CN112308298B - A multi-scenario performance index prediction method and system for semiconductor production lines - Google Patents

Abstract

The invention relates to a multi-scenario performance index prediction method and a multi-scenario performance index prediction system for a semiconductor production line, wherein the method comprises the following steps: the production scene quantitative division module is driven by data, quantitatively maps the product value, the average product processing period and the utilization rate of each device of the production line, and divides the production line into three scenes of light load, normal load and heavy load; the main prediction network construction module is used for combining a deep neural network algorithm and semiconductor production line performance prediction by taking the divided normal load data as sample data to construct a prediction network under a normal load scene; and the multi-scene prediction model building module is used for applying the idea of transfer learning to production line prediction and building networks under light load and heavy load scenes according to the built main prediction network so as to form the multi-scene prediction model. Compared with the prior art, the method quantitatively divides the production line scene, can more accurately predict the performance indexes of a plurality of production lines under different load scenes, and can be used on line.

Description

A multi-scenario performance index prediction method and system for semiconductor production lines

technical field

The invention relates to the technical field of intelligent manufacturing, in particular to a multi-scenario performance index prediction method and system for a semiconductor production line.

Background technique

Manufacturing is the main body of the national economy, the engine of rapid economic growth, and an important manifestation of comprehensive national strength. In order to ensure competitiveness, manufacturing enterprises need to quickly carry out dynamic scheduling management and rationally allocate resources. Real-time performance index prediction results can provide a basis for production decision-making and evaluation, so it is an inevitable requirement to quickly understand the performance index achieved by the production line in a certain scheduling mode.

At present, there are three main methods of predictive model modeling: mathematical modeling, simulation modeling, and machine learning modeling. First, semiconductor manufacturing is a complex process with high nonlinearity, multi-variable coupling, and random uncertainty, and the establishment of mathematical models is extremely difficult; second, traditional simulation modeling methods require a lot of investment and take a long time to predict , the ability to adapt to the dynamic environment is poor. In order to overcome the limitations of mathematical models and traditional simulation prediction models, experts and scholars have introduced machine learning technology to study data-driven production line prediction methods.

The Chinese patent "Performance Prediction Method for Dynamic Scheduling of Semiconductor Production Lines" (Patent Publication No.: CN103310285 A) invents a semiconductor production line scheduling system based on extreme learning machines, which collects historical data of semiconductor production lines to establish sample sets and test sample sets ; Use the extreme learning machine method to build a prediction model; use test samples to test the network performance of the prediction model, and output the prediction results after normalization. The Chinese patent "An LR-based Production Line Spare Part Damage Rate Prediction System" (Patent Publication No.: CN 108898254 A) invented a production line spare part damage rate prediction system. The method collects the operation records of the equipment through sensors, hourly current, voltage Average value, running time and other data; manually obtain the time from the installation of spare parts to replacement; obtain the model through LR training; predict the current loss rate of the corresponding category of equipment through the model.

It is not difficult to see that most of the existing method inventions are only aimed at a single production scenario, and there are many uncertain dynamic production factors in the semiconductor production line, such as changes in material feeding, changes in production line load, etc. The current identification and prediction methods are difficult to adapt to the dynamic production environment. , and most production line performance forecasts focus on a single performance indicator such as productivity, while there are fewer multi-performance indicators forecasting equipment utilization and productivity. To this end, a timely performance index prediction method that can adapt to a variety of production line states is required, including scene division, the construction of the main prediction network, and the construction of each scene network to form three components of a new model to improve the accuracy of production line performance prediction. and adaptability to provide stronger support for scheduling decisions. At present, there are no literatures and patents related to the above-mentioned production line performance prediction method.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a multi-scenario performance index prediction method and system for semiconductor production lines in order to overcome the above-mentioned defects in the prior art.

