Deep learning-based stretch film manufacturing process optimization method and system

Abstract

The invention relates to a deep learning-based stretch film manufacturing process optimization method and system, and the method comprises the following steps: S1, obtaining and preprocessing an original process environment, an equipment state and operation record data, and obtaining a multi-modal data set; s2, according to the preprocessed multi-source heterogeneous data set, performing multi-modal deep learning based on a deep learning model to obtain feature representation output by the deep learning model; s3, process-quality relation modeling is carried out based on feature representation output by the deep learning model, and a process-quality causal relation graph is obtained; s4, according to the process-quality causal relationship graph, the quality target and a preset constraint condition, intelligent optimization is carried out on process parameters under multi-target constraint, and an optimal process parameter combination is obtained; and S5, closed-loop control is carried out according to the optimal process parameter combination, and dynamic optimization and adaptive adjustment are carried out according to real-time data.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing, and in particular to a winding film manufacturing process optimization method and system based on deep learning. BACKGROUND

[0002] With the rapid development of new materials and intelligent manufacturing technology, as an important basic material in the packaging and logistics industry, winding film (stretch film) puts forward higher requirements on its physical properties, thickness uniformity and product consistency. The production process of winding film is highly complex, involving precise coordination and dynamic regulation of multiple material ratios and process parameters (such as extrusion temperature, stretching speed, cooling conditions, tension, etc.). Traditional production mainly relies on manual experience and simple process window setting, which is difficult to fully grasp the product quality fluctuations caused by multi-parameter interaction, process drift and raw material batch difference, and is prone to high energy consumption, raw material waste and unstable product performance. SUMMARY

[0003] In order to solve the above problems, the purpose of the present application is to provide a winding film manufacturing process optimization method and system based on deep learning, which effectively improves the product yield and consistency and reduces the consumption of raw materials and energy.

[0004] To achieve the above purpose, the present application adopts the following technical solutions:

[0005] A winding film manufacturing process optimization method based on deep learning, comprising the following steps:

[0006] S1: Obtain the original process environment, equipment state, operation record data, and pre-process to obtain a multi-modal data set;

[0007] S2: Based on the pre-processed multi-source heterogeneous data set, perform multi-modal deep learning based on a deep learning model to obtain feature representation output by the deep learning model;

[0008] S3: Based on the feature representation output by the deep learning model, perform process-quality relationship modeling to obtain a process-quality causal relationship graph;

[0009] S4: According to the process-quality causal relationship graph, the quality target and the preset constraint condition, adopt process parameter intelligent optimization under multi-objective constraint to obtain an optimal process parameter combination;

[0010] S5: According to the optimal process parameter combination, perform closed-loop control, and dynamically optimize and adaptively adjust according to real-time data.

[0011] Further, the original process environment, equipment state, and operation record data are obtained as follows:

[0012] Through the field environment monitoring system, real-time collection of environmental parameters in the production area, including temperature, humidity, cleanliness;

[0013] The running state of the extruder, die, stretching device, and winding device is collected, including temperature, pressure, speed, motor power, tension, speed ratio, and the device start-stop, maintenance, alarm, and die change record are obtained, reflecting the device health and working state;

[0014] The operator's process adjustment, parameter setting, abnormal intervention, and shift rotation operation behavior are recorded, capturing the influence of human factors on process fluctuations and abnormalities;

[0015] Collecting raw material batch information, incoming inspection results, and online and offline product quality inspection data (such as thickness, transparency, tensile strength, defect type, etc.);

[0016] The collected data is summarized, time-stamped, and aligned with various data streams to ensure data timing and source consistency; and the data is standardized and classified, distinguishing between process parameters, device status, raw material information, output quality, and different data blocks of operation activities.

[0017] Further, preprocessing, specifically as follows:

[0018] For mixed abnormal values and isolated points in real manufacturing site data (such as abnormal jumps caused by device failure, sensor failure, power failure, etc.), box plot method is used for basic threshold detection; at the same time, combined with local outlier factor detection algorithm, abnormal data points in multi-dimensional feature space are automatically identified; for known fault conditions and invalid conditions, event log is used for synchronous auxiliary removal;

[0019] For inevitable short packet loss in real-time data collection and data missing caused by instrument maintenance, first, linear interpolation is used to preliminarily complete single variable time series missing, and K nearest neighbor interpolation is used for multi-variable correlation structure to further improve data integrity;

[0020] For high-frequency noise interference, time series smoothing is used to restore the true trend and signal content of the data;

[0021] According to the process mechanism and analysis requirements, derived features are constructed from raw data, including mean, volatility, change rate, and device load rate;

[0022] Finally, the data is integrated into a unified structure, forming a multi-modal feature data set, which includes continuous time series data X seq , event parameters X cat and image signal stream X img , realizing comprehensive digital description of the entire manufacturing process.

