CN118618978B - Winding method and system based on artificial intelligence - Google Patents

Abstract

The invention discloses a winding method and a winding system based on artificial intelligence, comprising the following steps: s1: collecting film density data, film thickness data, roll diameter data and film tension data, and preprocessing the film tension data to obtain preprocessed film tension data; s2: constructing a tension prediction model, filling a film tension data missing value, and generating a roll diameter-tension curve; s3: calculating overshoot amplitude and fluctuation time generated by each time of tension adjustment of the PID system, and calculating an overshoot penalty item; s4: constructing a tension grade division model, and outputting the number of tension grades and corresponding tension according to the maximum coil diameter; s5: and controlling the film tension in the winding process through a PID system according to the number of tension grades and the corresponding tension. The invention considers the error accumulation caused by tension adjustment of the PID system, and solves the problem of poor winding effect caused by the fact that the traditional winding method does not fully consider the film characteristics, the winding diameter and the capability of the PID system.

Description

A rolling method and system based on artificial intelligence

Technical Field

The present invention relates to the field of battery separator preparation, and in particular to an artificial intelligence-based winding method and system.

Background Art

A winder is a device used to wind up roll products such as battery films. Its main function is to form a compact roll structure by winding up the film or sheet to facilitate subsequent processing and application.

Traditional winding methods usually do not fully consider factors such as film characteristics and roll diameter, resulting in unreasonable tension settings, which in turn causes various problems such as wrinkles, ribs, folds, and slippage in the film during the winding process, affecting the performance and quality of the battery separator.

In addition, although some solutions try to adjust the tension in real time, the PID (proportional-integral-differential) system used is difficult to accurately control the tension output. Setting too precise target tension may cause the tension output of the PID system in real time to be difficult to accurately follow the changes in the target tension, resulting in overshoot and fluctuations, which in turn affects the stability and quality of winding.

Summary of the invention

In view of this, the present invention aims to provide an artificial intelligence-based winding method and system, aiming to solve the problem that the traditional winding method does not fully consider the film characteristics, roll diameter and the capabilities of the PID system, resulting in poor winding effect.

A method for collecting rolls based on artificial intelligence comprises the following steps:

S1: collecting film density data, film thickness data, roll diameter data of the film roll during film winding, and film tension data during film winding, and preprocessing the film tension data to obtain preprocessed film tension data;

S2: Fusing the film density data, film thickness data and roll diameter data, and constructing a tension prediction model to fill in the missing values of film tension data and generate a roll diameter-tension curve;

The tension prediction model first proposes position missing coding and missing distance coding to mark the position of the missing value and the distance between the missing value and the true value respectively, and then integrates the position missing coding and missing distance coding with the film tension data, inputs them into the gated recurrent unit, and predicts and fills the film tension data;

S3: Calculate the overshoot amplitude and fluctuation time generated by each adjustment of the film tension by the PID system, and then calculate the overshoot penalty term;

S4: construct a tension level classification model, and output the number of tension levels and corresponding tension according to the maximum roll diameter;

The reward function of the tension level classification model uses a definite integral method to calculate the coil diameter-tension curve fitting degree, and the value function of the tension level classification model is calculated by a Q-learning method, and an ε-greedy strategy is used to update the tension level classification model;

S5: The film tension during the winding process is controlled by the PID system according to the number of tension levels and the corresponding tension.

Further, in the step S1, the roll diameter data is discrete time series data, and the film tension data corresponds to the roll diameter data one by one;

The density and thickness in the film data are the basic properties of lithium battery separators, which directly affect the tension;

The preprocessing of the film tension data comprises the following steps: using the spline interpolation method, performing a single interpolation on the film tension data under different roll diameters to obtain the interpolated film tension data, and the calculation method is:

;

;

;

;

in, is the interpolated film tension data output by the cubic spline function. is the input of the cubic spline function, i is the interval index, To solve for the coefficients, is the curl diameter of the left endpoint of the i-th interval, is the film tension data corresponding to the left endpoint of the i-th interval; is the first-order derivative of the cubic spline function at the left endpoint of the i-th interval, is the second-order derivative of the cubic spline function at the left endpoint of the ith interval, is the first-order derivative of the cubic spline function at the left endpoint of the i+1th interval, is the second-order derivative of the cubic spline function at the left endpoint of the i+1th interval.