The object of the present invention can be realized through the following technical solutions:

A multi-scenario performance index prediction method oriented to a semiconductor production line, the method includes the following steps:

Step 1: Quantitatively map the historical data of the production line, and divide the data according to the scenarios corresponding to the light load, normal load and heavy load scenarios respectively;

Step 2: Using the divided data under the normal load scenario as sample data, combine the deep neural network algorithm and the performance prediction of the semiconductor production line to construct a prediction network under the normal load scenario;

Step 3: Based on the prediction network under the normal load scenario, further construct the prediction network under the light load and heavy load scenarios, and form the prediction network under the light load, normal load and heavy load scenarios into a multi-scenario prediction model;

Step 4: After the online data of the production line is divided into different scenario results according to the threshold, the corresponding network is selected in the multi-scenario prediction model for prediction, and the performance index prediction result is obtained.

Further, the historical data of the production line includes input and output, wherein the input includes: sampling days, average product processing cycle, average daily moving steps, total work-in-progress, each buffer leader, buffer zone The total team leader and the total number of moving steps; the output includes: the average processing cycle of the product and the utilization rate of each equipment.

Further, the process of constructing the prediction network under the normal load scenario in the described step 2 includes the following sub-steps:

Step 201: Encode the scheduling rule symbol in the sample data, and use it as an input for normalization processing, limit the number within the [0,1] interval, and the normalization formula is [x-min(x _i ) ]/[max(x _i )-min(x _i )], where x refers to the input variable and i is the sample variable;

Step 202: Use a deep neural network to build a prediction network under a normal load scenario; combine the deep learning algorithm with the production line performance prediction, and use a grid search method to obtain the appropriate number of hidden layers, the number of neurons in each hidden layer, and the number of layers in each hidden layer. excitation function;

Step 203: use the test sample to test the network performance of the prediction network, compare the output value corresponding to the prediction result obtained by the test sample after inverse normalization processing with the output value of the test sample, and determine whether the accuracy requirement is met;

Step 204: If the prediction accuracy of the test result can meet the accuracy requirements, the prediction network under the normal load scenario is successfully established; if not, return to step 202, and reselect the number of hidden layers, the number of neurons in each hidden layer, and Each layer activates the function and trains the model again.

Further, the scheduling rule symbol in the step 201 is encoded by one-hot encoding, so that it can be received by the network.

Further, the input of the deep learning algorithm in the described step 202 is:

For a given training set of k different samples:

X ^k ＝{S ^k ,D ^k ,Ru ^k ,P ^k |S ^k ∈R ^∝ ,D ^k ∈R ^β ,Ru ^k ∈R ^λ ,P ^k ∈R ^m }

为智能车间系统状态；

为智能车间投料信息；

为智能车间调度规则；

为在当前智能车间系统状态S^k，当前投料信息D^k，当前调度规则Ru^k情况下1天后的性能指标；in,

is the status of the intelligent workshop system;

Feeding information for intelligent workshop;

Scheduling rules for smart workshops;

is the performance index 1 day after the current intelligent workshop system state S ^k , the current feeding information D ^k , and the current scheduling rule Ru ^k ;

The output of the deep learning algorithm in step 202 is represented by o _i , where i=1, . . ., m1, m1 refers to the dimension of the output value o, that is, there are m1 outputs.

Further, the precision requirement in the described step 204 uses the root mean square error as the benchmark, and the description formula is:

为输出变量样本数。where m is the number of output variables, P is the input performance index, Y is the predicted performance index,

is the number of samples of the output variable.

Further, the process of constructing the prediction network under the light-load and heavy-load scenarios in the described step 3 includes the following sub-steps:

Step 301: Assign the number of input layer neurons of the prediction network under the normal load scenario to the prediction network under the light load and heavy load scenarios;

Step 302: Assign the number of hidden layer neurons of the prediction network in the normal load scenario except the last hidden layer to the prediction network in the light load and heavy load scenarios as the neurons corresponding to the first n-1 layers. number;

Step 303: Assign the weight threshold information of each hidden layer except the last hidden layer of the prediction network under the normal load scenario to the prediction network under the light load and heavy load scenarios as initialization data;

Step 304: Assign the activation functions of each hidden layer except the last hidden layer of the prediction network under the normal load scenario to the prediction network under the light load and heavy load scenarios;

Step 305: Adjust the number of neurons in the hidden layer of the last layer through the network search method, and input the data in the light-load and heavy-load scenarios divided in step 1 into the prediction network under the light-load and heavy-load scenarios for training, After training, the prediction network in light-load and heavy-load scenarios is obtained.