[0023] Further, according to the pre-processed multi-source heterogeneous data set, multi-modal deep learning is carried out based on a deep learning model to obtain feature representation output by the deep learning model, as follows:

[0024] Different modal data adopts different neural network branches to realize deep feature extraction, including a time sequence branch, an event branch and an image branch, as follows:

[0025] The time sequence branch adopts LSTM to capture time sequence dependence of process parameters:

[0026]

[0027] Where x t is a parameter at the t time, h t is a hidden layer state, and z seq is a time sequence feature.

[0028] The event branch adopts a multi-layer perception MLP to embed a category variable into a low-dimensional space:

[0029] z cat =MLP(X cat );

[0030] The image branch adopts a CNN, and X img is feature extracted through a convolutional neural network CNN:

[0031] z img =CNN(X img );

[0032] Through splicing fusion, multi-modal feature fusion is carried out to obtain fused features X:

[0033] X=[z seq ;z cat ;z img ;]。

[0034] Further, based on the feature representation output by the deep learning model, process-quality relationship modeling is carried out to obtain a process-quality causal relationship graph, as follows:

[0035] The following causal structure equation is established: y j =f j (X,Q,E j )

[0036] Where y j is a jth quality index, Q is a device representation, and Ej is external noise.

[0037] LiNGAM structure learning algorithm is used to automatically mine the direct influence causal paths between features through conditional independence test and variable order arrangement;

[0038] For the dynamic relationship between process parameters and quality indicators, Granger causality test is used to determine the actual driving ability of the previous process variables on the subsequent quality results;

[0039] All process parameters, feature representations and quality indicators are regarded as nodes, and the significant causal relationships between nodes are sorted and confirmed using the above analysis method. The influence direction and strength are represented by directed edges, and a complete causal relationship structure diagram G=(V,E,W) is formed, where V is the node set, E is the directed edge set, and W is the weight set.

[0040] Further, LiNGAM structure learning algorithm is used to automatically mine the direct influence causal paths between features through conditional independence test and variable order arrangement, as follows:

[0041] Collect all variables to be analyzed in the form of observation data sample matrix: X=[x1,x2,...,x d ] T Where d is the total number of variables, and n is the sample number;

[0042] LiNGAM assumes that the relationship between variables is represented by the following linear non-Gaussian, acyclic structure equation:

[0043] X=BX+e where B is the causal coefficient matrix of the lower triangular matrix; e is the noise term with independent components and non-Gaussian;

[0044] Infer the order of a variable so that B is in the form of a lower triangular matrix, and e=e1,...,e d are independent of each other:

[0045] ICA is used to decompose the data to estimate the mixing matrix, thereby inferring the possible causal order of variables;

[0046] Through testing of the causal order, select the arrangement that meets the conditional independence assumption and has the best statistical score;

[0047] Estimate B using the least squares method, and test the conditional independence relationship in the variable and feature ordering process:

[0048] If x b has been arranged before x a and B ab ≠0, it means that x b directly affects x a , and if x a is added to a certain set S, x bIf the condition is independent, no directed edge is established.

[0049] Traverse all B ab , if B ab ≠0, then establish a directed edge b→a in the causal graph, and the edge weight is the size of B ab .

[0050] Further, for the dynamic relationship between process parameters and quality indicators, Granger causality test is used to determine the actual driving ability of the preceding process variables on the subsequent quality results, as follows:

[0051] Let X={x t} be the time series data of process parameters, Y={y t} be the corresponding quality indicator time series data, and t=1,2,…,T be different time points;

[0052] Two baseline prediction models are established: including an autoregressive model that only uses the quality indicator itself to predict, and an extended regression model that further uses the historical values of process parameters X as input variables to assist in predicting Y based on the autoregressive model;

[0053] Compare the prediction residuals of the two models. If the introduction of process parameters X reduces the prediction error value of Y by more than a threshold value, it is determined that X has a Granger causal relationship with Y;

[0054] All significant process parameters and quality indicators are used as nodes of the graph, and directed edges are added between them according to the Granger test results.