Film tension data is usually discrete and has fewer data points, which cannot fully reflect the subtle changes between tension and roll diameter. Therefore, using spline interpolation to interpolate discrete film tension data can obtain more data points and improve data density, which will help to obtain a more continuous and smooth tension-roll diameter curve in the future, and also provide sufficient data points for the subsequent tension prediction model for model training.

Furthermore, the S2 step includes:

S21: The film density data, film thickness data and roll diameter data are integrated and calculated as follows:

;

;

in, is the fused data, For splicing operation, They are film density data and film thickness data, is the fused feature, It is a fully connected layer operation;

Traditional methods often set the tension to a fixed value, or only consider the effect of the change in roll diameter on the tension; in contrast, step S21 combines the film density and film thickness, and then fuses them after feature expansion, taking into account more parameters, and can more comprehensively capture the change law of tension, so that the winder can more accurately control the change of tension when adjusting the tension, avoiding wrinkles, ribs and other problems during the winding process, and improving the winding effect of lithium battery separators; especially when facing films of different thicknesses or densities, this design can better adapt to different process requirements and ensure the stability and quality of the winding effect;

S22: Construct a tension prediction model to predict the missing values of the film tension data and fill in the missing values of the film tension data. The calculation method is:

;

;

;

;

;

in, is the input of the tension prediction model, is the missing code of the left endpoint of the i-th interval index, is the missing code of the left endpoint of the i-1th interval index, is the missing distance code for the left endpoint of the i-th interval index, is the missing distance code for the left endpoint of the i-1th interval index, is the attenuation code, is the prediction result of the missing value of the film tension data at the left endpoint of the i-th interval index, is the prediction result of the missing value of the film tension data at the left endpoint of the i-1th interval index, To take the exponential operation of a natural constant, To obtain the maximum value, is the attenuation parameter weight, is the attenuation bias vector, Operate for recurrent neural networks;

Compared with the traditional recurrent neural network, which only considers the temporal relationship between data, the present invention provides a signal to the recurrent neural network through explicit coding, so that the recurrent neural network pays attention to the missing position of the data and the time step between the missing value and the nearest non-missing value, so that the model can more easily use the temporal information of the film tension data, thereby improving the prediction accuracy and robustness of the model, and further improving the winding effect of the winder;

S23: According to the prediction results, the roll diameter data and film tension data at each moment are counted to generate a roll diameter-tension curve.

Furthermore, the S3 step includes:

S31: recording the overshoot amplitude generated by each adjustment of the film tension by the PID system;

S32: When the PID system output oscillates, the oscillation start time is recorded;

S33: When the oscillation amplitude returns to the stable range, the oscillation stop time is recorded and subtracted from the oscillation start time to obtain the fluctuation time of this adjustment;

S34: Calculate the overshoot penalty term, the calculation method is:

;

Among them, P is the overshoot penalty term, is the overshoot amplitude, is the fluctuation time, is the total time of this adjustment, is the overshoot amplitude weight, is the time weight.

Setting too precise target tension will make it difficult for the PID system to accurately follow the changes in target tension during real-time adjustment, and overshoot and fluctuations will often occur. These errors will accumulate as the number of adjustments increases, ultimately affecting the stability of the winding effect.

This step introduces an overshoot penalty term, which penalizes overshoot behavior by quantifying the overshoot amplitude and fluctuation time. For the tension level division model based on reinforcement learning in step S4, this penalty mechanism can effectively reduce the system error caused by frequent tension adjustment, making the tension level division model more inclined to a stable tension adjustment strategy rather than pursuing overly fine target tension, thereby reducing fluctuations in the tension adjustment process and improving the stability and quality of the winding process.

Furthermore, the S4 step includes:

S41: The state space of the tension level division model is designed as: the tension value sequence after the coil diameter-tension curve is segmented; the action space is: merging the current tension level and adding a new segmentation point to the current tension level; the tension value sequence takes 0 as the first value in the sequence and the maximum coil diameter as the last value in the sequence;

S42: Design the reward function of the tension level classification model, which is calculated as follows:

;

;

in, The reward value of the model divided into tension levels, is the influence weight of the overshoot penalty term, is the roll diameter-tension curve fitting degree, is the influence weight of the coil diameter-tension curve fitting, To take the absolute value, The tension is in the coil diameter from Change to The definite integral when

S43: The value function of the design tension level classification model is calculated as follows:

;

in, is the value function under the current state s and the current action a , is the learning rate, For the next state, For the next action, In order to make the next step corresponding to the maximum Q value, is the discount factor, To update the symbol;

S44: Adopting the ε-greedy strategy, updating the tension level division model, and outputting the number of tension levels and the corresponding tension of each tension level.

