of product processing can be attached, and the training effect and the prediction accuracy of the model are improved.

In addition, other means of predicting whether the machining result is acceptable may be employed.

For ease of understanding, fig. 5 is a schematic diagram of a self-encoder according to an embodiment of the present invention. The self-encoder may include an input layer, a hidden layer, and a reconstruction layer, among others.

2. And analyzing disqualification factors of the processing results.

After whether the processing result is acceptable or not is predictive-analyzed using the above-described embodiment, the cause associated with the processing result failure can be further analyzed.

For example, it is possible to analyze which processing parameters have a greater influence on whether the processing results are acceptable, and thus control accuracy of these processing parameters can be improved to improve the product yield.

The present embodiment is not limited to a specific manner of analyzing the factors.

In an alternative embodiment, for a self-encoder, encoding may be achieved by a hidden layer therein, extracting features for the actual processing parameters entered. While there is typically a difference in the extracted hidden features with respect to the actual processing parameters that have been determined to be either a pass or fail processing result, there is typically a difference in the distribution of hidden features. Thus, analysis can be performed in combination with hidden layer features in the self-encoder.

Thus, optionally, the above method flow may further include: inputting the actual processing parameters of a plurality of unqualified processing results and the actual processing parameters of a plurality of qualified processing results into a pre-trained self-encoder to obtain hidden layer characteristics; and determining the difference between the hidden layer characteristics of the qualified machining result and the hidden layer characteristics of the unqualified machining result so as to analyze the actual machining parameters associated with the unqualified machining result.

The embodiment is not limited to the specifically obtained hidden layer feature. It is understood that one or more concealment layers may be included in the self-encoder, which may be in particular an encoding layer or a decoding layer.

The pass and fail processing results may be determined in advance by prediction from the encoder or may be determined by actual inspection.

The present embodiment is not limited to the specific way of analyzing the hidden layer feature differences. Alternatively, the difference between the hidden layer features of the pass processing result and the hidden layer features of the fail processing result may be directly compared; the hidden layer characteristics of the qualified processing result can be subjected to statistical analysis, the mean value, the variance, the skewness, the kurtosis and the like are determined, the information such as the dispersion degree, the symmetry and the like of the characteristics is known, and further the hidden layer characteristics of the unqualified processing result are subjected to statistical analysis, so that the statistical analysis results are conveniently compared. The hidden layer features may also be visualized using dimension reduction techniques, such as PCA or t_sne, to observe the distribution of acceptable and unacceptable processing results in the hidden layer representation. Clustering analysis can also be performed on hidden layer features, wherein hidden layer features of qualified processing results are easily divided into the same cluster.

By analyzing the differences between hidden features, it is possible to easily determine differences in hidden features between acceptable and unacceptable processing results, and thus to easily derive input data associated with these differences in reverse, i.e., actual processing parameters.

For example, by analyzing differences in hidden layer characteristics, it is found that input data that causes these differences include at least a processing temperature and a processing humidity, so that it can be determined whether there is a large influence of the processing temperature and the processing humidity on the processing result.

Alternatively, the self-encoder generally contains more nodes, and data transmission can be performed between the nodes through weight parameters. For example, at the input layer node, the input features may be multiplied by weights and output to the hidden layer node. The data output by the connected plurality of input layer nodes can be summed at the hidden layer node. And these weight parameters may be adjusted during the training process.

Accordingly, the influence of the input actual processing parameters on whether the processing result is qualified or not can be conveniently analyzed and determined through the difference between hidden layer characteristics and the weight parameter distribution between the hidden layer nodes and the input layer nodes.

Thus, optionally, the above method flow may further include: inputting the actual processing parameters of a plurality of unqualified processing results and the actual processing parameters of a plurality of qualified processing results into a pre-trained self-encoder to obtain hidden layer characteristics; determining hidden layer nodes meeting preset screening conditions; the preset screening conditions may include: aiming at the qualified processing result and the unqualified processing result, the difference between the output hidden layer characteristics is larger than a preset difference threshold; weight parameters between the actual machining parameters and the determined hidden layer nodes are determined to facilitate analysis of the actual machining parameters associated with the failed machining results.

According to the hidden layer characteristic difference of the qualified processing result and the unqualified processing result in each hidden layer node, hidden layer nodes with larger characteristic difference can be determined, and for a subsequent output layer, distinguishing and identification can be conveniently carried out on the hidden layer characteristics, and the hidden layer nodes often cause different subsequent output results, so that the qualified processing result and the unqualified processing result can be conveniently distinguished and identified. Therefore, these hidden nodes have a greater impact on the subsequent output prediction results.

Correspondingly, if the hidden layer characteristics extracted by part of hidden layer nodes have smaller difference between the qualified processing result and the unqualified processing result, for the subsequent output layer, the hidden layer characteristics are difficult to distinguish and identify, and the influence on the output prediction result is smaller.

The present embodiment is not limited to the preset screening conditions and the preset variance threshold,

The preset screening conditions and the preset difference threshold values can be set for different hidden layer nodes respectively. This is because the numerical ranges of different hidden nodes may be different and thus need to be set separately.

The present embodiment is not limited to the manner of analyzing specifically according to the weight parameters between the actual processing parameters and the hidden layer nodes that are screened.

Optionally, the relationship between the hidden node and the input node may be further analyzed for hidden nodes with large feature differences on the pass and fail process results.

Different input nodes can respectively input different actual processing parameters. Any input node may input one or more actual process parameters. To facilitate attribution analysis, any input node may input an actual process parameter.

Whereas in a self-encoder there may be an association of weight parameters between the input node and the hidden node.

The present embodiment does not limit which hidden layer the analyzed hidden layer node is located in, wherein the hidden layer feature may be a feature output from any hidden layer in the encoder. In particular, it may be a hidden layer immediately adjacent to the input layer, or may be a hidden layer spaced from the input layer by one or more other hidden layers. However, with the analysis of the weight parameters, it can be conveniently determined that the weight parameters between the hidden node and which input nodes are larger, that is, which input nodes have a larger influence on the hidden node to be screened.

In other words, if the weight parameter of a part of input nodes connected with one hidden layer node is larger, the influence on the hidden layer node is larger, and the hidden layer characteristics output by the hidden layer node have larger difference between the qualified processing result and the unqualified processing result and influence on the subsequent output prediction result, so that the influence of the part of input nodes on the subsequent output prediction result can be determined to be larger.

Therefore, it can be determined that the actual machining parameters input by the input nodes have a greater influence on the output prediction result from the encoder and a greater correlation with the unqualified machining result.

In the subsequent strategy adjustment, the actual processing parameters can be controlled more strictly, so that the product yield is improved.

It will be appreciated that the analysis is more efficient for hidden layers immediately adjacent to the input layer.