[0055] Further, according to the process-quality causal relationship graph, the quality target and the preset constraint conditions, the intelligent optimization of process parameters under multi-objective constraints is used to obtain the optimal process parameter combination, as follows:

[0056] A. Extract the key path from the process-quality causal relationship graph to determine the main control process parameter subset P:

[0057] Initialize the key path set P, and for each quality indicator y, initialize all possible path set P y .

[0058] For each process parameter x∈V and x as an input node: use depth-first search (DFS) to find all paths from x to y; add the found path to P y , and calculate the total path weight w(p) for each path:

[0059] ;

[0060] According to the total path weight w(p), obtain the path importance score:

[0061]

[0062] Ranking P by path importance score in descending order y , select the top k paths with the largest weights to P;

[0063] B. Based on the key path set P, identify the master control process parameter subset X* that has the greatest impact on the quality indicators:

[0064] Initialize the parameter importance dictionary D = {}, for each key path p e P extract the process parameter nodes X p = {x | x e p and x e X} on the path;

[0065] For each x e X p , calculate the local importance I p (x) of x on the path p:

[0066] ;

[0067] Update D = I p (x) x w(p), rank D by importance in descending order;

[0068] Apply the Pareto principle, select X* according to the preset adjustment;

[0069] C. For all quality indicators, set the target value y j * and the allowable deviation range d j according to historical data, industry standards and actual needs; assign different weights w j according to the sensitivity of quality to customer or enterprise targets;

[0070] Within the physically feasible range of master control parameters , use Latin hypercube sampling method to randomly generate N groups of parameters to form the initial population;

[0071] D. According to the process-quality causal relationship diagram, substitute each group of process parameters X k to predict the corresponding quality indicators Y k ; and calculate the objective function; if there is a violation of the constraint, add a penalty;

[0072] The multi-objective comprehensive deviation is:

[0073] ;

[0074] E. Evaluate the objective function and constraint condition of each individual in each generation, and based on the particle swarm optimization algorithm, use the objective function and constraint condition of each parameter group to perform crossover and mutation operations on the population to generate new candidate parameter groups;

[0075] F. Iteration is continued, and if the optimal solution is converged or the maximum number of iterations is reached, the optimal parameter set is output.

[0076] Further, closed-loop control is performed according to the optimal process parameter combination, and dynamic optimization and adaptive adjustment are performed according to real-time data, as follows:

[0077] The optimal process parameter combination obtained in the multi-objective optimization link is used as the initial setting value of each process link and is deployed on the equipment or the automatic control system, and real-time acquisition of actual data of key process parameters and core quality indexes is performed.

[0078] In each production cycle or time slice, the actually measured quality data is compared with the pre-set target quality value, if it is found that all monitoring points are stable in the expected quality interval, the system only records and archives the data, and does not need to adjust the set parameters.

[0079] Online control charts are established for the key process parameters and the quality indexes, and the statistical distribution and the change trend of each data are dynamically tracked; by setting a reasonable alarm line, process drift, sudden abnormality or cumulative effect of multiple small amplitude deviations are captured; once an abnormal pattern of a key process parameter or a quality index is detected, a pre-warning is automatically triggered, and a process-quality causal relationship diagram is used to determine the main control parameter causing the deviation and calculate a compensation scheme.

[0080] A winding film manufacturing process optimization system based on deep learning includes a processor, a memory and a computer program stored on the memory, and when the processor executes the computer program, the steps of the winding film manufacturing process optimization method based on deep learning are specifically executed.

[0081] The application has the following beneficial effects:

[0082] 1. The application can deeply explore the nonlinear and dynamic correlation between process parameters and product quality by using a deep learning model to uniformly model multi-dimensional and multi-source real-time process data, greatly improves the accuracy of parameter optimization and quality prediction, and provides strong data support for scientific decision-making, significantly exceeding traditional experience and simple statistical analysis methods.

[0083] 2. In the process parameter optimization stage, not only is the key path and main control parameter screening based on the causal relationship introduced to improve the effectiveness of the optimization dimension, but also a multi-objective and multi-constraint evolutionary intelligent algorithm is used to consider multiple quality indexes and actual engineering constraint conditions, and the process parameter optimization efficiency and result feasibility are comprehensively improved.