Further, the multi-scene prediction model building module includes:

The light-load scene prediction network is constructed. Taking the light-load scene data as a sample, on the basis of the main prediction network, the number of neurons in the last hidden layer is changed to build a light-load scene prediction network;

The overloaded scene prediction network is constructed. Taking the overloaded scene data as a sample, on the basis of the main prediction network, the number of neurons in the last hidden layer is changed to build the overloaded scene prediction network;

Network combination, combining the three prediction networks. Form a multi-scenario prediction model.

Further, described step 1 comprises the following sub-steps:

Step 101: Under the same sampling days, collect historical data of the production line for quantitative mapping, and draw a line graph;

Step 102 : Perform data division according to the curve change law in each line graph and the scenario quantifications corresponding to the light load, normal load and heavy load scenarios respectively.

Further, in the light-load scenario in the step 102, in the corresponding curve, the starting point is the point corresponding to the total work-in-process value when the equipment utilization rate begins to enter a stable stage, and the end point is the total product when the average processing cycle curve slope is the maximum. The point corresponding to the WIP value;

In the normal load scenario, in the corresponding curve, the starting point is the point corresponding to the WIP value when the slope of the product average processing cycle curve is the largest, and the end point is the point corresponding to the WIP value when the product average processing cycle curve slope is the smallest;

In the heavy load scenario, in the corresponding curve, the starting point is the point corresponding to the total work-in-process value when the slope of the average processing cycle curve of the product is the smallest.

The present invention also provides a system that adopts the multi-scenario performance index prediction method for semiconductor production lines, the system comprising:

The main prediction network building module is used to use the divided data under normal load scenarios as sample data, and combine the deep neural network algorithm with the performance prediction of semiconductor production lines to construct a prediction network under normal load scenarios;

The multi-scenario prediction model building module is used to further construct the prediction network under the light load and heavy load scenarios based on the prediction network under the normal load scenario, and combine the prediction networks under the light load, normal load and heavy load scenarios into multi-scenario prediction Model;

The production scene quantitative division module is used to quantitatively map the historical data of the production line, and divide the data according to the quantitative scenes corresponding to the light load, normal load and heavy load scenarios respectively.

Compared with the prior art, the present invention has the following advantages:

1. The present invention divides the production environment into three scenarios of light load, normal load and heavy load by quantitatively dividing the production scenarios, so that the prediction method can be adapted to the dynamic production environment;

2. The present invention combines the deep learning algorithm with the prediction of the semiconductor production line, and uses the divided light load, normal load and heavy load data as sample data to establish a neural network prediction model, which can respond to real-time changes in the system in time. , reducing the need for rescheduling. The present invention regards the system as a black box, and provides a modeling method that utilizes the historical data and off-line data obtainable from the production line, mines useful knowledge therein, and realizes real-time on-line optimization control;

3. The present invention applies the transfer learning idea to the training process of the multi-scenario network, and builds the network under the light-load and heavy-load scenarios according to the main prediction network that has been constructed, so as to form a multi-scenario prediction model and reduce the network training time. time, improving the accuracy of production line performance prediction;

4. The present invention can predict the performance indicators of multiple production lines one day in the future, and provide more data support for production decision-making.

Description of drawings

Fig. 1 is the structural representation of the present invention;

Fig. 2 is the MiniFAB model structure diagram in the embodiment of the present invention;

FIG. 3 is a schematic diagram of the average processing cycle of products of EDD, EDD, SRPT, SRPT, and EDD, respectively, in the 26th day sampling Ma to Me equipment scheduling rules in the embodiment of the present invention;

4 is a schematic diagram of the number of work-in-progress of the equipment scheduling rules for sampling Ma to Me on the 26th day, respectively, EDD, EDD, SRPT, SRPT, EDD, and EDD in the embodiment of the present invention;

FIG. 5 is a schematic diagram of the utilization rate of each equipment of EDD, EDD, SRPT, SRPT, EDD, and EDD for sampling Ma to Me equipment scheduling rules on the 26th day in the embodiment of the present invention;

