Machine learning based adaptive control system for plastic extrusion lamination process

Abstract

The present application relates to a kind of machine learning-based plastic extrusion laminating process adaptive control system, specifically relates to the field of industrial process control and intelligent manufacturing technology, by source domain basic process kinetics model pre-training module extraction cross-line general process law, by the feature distribution deviation of old and new production line is eliminated to the module of antagonistic domain alignment, again by fast adaptive capacity training module gives complete model only needs one-step gradient update to adapt to the ability of new production line, finally by target production line fine-tuning and deployment module utilize a small amount of normal operation data to complete accurate fine-tuning and put into real-time control;Each module is connected in turn, so that complete model retains existing knowledge while quickly adapting to new conditions, significantly shortening the new production line commissioning cycle, reducing raw material waste and equipment wear and tear, improving film thickness control accuracy and production stability, realizing the automatic reuse and low-cost deployment of cross-line process knowledge.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and intelligent manufacturing technology, and more specifically, to an adaptive control system for plastic extrusion lamination process based on machine learning. Background Technology

[0002] Plastic extrusion lamination technology is widely used in the large-scale production of packaging materials, waterproof membranes, composite hoses, and other products. Its core lies in uniformly coating molten plastic onto the surface of a substrate to form a composite layer with specific thickness and adhesion strength. With the diversification of market demand and capacity expansion, large manufacturing enterprises often configure multiple extrusion lamination units within the same group. These units differ significantly in hardware parameters such as screw configuration, die width, number of heating zones, and cooling methods. For example, single-screw and twin-screw structures result in different thermal inertia and shear characteristics, and wide-width and narrow-width units exhibit different lateral film thickness distribution patterns. Simultaneously, the environmental temperature and humidity, supply... There are also significant differences in electrical stability and the parameter adjustment habits of operators. When a new production line is built and put into operation, it is difficult to directly train an effective adaptive control model for it due to the lack of sufficient historical operating data. If data is collected from scratch and gradually optimized, it usually takes several weeks or even months to debug. During this period, key parameters such as screw speed, melt temperature, and traction speed can only be adjusted repeatedly by the manual experience of senior engineers. This is not only inefficient, but also prone to uneven film thickness, insufficient tensile strength, or even film breakage accidents due to human error, resulting in a large amount of raw material waste and equipment wear, which seriously affects the company's production capacity ramp-up speed and market delivery capability.

[0003] To address the aforementioned issues, existing technologies primarily employ two solutions: First, relying entirely on expert on-site debugging, determining initial process parameters through empirical formulas and trial-and-error methods, and then gradually refining them based on product quality feedback. This method is highly dependent on personal experience and cannot guarantee consistency, with human resource bottlenecks being particularly prominent when multiple production lines are debugged in parallel. Second, attempting to directly migrate mature control models from existing production lines to new production lines. However, due to inherent differences in process characteristics between different units—for example, variations in screw geometry leading to changes in melt residence time distribution, and differences in heating power density causing temperature response curve shifts—there is a significant domain shift in the data distribution between the source and target domains. Directly reused models exhibit a sharp decline in performance on new production lines, even worse than manual adjustments. The debugging is even worse; in recent years, a few studies have introduced fine-tuning transfer learning methods, which update parameters using a small amount of labeled data in the target domain based on the pre-trained model in the source domain. However, fine-tuning usually requires dozens or even hundreds of labeled samples covering the complete working conditions to converge, and new production lines lack such high-quality labeled data in the early stages of debugging. Therefore, the actual effect of fine-tuning methods is still limited. In summary, there is an urgent need for a technical solution that can achieve rapid model adaptation under conditions of very few target domain samples. This solution should not only fully inherit the general process rules contained in the source domain production line, but also effectively overcome the differences in feature distribution between different production lines, so as to truly realize the reuse of cross-production line process knowledge and the rapid deployment of adaptive control systems. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an adaptive control system for plastic extrusion lamination process based on machine learning, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a source domain basic process dynamics model pre-training module, an adversarial domain alignment module, a rapid adaptability training module, and a target production line one-step fine-tuning and deployment module, which are connected in sequence via communication;

[0006] The source domain basic process dynamics model pre-training module collects historical source domain data from multiple existing extrusion lamination production lines. The historical source domain data includes multi-channel time series of melt temperature, die pressure, screw speed, traction speed, and film thickness measurements. The historical source domain data is segmented into source domain samples by a fixed time window. A Transformer-based temporal prediction network is used as the basic process dynamics model. The basic process dynamics model includes a feature extractor and a prediction head. The model is trained on the source domain samples with the total loss function consisting of prediction loss and feature independence loss weighted together as the objective, and a parameterized basic model is output.

[0007] Adversarial Domain Alignment Module: Receives the feature extractor from the base model and introduces a domain discriminator. Through adversarial training, it minimizes the difference between the hidden feature distribution of the source domain and the hidden feature distribution of the target domain. The target domain data is taken from unlabeled short-term running data during the no-load or low-load debugging phase of the new production line. After training, the domain discriminator is discarded, and the domain-aligned feature extractor is output.

[0008] Rapid Adaptability Training Module: A lightweight regression head is added to the domain-aligned feature extractor to form a complete model. The model-independent meta-learning algorithm is used to construct multiple virtual tasks from the historical data of the source domain. Each virtual task contains a support set and a query set. Through inner and outer loop optimization, the complete model can adapt to the new production line with only one gradient update and output the meta-initialization parameters of the complete model.

[0009] The target production line fine-tuning and deployment module collects half an hour of normal operation data after the new production line is put into operation. The normal operation data is divided into a support set and a validation set. The complete model with the initial values ​​of the meta-initialization parameters is updated once by gradient descent on the support set to obtain the fine-tuned complete model. The fine-tuned complete model is then deployed in the adaptive control of the new production line, and the correction amount of the process parameters is output in real time.

[0010] In a preferred embodiment, in the source domain basic process dynamics model pre-training module, the source domain historical data comes from multiple existing extrusion laminating production lines. The source domain historical data of each existing extrusion laminating production line is a multi-channel time series, which includes melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values. The sampling frequency is uniformly one hertz. Source domain samples are obtained by sliding and intercepting source domain historical data with a fixed time window length of sixty seconds. Each source domain sample contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values ​​for sixty consecutive seconds within the fixed time window. The values ​​of these five channels are stacked in chronological order to form an input tensor. The corresponding prediction target is the film thickness change in the next ten seconds. The film thickness change is defined as the difference between the film thickness measurement value ten seconds after the end of the fixed time window and the film thickness measurement value at the end of the fixed time window. The source domain samples also have production line identification tags.

[0011] In a preferred embodiment, the feature extractor adopts a Transformer-based encoder structure, including a multi-head self-attention layer and a feedforward network, for mapping the input tensor to hidden layer feature vectors, the hidden layer feature vectors having a dimension of 128; the prediction head is a two-layer fully connected network, for mapping the hidden layer feature vectors to predicted values ​​of film thickness changes.