[0084] 3、The present application effectively improves product yield and consistency, reduces raw material and energy consumption, reduces human intervention in the production process, and realizes cost reduction and efficiency increase through intelligent process regulation and continuous data-driven optimization. BRIEF DESCRIPTION OF DRAWINGS

[0085] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0086] The present application will be further described in detail below in combination with the drawings and specific embodiments:

[0087] Reference Figure 1 In the present embodiment, a winding film manufacturing process optimization method based on deep learning is provided, comprising the following steps:

[0088] S1: Obtain the original process environment, equipment state, operation record data, and pre-process to obtain a multi-modal data set;

[0089] S2: Based on the pre-processed multi-source heterogeneous data set, multi-modal deep learning is performed based on a deep learning model to obtain feature representation output by the deep learning model;

[0090] S3: Based on the feature representation output by the deep learning model, process-quality relationship modeling is performed to obtain a process-quality causal relationship graph;

[0091] S4: According to the process-quality causal relationship graph, quality target and preset constraint condition, process parameter intelligent optimization under multi-objective constraint is adopted to obtain an optimal process parameter combination;

[0092] S5: According to the optimal process parameter combination, closed-loop control is performed, and dynamic optimization and adaptive adjustment are performed according to real-time data.

[0093] Further, the original process environment, equipment state, and operation record data are obtained as follows:

[0094] Through the field environment monitoring system, the environmental parameters of the production area are collected in real time, including temperature, humidity, cleanliness;

[0095] The running state of the extruder, die, stretching device, and winding device is collected, including temperature, pressure, speed, motor power, tension, speed ratio, and the device start-stop, maintenance, alarm, and die change record are obtained at the same time, reflecting the device health and working state;

[0096] The operator's process adjustment, parameter setting, abnormal intervention, and shift rotation operation behavior are recorded, and the influence of human factors on process fluctuation and abnormality is captured;

[0097] Collecting raw material batch information, incoming inspection results, and online and offline product quality inspection data (such as thickness, transparency, tensile strength, defect type, etc.);

[0098] The obtained data is summarized, time-stamped, aligned with each data stream, and ensured to be consistent in terms of data timing and source. The data is then standardized and classified, with process parameters, device status, raw material information, output quality, and different data blocks of operation activities being distinguished.

[0099] In this embodiment, preprocessing is performed as follows:

[0100] For mixed abnormal values and outliers in real data in the manufacturing site (such as abnormal jumps caused by device failure, sensor failure, power failure, etc.), box plot method is used for basic threshold detection. At the same time, combined with the local outlier factor detection algorithm, abnormal data points in the multi-dimensional feature space are automatically identified. For known fault conditions and invalid conditions, event logs are used to assist in removing them.

[0101] For inevitable short packet loss in real-time data collection and data missing caused by instrument maintenance, linear interpolation is first used to preliminarily complete the single-variable time series missing. For multi-variable correlation structure, K-nearest neighbor interpolation is used to further improve the completeness of the data.

[0102] For high-frequency noise interference, time series smoothing is used to restore the true trend and signal content of the data.

[0103] According to the process mechanism and analysis requirements, derived features are constructed from the original data, including mean, volatility, change rate, and device load rate.

[0104] Finally, the data is integrated into a unified structure, forming a multi-modal feature dataset, which includes continuous time series data X seq (including temperature, pressure, speed, draw ratio, etc., arranged as a multi-dimensional time series matrix), event parameters X cat (including batch information, operation category, One-hot or embedding encoding, mapped to a low-dimensional vector), and image signal stream X img , achieving comprehensive digital description of the entire manufacturing process.

[0105] In this embodiment, based on the pre-processed multi-source heterogeneous data set, multi-modal deep learning is performed based on a deep learning model to obtain feature representations output by the deep learning model, as follows:

[0106] Different modal data uses different neural network branches to achieve deep feature extraction, including time series branch, event branch, and image branch, as follows:

[0107] The time sequence branch adopts LSTM to capture the time sequence dependence of process parameters:

[0108]

[0109] where x t is the parameter at the t-th moment, h t is the hidden layer state, and the final output is the state z seq at the last moment, which is the time sequence feature.

[0110] The event branch adopts a multi-layer perception network MLP to embed the category variable into a low-dimensional space:

[0111] z cat =MLP(X cat );

[0112] The image branch adopts CNN, and X img is extracted through a convolutional neural network CNN:

[0113] z img =CNN(X img );

[0114] Through splicing fusion, multi-modal feature fusion is performed to obtain the fused feature X:

[0115] X=[z seq ;z cat ;z img ;]。

[0116] In this embodiment, based on the feature representation output by the deep learning model, process-quality relationship modeling is performed to obtain a process-quality causal relationship graph, as follows:

[0117] The following causal structure equation is established: y j =f j (X,Q,E j )

[0118] where y j is the j-th quality index, Q is the device representation, and Ej is the external noise.