Of course, the above analysis methods may be combined with each other, and optionally, the above method flow may further include at least one of the following:

1) Inputting the actual processing parameters of a plurality of unqualified processing results and the actual processing parameters of a plurality of qualified processing results into a pre-trained self-encoder to obtain hidden layer characteristics; and determining the difference between the hidden layer characteristics of the qualified machining result and the hidden layer characteristics of the unqualified machining result so as to analyze the actual machining parameters associated with the unqualified machining result.

2) Inputting the actual processing parameters of a plurality of unqualified processing results and the actual processing parameters of a plurality of qualified processing results into a pre-trained self-encoder to obtain hidden layer characteristics; determining hidden layer nodes meeting preset screening conditions; the preset screening conditions comprise: aiming at the qualified processing result and the unqualified processing result, the difference between the output hidden layer characteristics is larger than a preset difference threshold; weight parameters between the actual machining parameters and the determined hidden layer nodes are determined to facilitate analysis of the actual machining parameters associated with the failed machining results.

In alternative embodiments, the analysis may be performed in other ways than the above-described way of performing the attribution analysis using hidden layer features.

Specifically, whether the machining result predicted from the encoder is acceptable may be further compared with the result of whether the machining result is actually acceptable.

According to the comparison result, the situation of correct prediction and the situation of incorrect prediction can be determined.

The following four cases may be specifically included:

1) And predicting that the machining result is qualified from the encoder, wherein the machining result is actually unqualified.

2) And predicting the qualified machining result from the encoder, wherein the machining result is actually qualified.

3) The self-encoder predicts that the machining result is not acceptable and that the machining result is actually acceptable.

4) The self-encoder predicts that the machining result is not acceptable and that the machining result is actually not acceptable.

The number of the 4 conditions is integrated, the indexes such as the accuracy, the precision, the recall rate, the F1-score and the like of the self-encoder can be conveniently analyzed, and further the self-encoder can be conveniently adjusted and improved.

The present embodiment is not limited to a specific manner of determining whether the machining result is actually acceptable, and may specifically be a manner of determining by quality inspection.

It can be understood that the distribution of the actual processing parameters under the four conditions can be analyzed, so that whether the actual processing parameters with larger influence on the processing result are qualified or not can be conveniently determined, and whether the actual processing parameters with larger influence on the self-encoder prediction are accurate or not can also be conveniently determined.

For example, if there are one or more actual processing parameters, in predicting a wrong processing result, a large difference in distribution from predicting a correct processing result, it may be determined that these actual processing parameters have a large influence on the prediction accuracy of the processing result.

For another example, if there are one or more actual processing parameters, in which the distribution difference is large in the processing results that are predicted to be acceptable versus the processing results that are predicted to be unacceptable, it may be determined whether these actual processing parameters have a large influence on the prediction of the processing results.

Thus, optionally, the method flow may further include: determining actual qualified detection results of a plurality of processing results; based on the determined actual pass detection results and the predicted results obtained by the self-encoder for the several machining results, a contribution of the actual machining parameters is determined in order to analyze the actual machining parameters associated with the reject machining results.

It can be appreciated that, in this embodiment, the contribution degree of the actual processing parameter may be determined according to the difference between the predicted qualified result and the actual qualified result, so as to facilitate analysis of the association between the actual processing parameter and whether the processing result is qualified, and in particular, the influence on the unqualified processing result.

The associated actual processing parameters can be adjusted or controlled later so as to improve the product yield.

The embodiment does not limit the mode of determining the contribution degree of the actual processing parameters, specifically, a correlation analysis algorithm can be adopted to determine the association between the actual processing parameters and unqualified processing results, and analysis can be performed by combining the indexes such as the accuracy, the precision, the recall, the F1-score and the like of the self-encoder, so that the analysis is convenient to determine whether the processing results are qualified or not, and the specific actual processing parameters with great influence on the processing results are determined.

The embodiment is not limited to a specific analysis mode and subsequent operation, and the product yield can be improved by adjusting or controlling actual processing parameters.

Further, there may be a plurality of product parameters for evaluation as to whether the processing result is acceptable. For example, the size, accuracy, color rendering, etc. of the product.

For convenience of finer analysis, when determining whether the processing result is actually acceptable or unacceptable, it may be finely divided into those that are unacceptable in terms of product parameters, i.e., which are not within a specified range or do not meet specified requirements.

And then, the method can be used for conveniently analyzing the actual processing parameters related to the specific product parameters by combining the predicted result and the actual result, namely whether each product parameter is qualified or not, and conveniently analyzing the actual processing parameters with larger influence on the specific product parameters.

Thus, optionally, the method flow may further include: determining actual qualified detection results of a plurality of processing results on the parameters of the target product; based on the determined actual pass detection results and the predicted results obtained by the self-encoder for the plurality of machining results, a contribution of the actual machining parameters to the target product parameters is determined so as to analyze the actual machining parameters associated with the machining results that are not pass by the target product parameters.

The target product parameter may specifically be any product parameter. The present embodiment is not limited. Specifically, the size, accuracy, color rendering degree, error, presence or absence of breakage of the processing result, and the like may be mentioned. It will be appreciated that the present embodiment may be performed separately for different product parameters for analysis.

The actual qualified detection result of the processing result on the target product parameter, that is, whether the processing result is qualified on the target product parameter, specifically, whether the target product parameter of the processing result is within a specified range or meets a specified requirement.

For example, if the size of the processing result is within the specified range, it may be determined that the processing result is acceptable in terms of the product parameter, which is the size. If a breakage defect exists in the processing result, it can be determined that the processing result is not qualified in the product parameter of "whether breakage exists".

And then, analyzing according to the difference between the predicted result and the actual result, and determining the influence of the actual processing parameter on the target product parameter, namely the contribution degree.

It can be appreciated that, in this embodiment, the contribution degree of the actual processing parameter to the target product parameter may be determined according to the difference between the predicted qualified result and the actual qualified result, so as to facilitate analysis of the association between the actual processing parameter and whether the target product parameter is qualified.

The embodiment does not limit the mode of determining the contribution degree of the actual processing parameter, specifically, a correlation analysis algorithm can be adopted to determine the correlation between the actual processing parameter and the target product parameter, and analysis can be performed by combining the indexes such as the accuracy, the precision, the recall, the F1-score and the like of the self-encoder, so that the analysis is convenient to determine whether the product is qualified or not, and the specific actual processing parameter with great influence is determined.

Of course, the above analysis methods may be combined with each other. Optionally, the above method flow may further include at least one of:

1) Determining actual qualified detection results of a plurality of processing results on the parameters of the target product; and determining the contribution degree of the actual processing parameters to the target product parameters based on the actual qualification detection results and the prediction results obtained by the self-encoder aiming at a plurality of processing results so as to analyze the actual processing parameters related to the processing results unqualified by the target product parameters.

2) Determining actual qualified detection results of a plurality of processing results; and determining the contribution degree of the actual processing parameters based on the actual qualified detection result and the prediction results obtained by the self-encoder for a plurality of processing results so as to analyze the actual processing parameters related to the unqualified processing results.