[0012] In a preferred embodiment, the training objective of the source domain basic process dynamics model pre-training module is composed of two weighted components: prediction loss and feature independence loss. The prediction loss uses mean square error to measure the deviation between the predicted value and the measured value of the film thickness change. The feature independence loss uses the Hilbert-Schmidt independence criterion to measure the statistical dependence between the hidden layer feature vector and the production line identity label. The specific calculation process is as follows:

[0013] For a mini-batch, which consists of multiple source domain samples, we first calculate the Gaussian kernel matrix of the hidden layer feature vectors and the Kronecker kernel matrix of the production line identity labels. The bandwidth parameter of the Gaussian kernel matrix is ​​the median of the Euclidean distance between any two pairs of hidden layer feature vectors in the mini-batch. The Kronecker kernel matrix is ​​generated by the Kronecker function. When two samples have the same production line identity label, the corresponding position is 1, and when they do not have the same label, the value is 0.

[0014] Subsequently, a centering matrix is ​​introduced to center the Gaussian kernel matrix and the Kronecker kernel matrix. The trace of the product of the two kernel matrices after processing is calculated, and the result is divided by the square of the number of source domain samples in the mini-batch minus one. The final value obtained is the feature independence loss. The sum of the prediction loss and the feature independence loss constitutes the total loss function, where the feature independence loss is multiplied by a balance coefficient, which is determined to be 0.1 through cross-validation. During training, the Adam optimizer is used with an initial learning rate of 0.0001 and a batch size of 64. All source domain samples are traversed for 200 iterations. In each mini-batch, the prediction loss and the feature independence loss are calculated simultaneously, and all parameters of the feature extractor and the prediction head are updated through the backpropagation algorithm. After training, the parameters of the feature extractor and the prediction head are saved, and the two together constitute the parameterized basic model.

[0015] In a preferred embodiment, the adversarial domain alignment module receives the feature extractor of the base model and freezes its structure as the starting point for adversarial training; collects target domain data, which is unlabeled short-term operating data collected during the no-load or low-load debugging phase of the new production line, with a duration of not less than 30 minutes and a sampling frequency of 1 Hz; slides the target domain data according to a fixed time window length of 60 seconds to obtain a target domain sample set, each target domain sample containing time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values ​​for 60 consecutive seconds within the fixed time window, and stacks the values ​​of these five channels in chronological order to form an input tensor; simultaneously, randomly selects source domain samples equal in number to the target domain samples from the source domain historical data, uses only the input tensor portion of the source domain samples, and discards the film thickness change label to construct unsupervised adversarial training pairs;

[0016] The domain discriminator employs a four-layer fully connected network with 256, 128, 64, and 1 neurons in each layer, respectively. The last layer does not use an activation function to output a real scalar. The objective function for adversarial training consists of three terms: Wasserstein adversarial loss, gradient penalty term, and feature distribution regularization term. The Wasserstein adversarial loss measures the Wasserstein distance between the hidden feature distributions of the source and target domains. It is calculated by subtracting the expected value of the score output by the domain discriminator for the hidden feature vectors of the source samples from the expected value of the score output by the domain discriminator for the hidden feature vectors of the target samples. The difference is the Wasserstein adversarial loss. The feature extractor aims to reduce the Wasserstein adversarial loss, while the domain discriminator aims to increase it.

[0017] The gradient penalty term is calculated as follows: one sample is randomly selected from the source domain sample and one sample is randomly selected from the target domain sample. Interpolation points are obtained by uniformly sampling along the hidden layer feature connection between the two samples. The gradient norm 2 of the domain discriminator at the interpolation point is calculated, and the square of the difference between the gradient norm 2 and the value 1 is used as the penalty. The gradient penalty term is only used to update the parameters of the domain discriminator.

[0018] The specific calculation process of the feature scatter regularization term is as follows: For a mini-batch containing multiple source domain samples or target domain samples, extract the hidden layer feature vectors obtained after all samples have passed through the feature extractor, and arrange all the hidden layer feature vectors in rows to form a two-dimensional feature matrix; calculate the average value of each column of the two-dimensional feature matrix to obtain a mean vector; subtract the mean vector element of the corresponding column from each element of the two-dimensional feature matrix to obtain the mean-free feature matrix; calculate the product of the transpose of the mean-free feature matrix and the mean-free feature matrix, and then divide by the number of mini-batch samples minus one to obtain a covariance matrix; calculate the sum of the absolute values ​​of the differences between corresponding elements of the covariance matrix and the identity matrix with the same dimension, and then divide the sum of the absolute values ​​by the total number of dimensions of the covariance matrix to obtain a scalar value, which is used as the feature scatter regularization term. The feature scatter regularization term is also used to update the feature extractor parameters.

[0019] In a preferred embodiment, the total loss function of the feature extractor is the Wasserstein adversarial loss plus a feature scattering regularization term multiplied by a first weight coefficient, and the total loss function of the domain discriminator is the negative Wasserstein adversarial loss plus a gradient penalty term multiplied by a second weight coefficient; wherein the first weight coefficient is determined to be 0.01 within the range of 0.001 to 0.1 through grid search, and the second weight coefficient is 10; during training, the RMSprop optimizer is used, the learning rate of the domain discriminator is set to 5 times 10 to the power of -5, and the learning rate of the feature extractor is set to 1 times 10 to the power of -4; in each iteration, the feature extractor is fixed first, and the domain discriminator is updated five times with a mini-batch size of 64 for each update, and then the domain discriminator is fixed again, and the feature extractor is updated once; a total of 500 iterations are performed; after training, all parameters of the domain discriminator are discarded, and only the updated parameters of the feature extractor are retained to obtain the domain-aligned feature extractor.

[0020] In a preferred embodiment, the rapid adaptation training module divides the historical data of the source domain into multiple subsets based on the production line identity tags. Each subset corresponds to an existing extrusion lamination production line. For each subset corresponding to an existing extrusion lamination production line, five source domain samples are randomly selected as a support set, and fifteen source domain samples mutually exclusive with the support set are selected as a query set. Each virtual task consists of a support set and a query set corresponding to an existing extrusion lamination production line. The number of all virtual tasks is equal to the total number of existing extrusion lamination production lines participating in the training, and the total number ranges from five to ten.

[0021] The complete model is constructed by adding a lightweight regression head to the domain-aligned feature extractor. The regression head is a two-layer fully connected network. The first layer maps the 128-dimensional hidden features to 64 dimensions with the ReLU activation function. The second layer maps the 64-dimensional features to one dimension without an activation function. The parameters of the complete model include the feature extractor parameters and the regression head parameters. The feature extractor parameters are initialized to the domain-aligned feature extractor parameters, while the regression head parameters are randomly initialized.