In step S4, the number of segmentation points and the tension value sequence after segmentation are calculated by reinforcement learning under the condition of a given maximum coil diameter. The difference in the absolute value of the area in the curve fitting degree is used to indicate the degree of fit between the adjusted curve and the original coil diameter-tension curve. However, blindly pursuing the degree of fit will lead to too many segmentation points, and each segmentation will produce overshoot and fluctuation errors when the PID system is controlled in the future. Therefore, an overshoot penalty term is also designed in the reward function, which will cause the model to receive a larger penalty when selecting more segmentation times.

The overshoot penalty term can constrain the system's adjustment process, avoiding frequent adjustments and error accumulation, while the curve fitting degree can evaluate the matching degree between the tension curve and the change in roll diameter, further optimizing the design of the reward function. The two work together to improve the performance of the tension level division model, thereby achieving a more intelligent and practical tension control method.

Furthermore, the step S5 includes:

S51: Initialize the proportional, integral and differential parameters of the PID system;

S52: Real-time monitoring of film tension and roll diameter during the winding process;

S53: According to the number of tension levels and the corresponding tension of each tension level, the film tension during the winding process is adjusted through the PID system.

In order to realize the above-mentioned artificial intelligence-based winding method, the present invention also provides an artificial intelligence-based winding system, comprising:

Data acquisition and preprocessing module: collects film density data, film thickness data, film roll diameter data during film winding, and film tension data during film winding, and preprocesses the film tension data to obtain preprocessed film tension data;

Tension prediction model building module: integrates film density data, film thickness data and roll diameter data, builds a tension prediction model, fills in missing values of film tension data, and generates a roll diameter-tension curve;

Overshoot penalty calculation module: calculates the overshoot amplitude and fluctuation time generated by each adjustment of the film tension by the PID system, and then calculates the overshoot penalty term;

Tension level classification model construction module: constructs a tension level classification model, and outputs the number of tension levels and corresponding tensions according to the maximum roll diameter;

PID control module: According to the number of tension levels and the corresponding tension, the film tension during the winding process is controlled by the PID system.

Compared with the prior art, the advantages of the present invention are:

(1) In terms of the overall solution, the present invention constructs a tension prediction model, combines the film thickness data and the film density data, completes the missing values of the film tension data, and calculates the roll diameter-tension curve; secondly, through the reinforcement learning method, a tension level classification model is constructed to calculate the appropriate number of tension levels and the corresponding tension of each level, and then the actual tension is adjusted based on this to achieve a balance between the number of tension adjustments and the effect. According to different roll diameters, a reasonable number of tension adjustments is calculated to control the winder, thereby improving the lithium battery winding effect and reducing wrinkles, ribs and other problems during the winding process.

(2) In terms of algorithm improvement, the present invention first expands the features of film thickness data, film density data and film tension data to enhance the representation ability of vectors; secondly, based on the recurrent neural network, the missing position, the time step between the missing value and the adjacent non-missing value, and the attenuation method are explicitly modeled, thereby improving the effect of filling the missing values of the film tension data and obtaining the roll diameter-tension curve.

(3) The present invention takes into account the error accumulation caused by each tension adjustment of the PID system in actual situations, introduces the idea of reinforcement learning, and calculates the reasonable number of tension segments and the corresponding target tension according to the maximum roll diameter, so that the tension segment control scheme is more reasonable, reduces the system error caused by frequent tension adjustment, and thus reduces the fluctuation in the tension adjustment process, thereby improving the stability and quality of the winding process.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of the artificial intelligence-based winding method provided by the present invention.

FIG. 2 is a schematic diagram of the algorithm structure of the tension prediction model provided by the present invention.