7. An embodiment of a display panel.

The embodiment of the invention also provides a specific embodiment of the display panel.

In the method embodiment of the product data analysis method, at least one pre-processing step and a current processing step are sequentially performed in the process of processing the display panel product; at least one pre-processing step and a current processing step may be included in the display panel product manufacturing process. The current processing step is performed on the basis of the processing results of at least one preceding processing step. The processing of the display panel product may include several processing steps. After each processing step is executed, each processing result can be obtained respectively.

The method may comprise the steps of:

S201: for any one of the at least one pre-processing step, the location of the pre-defect in the processing result for the pre-processing step is determined.

S202: and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the determined distribution condition of the pre-defect positions.

S203: and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

Optionally, according to the determined distribution situation of the pre-defect positions, determining a pre-defect high-incidence area corresponding to the pre-processing step may specifically be: determining a region where the pre-defect positions are gathered according to a preset anomaly detection algorithm based on the determined pre-defect positions; and predicting a pre-defect high-incidence area corresponding to the pre-processing step based on the determined pre-defect position gathered area.

Optionally, the method flow may further include any one of the following:

1) Determining an actual defect high-incidence area corresponding to the current processing step; and updating parameters of a preset abnormality detection algorithm to reduce the difference between the determined predicted defect high incidence area and the determined actual defect high incidence area.

2) Determining the actual defect position in the processing result of the current processing step; and updating parameters of a preset abnormality detection algorithm to increase the number of the determined actual defect positions contained in the determined predicted defect high incidence area.

Optionally, based on the determined pre-defect positions, determining a region where the pre-defect positions are gathered according to a preset anomaly detection algorithm, including at least one of:

Optionally, for any one of the at least one pre-processing step, determining a pre-defect position in a processing result of the pre-processing step, and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the determined distribution of the pre-defect positions, where the pre-defect high-incidence area may specifically be: for different pre-processing steps, respectively executing: and determining the pre-defect positions in the processing result of the pre-processing step, and determining the pre-defect high-incidence area corresponding to the pre-processing step according to the distribution condition of the determined pre-defect positions.

Optionally, the predicting the predicted defect high incidence area corresponding to the current processing step may specifically be based on the pre-defect high incidence area corresponding to the pre-processing step: and predicting the predicted defect high-incidence areas corresponding to the current processing step based on the pre-defect high-incidence areas corresponding to the different pre-processing steps.

Optionally, predicting the predicted defect high incidence area corresponding to the current processing step based on the pre-defect high incidence areas respectively corresponding to the different pre-processing steps includes any one of the following:

Optionally, the method flow may further include: determining a historical defect position in a historical processing result of the current processing step; according to the determined distribution condition of the historical defect positions, determining a high incidence area of the historical defects;

Optionally, the predicting the predicted defect high incidence area corresponding to the current processing step may specifically be based on the pre-defect high incidence area corresponding to the pre-processing step: and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step and the determined historical defect high-incidence area.

Optionally, for any one of the at least one pre-processing steps, determining the pre-defect location in the processing result of the pre-processing step for which it is intended may specifically be: predicting, for any one of the at least one pre-processing step, a failure processing result from the processing results of the pre-processing step for which it is aimed; and checking the predicted unqualified machining result to determine the position of the pre-defect.

Optionally, the method flow may further include: predicting whether any processing result of any processing step in the processing process of the display panel product is qualified.

Optionally, the prediction method for predicting whether the machining result is qualified includes: inputting the actual processing parameters of any processing result into a pre-trained self-encoder to obtain corresponding reconstruction processing parameters output by the self-encoder; and under the condition that the difference between the actual processing parameters and the corresponding reconstructed processing parameters is larger than a preset difference threshold value, predicting any processing result as a disqualified processing result.

Optionally, the training process of the self-encoder includes:

Determining an initial self-encoder;

determining actual processing parameters of qualified processing results;

Optionally, the above method flow may further include at least one of:

1) Determining actual qualified detection results of a plurality of processing results on the parameters of the target product; based on the determined actual pass detection results and the predicted results obtained by the self-encoder for the plurality of machining results, a contribution of the actual machining parameters to the target product parameters is determined so as to analyze the actual machining parameters associated with the failed machining results.

2) Determining actual qualified detection results of a plurality of processing results; based on the determined actual pass detection results and the predicted results obtained by the self-encoder for the several machining results, a contribution of the actual machining parameters is determined in order to analyze the actual machining parameters associated with the reject machining results.

A specific explanation of the present embodiment can be found in the above-described embodiments.

According to the embodiment, the data analysis can be carried out on the display panel, so that the area where the defect is easy to occur in the current processing step can be conveniently determined, the adjustment and the improvement are convenient, and the product yield is improved.

8. A method for detecting qualified products.

It is understood that the method of qualification detection for the processing result can also be applied to products.

Therefore, the embodiment of the invention also provides an embodiment of a method for detecting the qualified products.

It will be appreciated that the data analysis may be performed directly for the product processing process, and specifically, the data analysis may be performed in combination with the processing steps and the processing results in the foregoing embodiments. The above embodiments may not be combined, and the detection is mainly performed for whether the product is qualified.

Fig. 6 is a flow chart of another method for analyzing product data according to an embodiment of the invention.

The product data analysis method specifically comprises the following steps:

S301: inputting the actual processing parameters of the product to be detected into a pre-trained self-encoder to obtain corresponding reconstruction processing parameters output by the self-encoder.

S302: and under the condition that the difference between the actual processing parameter and the corresponding reconstruction processing parameter is larger than a preset difference threshold, predicting the product to be detected as a disqualified product.

The product to be detected may be any product, and the embodiment is not particularly limited.

The present embodiment is not limited to a specific product type, and may be specifically a display panel, a toy, a cup, a screen, or the like.

It will be appreciated that this embodiment is illustrative of analysis for a single product of a single type, and that analysis for multiple products of multiple types may be performed using this embodiment and the following explanation, respectively.

Optionally, the training process of the self-encoder may specifically include:

Determining an initial self-encoder;

Determining actual processing parameters of qualified products;

Optionally, the method flow may further include at least one of the following:

1) Inputting the actual processing parameters of a plurality of unqualified products and the actual processing parameters of a plurality of qualified products into a pre-trained self-encoder to obtain hidden layer characteristics; the differences between the hidden features of the acceptable product and the hidden features of the unacceptable product are determined to facilitate analysis of actual processing parameters associated with the unacceptable product.

2) Inputting the actual processing parameters of a plurality of unqualified products and the actual processing parameters of a plurality of qualified products into a pre-trained self-encoder to obtain hidden layer characteristics; determining hidden layer nodes meeting preset screening conditions; the preset screening conditions may include: aiming at qualified products and unqualified products, the difference between the output hidden layer characteristics of the hidden layer nodes is larger than a preset difference threshold; weight parameters between the actual process parameters and the determined hidden layer nodes are determined to facilitate analysis of the actual process parameters associated with the rejected product.