[0022] The parameter update process of the complete model consists of two stages: an inner loop and an outer loop. In the inner loop, for each virtual task, the prediction loss is calculated using the support set. The prediction loss is expressed as mean squared error. The calculation process is as follows: subtract the corresponding film thickness change from the predicted values ​​of film thickness changes for all samples in the support set, square the difference, sum all the squared differences, and divide the sum by the number of samples in the support set. The quotient is the prediction loss. Subsequently, a gradient update is performed on the parameters of the complete model. The updated parameters are called temporary parameters. The gradient update calculation process is as follows: subtract the inner loop learning rate multiplied by the gradient of the prediction loss with respect to the parameters of the complete model from the parameters of the complete model. The difference is the temporary parameter. The inner loop learning rate is set to 0.01.

[0023] In a preferred embodiment, the learning rate of the outer loop is set to 0.001. The outer loop uses the query set of each virtual task to calculate the query loss of the temporary parameters. The query loss is also in the form of mean squared error. The calculation process is as follows: subtract the corresponding membrane thickness change from the predicted value of the membrane thickness change of all samples in the query set, square the difference, sum all the squared differences, and divide the sum by the number of samples in the query set. The quotient is the query loss. Sum the query losses of all virtual tasks and use the sum as the meta-target. Calculate the gradient of the meta-target with respect to the parameters of the complete model using the backpropagation algorithm, and use this gradient as the meta-gradient to update the parameters of the complete model.

[0024] The calculation process of meta-gradient update is as follows: subtract the outer loop learning rate multiplied by the meta-gradient from the parameters of the complete model, and the difference is the updated parameters; the inner loop and outer loop are executed alternately for a total of 300 rounds, and source domain samples in the support set and query set are randomly re-sampled in each round to enhance task diversity; after training, the parameters of the complete model are saved as meta-initialization parameters.

[0025] In a preferred embodiment, the process of obtaining the fine-tuned complete model in the target production line one-step fine-tuning and deployment module is specifically as follows:

[0026] The normal operation data was randomly divided into a support set and a validation set in a ratio of 7:3. The number of samples in the support set was 70% of the total data in half an hour. The total data in half an hour was the total number of samples corresponding to a collection time of 1800 seconds and a sampling frequency of 1 Hz.

[0027] During fine-tuning, a regularization term based on elastic weight consolidation is introduced. The calculation process of the regularization term is as follows: For each parameter of the complete model, the gradient of the log-likelihood of the parameter with respect to each sample in the historical data of the source domain is calculated on the historical data of the source domain. The gradient is squared and then averaged over all samples in the source domain. The result is used as the importance measure of the parameter. Subsequently, the importance measure of each parameter is multiplied by the square of the difference between the parameter and the corresponding parameter in the meta-initialization parameters, and then multiplied by the regularization strength coefficient, which is set to 0.1. Finally, the product of all parameters is summed to obtain the regularization term.

[0028] The total loss function for fine-tuning consists of the mean squared error loss on the support set and the regularization term. The mean squared error loss is calculated as follows: subtract the corresponding film thickness change from the predicted value of the film thickness change of all samples in the support set, square the difference, sum all the squared differences, and then divide the sum by the number of samples in the support set. The quotient is the mean squared error loss.

[0029] During fine-tuning, the initial values ​​of the meta-initialization parameters are used as the initial values. The gradient of the total loss function with respect to the parameters of the complete model is calculated on the support set, and a gradient descent update is performed: the parameters of the complete model are subtracted from the learning rate multiplied by the gradient. The learning rate is set to 0.001, and the fine-tuned complete model is obtained.

[0030] In a preferred embodiment, the learning rate for the step gradient descent update is set to 0.001, and the optimizer uses stochastic gradient descent with momentum, and the momentum coefficient is set to 0.9.

[0031] The finely tuned complete model is deployed in the adaptive control of the new production line. The complete model receives the five-channel sensor sequence from 60 seconds prior to the current moment in real time. The five-channel sensor sequence contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values, forming an input tensor and outputting a predicted value of the film thickness change for the next 10 seconds. The predicted value of the film thickness change is converted into screw speed compensation and temperature setpoint correction through a predefined process mapping table and sent to the underlying PID controller for execution.

[0032] Simultaneously, the prediction error of the fine-tuned complete model is continuously monitored using the validation set. The prediction error is expressed as root mean square error (RMSE). The calculation process is as follows: subtract the corresponding film thickness change from the predicted value of the film thickness change of all samples in the validation set, square the difference, sum all the squared differences, divide the sum by the number of samples in the validation set, and finally square root the quotient to obtain the RMSE. If the RMSE exceeds a preset threshold (0.8%), a re-fine-tuning process is triggered: collect half an hour of normal operation data, randomly divide the normal operation data into a support set and a validation set at a ratio of 7:3, use the initialization parameters as the initial values, and repeat the fine-tuning steps. The fine-tuning steps include calculating the regularization term, the total loss function, and performing a gradient descent update until the RMSE does not exceed 0.8% or the preset maximum number of re-fine-tuning times is reached.

[0033] The beneficial effects of this invention are as follows: A pre-training module for the source domain basic process dynamics model extracts general process rules across production lines; an adversarial domain alignment module eliminates the characteristic distribution offset between new and old production lines; a rapid adaptability training module endows the complete model with the ability to adapt to new production lines with only one gradient update; and finally, a target production line fine-tuning and deployment module completes precise fine-tuning using a very small amount of normal operating data and puts it into real-time control. These modules are sequentially connected, enabling the complete model to quickly adapt to new operating conditions while retaining existing knowledge, significantly shortening the debugging cycle of new production lines, reducing raw material waste and equipment wear, improving film thickness control accuracy and production stability, and realizing automated reuse and low-cost deployment of process knowledge across production lines. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention;

[0035] Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0038] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0039] Example 1: This example provides the following... Figure 1-2 The diagram illustrates an adaptive control system for plastic extrusion lamination process based on machine learning, specifically comprising: a source domain basic process dynamics model pre-training module, an adversarial domain alignment module, a rapid adaptation capability training module, and a target production line one-step fine-tuning and deployment module, which are sequentially connected via communication.

[0040] The source domain basic process dynamics model pre-training module collects historical source domain data from multiple existing extrusion lamination production lines. The historical source domain data includes multi-channel time series of melt temperature, die pressure, screw speed, traction speed, and film thickness measurements. The historical source domain data is segmented into source domain samples by a fixed time window. A Transformer-based temporal prediction network is used as the basic process dynamics model. The basic process dynamics model includes a feature extractor and a prediction head. The model is trained on the source domain samples with the total loss function consisting of prediction loss and feature independence loss weighted together as the objective, and a parameterized basic model is output.

[0041] Adversarial Domain Alignment Module: Receives the feature extractor from the base model and introduces a domain discriminator. Through adversarial training, it minimizes the difference between the hidden feature distribution of the source domain and the hidden feature distribution of the target domain. The target domain data is taken from unlabeled short-term running data during the no-load or low-load debugging phase of the new production line. After training, the domain discriminator is discarded, and the domain-aligned feature extractor is output.