6 is a schematic diagram showing the comparison of the root mean square error in productivity of each scenario model, a scenario-independent model, and a multi-scenario multi-performance index prediction model in an embodiment of the present invention;

7 is a schematic diagram illustrating the comparison of the root mean square error of each scenario model, a scenario-independent model, and a multi-scenario multi-performance index prediction model on Ma equipment utilization in an embodiment of the present invention;

8 is a schematic diagram showing the comparison of the root mean square error of each scenario model, a scenario-independent model, and a multi-scenario multi-performance index prediction model on Mb device utilization in an embodiment of the present invention;

9 is a schematic diagram illustrating the comparison of the root mean square error of each scenario model, a scenario-independent model, and a multi-scenario multi-performance index prediction model on Mc equipment utilization in an embodiment of the present invention;

10 is a schematic diagram showing the comparison of the root mean square error of each scenario model, a scenario-independent model, and a multi-scenario multi-performance index prediction model on Md equipment utilization in an embodiment of the present invention;

FIG. 11 is a schematic diagram showing the comparison of the root mean square error of the Me device utilization rate of each scenario model, a scenario-independent model, and a multi-scenario multi-performance index prediction model in an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

specific embodiment

As shown in FIG. 1 , the present invention provides a multi-scenario multi-performance index prediction method for semiconductor production lines, including a production scenario quantitative division module, a main prediction network building module, and a multi-scenario prediction model building module. The quantitative division module of production scenarios is driven by data, and quantitatively maps the value of the production line, the average processing cycle of products, and the utilization rate of each equipment, and divides the production line into three scenarios: light load, normal load, and heavy load; the main prediction network building module , taking the divided normal load data as sample data, combining the deep neural network algorithm with the performance prediction of semiconductor production lines to build a prediction network under normal load scenarios; multi-scenario prediction model building module: applying the idea of transfer learning to production line prediction In , according to the main prediction network that has been constructed, the network under light-load and heavy-load scenarios is constructed to form a multi-scenario prediction model.

This embodiment takes the MiniFab model as an example for description. MiniFab is an experimental simulation model of a semiconductor production process proposed by Dr. Karl Kempf, chief scientist of Intel's manufacturing system. It consists of 5 devices and 6 processes. The model structure is shown in Figure 2.

Using the MiniFab model as the implementation object, the working process of the above-mentioned multi-scenario and multi-performance index prediction method for semiconductor production lines is as follows:

描述智能车间系统状态；

描述智能车间投料信息；

描述智能车间调度规则；

描述在当前状态S^k，当前投料方式D^k，当前调度规则Ru^k下1天后的性能指标。选取车间生产率PROD、设备Ma利用率U_a、设备Mb利用率U_b、设备Mc利用率U_c、设备Md利用率U_d、设备Me利用率U_e这六个性能指标作为预测目标，样本信息中的生产线系统状态S^k、投料信息D^k、调度规则Ru^k定义见表1、表2和表3。选取10％作为测试样本集；Step 1, define sample information X ^k ={S ^k ,D ^k ,Ru ^k ,P ^k |S ^k ∈R ^∝ ,D ^k ∈R ^β ,Ru ^k ∈R ^λ ,P ^k ∈R ^m }, where

Describe the status of the smart workshop system;

Describe the material feeding information of the intelligent workshop;

Describe intelligent workshop scheduling rules;

Describe the performance index after 1 day under the current state S ^k , the current feeding mode D ^k , and the current scheduling rule Ru ^k . Six performance indicators of workshop productivity PROD, equipment Ma utilization rate U _a , equipment Mb utilization rate U _b , equipment Mc utilization rate U _c , equipment Md utilization rate U _d , equipment Me utilization rate U _e are selected as prediction targets, and the sample information The definitions of the production line system state ^Sk , the feeding information D ^k , and the scheduling rule Ru ^k are shown in Table 1, Table 2 and Table 3. Select 10% as the test sample set;