Optionally, the method flow may further include at least one of the following:

1) Determining actual qualified detection results of a plurality of products on target product parameters; based on the determined actual pass detection results and the predicted results obtained by the self-encoder for the plurality of products, a contribution of the actual processing parameters to the target product parameters is determined so as to analyze the actual processing parameters associated with the products whose target product parameters are not pass.

2) Determining actual qualified detection results of a plurality of products; based on the determined actual pass detection results and the predicted results obtained by the self-encoders for the several products, a contribution of the actual processing parameters is determined in order to analyze the actual processing parameters associated with the reject product.

The method and the device can automatically predict the qualification of the product, and can improve the comprehensiveness and efficiency of the qualification prediction of the product.

Of course, it is understood that the product in this embodiment may also be an intermediate product produced during processing, that is, the processing result in the above embodiment.

9. Application example.

For ease of understanding, the embodiments of the present invention also provide application embodiments.

Application embodiment one.

In the display panel industry, the complete production process is very complex, often involving tens or even hundreds of processes.

In addition to the quality inspection station of the final product, after some important procedures are completed, a quality inspection station is also set to inspect the product at the stage, find defects in time and rework or repair.

Most of the quality inspection sites in the earlier process are bad in Map overlapping of manual identification, labor is consumed, the situation that defect aggregation is not obvious cannot be effectively identified and intercepted, bad identification in the later process is mainly judged based on personnel experience, timeliness and accuracy of data analysis cannot be guaranteed, certain leakage exists, and backtracking is difficult.

The embodiment provides an anomaly detection and position location method based on product bad aggregation analysis in an intelligent manufacturing scene.

During the production and manufacturing process, each manufacturing site records the quality inspection result data and synchronizes the quality inspection result data to the Cl ickHouse database in real time.

And the data warehouse regularly fuses the site data with the front flow according to the product ID, the distribution area and the defect level.

The algorithm service system performs abnormal aggregation analysis on the fusion data, and predicts a high incidence defective area which still exists in the subsequent flow site, namely a high incidence defective area.

And (3) confirming an abnormal threshold value of algorithm parameters by adopting an abnormal detection algorithm, and periodically synchronizing the abnormal threshold value into a Cl ickHouse database.

The subsequent production station periodically acquires the latest production process anomaly information, which may be quality inspection result data, from Cl ickHouse. Before the subsequent production site, the method assists the factory personnel to analyze the high-incidence bad area of the product which is about to pass through the subsequent production site; if the information is abnormal, synchronizing the information to Cl ickHouse in real time. The determination of abnormality may be specifically a manual rescreening of the high-occurrence defective region.

When each quality inspection station processes a new batch of products, abnormal information of the batch of products is obtained from Cl ickHouse, specifically, the abnormal information can comprise a high-incidence bad area, and an abnormal detection strategy can be dynamically adjusted according to the information, specifically, the quality inspection strategy can be adjusted.

By applying highly automated processing and machine learning algorithms to the data, potential anomalies can be identified early and problem location information located.

The overall scheme is described first.

As shown in fig. 7, fig. 7 is a schematic diagram illustrating a product data analysis according to an embodiment of the present invention.

In the product manufacturing process, each production site records the production process data and synchronizes the production process data to a database in real time.

Each quality inspection site records its quality inspection result data and synchronizes to Kafka in real time.

Wherein the production process data includes, but is not limited to: product information, equipment parameters, important time node information, product batch information, product model information and the like.

Quality inspection result data includes, but is not limited to: product information, whether or not a defect has occurred, type of defect, grade of defect, defect code, defect occurrence area, defect occurrence coordinates, and the like. The defect here may be a defect in particular.

In addition, the overall system may also include a production database, a big data computing platform, and an algorithm system.

The data warehouse in the big data computing platform can regularly fuse the historical process data with the quality inspection data according to the product ID.

The algorithm service system can conduct abnormal aggregation analysis on the fusion data, and predicts high-incidence bad areas still existing in subsequent flow sites. And (3) confirming an abnormal threshold value of algorithm parameters by adopting an abnormal detection algorithm, and periodically synchronizing the abnormal threshold value into a Cl ickHouse database. The Cl ickHouse database may be in the production database.

The subsequent production site periodically acquires the latest abnormal information of the production process from Cl ickHouse, and before the subsequent production site, the method assists factory personnel in decision-making and analyzing the high-incidence defective area of the product to be passed through the site; if the information is abnormal, synchronizing the information to Cl ickHouse in real time.

And when each quality inspection station processes a new batch of products, acquiring abnormal information of the batch of products from Cl ickHouse, and dynamically adjusting an abnormal detection strategy according to the information. Through the application of highly automated processing and machine learning algorithms to the data, potential anomalies are identified early, problem location information is located, and losses are recovered.

With respect to big data computing platforms.

The method can be used for constructing a big data computing platform and analyzing exploratory data. And cleaning station data and fusing quality inspection data to calculate bad information of the single product, wherein the bad information comprises defect areas, defect coordinates, defect grades, defect codes and the like. According to factory screening rules, products are classified and calculated according to batches, models, IDs, defect types, defect areas and defect quantity, and processed data are sent to a subsequent algorithm module.

With respect to anomaly detection algorithms.

And determining a potential high-incidence defective area by adopting an anomaly detection algorithm based on each related parameter taking the product as a unit and a data set of the detection result. Anomaly detection algorithms include, but are not limited to:

2) For each relevant parameter, a self-codec model is constructed. And respectively constructing a four-layer encoder and a two-layer decoder, wherein the activation function is Re l u, the defined loss function is MSE (Mean Square Error), and the learning rate alpha of the optimizer is set to be 0.001. Based on the comparison of the reconstructed data and the high incidence bad region in the subsequent flow, calculating the accuracy of the model, and judging whether the model is abnormal or not according to the model.

Regardless of the manner, the recall and accuracy are calculated based on the actual result label of the dataset, thereby measuring the accuracy of the model. The recall rate and accuracy calculation formula is:

Recall = number of samples of correct judgment bad region/number of all samples of true bad region X100%.

Accuracy = number of samples of correctly judged bad areas/number of all samples of judged bad areas X100%.

With respect to alert services.

Each station obtains the latest product alarm data information from Cl ickHouse each time a new batch of products is processed. With the regional information in the alarm data as assistance and in combination with business experience, factory personnel can efficiently judge potential high-incidence bad regions. By means of the advanced judging processing mode, the high probability of poor occurrence in the subsequent site flow is reduced.

In the present embodiment, at least the following effects can be achieved.

1. The big data computing platform successfully implements an anomaly detection algorithm, monitors the system performance in real time, and immediately alarms online when a problem is found, so that an efficient closed-loop management system is formed.

2. According to actual business requirements of a factory, a plurality of abnormality detection methods such as machine learning, a rule engine and a statistical model are combined, an automatic defect Map centralized recognition and high-incidence bad abnormality recognition algorithm is developed for the factory, a matched alarm service is provided, and the method has strong interpretability and ensures timeliness and accuracy.