[0042] Rapid Adaptability Training Module: A lightweight regression head is added to the domain-aligned feature extractor to form a complete model. The model-independent meta-learning algorithm is used to construct multiple virtual tasks from the historical data of the source domain. Each virtual task contains a support set and a query set. Through inner and outer loop optimization, the complete model can adapt to the new production line with only one gradient update and output the meta-initialization parameters of the complete model.

[0043] The target production line fine-tuning and deployment module collects half an hour of normal operation data after the new production line is put into operation. The normal operation data is divided into a support set and a validation set. The complete model with the initial values ​​of the meta-initialization parameters is updated once by gradient descent on the support set to obtain the fine-tuned complete model. The fine-tuned complete model is then deployed in the adaptive control of the new production line, and the correction amount of the process parameters is output in real time.

[0044] In this embodiment, it is specifically necessary to explain that in the pre-training module of the source domain basic process dynamics model, the source domain historical data comes from multiple existing extrusion laminating production lines. The source domain historical data of each existing extrusion laminating production line is a multi-channel time series, which includes melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values. The sampling frequency is uniformly one Hz. The source domain historical data is slidably intercepted with a fixed time window length of sixty seconds to obtain source domain samples. Each source domain sample contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values ​​for a continuous sixty seconds within the fixed time window. The values ​​of these five channels are stacked in chronological order to form an input tensor. The corresponding prediction target is the film thickness change in the next ten seconds. The film thickness change is defined as the difference between the film thickness measurement value ten seconds after the end of the fixed time window and the film thickness measurement value at the end of the fixed time window. The source domain samples are also accompanied by production line identification tags, which use integer numbers to identify the production line to which the sample belongs.

[0045] Specifically, the five parameters of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement were chosen because they constitute the key physical quantities that most directly affect the uniformity of film thickness in the extrusion lamination process, and there is a strong coupled nonlinear relationship between them. The sampling frequency is uniformly set to one Hz, which is sufficient to capture dynamic changes in the process (the melt temperature change cycle is usually several seconds to tens of seconds) while avoiding data redundancy and storage pressure caused by excessively high sampling. The fixed time window length of sixty seconds is selected based on the process response time constant: the thermal inertia and flow delay of the extrusion lamination system are usually between thirty and ninety seconds, so sixty seconds... The window can completely cover the entire process of a typical disturbance from occurrence to decay; the prediction target is set as the film thickness change in the next ten seconds because ten seconds is exactly in the golden response range of process adjustment - too short and noise dominates, too long and the prediction uncertainty increases dramatically; the input tensor is stacked as follows: the five channel values ​​of each time step are arranged along the time axis to form a two-dimensional matrix of shape (sixty, five), and then the batch dimension is expanded and fed into the network; the production line identity label is used for the subsequent calculation of feature independence loss, and its integer number range is 1 to K, where K is the total number of existing production lines participating in the training, and each number uniquely corresponds to one production line;

[0046] The feature extractor adopts a Transformer-based encoder structure, which includes a multi-head self-attention layer and a feedforward network to map the input tensor into hidden layer feature vectors. The hidden layer feature vectors have a dimension of 128. The prediction head is a two-layer fully connected network to map the hidden layer feature vectors into predicted values ​​of membrane thickness changes.

[0047] Specifically, the Transformer encoder consists of four identical stacked coding layers. Each coding layer contains a multi-head self-attention sublayer and a feedforward neural network sublayer. Each sublayer is followed by a normalization layer and a residual connection. The multi-head self-attention layer has eight heads, with each head having an attention dimension of sixteen, thus capturing dependencies at different time scales while maintaining computational efficiency. The feedforward network has a hidden layer dimension of 256, using ReLU as the activation function. The hidden layer feature vector dimension is chosen to be 128, a trade-off between model capacity and generalization ability: too low a dimension makes it difficult to express complex coupling relationships, while too high a dimension easily leads to overfitting and increases the computational burden of subsequent domain alignment. The prediction head consists of two fully connected layers. The first layer has an output dimension of 64, with ReLU as the activation function, while the second layer has an output dimension of one, representing the predicted value of the membrane thickness change, with no activation function (linear output), to accommodate the unlimited range of regression tasks.

[0048] The training objective of the pre-training module for the source domain basic process dynamics model consists of a weighted average of prediction loss and feature independence loss. The prediction loss uses mean squared error to measure the deviation between the predicted and measured values ​​of film thickness variation. The feature independence loss uses the Hilbert-Schmidt independence criterion to measure the statistical dependence between the hidden layer feature vector and the production line identity label. The specific calculation process is as follows:

[0049] For a mini-batch, which consists of multiple source domain samples, we first calculate the Gaussian kernel matrix of the hidden layer feature vectors and the Kronecker kernel matrix of the production line identity labels. The bandwidth parameter of the Gaussian kernel matrix is ​​the median of the Euclidean distance between any two pairs of hidden layer feature vectors in the mini-batch. The Kronecker kernel matrix is ​​generated by the Kronecker function. When two samples have the same production line identity label, the corresponding position is 1, and when they do not have the same label, the value is 0.

[0050] Subsequently, a centering matrix is ​​introduced to center the Gaussian kernel matrix and the Kronecker kernel matrix. The trace of the product of the two kernel matrices after processing is calculated, and the result is divided by the square of the number of source domain samples in the mini-batch minus one. The final value obtained is the feature independence loss. The smaller the feature independence loss value, the stronger the independence between the hidden layer feature vector and the product line identity. The sum of the prediction loss and the feature independence loss constitutes the total loss function, where the feature independence loss is multiplied by a balance coefficient, which is determined to be 0.1 through cross-validation. During training, the Adam optimizer is used with an initial learning rate of 0.0001 and a batch size of 64. All source domain samples are traversed for 200 iterations. In each mini-batch, the prediction loss and feature independence loss are calculated simultaneously, and all parameters of the feature extractor and prediction head are updated through the backpropagation algorithm. After training, the parameters of the feature extractor and the prediction head are saved, and the two together constitute the parameterized basic model.