DETAILED DESCRIPTION

The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

Embodiment 1:

An artificial intelligence-based winding method, as shown in FIG1 , comprises the following steps:

S1: collecting film density data, film thickness data, roll diameter data of the film roll during film winding, and film tension data during film winding, and preprocessing the film tension data to obtain preprocessed film tension data;

The roll diameter data is discrete time series data, and the film tension data corresponds to the roll diameter data one by one; the density and thickness in the film data are the basic properties of the lithium battery separator, which directly affect the tension;

The preprocessing of the film tension data comprises the following steps: using the spline interpolation method, performing a single interpolation on the film tension data under different roll diameters to obtain the interpolated film tension data, and the calculation method is:

;

;

;

;

in, is the interpolated film tension data output by the cubic spline function. is the input of the cubic spline function, i is the interval index, To solve for the coefficients, is the curl diameter of the left endpoint of the i-th interval, is the film tension data corresponding to the left endpoint of the i-th interval; is the first-order derivative of the cubic spline function at the left endpoint of the i-th interval, is the second-order derivative of the cubic spline function at the left endpoint of the ith interval, is the first-order derivative of the cubic spline function at the left endpoint of the i+1th interval, is the second-order derivative of the cubic spline function at the left endpoint of the i+1th interval.

Film tension data is usually discrete and has fewer data points, which cannot fully reflect the subtle changes between tension and roll diameter. Therefore, using spline interpolation to interpolate discrete film tension data can obtain more data points and improve data density, which will help to obtain a more continuous and smooth tension-roll diameter curve in the future, and also provide sufficient data points for the subsequent tension prediction model for model training.

For example,

Before interpolation:

After interpolation:

Interpolation does not mean to fill in all the data, but to preliminarily process the trends between known data points through the fitting ability of the cubic spline function, so as to provide more data points for the subsequent tension prediction model, thereby enhancing the prediction performance and robustness of the model; in terms of real-time and ease of use, the interpolation method has obvious advantages over other missing value filling methods.

S2: Fusing the film density data, film thickness data and roll diameter data, and constructing a tension prediction model to fill in the missing values of the film tension data and generate a roll diameter-tension curve, as shown in FIG2 ;

The tension prediction model first proposes position missing coding and missing distance coding to mark the position of the missing value and the distance between the missing value and the true value respectively, and then integrates the position missing coding and missing distance coding with the film tension data and inputs them into the gated recurrent unit to predict and fill the film tension data.

Step S2 includes:

S21: The film density data, film thickness data and roll diameter data are integrated and calculated as follows:

;

;

in, is the fused data, For splicing operation, They are film density data and film thickness data, is the fused feature, It is a fully connected layer operation;

Traditional methods often set the tension to a fixed value, or only consider the effect of the change in roll diameter on the tension; in contrast, step S21 combines the film density and film thickness, and then fuses them after feature expansion, taking into account more parameters, and can more comprehensively capture the change law of tension, so that the winder can more accurately control the change of tension when adjusting the tension, avoiding wrinkles, ribs and other problems during the winding process, and improving the winding effect of lithium battery separators; especially when facing films of different thicknesses or densities, this design can better adapt to different process requirements and ensure the stability and quality of the winding effect;

S22: Construct a tension prediction model to predict the missing values of the film tension data and fill in the missing values of the film tension data. The calculation method is:

;

;

;

;

;

in, is the input of the tension prediction model, is the missing code of the left endpoint of the i-th interval index, is the missing code of the left endpoint of the i-1th interval index, is the missing distance code for the left endpoint of the i-th interval index, is the missing distance code for the left endpoint of the i-1th interval index, is the attenuation code, is the prediction result of the missing value of the film tension data at the left endpoint of the i-th interval index, is the prediction result of the missing value of the film tension data at the left endpoint of the i-1th interval index, To take the exponential operation of a natural constant, To obtain the maximum value, is the attenuation parameter weight, is the attenuation bias vector, Operate for recurrent neural networks;

Compared with the traditional recurrent neural network, which only considers the temporal relationship between data, the present invention provides a signal to the recurrent neural network through explicit coding, so that the recurrent neural network pays attention to the missing position of the data and the time step between the missing value and the nearest non-missing value, so that the model can more easily use the temporal information of the film tension data, thereby improving the prediction accuracy and robustness of the model, and further improving the winding effect of the winder;

For example,

The input of a traditional recurrent neural network is a single-layer vector:

[13.5N, 14.755N, _, 15.525N, 16.290N, _, _, 16.875N]

The input after explicit encoding is a three-layer matrix:

[13.5N, 14.755N, _, 15.525N, 16.290N, _, _, 16.875N]

[1, 1, 0, 1, 1, 0, 0, 1]

[0, 0, 1, 0, 0, 1, 2, 0]

The second and third layers explicitly encode the missing value position and the distance between the missing value and the previous non-missing value, respectively, so that the recurrent neural network can more clearly know which positions should be focused on, thereby improving the effect of missing value filling in film tension data.