Table 1 Definition of system state ^Sk in MiniFab sample information

field name describe type of data Day Sampling days Integer MCT Average product cycle time floating point MDayMOV Average daily steps Integer WIP Work in process Integer Mab_Queue MaMb buffer captain Integer Mcd_Queue Captain McMd Buffer Integer Me_Queue Captain Me Buffer Integer Total_Queue buffer chief Integer Total_MOV total moving steps Integer

Table 2 Definition of feed information D ^k in MiniFab sample information

field name describe type of data x_Release Average daily feed for product x Integer y_Release Average daily feed for product y Integer z_Release Average daily feed for product z Integer

Table 3 Scheduling rule Ru ^k definition in MiniFab sample information

field name describe type of data Ma_Rule Scheduling rules for equipment Ma character type Mb_Rule Scheduling rules for device Mb character type Mc_Rule Scheduling rules for device Mc character type Md_Rule Scheduling rules for equipment Md character type Me_Rule Scheduling rules for equipment Me character type

In this embodiment, the number of sampling days is 1 to 30, the daily average feed of product a, product b, and product c is increased from 1 to 19, respectively, EDD or SRPT is selected for the scheduling rules of Ma and Mb, and EDD or SRPT is selected for the scheduling rules of Mc and Md. Or select EDD or SRPT for CR, Me scheduling rules, and obtain the utilization rate and productivity performance indicators of each equipment after 1 day.

Step 2: Pass through the average product processing cycle MCT, the number of WIP, the utilization rate U a of the equipment Ma, the utilization rate U _b of the equipment Mb, the utilization rate U _c of the equipment M _{d, the utilization rate U d of} the equipment Me, and the utilization rate U _e of the equipment Me. Map the number of sampling days and the average daily feeding total of the product, observe the corresponding relationship between the curve changes in the figure, and quantitatively divide the light load, normal load and heavy load scenarios of the production line.

Scenario 1: Light load scenario. In this scenario, the device Ma utilization rate U _a , the device Mb utilization rate U _b , the device Mc utilization rate U _c , the device Md utilization rate U _d , and the device Me utilization rate U _e tend to be stable, and the starting point of this scenario is MCT The point at which the equipment utilization begins to enter a stable stage, the end point of the scene is the point with the maximum slope of the MCT curve, and the light load scene is quantitatively divided according to the start point and end point of the scene corresponding to the WIP curve.

Scenario 2: Normal scenario. In this scenario, the utilization rate of each device is relatively stable, and the MCT fluctuates. The starting point of the scenario is the point with the maximum slope of the MCT curve, and the end point of the scenario is the point with the minimum slope of the MCT curve. Quantitatively divide normal scenes.

Scenario 3: Reload scenario. In this scenario, the utilization rate of each device is relatively stable, the MCT curve keeps rising and does not drop, and the WIP curve rises. The starting point of this scenario is the minimum slope of the MCT curve. According to the WIP curve corresponding to the starting and ending points of the scenario, quantitatively divide the load the scene.

According to the starting abscissas corresponding to the three scenarios, the WIP values of the three scenarios are quantitatively obtained, which lays the foundation for the subsequent online data to divide the scenarios according to the specific WIP values.

1a) Draw a map: Under the same sampling days, select the average product processing cycle MCT, the number of WIP, the utilization rate U a of equipment Ma, the utilization rate U _b of equipment Mb, and the utilization rate U _c of equipment Mc under several scheduling rules _. , the utilization rate U _d of the equipment Md and the utilization rate U _e of the equipment Me to draw a line graph. Here, taking the sampling days of 26, Ma, Mb, Mc, Md, and Me equipment scheduling rules as EDD, EDD, SRPT, SRPT, EDD, and EDD as an example, Figure 3, Figure 4, and Figure 5 are drawn;

1b) Quantitative division of scenarios: when the sample id is 6, the equipment utilization in Figure 5 tends to be stable, and the WIP value in Figure 4 is roughly 60; when the sample id is 8, the slope of the MCT fluctuation phase in Figure 3 is the largest, At this time, the WIP in Figure 4 fluctuates around 200; when the sample id is 16, the MCT only rises and does not fall, and the WIP fluctuates around 750. The definition index of scene division is WIP. Therefore, when WIP is in [60, 200), it is determined as a light-loaded scene, when WIP is in [200, 750), it is determined as a normal scene, and when WIP is in [750, ∞), it is determined as an overloaded scene.