[0051] Specifically, the Hilbert-Schmidt independence criterion is adopted because it is a non-parametric independence measure that does not require assumptions about the specific form of the feature distribution, making it suitable for scenarios with high-dimensional hidden features. The bandwidth of the Gaussian kernel, taken as the median of the Euclidean distance within the mini-batch, is an adaptive strategy that ensures the kernel function has reasonable sensitivity to feature distributions at different scales, avoiding manual parameter tuning. The centralization matrix is ​​formed by subtracting a matrix from the identity matrix where all elements represent one-tenth of the mini-batch sample size; this removes the row and column mean of the kernel matrix, ensuring the independence measure is unaffected by translation. The balance coefficient of 0.1 is determined through five-fold cross-validation: a validation set is partitioned on the source domain data, and the harmonic index between the prediction loss and the feature independence loss on the validation set is used as the benchmark. The criterion is to search within the range of [0.01, 0.5] to find the optimal value of 0.1; the parameters of the Adam optimizer are set to β_1=0.9, β_2=0.999, and ε=1e-8, which is a common configuration in time series prediction tasks; the 200 iterations are based on the observation of early stopping: the loss usually tends to plateau around 150 iterations, and 200 iterations ensure sufficient convergence without overfitting; the batch size of 64 takes into account both gradient stability and memory consumption; the parameters saved after training are the starting point for the initialization of subsequent modules of the basic model. Among them, the feature extractor parameters will be directly loaded by the adversarial domain alignment module, and although the prediction head parameters are no longer used in subsequent modules, their contribution to the backpropagation of the feature extractor during training has been solidified in the feature extractor parameters.

[0052] In this embodiment, it is specifically necessary to explain that in the adversarial domain alignment module, the feature extractor of the base model is received, and its structure is frozen as the starting point for adversarial training. Target domain data is collected. The target domain data consists of unlabeled short-term operating data collected during the no-load or low-load debugging phase of the new production line, with a duration of no less than thirty minutes and a sampling frequency consistent with the source domain, i.e., one hertz. The target domain data is truncated by sliding according to a fixed time window length of sixty seconds, the same as that of the source domain, to obtain a target domain sample set. Each target domain sample contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values ​​for sixty consecutive seconds within the fixed time window. The values ​​of these five channels are stacked in chronological order to form an input tensor. Simultaneously, source domain samples with an equal number of target domain samples are randomly selected from the historical data of the source domain. Only the input tensor portion of the source domain samples is used, and the film thickness change label is discarded to construct unsupervised adversarial training pairs.

[0053] The domain discriminator employs a four-layer fully connected network with 256, 128, 64, and 1 neurons in each layer, respectively. The last layer does not use an activation function and outputs a real scalar. Specifically, the input to the domain discriminator is the hidden layer feature vector, which has a dimension of 128. After passing through the first fully connected layer, it is mapped to 256 dimensions, using LeakyReLU with a slope of 0.2. The second layer reduces the dimension from 256 to 128, again using LeakyReLU. The third layer reduces the dimension from 128 to 64, using LeakyReLU. The fourth layer reduces the dimension from 64 to 1, with no activation function, and outputs a real scalar representing the probability score of a sample belonging to the source domain. The design of the domain discriminator allows it to distinguish the feature distributions of the source and target domains sufficiently, while ensuring Lipschitz continuity through a gradient penalty term. The objective function for adversarial training consists of Wasserstein adversarial loss, a gradient penalty term, and features. The scattering regularization term consists of three terms. The Wasserstein adversarial loss measures the Wasserstein distance between the hidden feature distributions of the source and target domains. Its calculation is as follows: the expected value of the score output by the domain discriminator for the hidden feature vectors of the source domain samples is subtracted from the expected value of the score output by the domain discriminator for the hidden feature vectors of the target domain samples. The difference is the Wasserstein adversarial loss. The feature extractor aims to reduce the Wasserstein adversarial loss, while the domain discriminator aims to increase it. In actual calculations, the expected value is approximated by the average value within a mini-batch. That is, the average score of the source domain samples in a mini-batch is subtracted from the average score of the target domain samples in the same batch. The difference is the estimated value of the Wasserstein adversarial loss. The larger this loss, the greater the distance between the two distributions. The feature extractor attempts to compress this distance, while the domain discriminator attempts to widen it, forming a minimax game.

[0054] The gradient penalty term is used to constrain the domain discriminator to satisfy Lipschitz continuity. Its calculation process is as follows: One sample is randomly selected from both the source and target domain samples. Interpolation points are obtained by uniformly sampling along the hidden feature vectors connecting the two samples. The gradient norm 2 of the domain discriminator at the interpolation point is calculated, and the square of the difference between the gradient norm 2 and the value of 1 is used as the penalty. The gradient penalty term is only used to update the domain discriminator parameters. Specifically, the interpolation points are generated by randomly selecting a mixing ratio between the hidden feature vectors of the source and target domain samples. This ratio follows a uniform distribution. The two feature vectors are then linearly combined according to this ratio to obtain the interpolated feature vector. The gradient of the domain discriminator at this interpolation point is calculated through backpropagation. Its norm 2 is taken, subtracted by 1, and then squared to obtain the value of the gradient penalty term. This penalty term forces the gradient norm of the domain discriminator to be close to 1 everywhere, thus ensuring the approximate accuracy of the Wasserstein distance.

[0055] The feature dispersion regularization term is used to prevent adversarial training from causing the hidden layer features to collapse into a low-rank subspace, encouraging features to remain dispersed across all dimensions. The specific calculation process is as follows: For a mini-batch containing multiple source or target domain samples, extract the hidden layer feature vectors obtained after all samples have passed through the feature extractor. Arrange all hidden layer feature vectors in rows to form a two-dimensional feature matrix. Calculate the mean of each column of the two-dimensional feature matrix to obtain a mean vector. Subtract the mean vector element of the corresponding column from each element of the two-dimensional feature matrix to obtain the mean-reduced feature matrix. Calculate the product of the transpose of the mean-reduced feature matrix and the mean-reduced feature matrix, then divide by the mini-batch sample size minus one to obtain a covariance matrix. Calculate the covariance... The sum of the absolute values ​​of the differences between corresponding elements of the difference matrix and the identity matrix with the same dimensions, divided by the total number of dimensions of the covariance matrix, yields a scalar value. This scalar value is used as the feature scatter regularization term, which is also used to update the feature extractor parameters. The design of this regularization term is inspired by covariance matrix regularization, and its physical meaning is: ideally, each dimension of the hidden layer features should have unit variance and be uncorrelated with each other, i.e., the covariance matrix is ​​close to the identity matrix. By minimizing the difference between the covariance matrix and the identity matrix, we can prevent the features from becoming consistent across all samples, thus preserving sufficient expressive power for subsequent domain alignment and meta-learning. The total number of dimensions is 128, which is the dimension of the hidden layer feature vector.