S23: According to the prediction results, the roll diameter data and film tension data at each moment are counted to generate a roll diameter-tension curve.

S3: Calculate the overshoot amplitude and fluctuation time generated by each tension adjustment of the PID system, and then calculate the overshoot penalty term;

S31: recording the overshoot amplitude generated by each adjustment of the film tension by the PID system;

S32: When the PID system output oscillates, the oscillation start time is recorded;

S33: When the oscillation amplitude returns to the stable range, the oscillation stop time is recorded and subtracted from the oscillation start time to obtain the fluctuation time of this adjustment;

S34: Calculate the overshoot penalty term, the calculation method is:

;

Among them, P is the overshoot penalty term, is the overshoot amplitude, is the fluctuation time, is the total time of this adjustment, is the overshoot amplitude weight, is the time weight.

Setting too precise target tension will make it difficult for the PID system to accurately follow the changes in target tension during real-time adjustment, and overshoot and fluctuations will often occur. These errors will accumulate as the number of adjustments increases, ultimately affecting the stability of the winding effect.

This step introduces an overshoot penalty term, which penalizes overshoot behavior by quantifying the overshoot amplitude and fluctuation time. For the tension level division model based on reinforcement learning in step S4, this penalty mechanism can effectively reduce the system error caused by frequent tension adjustment, making the tension level division model more inclined to a stable tension adjustment strategy rather than pursuing overly fine target tension, thereby reducing fluctuations in the tension adjustment process and improving the stability and quality of the winding process.

S4: construct a tension level classification model, and output the number of tension levels and corresponding tension according to the maximum roll diameter;

The reward function of the tension level classification model uses a definite integral method to calculate the coil diameter-tension curve fitting degree, and the value function of the tension level classification model is calculated using the Q-learning method, and the ε-greedy strategy is used to update the tension level classification model.

Step S4 includes:

S41: The state space of the tension level division model is designed as: the tension value sequence after the coil diameter-tension curve is segmented; the action space is: merging the current tension level and adding a new segmentation point to the current tension level; the tension value sequence takes 0 as the first value in the sequence and the maximum coil diameter as the last value in the sequence;

S42: Design the reward function of the tension level classification model, which is calculated as follows:

;

;

in, The reward value of the model divided into tension levels, is the influence weight of the overshoot penalty term, is the roll diameter-tension curve fitting degree, is the influence weight of the coil diameter-tension curve fitting, To take the absolute value, The tension is in the coil diameter from Change to For example, if the maximum roll diameter is 60 cm, the number of segmentation points is set to 100 at the beginning of the model training. That is, the PID system has to adjust the target tension once for every 0.6 cm change in roll diameter. The frequency is too high. Therefore, according to the overshoot penalty term, the model will be punished more severely and the number of segmentation points will be reduced.

In the middle of the training, the model sets the number of segmentation points to 6, that is, the PID system adjusts the target tension once for every 10 cm change in the roll diameter. The frequency is too low, and the absolute value of the area of the roll diameter-tension curve is large, and the degree of fit is poor. Therefore, according to the curve fit, the model will still be penalized and the number of segmentation points will be increased;

In the later stage of training, after two-way punishment, the model calculated that the number of segmentation points was more appropriate to be 20, that is, the PID system adjusted the target tension once for every 3 cm change in the roll diameter, with a moderate frequency, ensuring the degree of fitting of the roll diameter-tension curve without excessive overshoot and fluctuation.

S43: The value function of the design tension level classification model is calculated as follows:

;

in, is the value function under the current state s and the current action a , is the learning rate, For the next state, For the next action, In order to make the next step corresponding to the maximum Q value, is the discount factor, To update the symbol;

S44: Adopting the ε-greedy strategy, updating the tension level division model, and outputting the number of tension levels and the corresponding tension of each tension level.