[0119] The LiNGAM structure learning algorithm is used to automatically mine the direct influence causal paths between features through conditional independence test and variable order arrangement.

[0120] For the dynamic relationship between process parameters and quality indicators, Granger causality test is used to determine the actual driving ability of the preceding process variables on the subsequent quality results.

[0121] Taking all process parameters, feature characterizations, and quality indicators as nodes, using the above analysis method to sort out and confirm the significant causal relationships between nodes, and representing the influence direction and strength with directed edges to form a complete causal relationship structure diagram G = (V, E, W), where V is the set of nodes, E is the set of directed edges, and W is the set of weights.

[0122] In this embodiment, the LiNGAM structure learning algorithm is used to automatically mine the direct influence causal paths between features through conditional independence testing and variable order arrangement, as follows:

[0123] Collect all variables to be analyzed (process parameters, depth features, equipment / environment factors, quality indicators, etc.) and organize them in the form of an observed data sample matrix: X = [x1, x2,..., x d T where d is the total number of variables and n is the number of samples;

[0124] LiNGAM assumes that the relationship between variables is represented by the following linear non-Gaussian, acyclic structural equation:

[0125] X = BX + e where B is a lower triangular causal coefficient matrix (i.e., B ab ≠0 only when b < a); e is a noise term with independent and non-Gaussian components;

[0126] Infer a variable permutation order (causal order) such that B is in lower triangular form and e = e1,..., e d are mutually independent:

[0127] Use ICA to decompose the data, estimate the mixing matrix, and thus infer the possible causal order of variables;

[0128] Through testing the causal order, select the permutation that satisfies the conditional independence assumption and has the optimal statistical score;

[0129] Estimate B using the least squares method. During the process of variable and feature sorting, test the conditional independence relationship (such as x a ⊥x b ∣ S):

[0130] If x b has been ranked before x a and B ab ≠0, it means that x b directly affects x a . If after adding a certain set S, x a and x b are conditionally independent, then no directed edge is established;

[0131] Traverse all B ab , if B​ab ≠0, then a directed edge from b to a is established in the causal graph, and the edge weight is B ab .

[0132] In this embodiment, for the dynamic relationship between process parameters and quality indicators, Granger causality test is used to determine the actual driving ability of the preceding process variables on the subsequent quality results, as follows:

[0133] Let X={x t} be the time series data of process parameters, Y={y t} be the corresponding time series data of quality indicators, and t=1,2,…,T be different time points;

[0134] Two baseline prediction models are established: an autoregressive model that only uses the quality indicators themselves to predict, and an extended regression model that further uses the historical values of process parameters X as input variables to assist in predicting Y based on the autoregressive model;

[0135] Compare the prediction residuals of the two models. If the introduction of process parameters X reduces the prediction error of Y by more than a threshold value, it is determined that X has a Granger causal relationship with Y. Calculate the p-value through Granger F-value test. If p<0.05, the causality is significantly established;

[0136] All significant process parameters and quality indicators are used as nodes of the graph, and directed edges are added between them according to the Granger test results. Each directed edge represents the dynamic causal driving force of the preceding process parameters on the subsequent quality results. The weight of the edge can be the sum of the Granger F-value or the regression coefficient, reflecting the strength of the causal influence.

[0137] In this embodiment, according to the process-quality causal relationship graph, the quality target and the preset constraint conditions, the intelligent optimization of process parameters under multi-objective constraints is used to obtain the optimal process parameter combination, as follows:

[0138] A. Extract the critical path from the process-quality causal relationship graph to determine the main control process parameter subset P:

[0139] Initialize the critical path set P, and for each quality indicator y, initialize the set of all possible paths P y ;

[0140] For each process parameter x∈V and x as an input node: use depth-first search (DFS) to find all paths from x to y; add the found paths to P y , and calculate the total weight w(p) of each path:

[0141] ;

[0142] According to the total weight of the path w(p), the path importance score is obtained:

[0143]

[0144] The P is arranged in descending order of path importance score y , and the top k paths with the largest weights are selected to join P;

[0145] B. Based on the key path set P, the master control process parameter subset X* with the greatest impact on the quality index is identified:

[0146] Initialize the parameter importance dictionary D = {}, and for each key path p e P, extract the process parameter nodes X p on the path = {x | x e p and x e X};

[0147] For each x e X p , calculate the local importance I p (x) of x on the path p:

[0148] ;

[0149] Update D = I p (x) x w(p), and arrange D in descending order of importance;

[0150] Apply the Pareto rule, and according to the preset adjustment, select X* (the top 20% parameters with the greatest contribution or the parameters with cumulative importance reaching 80% as X*).