Application example two.

Technical problems: in the production process of manufacturing factories, such as OLED screen production, the related procedures of products are complex, the related process parameter dimension is high, and the bad data of the products is small in proportion. How to analyze the production data with high dimension and extremely unbalanced positive and negative sample distribution, detect abnormal conditions, mine the cause of the abnormality, and have important significance for improving the production yield and shortening the analysis time of the cause of the abnormality.

Technical benefit: the yield analysis method based on the self-encoder can help analyze relevant factors of product abnormality, so that the yield of the product is improved.

The embodiment relates to a yield analysis method based on a self-encoder.

Firstly, collecting detailed data of a production process, including various sensor data, technological parameters, production conditions and the like, and information of good products and defective products, and arranging data to construct a data sample.

And secondly, constructing a self-encoder network according to the data samples, and performing model training and verification.

Then, the coding layer characteristics extracted from the encoder are analyzed for characteristic correlation and the like, and data points inconsistent with the normal production process are identified according to the reconstruction errors.

Finally, visual explanation and interpretation of the analysis result are carried out.

The embodiment can help identify abnormal states occurring in the production process and determine key factors causing low yield in the production process of products.

1. Data preparation:

Firstly, collecting detailed data of OLED production process, including various sensor data, technological parameters, production process data and the like, and meanwhile collecting labeling information of good products and defective products, and finishing a data set required by a forming algorithm.

The OLED production process comprises a plurality of working procedures, and after the staged working procedures are all completed, the staged quality inspection is carried out by quality inspection equipment, so that parameter evaluation results with multiple dimensions are obtained.

In this embodiment, taking the production of the BP process section as an example, the quality inspection stage mainly involves measurement of the electrical related parameters. Specific quality inspection cases are shown in table 1 below.

Table 1 quality inspection Condition Table

Quality inspection	Quality inspection results	Specification of specification
			Parameter 1	2.0	[1.0，5.0]
Parameter 2	150	[120,180]
			…	…	…

The sensor data mainly includes: temperature, humidity, pressure value, gas concentration, etc

Other process and production parameters include: evaporation rate, flow rate, water resistance, number of washes, waiting time, treatment time, etc.

For the production stage, a glass substrate may flow through multiple production facilities, such as evaporation facilities requiring high temperatures, cleaning facilities requiring multiple cooling and rinsing, and the like, where some facilities may have multiple chambers, each facility may have multiple process parameter settings and may be capable of monitoring multiple environmental parameters, so that thousands of process related features may be collected during the production process.

After the quality inspection is finished, powering on the screen, measuring the results of various key parameters, and marking whether the quality inspection of each dimension parameter passes or not according to the industry standard specification. The same screen may pass all the quality checks, or there may be cases where individual parameter items fail quality checks.

One piece of glass is a sample unit and the data is shown in table 2 below:

TABLE 2 sample characteristics and quality control results table

2. Self-encoder training.

In general, for abnormality detection of production, it is common to determine from quality inspection measurement results at a quality inspection stage after a production process. This way is reliable but with a certain hysteresis.

In the production process, the abnormal occurrence ratio is very small, and in this case, the self-encoder can be used for detecting the abnormality, so that the abnormal condition can be found in time in the production process, and the abnormal product can be alarmed and processed in advance. The self-encoder can analyze all sub-processes, and can analyze some sub-processes in a targeted manner.

The self-encoder is a network model with unsupervised properties, which can map the input high-dimensional data to a low-dimensional space, obtain hidden features of the high-dimensional data, obtain a compressed representation of the data, and then complete data reconstruction.

In anomaly detection, the anomaly is much smaller than the normal, so it is considered that the original input data is anomalous if the difference between the output reconstructed from the encoder and the original input exceeds a certain threshold.

Typically, a basic self-encoder network has a three-layer structure, as shown in fig. 8, and fig. 8 is a schematic diagram of another self-encoder according to an embodiment of the present invention.

The method comprises the following steps: the three layers of the input layer, the hidden layer and the reconstruction layer realize two important conversion, namely coding and decoding conversion.

The self-encoder can be seen in particular as consisting of two concatenated networks. The first network is an encoder, which is responsible for receiving an input x and transforming the input into a signal y by a function h:

y＝h(x)

The second network takes as its input the encoded signal y, and obtains the reconstructed output by a function f:

x′＝f(y)＝f(h(x))

The encoder converts the large-scale data set into input vectors with different dimensions, then the input vectors are recombined into output vectors through the decoder, and a reconstruction error between the input vectors and the output vectors is calculated, wherein the calculation mode of the reconstruction error is as follows:

e＝∑|x′-x|²

an appropriate threshold is selected to determine anomalies based on the reconstruction errors of the training data. The determination method can adopt a statistical method, such as adding or subtracting 3 times standard deviation from the mean value.

In the actual process, the number of layers of the encoder and the decoder can be properly increased according to the dimension of the original feature, so that more proper compression feature representation is obtained.

3. And (5) extracting characteristics.

Coding layer features extracted from the encoder, which are compressed representations of the input data. These features may cover critical information in the production process. Therefore, more dimensionality analysis can be performed on hidden layer features, and model optimization and anomaly cause analysis are facilitated. Including but not limited to the following:

(1) Statistical analysis: and carrying out statistical analysis on the layer hidden characteristics, such as mean value, variance, skewness, kurtosis and the like, and knowing the information of the dispersion degree, symmetry and the like of the characteristics.

(2) Using dimension reduction techniques, such as PCA or t_sne, etc., the hidden layer features are visualized, observing the distribution of normal and abnormal data in the hidden layer representation.

(3) Using a clustering algorithm, such as K-means, to cluster hidden layer features into different clusters, the outlier data may form independent clusters.

(4) And (3) back-pushing analysis: and analyzing weight parameters of the self-encoder, and analyzing which input features have great weight influence on hidden layer features, namely generating the contribution to be the largest. Data anomalies may have strong correlations with these more contributing features.

(5) Comparison analysis: the data and the anomaly data are input from the encoder, respectively, and their differences in the normal hidden layer representation are compared.

4. Abnormality detection

And sending the data to be tested into the trained self-encoder model to obtain reconstruction output, and calculating a reconstruction error. If the reconstruction error is greater than the set threshold, the data is considered to be anomalous.

And labeling positive and negative samples of each parameter for the actual measurement result of the quality inspection stage, and generating an actual result of each quality inspection parameter.

By comparing the result of the self-encoder with the actual result of each parameter, indexes such as accuracy, recall, F1 score and the like of the model judging result and the actual result of the quality inspection parameter are calculated according to the measurement rule of the classification model.

If the values of the accuracy, the recall and the F1 are closer to 1, the extracted features are considered to have stronger correlation with the current quality inspection parameters. If the value is close to 0, the extracted features are considered to be uncorrelated with the quality inspection parameters. The features extracted here may in particular be the actual processing parameters of the product.