[0056] The total loss function of the feature extractor is the Wasserstein adversarial loss plus the feature spread regularization term multiplied by the first weight coefficient. The total loss function of the domain discriminator is the negative Wasserstein adversarial loss plus the gradient penalty term multiplied by the second weight coefficient. The first weight coefficient is determined to be 0.01 within the range of 0.001 to 0.1 through grid search, and the second weight coefficient is set to 10. During training, the RMSprop optimizer is used. The learning rate of the domain discriminator is set to 5 x 10^-5, and the learning rate of the feature extractor is set to 1 x 10^-4. In each iteration, the feature extractor is fixed, and five gradient updates are performed on the domain discriminator, with a mini-batch size of 64 for each update. Then, the domain discriminator is fixed again, and one gradient update is performed on the feature extractor. A total of 500 iterations are performed. After training, all parameters of the domain discriminator are discarded, and only the updated feature extractor parameters are retained, resulting in a domain-aligned feature extractor. Specifically, the RMSprop optimizer is chosen instead of Adam because Adam's momentum mechanism is more efficient than Wasserstein's. While Wasserstein adversarial training can lead to oscillations and instability, RMSprop, with its adaptive learning rate and lack of momentum, is better suited for minimax games. The domain discriminator has a lower learning rate (five times ten to the power of -5) to stabilize its training, while the feature extractor has a slightly higher learning rate (one times ten to the power of -4) to accelerate domain alignment. The alternating ratio of five updates for the domain discriminator and one update for the feature extractor in each iteration is based on the rule of thumb: the domain discriminator needs to more fully approximate the Wasserstein distance, while the feature extractor only needs to adjust slowly following the changes in the domain discriminator. The first weight coefficient of 0.01 is determined through grid search and selected after evaluating the domain alignment effect on the validation set (e.g., the overlap between the hidden features of the target domain samples and the distribution of the source domain samples). The second weight coefficient is the recommended value for standard WassersteinGAN, which effectively constrains the gradient norm. After training, the domain-aligned feature extractor parameters are saved and directly used as the initialization of the feature extractor for the complete model in the subsequent fast adaptation training module, thus achieving seamless integration across modules.

[0057] In this embodiment, it is specifically noted that in the rapid adaptation training module, the historical data of the source domain is divided into multiple subsets based on the production line identity labels. Each subset corresponds to an existing extrusion lamination production line. For each subset corresponding to an existing extrusion lamination production line, five source domain samples are randomly selected as the support set, and fifteen source domain samples mutually exclusive with the support set are selected as the query set. Each virtual task consists of the support set and the query set corresponding to an existing extrusion lamination production line. The support set is used to simulate the very limited amount of labeled data available for the new production line, and the query set is used to evaluate the prediction performance of the complete model after one gradient update on the corresponding production line. The number of all virtual tasks is equal to the number of tasks participating in the training. The total number of existing extrusion lamination production lines is between five and ten. Specifically, the support set sample size is set to five, based on the realistic constraint that new production lines can only provide a very small amount of labeled data in the initial stage of actual debugging. At the same time, five samples are sufficient to generate meaningful parameter adjustment directions in one gradient update, avoiding overfitting. The query set sample size is set to fifteen to ensure sufficient statistical reliability when evaluating the adaptation effect, while maintaining an appropriate ratio with the support set so that the meta-learning objective can reflect the generalization ability of the model after adaptation with a small number of samples. The total number of virtual tasks is equal to the number of existing production lines to ensure that the process characteristics of each production line are included in the meta-learning training, avoiding bias caused by the omission of some production lines.

[0058] A lightweight regression head is added to the domain-aligned feature extractor to form the complete model. The regression head is a two-layer fully connected network. The first layer maps the 128-dimensional hidden features to 64 dimensions with the ReLU activation function, and the second layer maps the 64-dimensional features to one dimension without an activation function. The parameters of the complete model include the feature extractor parameters and the regression head parameters. The feature extractor parameters are initialized with the domain-aligned feature extractor parameters, while the regression head parameters are randomly initialized. Specifically, the lightweight regression head is designed according to the principle of "small and shallow head". Since the domain-aligned feature extractor already provides high-quality feature representations, the regression head only needs to learn a simple linear mapping from features to membrane thickness changes. Two layers are sufficient to capture the remaining nonlinear relationships. The choice of 64 output dimensions for the first layer is to maintain sufficient capacity while avoiding the introduction of too many parameters that would make fine-tuning difficult. The regression head parameters are randomly initialized using a normal distribution with a mean of zero and a standard deviation of 0.01. This small-range initialization helps to maintain gradient stability in the early stages of meta-learning and prevents large updates from destroying the domain-aligned feature space.

[0059] The parameter update process of the complete model consists of two stages: an inner loop and an outer loop. In the inner loop, for each virtual task, the prediction loss is calculated using the support set. The prediction loss is expressed as mean squared error. The calculation process is as follows: subtract the corresponding membrane thickness change from the predicted value of the membrane thickness change for all samples in the support set, square the difference, sum all the squared differences, and divide the sum by the number of samples in the support set. The quotient is the prediction loss. Subsequently, a gradient update is performed on the parameters of the complete model. The updated parameters are recorded as temporary parameters. The gradient update calculation process is as follows: subtract the learning rate of the inner loop from the parameters of the complete model and multiply the prediction loss by the prediction loss. The difference between the gradients of the parameters of the complete model and the gradients of the parameters is the temporary parameter; the learning rate of the inner loop is set to 0.01. Specifically, the choice of the inner loop learning rate of 0.01 is based on the following considerations: this value is large enough to produce a significant parameter shift with only five samples, but not so large that the gradient update would overshoot the optimal region; the one-step gradient update design simulates the real-world constraint of only being able to make one fine-tuning in the debugging of a new production line, forcing the meta-learning process to learn how to achieve the best results within a very limited number of update steps; the temporary parameter is only used for query set evaluation and is not persisted, thus ensuring that the adaptation process of each virtual task is independent of each other;

[0060] The learning rate of the outer loop is set to 0.001. The query loss of the temporary parameters is calculated using the query set of each virtual task in the outer loop. The query loss is also in the form of mean squared error. The calculation process is as follows: subtract the corresponding membrane thickness change from the predicted value of the membrane thickness change of all samples in the query set, square the difference, sum all the squared differences, and divide the sum by the number of samples in the query set. The quotient is the query loss. The query losses of all virtual tasks are summed, and the sum is used as the meta-target. The gradient of the meta-target with respect to the parameters of the complete model is calculated through the backpropagation algorithm, and this gradient is used as the meta-gradient. The parameters of the complete model are updated using the meta-gradient.