In step S4, the number of segmentation points and the tension value sequence after segmentation are calculated by reinforcement learning under the condition of a given maximum coil diameter. The difference in the absolute value of the area in the curve fitting degree is used to indicate the degree of fit between the adjusted curve and the original coil diameter-tension curve. However, blindly pursuing the degree of fit will lead to too many segmentation points, and each segmentation will produce overshoot and fluctuation errors when the PID system is controlled in the future. Therefore, an overshoot penalty term is also designed in the reward function, which will cause the model to receive a larger penalty when selecting more segmentation times.

The overshoot penalty term can constrain the system's adjustment process, avoiding frequent adjustments and error accumulation, while the curve fitting degree can evaluate the matching degree between the tension curve and the change in roll diameter, further optimizing the design of the reward function. The two work together to improve the performance of the tension level division model, thereby achieving a more intelligent and practical tension control method.

S5: Control the film tension during the winding process through the PID system according to the number of tension levels and the corresponding tension;

S51: Initialize the proportional, integral and differential parameters of the PID system;

S52: Real-time monitoring of film tension and roll diameter during the winding process;

S53: According to the number of tension levels and the corresponding tension of each tension level, the film tension during the winding process is adjusted through the PID system.

The present invention is particularly suitable for production scenarios with high requirements for winding effects, especially when frequent tension adjustment is required; the present invention can fully consider the film characteristics, roll diameter and the capabilities of the PID system, solve the problems of error accumulation, overshoot and fluctuation that occur in the traditional winding method during the tension adjustment process, and improve the winding effect.

Example 2

An artificial intelligence-based winding system, used to implement the artificial intelligence-based winding method, comprises:

Data acquisition and preprocessing module: collects film density data, film thickness data, film roll diameter data during film winding, and film tension data during film winding, and preprocesses the film tension data to obtain preprocessed film tension data;

Tension prediction model building module: integrates film density data, film thickness data and roll diameter data, builds a tension prediction model, fills in missing values of film tension data, and generates a roll diameter-tension curve;

Overshoot penalty calculation module: calculates the overshoot amplitude and fluctuation time generated by each adjustment of the film tension by the PID system, and then calculates the overshoot penalty term;

Tension level classification model construction module: constructs a tension level classification model, and outputs the number of tension levels and corresponding tensions according to the maximum roll diameter;

PID control module: According to the number of tension levels and the corresponding tension, the film tension during the winding process is controlled through the PID system.

It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims (7)