True category	Predicted to be positive	Predicted as inverse
			Positive example	Real example TP	False counter example FN
Counterexample	False positive FP	True counterexample TN

The prediction may be a prediction of pass, and the positive example may be an actual pass product. The prediction may be a prediction of disqualification, and the counterexample may be an actual disqualification product. The true cases may be products that are predicted to be acceptable and predicted to be correct, and the TP may be the number of true cases. The false positive may be a product that is predicted to be acceptable and mispredicted, and FP may be the number of false positive. The false counter may be a product that is predicted to be unacceptable and that is predicted to be incorrect, and FN may be the number of false counter. The true counter may be the product that is predicted to be unacceptable and the correct, and TN may be the number of true counter.

Accuracy accuracy = (tp+tn)/(tp+fn+fp+tn), which indicates that the number of correctly classified samples is the total number of samples.

Accuracy precis ion =tp/(tp+fp), and the predicted correct positive example data is a proportion of the predicted positive example data.

Recall reca l l =tp/(tp+fn), which represents the proportion of "predicted correct positive example data TP" to all "actual positive examples (tp+fn").

F1＝(precis ion*reca l l*2)/(precesion+reca l l)。

The closer the values of the accuracy rate, the recall rate and F1 are to 1, the more accurate the prediction can be determined, so that the larger the correlation between the actual processing parameters of the current batch and the current quality inspection parameters can be analyzed, and the actual processing parameter distribution condition in the current batch of products can be specifically analyzed.

Specifically, as shown in fig. 9, fig. 9 is a schematic diagram illustrating an abnormality detection principle for determining correlation of quality inspection parameters according to an embodiment of the present invention.

The correlation between the actual processing parameter and the current quality inspection parameter can be judged according to the prediction result and the quality inspection result of the self-encoder.

Specifically, the method may be to compare the input characteristic and the reconstruction characteristic of the self-encoder, that is, the input actual processing parameter and the reconstructed processing parameter, determine a reconstruction error, and determine whether the prediction is qualified according to whether the reconstruction error is greater than a preset error threshold. If the reconstruction error is smaller than a preset error threshold, determining that the predicted product is qualified; if the reconstruction error is greater than a preset error threshold, it may be determined that the predicted product is unacceptable.

Accordingly, matching and comparison can be performed in quality inspection parameter specifications according to quality inspection results. Specifically, the quality is checked to obtain the size of the product, and whether the actual size detected by the quality is within the size specification range is further judged. If the actual size of the quality inspection is within the size specification range, the actual product can be determined to be qualified in size. If the actual size of the quality inspection is not within the size specification range, it can be determined that the actual product is not qualified in size.

And then, according to the difference between the predicted result and the actual result, judging the correlation between the actual processing parameter and the current quality inspection parameter.

5. Visual interpretation.

The analysis associated in step 3 is presented in the form of a visual component.

This embodiment can achieve at least the following effects.

1) According to the yield analysis method based on the self-encoder, key features are extracted from high-dimensional production process data through the self-encoder network, and data reconstruction is carried out, so that key influence factors causing abnormality and bad are analyzed.

2) And determining the correlation between the current relevant characteristic data and a certain bad through the coincidence degree of the abnormal detection marking result and the actual sample.

The various technical features of the above embodiments may be arbitrarily combined as long as there is no conflict or contradiction between the features, but are not described in detail, and therefore, the arbitrary combination of the various technical features of the above embodiments is also within the scope of the present disclosure.

The embodiment of the invention also provides a device embodiment corresponding to the embodiment of the method.

As shown in fig. 10, fig. 10 is a schematic structural view of a product data analysis device according to an embodiment of the present invention.

In the process of processing any type of product, at least one pre-processing step and the current processing step are sequentially carried out; at least one pre-processing step and a current processing step are included in the product processing process; the current processing step is performed on the basis of the processing results of at least one preceding processing step.

The apparatus may include the following units.

A position unit 401 for: for any one of the at least one pre-processing step, the location of the pre-defect in the processing result for the pre-processing step is determined.

A summarizing unit 402, configured to: and determining a pre-defect high-incidence area corresponding to the pre-processing step according to the determined distribution condition of the pre-defect positions.

A prediction unit 403, configured to: and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step.

Optionally, the summarizing unit 402 is configured to:

Determining a region where the pre-defect positions are gathered according to a preset anomaly detection algorithm based on the determined pre-defect positions;

and predicting a pre-defect high-incidence area corresponding to the pre-processing step based on the determined pre-defect position gathered area.

Optionally, the summarizing unit 402 is further configured to perform any one of the following:

Determining an actual defect high-incidence area corresponding to the current processing step; updating parameters of a preset anomaly detection algorithm to reduce the difference between the determined predicted defect high incidence area and the determined actual defect high incidence area;

Determining the actual defect position in the processing result of the current processing step; and updating parameters of a preset abnormality detection algorithm to increase the number of the determined actual defect positions contained in the determined predicted defect high incidence area.

Optionally, the summarizing unit 402 is configured to perform at least one of:

Clustering is carried out according to a preset clustering algorithm aiming at the determined pre-defect position, so as to obtain a clustering result; determining an area where the pre-defect positions are gathered based on clusters, wherein the number of the clusters including the pre-defect positions meets the preset number condition, in the clustering result;

inputting the determined pre-defect position into a pre-trained self-encoder to obtain a corresponding reconstruction position output by the self-encoder; determining a region where the pre-defect positions are aggregated based on the pre-defect positions having a distance to the corresponding reconstructed position less than a preset distance threshold;

Determining the position of the outlier pre-defect according to an isolated forest algorithm aiming at the determined position of the pre-defect; the area in which the pre-defect locations are clustered is determined based on other pre-defect locations than the outlier pre-defect locations.

Optionally, the location unit 401 is configured to:

the summarizing unit 402 is configured to:

The prediction unit 403 is configured to:

And predicting the predicted defect high-incidence areas corresponding to the current processing step based on the pre-defect high-incidence areas corresponding to the different pre-processing steps.

Optionally, the prediction unit 403 is configured to perform any one of the following:

Aiming at the preposed defect high-incidence areas corresponding to different preposed processing steps respectively, determining a union area, and predicting a predicted defect high-incidence area corresponding to the current processing step based on the union area;

and determining an intersection area aiming at the pre-defect high-incidence areas corresponding to different pre-processing steps respectively, and predicting the predicted defect high-incidence area corresponding to the current processing step based on the intersection area.

Optionally, the summarizing unit 402 is further configured to:

determining a historical defect position in a historical processing result of the current processing step; according to the determined distribution condition of the historical defect positions, determining a high incidence area of the historical defects;

The prediction unit 403 is configured to:

and predicting the predicted defect high-incidence area corresponding to the current processing step based on the pre-defect high-incidence area corresponding to the pre-processing step and the determined historical defect high-incidence area.