[0061] The meta-gradient update process is as follows: subtract the outer loop learning rate multiplied by the meta-gradient from the parameters of the complete model; the difference is the updated parameters. The meta-gradient calculation in the outer loop uses a first-order approximation, ignoring the second-order derivative term. The inner and outer loops are executed alternately for a total of 300 rounds, with source domain samples randomly re-sampled from the support and query sets in each round to enhance task diversity. After training, the parameters of the complete model are saved as meta-initialization parameters. Specifically, the outer loop learning rate of 0.001 is significantly smaller than the inner loop learning rate because the outer loop is responsible for optimizing the global initialization parameters, requiring slow and robust adjustments to avoid disrupting the rapid adaptation ability learned by the inner loop. The meta-objective is obtained by summing the query losses of all virtual tasks, ensuring that the optimization direction comprehensively considers the adaptation effects of all production lines, rather than favoring a single production line. The first-order approximation (i.e., ignoring the second-order derivative term)... The implementation is as follows: when calculating the meta-gradient, the derivative of the temporary parameter with respect to the initial parameter is treated as a constant, and the higher-order chain rule is not expanded. This approximation has been proven in the MAML literature to significantly reduce computational overhead while maintaining performance, and is especially suitable for parallel training of hundreds of virtual tasks in this scenario. The number of iterations of 300 rounds is based on experimental observation: the meta-loss usually tends to stabilize around 200 rounds, and 300 rounds ensure sufficient convergence without overfitting. In each round, the support set and query set are randomly resampled, which is equivalent to resampling the task distribution, forcing the model to face various possible data combinations, thereby enhancing the robustness of the learned initialization parameters. The meta-initialization parameters saved after training contain the prior knowledge of "how to adapt quickly" learned by the model from all existing production lines. These parameters will be directly loaded by the target production line's one-step fine-tuning and deployment module as the starting point for fine-tuning the new production line.

[0062] In this embodiment, it is specifically necessary to explain the process of obtaining the complete model after fine-tuning in the target production line one-step fine-tuning and deployment module as follows:

[0063] The normal operating data was randomly divided into a support set and a validation set in a 7:3 ratio. The support set was used for fine-tuning, and the validation set was used for performance evaluation and online monitoring after fine-tuning. The number of samples in the support set was 70% of the total data for half an hour, and the total data for half an hour was the total number of samples corresponding to a collection time of 1800 seconds and a sampling frequency of 1 Hz. Specifically, the number of samples in the support set was approximately 1260, which was sufficient to drive one effective gradient update. The number of samples in the validation set was approximately 540, which was used to evaluate the generalization error of the fine-tuned model on the new production line. The 7:3 division ratio is a common practice in the field of machine learning, which ensures sufficient fine-tuning data while leaving enough validation samples to monitor the risk of overfitting.

[0064] During fine-tuning, a regularization term based on elastic weight consolidation is introduced. The calculation process for the regularization term is as follows: For each parameter of the complete model, the gradient of the parameter with respect to the log-likelihood of each sample in the source domain historical data is calculated. This gradient is squared and then averaged over all source domain samples. The result is used as the importance measure of the parameter. Subsequently, the importance measure of each parameter is multiplied by the square of the difference between the parameter and the corresponding parameter in the meta-initialization parameters, and then multiplied by the regularization strength coefficient, which is set to 0.1. Finally, the products of all parameters are summed to obtain the regularization term. Specifically, the log-likelihood of each parameter with elastic weight consolidation is calculated. However, based on the Gaussian distribution assumption: given an input tensor, the predicted value of the membrane thickness variation follows a Gaussian distribution with the true value as the mean and a fixed variance. In this case, the log-likelihood is equivalent to the negative mean squared error loss multiplied by a constant. Therefore, the average of the squared gradients is the diagonal element of the Fisher information matrix, which quantifies the contribution of each parameter to the prediction accuracy in the source domain. The regularization strength coefficient of 0.1 is determined in the source domain simulation through cross-validation, which can achieve a balance between preserving the meta-initialization knowledge and adapting to new production line data. The role of the regularization term is to penalize parameters that are important to the source domain task and cause large changes, thereby preventing catastrophic forgetting.

[0065] The total loss function for fine-tuning consists of the mean squared error loss on the support set and a regularization term. The mean squared error loss is calculated as follows: the predicted value of the film thickness change for all samples in the support set is subtracted from the corresponding film thickness change, the difference is squared, all squared differences are summed, and the sum is divided by the number of samples in the support set. The quotient is the mean squared error loss. This loss drives the model to fit the local process characteristics of the new production line, ensuring that the fine-tuned model has high accuracy on the new production line.

[0066] During fine-tuning, the initial values ​​of the meta-initialization parameters are used as the initial values. The gradient of the total loss function with respect to the parameters of the complete model is calculated on the support set, and a gradient descent update is performed: the parameters of the complete model are subtracted from the learning rate multiplied by the gradient. The learning rate is set to 0.001, resulting in the fine-tuned complete model. The learning rate of 0.001 is significantly smaller than the learning rate of the meta-learning inner loop (0.01) to ensure the robustness of the fine-tuning process and avoid large updates that would destroy the fast adaptation direction implied in the meta-initialization. The optimizer uses stochastic gradient descent with momentum (momentum coefficient 0.9) because SGD with momentum is more stable than Adam on small batch data and is less prone to divergence.

[0067] The learning rate for one-step gradient descent update is set to 0.001. The optimizer uses stochastic gradient descent with momentum and the momentum coefficient is set to 0.9. Stochastic gradient descent with momentum can smoothly update the direction with a small amount of data by accumulating historical gradient directions and avoiding oscillations. It is particularly suitable for situations in this scenario where there is only one large step of fine adjustment.

[0068] The finely tuned complete model is deployed in the adaptive control of the new production line. The complete model receives a five-channel sensor sequence from 60 seconds prior to the current moment. This sequence contains time-series data on melt temperature, die pressure, screw speed, traction speed, and film thickness measurements, forming an input tensor and outputting a predicted film thickness change for the next 10 seconds. The predicted film thickness change is converted into screw speed compensation and temperature setpoint correction values ​​using a predefined process mapping table and sent to the underlying PID controller for execution. Specifically, the process mapping table is pre-established based on historical process data and expert experience. For example, when the predicted film thickness change is positive (indicating a future film thickness that will be too high), the mapping table recommends reducing the screw speed by 0.5% or increasing the die temperature by 1°C to reduce the film thickness; when the predicted value is negative, the opposite adjustment is made. The mapping table is calibrated through offline simulation and on-site debugging to ensure that the converted correction values ​​are within a safe range. The underlying PID controller receives these correction values ​​as offsets for the setpoint and adjusts the actuators in real time.

[0069] Simultaneously, the prediction error of the fine-tuned complete model is continuously monitored using the validation set. The prediction error is expressed as root mean square error (RMSE), calculated as follows: subtract the corresponding film thickness change from the predicted film thickness change of all samples in the validation set, square the difference, sum all the squared differences, divide the sum by the number of samples in the validation set, and finally square root the quotient to obtain the RMSE. If the RMSE exceeds a preset threshold (0.8%), a re-fine-tuning process is triggered: collect half an hour of normal operating data, randomly divide the normal operating data into a support set and a validation set at a ratio of 7:3, use the initialization parameters as the initial values, and repeat the fine-tuning steps. The fine-tuning steps include calculating the regularization term, the total loss function, and executing a step-by-step process. Gradient descent updates continue until the root mean square error (RMSE) does not exceed 0.8% or the preset maximum number of re-fine-tuning attempts is reached. The preset threshold of 0.8% represents the percentage of the RMS error relative to the film thickness measurement. This value is set according to process quality requirements to ensure that the film thickness control accuracy is within the allowable range. The preset maximum number of re-fine-tuning attempts is set to 5 times to prevent infinite loops. The validation set can be monitored every hour or after each roll of material is produced. If the error exceeds the limit, re-fine-tuning is triggered immediately. During re-fine-tuning, the original initialization parameters are still used as the starting point, rather than the parameters after the previous fine-tuning, to avoid error accumulation. This mechanism ensures that the model can automatically adapt to the slow drift of production line conditions (such as heater aging and raw material batch changes) during long-term operation, maintaining stable control performance.