1. A winding method based on artificial intelligence, characterized in that it comprises the following steps: S1: collecting film density data, film thickness data, roll diameter data of the film roll during film winding, and film tension data during film winding, and preprocessing the film tension data to obtain preprocessed film tension data; S2: Fusing the film density data, film thickness data and roll diameter data, and constructing a tension prediction model to fill in the missing values of film tension data and generate a roll diameter-tension curve; The tension prediction model first proposes position missing coding and missing distance coding to mark the position of the missing value and the distance between the missing value and the true value respectively, and then integrates the position missing coding and missing distance coding with the film tension data, inputs them into the gated recurrent unit, and predicts and fills the film tension data; S3: Calculate the overshoot amplitude and fluctuation time generated by each adjustment of the film tension by the PID system, and then calculate the overshoot penalty term; S4: construct a tension level classification model, and output the number of tension levels and corresponding tension according to the maximum roll diameter; The reward function of the tension level classification model uses a definite integral method to calculate the coil diameter-tension curve fitting degree, and the value function of the tension level classification model is calculated by a Q-learning method, and an ε-greedy strategy is used to update the tension level classification model; S5: The film tension during the winding process is controlled by the PID system according to the number of tension levels and the corresponding tension. 2. The artificial intelligence-based winding method according to claim 1, characterized in that, in the step S1, the roll diameter data is discrete time series data, and the film tension data corresponds to the roll diameter data one by one; The preprocessing of the film tension data comprises the following steps: using the spline interpolation method, performing a single interpolation on the film tension data under different roll diameters to obtain the interpolated film tension data, and the calculation method is: ; ; ; ; in, is the interpolated film tension data output by the cubic spline function. is the input of the cubic spline function, i is the interval index, To solve for the coefficients, is the curl diameter of the left endpoint of the i-th interval, is the film tension data corresponding to the left endpoint of the i-th interval; is the first-order derivative of the cubic spline function at the left endpoint of the i-th interval, is the second-order derivative of the cubic spline function at the left endpoint of the ith interval, is the first-order derivative of the cubic spline function at the left endpoint of the i+1th interval, is the second-order derivative of the cubic spline function at the left endpoint of the i+1th interval. 3. The artificial intelligence-based winding method according to claim 2, characterized in that step S2 comprises: S21: The film density data, film thickness data and roll diameter data are integrated and calculated as follows: ; ; in, is the fused data, For splicing operation, They are film density data and film thickness data, is the fused feature, It is a fully connected layer operation; S22: Construct a tension prediction model to predict the missing values of the film tension data and fill in the missing values of the film tension data. The calculation method is: ; ; ; ; ; in, is the input of the tension prediction model, is the missing code of the left endpoint of the i-th interval index, is the missing code of the left endpoint of the i-1th interval index, is the missing distance code for the left endpoint of the i-th interval index, is the missing distance code for the left endpoint of the i-1th interval index, is the attenuation code, is the prediction result of the missing value of the film tension data at the left endpoint of the i-th interval index, is the prediction result of the missing value of the film tension data at the left endpoint of the i-1th interval index, To take the exponential operation of a natural constant, To obtain the maximum value, is the attenuation parameter weight, is the attenuation bias vector, Operate for recurrent neural networks; S23: According to the prediction results, the roll diameter data and film tension data at each moment are counted to generate a roll diameter-tension curve. 4. The artificial intelligence-based winding method according to claim 3, characterized in that the step S3 comprises: S31: recording the overshoot amplitude generated by each adjustment of the film tension by the PID system; S32: When the PID system output oscillates, the oscillation start time is recorded; S33: When the oscillation amplitude returns to the stable range, the oscillation stop time is recorded and subtracted from the oscillation start time to obtain the fluctuation time of this adjustment; S34: Calculate the overshoot penalty term, the calculation method is: ; Among them, P is the overshoot penalty term, is the overshoot amplitude, is the fluctuation time, is the total time of this adjustment, is the overshoot amplitude weight, is the time weight. 5. The artificial intelligence-based winding method according to claim 4, characterized in that the step S4 comprises: S41: The state space of the tension level division model is designed as: the tension value sequence after the coil diameter-tension curve is segmented; the action space is: merging the current tension level and adding a new segmentation point to the current tension level; the tension value sequence takes 0 as the first value in the sequence and the maximum coil diameter as the last value in the sequence; S42: Design the reward function of the tension level classification model, which is calculated as follows: ; ; in, The reward value of the model divided into tension levels, is the influence weight of the overshoot penalty term, is the roll diameter-tension curve fitting degree, is the influence weight of the coil diameter-tension curve fitting, To take the absolute value, The tension is in the coil diameter from Change to The definite integral when S43: The value function of the design tension level classification model is calculated as follows: ; in, is the value function under the current state s and the current action a , is the learning rate, For the next state, For the next action, In order to make the next step corresponding to the maximum Q value, is the discount factor, To update the symbol; S44: Adopting the ε-greedy strategy, updating the tension level division model, and outputting the number of tension levels and the corresponding tension of each tension level. 6. The artificial intelligence-based winding method according to claim 5, characterized in that the step S5 comprises: S51: Initialize the proportional, integral and differential parameters of the PID system; S52: Real-time monitoring of film tension and roll diameter during the winding process; S53: According to the number of tension levels and the corresponding tension of each tension level, the film tension during the winding process is adjusted through the PID system. 7. An artificial intelligence-based winding system, characterized by comprising: Data acquisition and preprocessing module: collects film density data, film thickness data, film roll diameter data during film winding, and film tension data during film winding, and preprocesses the film tension data to obtain preprocessed film tension data; Tension prediction model building module: integrates film density data, film thickness data and roll diameter data, builds a tension prediction model, fills in missing values of film tension data, and generates a roll diameter-tension curve; Overshoot penalty calculation module: calculates the overshoot amplitude and fluctuation time generated by each adjustment of the film tension by the PID system, and then calculates the overshoot penalty term; Tension level classification model construction module: constructs a tension level classification model, and outputs the number of tension levels and corresponding tensions according to the maximum roll diameter; PID control module: According to the number of tension levels and the corresponding tension, the film tension during the winding process is controlled by the PID system; To realize the artificial intelligence-based winding method as described in any one of claims 1-6.