Optionally, the summarizing unit 402 is configured to:

predicting, for any one of the at least one pre-processing step, a failure processing result from the processing results of the pre-processing step for which it is aimed;

and checking the predicted unqualified machining result to determine the position of the pre-defect.

Optionally, the apparatus further comprises:

And the qualification unit 404 is used for predicting whether any processing result of any processing step in the product processing process is qualified.

Optionally, the prediction method for predicting whether the machining result is qualified includes:

Inputting the actual processing parameters of any processing result into a pre-trained self-encoder to obtain corresponding reconstruction processing parameters output by the self-encoder;

Optionally, the training process of the self-encoder includes:

Determining an initial self-encoder;

determining actual processing parameters of qualified processing results;

Optionally, the apparatus further comprises an analysis unit 405.

Optionally, the analysis unit 405 is configured to perform at least one of the following:

Optionally, the analysis unit 405 is configured to perform at least one of:

1) Determining actual qualified detection results of a plurality of processing results on the parameters of the target product; based on the determined actual pass detection results and the predicted results obtained by the self-encoder for the plurality of machining results, a contribution of the actual machining parameters to the target product parameters is determined so as to analyze the actual machining parameters associated with the machining results that are not pass by the target product parameters.

Specific explanation can be found in the method examples described above.

The embodiment of the invention also provides another embodiment of the product data analysis device.

The device comprises:

The reconstruction unit 501 is configured to: inputting the actual processing parameters of the product to be detected into a pre-trained self-encoder to obtain corresponding reconstruction processing parameters output by the self-encoder.

The difference unit 502 is configured to: and under the condition that the difference between the actual processing parameter and the corresponding reconstruction processing parameter is larger than a preset difference threshold, predicting the product to be detected as a disqualified product.

Specific explanation can be found in the method examples described above.

The embodiment of the invention also provides computer equipment, which at least comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the program to realize any method embodiment.

The embodiment of the invention also provides electronic equipment, which comprises: at least one processor; and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the one processor to enable the at least one processor to perform any one of the method embodiments described above.

FIG. 11 is a schematic diagram of a hardware architecture of a computer device for configuring a method according to an embodiment of the present invention, where the device may include: a processor 1010, a memory 1020, an input/output interface 1030, a communication interface 1040, and a bus 1050. Wherein processor 1010, memory 1020, input/output interface 1030, and communication interface 1040 implement communication connections therebetween within the device via a bus 1050.

The processor 1010 may be implemented by a general-purpose CPU (Centra l Process ing Unit ), a microprocessor, an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits, etc. for executing related programs to implement the technical solutions provided by the embodiments of the present invention.

Memory 1020 may be implemented in the form of ROM (Read On ly Memory ), RAM (Random Access Memory, random access memory), static storage, dynamic storage, and the like. Memory 1020 may store an operating system and other application programs, and when the embodiments of the present invention are implemented in software or firmware, the associated program code is stored in memory 1020 and executed by processor 1010.

The input/output interface 1030 is used to connect with an input/output module for inputting and outputting information. The input/output module may be configured as a component in a device (not shown) or may be external to the device to provide corresponding functionality. Wherein the input devices may include a keyboard, mouse, touch screen, microphone, various types of sensors, etc., and the output devices may include a display, speaker, vibrator, indicator lights, etc.

Communication interface 1040 is used to connect communication modules (not shown) to enable communication interactions of the present device with other devices. The communication module may implement communication through wired mode (such as USB, network cable, etc.), or may implement communication through wireless mode (such as mobile network, WI FI, bluetooth, etc.).

Bus 1050 includes a path for transferring information between components of the device (e.g., processor 1010, memory 1020, input/output interface 1030, and communication interface 1040).

It should be noted that although the above-described device only shows processor 1010, memory 1020, input/output interface 1030, communication interface 1040, and bus 1050, in an implementation, the device may include other components necessary to achieve proper operation. Furthermore, it will be understood by those skilled in the art that the above-described apparatus may include only the components necessary for implementing the embodiments of the present invention, and not all the components shown in the drawings.

The present invention also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor implements any of the method embodiments described above.

The present invention also provides a computer readable storage medium storing a computer program which, when executed by a processor, implements any of the method embodiments described above.

Computer readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of storage media for a computer include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Disks (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium which can be used to store information that can be accessed by a computing device. Computer-readable media, as defined herein, does not include transitory computer-readable media (trans itory media), such as modulated data signals and carrier waves.

Embodiments of the present invention also provide a computer program product comprising a computer program/instruction which, when executed by a processor, implements any of the method embodiments described above.

From the foregoing description of the embodiments, it will be apparent to those skilled in the art that embodiments of the present invention may be implemented in software plus a necessary general purpose hardware platform. Based on such understanding, the technical solution of the embodiments of the present invention may be embodied essentially or in contributing parts in the form of a software product, which may be stored in a storage medium, such as a ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the method described in the embodiments or some parts of the embodiments of the present invention.

The system, apparatus, module or unit set forth in the above embodiments may be implemented in particular by a computer chip or entity, or by a product having a certain function. A typical implementation device is a computer, which may be in the form of a personal computer, laptop computer, cellular telephone, camera phone, smart phone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or a combination of any of these devices.

In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for the device embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, and reference is made to the description of the method embodiments for relevant points. The apparatus embodiments described above are merely illustrative, in which the modules illustrated as separate components may or may not be physically separate, and the functions of the modules may be implemented in the same piece or pieces of software and/or hardware when implementing embodiments of the present invention. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

The foregoing is merely illustrative of the principles of this invention and it will be appreciated by those skilled in the art that numerous modifications and variations could be made without departing from the principles of this invention.

In the present invention, the terms "first", "second" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance. The term "plurality" refers to two or more, unless explicitly defined otherwise.

Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

It is to be understood that the invention is not limited to the precise arrangements and instrumentalities shown in the drawings, which have been described above, and that various modifications and changes may be effected without departing from the scope thereof. The scope of the invention is limited only by the appended claims.

Claims

1. A product data analysis method, characterized in that the product processing process includes at least one pre-processing step and a current processing step; the current processing step is executed based on the processing result of the at least one pre-processing step;

The method comprises:

For any one of the at least one pre-processing step, determining a pre-defect position in a processing result of the pre-processing step;

Determine a high-incidence area of pre-defects corresponding to the pre-processing step according to the distribution of the pre-defect positions;

Based on the previous defect high-incidence area corresponding to the previous processing step, the predicted defect high-incidence area corresponding to the current processing step is predicted.

2. The method according to claim 1, characterized in that the step of determining the high-incidence area of the pre-defects corresponding to the pre-processing step according to the distribution of the pre-defect positions comprises:

Based on the preceding defect positions, and according to a preset anomaly detection algorithm, determining an area where the preceding defect positions are concentrated;

Based on the area where the positions of the pre-defects are concentrated, a high-incidence area of pre-defects corresponding to the pre-processing step is predicted.