Step 3: Encode the scheduling rule symbol so that it can be received by the network, and normalize the original data through min-max, remove its dimension, and change the original data to a decimal between [0,1] . The coding table is shown in Table 4.

Table 4 Scheduling rule coding table

scheduling rules coding EDD [0,1] SRPT [1,0] CR [1,1]

Step 4: Use the DNN algorithm to divide the data into normal load scenarios, and use the DNN algorithm to calculate the feed information D, the MaMb scheduling rule, the McMd scheduling rule, the Me scheduling rule, the number of work-in-progress in the system state S, the MaMb buffer length, and the McMd buffer. The captain, Me buffer captain, total buffer captain, and total moving steps are input, and P, which represents the performance index, is the output, and the matching relationship between feeding information, scheduling rules, system status and performance indicators is obtained, that is, the main prediction model.

The specific process of using the deep neural network algorithm to learn and obtain the main prediction model is as follows:

2a) Initialize the weight matrices W ^η , V and threshold matrices θ ^q , γ of the network, and the values are random numbers between [-1, 1]. Set the training times counter h.

2b) Select the kth pair of data samples, and calculate the output results of each hidden layer:

2c) Determine the output of the output layer:

2d) Calculate the correction error of each neuron in the output layer and the hidden layer

2e) Calculate the error of each layer of the nth hidden layer

2f) According to the selected gradient descent optimizer is Adadelta Optimizer

2g) Set the loss function:

为输出变量样本数。where m is the number of output variables, P is the input performance index, Y is the predicted performance index,

is the number of samples of the output variable.

2h) Set the learning rate to 0.05 and set the network learning times to 50,000 times;

2i) Calculate the total output error according to the loss function defined in the network. If the loss function value is less than the set value, jump out of the training program. If not, adjust the weights and thresholds of each layer and continue training.

2j) Check whether the training times counter h reaches the network setting value, if not, then the training times counter h+1; if it does, jump out of the program and reset the number of hidden layers, the number of neurons and the excitation function.

Finally, the grid search method is used to determine that the network layer of network A is 6 layers, the network structure is 10*10*10*9*8*6, the activation function of the hidden layers 1-4 is tanh, and the excitation function of the output layer is sigmoid.

Step 5: According to the network information of the main prediction network, a light-load scenario network and a heavy-load scenario network are constructed by the method of transfer learning. The steps are as follows:

Step K1: Assign the number of neurons in the input layer of the main prediction network to the new light load/heavy load network;

Step K2: Assign the number of neurons in each hidden layer of the main prediction network except the last hidden layer to the new light-load/heavy-load network as the number of neurons in the first n-1 layers;

Step K3: Assign the weight threshold information of each hidden layer of the main prediction network except the last hidden layer to the new light load/heavy load network as initialization data;

Step K4: Assign the activation function of each hidden layer of the main prediction network except the last hidden layer to the new light load/heavy load network;

Step K5: Adjust the number of neurons in the last hidden layer by grid search method, and input the light load/heavy load scene data for training.

In this embodiment, the main prediction network A has a total of 4 hidden layers, the first 3 hidden layers of the network A are migrated to the first 3 hidden layers of the network B and the network C, and the number of neurons in the last hidden layer is adjusted by adjusting the number of neurons in the last hidden layer. , train to obtain network B and network C, and finally determine that the number of network layers of network B is 6 layers, the network structure is 10*10*10*9*9*6, the hidden layer 1-4 layer excitation function is tanh, and the output layer excitation The function is sigmoid; the network layer of network C is 6 layers, the network structure is 10*10*10*9*7*6, the activation function of the hidden layers 1-4 is tanh, and the excitation function of the output layer is sigmoid. The three networks are combined into one to form a multi-scenario multi-index prediction model. The model structure is shown in Table 5.