[0151] C. For all quality indicators, set target values y j * and allowable deviation ranges d j according to historical data, industry standards and actual needs; and assign different weights w j according to the sensitivity of quality to customer or enterprise targets;

[0152] Within the physically feasible range of master control parameters , use Latin hypercube sampling method to randomly generate N groups of parameters to form the initial population;

[0153] D. According to the process-quality causal relationship diagram, substitute each group of process parameters X k to predict the corresponding quality indicators Y k ; and calculate the objective function; if there is a violation of the constraints (such as parameter or quality out of bounds), add a penalty;

[0154] The multi-objective comprehensive deviation is:

[0155] ;

[0156] E. Evaluate the objective function and constraint condition of each individual in each generation, and based on the particle swarm optimization algorithm, use the objective function and constraint condition of each parameter set to perform crossover and mutation operations on the population to generate new candidate parameter sets;

[0157] F. Continue iteration, and if the optimal solution is converged or the maximum number of iterations is reached, output the optimal parameter set.

[0158] In this embodiment, closed-loop control is performed according to the optimal process parameter combination, and dynamic optimization and adaptive adjustment are performed according to real-time data, as follows:

[0159] The optimal process parameter combination obtained in the multi-objective optimization link is used as the initial setting value of each process link and is deployed on the equipment or automatic control system. Real-time acquisition of key process parameters (such as temperature, pressure, stirring speed, etc.) and core quality indicators (such as size, strength, purity, etc.) is performed. These real-time data are continuously transmitted to the quality monitoring and optimization module through industrial Internet of Things or automatic acquisition interface.

[0160] In each production cycle or time slice, the actual measured quality data are compared with the pre-set target quality value. If it is found that all monitoring points are stable in the expected quality interval, the system only records and archives the data, and does not need to adjust the set parameters.

[0161] Online control charts (such as mean control chart, range chart, CUSUM chart, etc.) are established for key process parameters and quality indicators, and the statistical distribution and change trend of each data are dynamically tracked. By setting a reasonable alarm line (such as ±3σ), process drift, sudden abnormality, or cumulative effect of multiple small deviations are captured. Once an abnormal pattern (for example, exceeding the control limit point, continuous points showing unilateral deviation, etc.) of a key process parameter or quality indicator is detected, a pre-warning is automatically triggered, and a process-quality causal relationship chart is used to determine the main control parameter causing the deviation and calculate a compensation scheme.

[0162] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) containing computer-usable program code.

[0163] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0164] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0165] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0166] The above description is only preferred embodiments of the present application, and is not intended to limit the present application to other forms described above. Any person skilled in the art can make modifications or improvements to the above-mentioned disclosed technical content as equivalent embodiments. However, any simple modification, equivalent change and modification of the above-mentioned embodiments without departing from the technical solution of the present application, according to the technical essence of the present application, still belongs to the protection scope of the present application.

Claims

1. A deep learning-based method for optimizing the manufacturing process of stretch film, characterized in that, Includes the following steps: S1: Acquire raw process environment, equipment status, and operation record data, and preprocess them to obtain a multimodal dataset; S2: Based on the preprocessed multi-source heterogeneous dataset, perform multimodal deep learning based on the deep learning model to obtain the feature representation output by the deep learning model; S3: Based on the feature representation output by the deep learning model, perform process-quality relationship modeling and obtain the process-quality causal relationship graph; S4: Based on the process-quality cause-and-effect relationship diagram, quality objectives and preset constraints, intelligent optimization of process parameters under multi-objective constraints is adopted to obtain the optimal combination of process parameters; S5: Perform closed-loop control based on the optimal combination of process parameters, and dynamically optimize and adaptively adjust based on real-time data.

2. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 1, characterized in that, The acquisition of the original process environment, equipment status, and operation record data is as follows: The on-site environmental monitoring system collects environmental parameters of the production area in real time, including temperature, humidity, and cleanliness. For extruders, dies, stretching devices, and winding devices, collect their operating status, including temperature, pressure, speed, motor power, tension, and speed ratio. At the same time, acquire equipment start-up and shutdown, maintenance, alarm, and die change records; record operator operation behaviors such as process adjustments, parameter settings, abnormal interventions, and shift rotations. Collect raw material batch information, incoming inspection results, and various online and offline finished product quality inspection data; The acquired data is aggregated, timestamps are unified, and data streams are aligned to ensure consistency in data sequence and source. The data is then standardized, coded, and categorized into different data modules, including process parameters, equipment status, raw material information, output quality, and operational activities.

3. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 2, characterized in that, The preprocessing is as follows: To address outliers and isolated points mixed in with real manufacturing data, a box plot method is used for basic threshold detection. At the same time, a local outlier detection algorithm is combined to automatically identify abnormal data points in the multidimensional feature space. For known fault conditions and invalid conditions, event log synchronization is used to assist in their removal. To address the short-term packet loss during real-time data acquisition and data gaps caused by instrument maintenance, linear interpolation is first used to preliminarily fill in the missing time series data for single variables. For multivariate correlation structures, K-nearest neighbor interpolation is used. For high-frequency noise interference, time-series smoothing is used to restore the true trend of data and signal content; Based on the process mechanism and analysis requirements, derived features are constructed from the raw data, including mean, volatility, rate of change, and equipment load rate; Finally, the data is integrated into a unified structure to form a multimodal feature dataset, which contains continuous time-series data X. seq Event parameter X cat and image signal stream X img .

4. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 3, characterized in that, The process involves performing multimodal deep learning based on a deep learning model using the preprocessed multi-source heterogeneous dataset to obtain the feature representation output by the deep learning model, as detailed below: Different neural network branches are used to achieve deep feature extraction for different modal data, including time-series branches, event branches, and image branches, as detailed below: Timing branches use LSTM to capture the timing dependencies of process parameters: ； Where, x t h is the parameter at time t. t The hidden state is given, and the final output is the state z at the last moment. seq It is a time-series feature; Event branching uses a multilayer perceptron (MLP) to embed categorical variables into a low-dimensional space: With cat =MLP(X cat ); Image branching uses CNN, X img Feature extraction using a convolutional neural network (CNN): z img =CNN(X img ); By splicing and fusing, multimodal feature fusion is performed to obtain the fused feature X: X=[z seq ;z cat ;z img ;]。 5. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 4, characterized in that, The feature representation based on the output of the deep learning model is used to model the process-quality relationship and obtain a process-quality causal relationship graph, as detailed below: Establish the following causal structure equation: y j =f j (X,Q,E j ) Where y j Let Q be the j-th quality indicator, and E be the equipment characteristic. j External noise; By employing the LiNGAM structure learning algorithm, and through conditional independence testing and variable ordering, the direct causal paths of influence between features are automatically discovered. To determine the dynamic relationship between process parameters and quality indicators, Granger causality tests are used to assess the actual driving force of preceding process variables on subsequent quality results. All process parameters, characteristic features, and quality indicators are treated as nodes. The above analysis method is used to sort out and confirm the significant causal relationships between the nodes. Directed edges are used to represent the direction and intensity of the influence, forming a complete causal relationship structure graph G=(V,E,W), where V is the set of nodes, E is the set of directed edges, and W is the set of weights.

6. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 5, characterized in that, The LiNGAM structure learning algorithm is used to automatically discover the causal paths of direct influence between features through conditional independence tests and variable ordering, as detailed below: Collect all variables to be analyzed and organize them in the form of an observed data sample matrix: X=[x1,x2,...,x d ] T in, d represents the total number of variables, and n represents the number of samples; LiNGAM assumes that the relationships between the variables are represented by the following linear non-Gaussian, acyclic structure equations: X = BX + e, where B is a lower triangular causality coefficient matrix; e is a noise term whose components are independent and non-Gaussian. Infer the order of variables such that B is a lower triangle and e = e1,...,e d Independent of each other: By using ICA to decompose the data and estimate the mixture matrix, the possible causal order of the variables can be inferred. By testing the causal order, the arrangement that satisfies the conditional independence assumption and has the optimal statistical score is selected; Estimate B using the least squares method, and test the conditional independence relationship during the variable and feature ranking process: If x b Already ranked in x a Before and B ab ≠0, indicating that x b Directly affects x a If x is added to a set S, then x a With x b If the conditions are independent, then no directed edge is created; Traverse all B ab If B ab If the weight of the edge is not equal to 0, then a directed edge from b to a is constructed in the causal graph, and the weight of the edge is B. ab Size.

7. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 5, characterized in that, Regarding the dynamic relationship between process parameters and quality indicators, the Granger causality test is used to determine the actual driving force of preceding process variables on subsequent quality results, as detailed below: Let X = {x} t } represents the time-series data of the process parameters, Y={y t } represents the time series data of the corresponding quality indicators, where t=1,2,…,T represents different time points; Two baseline prediction models were established: an autoregressive model that uses only the quality indicators themselves for prediction, and an extended regression model that further incorporates the historical values ​​of process parameter X as input variables to assist in the prediction of Y based on the autoregressive model. By comparing the prediction residuals of the two models, if the prediction error of Y caused by the reduction of process parameter X exceeds the threshold, then X is determined to have Granger causality with Y. All significant process parameters and quality indicators are treated as nodes in the graph, and directed edges are added between them based on the Granger test results.

8. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 5, characterized in that, Based on the process-quality causal relationship diagram, quality objectives, and preset constraints, intelligent optimization of process parameters under multi-objective constraints is employed to obtain the optimal combination of process parameters, as detailed below: A. Extract the critical path from the process-quality cause-effect diagram and determine the subset P of key process parameters: Initialize the critical path set P, and for each quality metric y, initialize the set of all possible paths P. y ; For each process parameter x∈V and x is an input node: use Depth-First Search (DFS) to find all paths from x to y; add the found paths to P. y Calculate the total path weight w(p) for each path: ； Obtain the path importance score based on the total path weight w(p): ； Ranked in descending order of path importance score. y Select the k paths with the largest weights and add them to P; B. Based on the critical path set P, identify the subset X* of key process parameters that have the greatest impact on quality indicators: Initialize the parameter importance dictionary D={}, and for each critical path p∈P, extract the process parameter nodes X on the path. p ={x | x∈p and x∈X}; For each x∈X p Calculate the local importance I of x on path p. p (x): ； Update D= I p (x) × w(p), arranged in descending order of importance, D; Apply the Pareto principle and adjust the selection of X* according to the preset; C. Set target values ​​y for all quality indicators, taking into account historical data, industry standards, and actual needs. j * and allowable deviation range δ j ; assign different weights according to the sensitivity of quality to customer or corporate objectives. j ; Within the physically feasible range of the master control parameters Within the population, N sets of parameters are randomly generated using the Latin hypercube sampling method to form the initial population; D. Based on the process-quality cause-and-effect diagram, assign each set of process parameters X... k Substitute the values ​​and predict the corresponding quality index Y. k ; and calculate the objective function; if there is a violation of constraints, increase the penalty points; The multi-objective overall bias is: ； E. Evaluate the objective function and constraints of all individuals in each generation, and based on the particle swarm optimization algorithm, use the objective function and constraints of each parameter group to perform crossover and mutation operations on the population to generate new candidate parameter groups; F. Continue iterating. If the solution converges to the optimal solution or the maximum number of iterations is reached, the optimal parameter set is output.

9. The deep learning-based method for optimizing the manufacturing process of stretch film according to claim 1, characterized in that, The closed-loop control based on the optimal combination of process parameters, and the dynamic optimization and adaptive adjustment based on real-time data, are described in detail below: The optimal combination of process parameters obtained in the multi-objective optimization process is used as the initial setpoint for each process step and deployed on the equipment or automated control system to collect the actual data of key process parameters and core quality indicators in real time. In each production cycle or time slice, the actual measured quality data is compared with the preset target quality value. If all monitoring points are found to be stable within the expected quality range, the system only records and archives the data without adjusting the settings. Online control charts are established for key process parameters and quality indicators to dynamically track the statistical distribution and changing trends of various data. By setting reasonable alarm thresholds, the cumulative effects of process drift, sudden anomalies, or multiple small deviations are captured. Once an abnormal mode is detected in a key process parameter or quality indicator, an early warning is automatically triggered. At the same time, the process-quality causal relationship diagram is used to determine the main control parameter causing the deviation and calculate the compensation scheme.

10. A deep learning-based system for optimizing the manufacturing process of stretch film, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the deep learning-based method for optimizing the manufacturing process of a stretch film as described in any one of claims 1-9.

Images