3. The method according to claim 2, further comprising any one of the following:

Determining an actual high-defect incidence area corresponding to the current processing step; updating parameters of the preset anomaly detection algorithm to reduce the difference between the predicted high-defect incidence area and the actual high-defect incidence area;

Determine the actual defect location in the processing result of the current processing step; update the parameters of the preset anomaly detection algorithm to increase the number of the actual defect locations included in the predicted defect high-incidence area.

4. The method according to claim 2, characterized in that the step of determining the area where the preceding defect positions are concentrated based on the preceding defect positions and according to a preset anomaly detection algorithm comprises at least one of the following:

For the front defect positions, clustering is performed according to a preset clustering algorithm to obtain a clustering result; based on the clusters in which the number of front defect positions in the clustering result meets a preset number condition, an area where the front defect positions are clustered is determined;

Inputting the preceding defect position into a pre-trained autoencoder to obtain a corresponding reconstructed position output by the autoencoder; determining an area where the preceding defect positions are concentrated based on preceding defect positions whose distance from the corresponding reconstructed position is less than a preset distance threshold;

For the pre-defect positions, outlier pre-defect positions are determined according to an isolation forest algorithm; and based on other pre-defect positions other than the outlier pre-defect positions, an area where the pre-defect positions are concentrated is determined.

5. The method according to claim 1, characterized in that, for any pre-processing step of the at least one pre-processing step, determining the pre-defect position in the processing result of the pre-processing step, and determining the pre-defect high-incidence area corresponding to the pre-processing step according to the distribution of the pre-defect position, comprises:

For different pre-processing steps, respectively: determining the pre-defect position in the processing result of the pre-processing step, and determining the pre-defect high-incidence area corresponding to the pre-processing step according to the distribution of the pre-defect position;

The predicting the predicted defect high-incidence area corresponding to the current processing step based on the previous defect high-incidence area corresponding to the previous processing step includes:

Based on the preceding defect high-incidence areas corresponding to the different preceding processing steps, the predicted defect high-incidence area corresponding to the current processing step is predicted.

6. The method according to claim 5, characterized in that the predicting defect high-incidence area corresponding to the current processing step based on the previous defect high-incidence areas corresponding to the different previous processing steps respectively comprises any one of the following:

Determine a union region for the pre-defect high-incidence regions corresponding to the different pre-processing steps, and predict the predicted defect high-incidence region corresponding to the current processing step based on the union region;

An intersection region is determined for the pre-processing high-defect regions corresponding to the different pre-processing steps, and based on the intersection region, a predicted defect high-incidence region corresponding to the current processing step is predicted.

7. The method according to claim 1, further comprising:

Determine the historical defect positions in the historical processing results of the current processing step; determine the high-incidence area of historical defects according to the distribution of the historical defect positions;

Based on the previous defect high-incidence area corresponding to the previous processing step and the historical defect high-incidence area, the predicted defect high-incidence area corresponding to the current processing step is predicted.

8. The method according to claim 1, characterized in that, for any one of the at least one pre-processing step, determining the position of the pre-processing defect in the processing result of the pre-processing step comprises:

For any one of the at least one pre-processing step, predicting an unqualified processing result from the processing results of the pre-processing step;

The predicted unqualified processing results are inspected to determine the location of the preceding defects.

9. The method according to claim 1, further comprising:

Predict whether any processing result of any processing step in the product processing process is qualified.

10. The method according to claim 8 or 9, characterized in that the prediction method for predicting whether the processing result is qualified comprises:

Inputting the actual processing parameters of any processing result into a pre-trained autoencoder to obtain the corresponding reconstructed processing parameters output by the autoencoder;

When it is determined that the difference between the actual processing parameter and the corresponding reconstructed processing parameter is greater than a preset difference threshold, any processing result is predicted as an unqualified processing result.

11. The method according to claim 10, characterized in that the training process of the autoencoder comprises:

Determine the initial autoencoder;

Determine the actual processing parameters for acceptable processing results;

The following steps are executed repeatedly until the preset training stop condition is met:

Inputting the determined actual processing parameters into the current initial autoencoder to obtain the corresponding reconstruction processing parameters output by the current initial autoencoder;

Updating the current initial autoencoder to reduce the difference between the input actual processing parameters and the corresponding reconstructed processing parameters, and re-determining the updated initial autoencoder as the current initial autoencoder;

After the cycle ends, the current initial autoencoder is determined as the trained autoencoder.

12. The method according to claim 10, further comprising at least one of the following:

Inputting actual processing parameters of a number of unqualified processing results and actual processing parameters of a number of qualified processing results into a pre-trained autoencoder to obtain hidden features; determining the difference between the hidden features of the qualified processing results and the hidden features of the unqualified processing results to facilitate analysis of the actual processing parameters associated with the unqualified processing results;

The actual processing parameters of a number of unqualified processing results and the actual processing parameters of a number of qualified processing results are input into a pre-trained autoencoder to obtain hidden layer features; hidden layer nodes that meet preset screening conditions are determined; the preset screening conditions include: the difference between the hidden layer features output by the hidden layer nodes for the qualified processing results and the unqualified processing results is greater than a preset difference threshold; and weight parameters between the actual processing parameters and the determined hidden layer nodes are determined to facilitate analysis of the actual processing parameters associated with the unqualified processing results.

13. The method according to claim 10, further comprising at least one of the following:

Determine actual qualified test results of a plurality of processing results on target product parameters; determine the contribution of actual processing parameters to the target product parameters based on the actual qualified test results and the prediction results obtained by the autoencoder for the plurality of processing results, so as to analyze the actual processing parameters associated with the processing results that fail to meet the target product parameters;

Determine actual qualified test results of several processing results; based on the actual qualified test results and the prediction results obtained by the autoencoder for several processing results, determine the contribution of actual processing parameters to facilitate analysis of actual processing parameters associated with unqualified processing results.

14. A product data analysis device, characterized in that the product processing process includes at least one pre-processing step and a current processing step; the current processing step is executed based on the processing result of the at least one pre-processing step;

The device comprises:

A location unit, used to: determine, for any one of the at least one pre-processing step, a location of a pre-defect in a processing result of the pre-processing step;

A summarizing unit, used to: determine a high-incidence area of pre-defects corresponding to a pre-processing step according to the distribution of the pre-defect positions;

The prediction unit is used to predict the predicted defect high-incidence area corresponding to the current processing step based on the previous defect high-incidence area corresponding to the previous processing step.

15. An electronic device, comprising:

at least one processor; and,

a memory communicatively connected to the at least one processor; wherein,

The memory stores instructions executable by the one processor, and the instructions are executed by the at least one processor so as to enable the at least one processor to perform the method according to any one of claims 1 to 13.

16. A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to any one of claims 1 to 13 when executed by a processor.

17. A computer program product, comprising a computer program/instruction, wherein the computer program/instruction implements the steps of the method according to any one of claims 1 to 13 when executed by a processor.