Table 5 Multi-scenario multi-index prediction model structure

Before the online data enters the model, the scene is identified according to the threshold value. If it is identified as a normal scene, the main prediction network in the model is selected for prediction; if it is identified as a light-load scene, the light-load scene prediction network in the model is selected for prediction. Prediction; if it is identified as an overloaded scene, select the overloaded scene prediction network in the model for prediction.

The scenario-independent model, the single-scenario model and the multi-scenario multi-performance index prediction model are respectively applied to the MiniFAB model, and the prediction results of the scenario-independent network, the single-scenario model and the multi-scenario multi-performance indicator prediction model are compared.

Table 6 shows the comparison between the output of the scenario-independent model and the multi-scenario multi-index prediction model and the average value of the test sample.

Table 6. Comparison between the output of the non-scenario model and the multi-scenario multi-index prediction model and the average value of the test sample

In the scenario-independent model, that is, the data is mixed to train to obtain an overall network model, the highest relative error is 3.4%, and the lowest is 1.0%. In the multi-scenario and multi-index prediction model, the highest relative error is 1.5%, and the lowest relative error is 0.15%. From the perspective of the relative error of the output average, the multi-scenario multi-index prediction model is better than the no-scenario model.

The average output root mean square error of the single-scenario model, the no-scenario model and the multi-scenario multi-index prediction model is shown in Table 7.

Table 7 The average value of the root mean square error of each model

On the whole, the multi-scenario multi-index prediction model has higher accuracy and better quality.

Figures 6 to 11 show the root mean square error of productivity and the root mean square error of each equipment utilization rate for the single-scenario model, the no-scenario model, and the multi-scenario multi-index prediction model.

It can be seen from Figure 6 to Figure 11 that the light-load single-scenario model has the largest prediction error for the heavy-load scenario, and the smallest prediction error for the light-load scenario; the heavy-load single-scenario model has the smallest prediction error for the heavy-load scenario, and the prediction error for the light-load scenario is the smallest. The prediction error is the largest; the prediction error of the normal single scene is the smallest for the normal scene, and the prediction error for the overloaded scene is the largest; the model regardless of the scene has the largest error for the overloaded scene, and the prediction error for the normal scene is the smallest; The prediction errors of the multi-index prediction models for the three scenarios are almost the same, and they are all in the optimal position of the four models.

The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited to this. Any person skilled in the art can easily think of various equivalents within the technical scope disclosed by the present invention. Modifications or substitutions should be included within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims (4)

1. A multi-scenario performance index prediction method for a semiconductor production line is characterized by comprising the following steps:

step 1: carrying out quantitative mapping on historical data of a production line, and carrying out data division according to scene quantification respectively corresponding to light load scenes, normal load scenes and heavy load scenes;

step 2: taking the divided data under the normal load scene as sample data, and combining a deep neural network algorithm with the performance prediction of a semiconductor production line to construct a prediction network under the normal load scene;

and step 3: based on the prediction network under the normal load scene, the prediction networks under the light load scene and the heavy load scene are further constructed, and the prediction networks under the light load scene, the normal load scene and the heavy load scene form a multi-scene prediction model;

and 4, step 4: after dividing the online data of the production line into different scene results according to a threshold value, selecting a corresponding network from the multi-scene prediction model for prediction to obtain a performance index prediction result;

the process of constructing the prediction network under the normal load scene in the step 2 comprises the following sub-steps:

step 201: coding a scheduling rule symbol in the sample data, and performing normalization processing by taking the scheduling rule symbol as input;

step 202: adopting a deep neural network to construct a prediction network under a normal load scene; combining a deep learning algorithm with the performance prediction of a production line, and obtaining the number of proper hidden layers, the number of neurons of each hidden layer and excitation functions of each layer by adopting a grid search method;

step 203: testing the network performance of the prediction network by adopting a test sample, comparing an output value corresponding to a prediction result obtained by the test sample after reverse normalization processing with an output value of the test sample, and judging whether the accuracy requirement is met;

step 204: if the prediction precision of the test result can meet the precision requirement, the prediction network under the normal load scene is successfully established, if the prediction precision of the test result cannot meet the precision requirement, the step 202 is returned, the number of hidden layers, the number of neurons of each hidden layer and excitation functions of each layer are selected again, and the model is trained again;

the deep learning algorithm in step 202 has the following inputs:

for a given training set of k different samples:

X^k＝{S^k，D^k，Ru^k，P^k|S^k∈R^∝，D^k∈R^β，Ru^k∈R^λ，P^k∈R^m}

wherein,

the state is the state of an intelligent workshop system;

feeding information for an intelligent workshop;

scheduling rules for the intelligent workshop;

for the current state S of the intelligent workshop system^kCurrent feeding information D^kCurrent scheduling rule Ru^kPerformance index after 1 day under the circumstances;

the output of the deep learning algorithm in step 202 is o_iWhere i is 1, …, m1, m1 indicates the dimension of the output value o, i.e. there are m1 outputs;

the accuracy requirement in step 204 is based on the root mean square error, and the description formula is as follows:

where m is the number of output variables, P is the input performance index, Y is the predicted performance index,

is the output variable sample number;

the process of constructing the prediction network under the light load scene and the heavy load scene in the step 3 comprises the following sub-steps:

step 301: assigning the number of neurons of an input layer of the prediction network under a normal load scene to the prediction network under light load and heavy load scenes;

step 302: assigning the number of neurons of each hidden layer except the last hidden layer of the prediction network under the normal load scene to the prediction network under the light load scene and the heavy load scene as the number of neurons of the corresponding front n-1 layer;

step 303: assigning the weight threshold information of each hidden layer except the last hidden layer of the prediction network under the normal load scene to the prediction network under the light load scene and the heavy load scene as initialization data;

step 304: assigning excitation functions of all hidden layers except the last hidden layer of the prediction network under a normal load scene to the prediction network under light load and heavy load scenes;

step 305: adjusting the number of neurons in the last hidden layer by a network search method, inputting the data under the light load and heavy load scenes divided in the step 1 into a prediction network under the light load and heavy load scenes for training, and obtaining the prediction network under the light load and heavy load scenes after the training is finished;

the step 1 comprises the following sub-steps:

step 101: collecting historical data of a production line for quantitative mapping under the same sampling days, and drawing a line graph;

step 102: dividing data according to curve change rules in each line graph and scene quantification corresponding to light load scenes, normal load scenes and heavy load scenes respectively, wherein in the corresponding curves, a starting point is a point corresponding to a total product value when the equipment utilization rate starts to enter a stable stage, and an end point is a point corresponding to the total product value when the slope of the curve of the average processing period of the product is maximum;

under a normal load scene, in the corresponding curve, the starting point is the point corresponding to the product value when the slope of the average processing period curve of the product is maximum, and the end point is the point corresponding to the product value when the slope of the average processing period curve of the product is minimum;

under a heavy load scene, the starting point in the corresponding curve is the point corresponding to the total product value when the slope of the average processing cycle curve of the product is minimum.

2. The method as claimed in claim 1, wherein the production line history data includes input quantity and output quantity, wherein the input quantity includes: sampling days, average processing period of products, average daily moving steps, total number of products in process, queue length of each buffer area, total queue length of the buffer areas and total moving steps; the output quantity comprises: average processing period of products and utilization rate of each device.

3. The method as claimed in claim 1, wherein the scheduling rule symbol in step 201 is encoded by using a one-hot encoding method so that it can be received by a network.

4. A system using the semiconductor production line oriented multi-scenario performance index prediction method of claim 1, characterized in that the system comprises:

the main prediction network construction module is used for combining a deep neural network algorithm and semiconductor production line performance prediction by taking the divided data under the normal load scene as sample data to construct a prediction network under the normal load scene;

the multi-scene prediction model construction module is used for further constructing prediction networks under light load and heavy load scenes based on the prediction networks under normal load scenes, and forming the prediction networks under the light load scenes, the normal load scenes and the heavy load scenes into a multi-scene prediction model;

and the production scene quantitative division module is used for carrying out quantitative mapping on the historical data of the production line and carrying out data division according to the scene quantification respectively corresponding to the light load scene, the normal load scene and the heavy load scene.

Images