[0070] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0071] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0072] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0073] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0075] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0076] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An adaptive control system for a plastic extrusion lamination process based on machine learning, characterized in that, Specifically, it includes: The module consists of a source domain basic process dynamics model pre-training module, an adversarial domain alignment module, a rapid adaptation capability training module, and a target production line one-step fine-tuning and deployment module, which are connected sequentially via communication. The source domain basic process dynamics model pre-training module collects historical source domain data from multiple existing extrusion lamination production lines. The historical source domain data includes multi-channel time series of melt temperature, die pressure, screw speed, traction speed, and film thickness measurements. The historical source domain data is segmented into source domain samples by a fixed time window. A Transformer-based temporal prediction network is used as the basic process dynamics model. The basic process dynamics model includes a feature extractor and a prediction head. The model is trained on the source domain samples with the total loss function consisting of prediction loss and feature independence loss weighted together as the objective, and a parameterized basic model is output. Adversarial Domain Alignment Module: Receives the feature extractor from the base model and introduces a domain discriminator. Through adversarial training, it minimizes the difference between the hidden feature distribution of the source domain and the hidden feature distribution of the target domain. The target domain data is taken from unlabeled short-term running data during the no-load or low-load debugging phase of the new production line. After training, the domain discriminator is discarded, and the domain-aligned feature extractor is output. Rapid Adaptability Training Module: A lightweight regression head is added to the domain-aligned feature extractor to form a complete model. The model-independent meta-learning algorithm is used to construct multiple virtual tasks from the historical data of the source domain. Each virtual task contains a support set and a query set. Through inner and outer loop optimization, the complete model can adapt to the new production line with only one gradient update and output the meta-initialization parameters of the complete model. The target production line fine-tuning and deployment module collects half an hour of normal operation data after the new production line is put into operation. The normal operation data is divided into a support set and a validation set. The complete model with the initial values ​​of the meta-initialization parameters is updated once by gradient descent on the support set to obtain the fine-tuned complete model. The fine-tuned complete model is then deployed in the adaptive control of the new production line, and the correction amount of the process parameters is output in real time.

2. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 1, characterized in that: In the pre-training module of the source domain basic process dynamics model, the source domain historical data comes from multiple existing extrusion laminating production lines. The source domain historical data of each existing extrusion laminating production line is a multi-channel time series, which includes melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values. The sampling frequency is uniformly one Hz. Source domain samples are obtained by sliding the source domain historical data with a fixed time window length of sixty seconds. Each source domain sample contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values ​​for sixty consecutive seconds within the fixed time window. The values ​​of these five channels are stacked in chronological order to form an input tensor. The corresponding prediction target is the film thickness change in the next ten seconds. The film thickness change is defined as the difference between the film thickness measurement value ten seconds after the end of the fixed time window and the film thickness measurement value at the end of the fixed time window. The source domain samples also have production line identification tags.

3. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 2, characterized in that: The feature extractor adopts a Transformer-based encoder structure, which includes a multi-head self-attention layer and a feedforward network to map the input tensor into hidden layer feature vectors. The hidden layer feature vectors have a dimension of 128. The prediction head is a two-layer fully connected network to map the hidden layer feature vectors into predicted values ​​of the membrane thickness change.

4. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 3, characterized in that: The training objective of the pre-training module for the source domain basic process dynamics model consists of a weighted average of prediction loss and feature independence loss. The prediction loss uses mean squared error to measure the deviation between the predicted and measured values ​​of film thickness variation. The feature independence loss uses the Hilbert-Schmidt independence criterion to measure the statistical dependence between the hidden layer feature vector and the production line identity label. The specific calculation process is as follows: For a mini-batch, which consists of multiple source domain samples, we first calculate the Gaussian kernel matrix of the hidden layer feature vectors and the Kronecker kernel matrix of the production line identity labels. The bandwidth parameter of the Gaussian kernel matrix is ​​the median of the Euclidean distance between any two pairs of hidden layer feature vectors in the mini-batch. The Kronecker kernel matrix is ​​generated by the Kronecker function. When two samples have the same production line identity label, the corresponding position is 1, and when they do not have the same label, the value is 0. Subsequently, a centering matrix is ​​introduced to center the Gaussian kernel matrix and the Kronecker kernel matrix. The trace of the product of the two kernel matrices after processing is calculated, and the result is divided by the square of the number of source domain samples in the mini-batch minus one. The final value obtained is the feature independence loss. The sum of the prediction loss and the feature independence loss constitutes the total loss function, where the feature independence loss is multiplied by a balance coefficient, which is determined to be 0.1 through cross-validation. During training, the Adam optimizer is used with an initial learning rate of 0.0001 and a batch size of 64. All source domain samples are traversed for 200 iterations. In each mini-batch, the prediction loss and the feature independence loss are calculated simultaneously, and all parameters of the feature extractor and the prediction head are updated through the backpropagation algorithm. After training, the parameters of the feature extractor and the prediction head are saved, and the two together constitute the parameterized basic model.

5. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 4, characterized in that: In the adversarial domain alignment module, the feature extractor of the base model is received, and its structure is frozen as the starting point for adversarial training. Collect target domain data, which consists of unlabeled short-term operating data collected during the no-load or low-load commissioning phase of the new production line, with a duration of no less than 30 minutes and a sampling frequency of 1 Hz. Slide the target domain data according to a fixed time window of 60 seconds to obtain a target domain sample set. Each target domain sample contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurements for 60 consecutive seconds within the fixed time window. Stack the values ​​of these five channels in chronological order to form an input tensor. At the same time, randomly select source domain samples with an equal number of target domain samples from the historical source domain data. Use only the input tensor part of the source domain samples and discard the film thickness change label to construct unsupervised adversarial training pairs. The domain discriminator employs a four-layer fully connected network with 256, 128, 64, and 1 neurons in each layer, respectively. The last layer does not use an activation function to output a real scalar. The objective function for adversarial training consists of three terms: Wasserstein adversarial loss, gradient penalty term, and feature distribution regularization term. The Wasserstein adversarial loss measures the Wasserstein distance between the hidden feature distributions of the source and target domains. It is calculated by subtracting the expected value of the score output by the domain discriminator for the hidden feature vectors of the source samples from the expected value of the score output by the domain discriminator for the hidden feature vectors of the target samples. The difference is the Wasserstein adversarial loss. The feature extractor aims to reduce the Wasserstein adversarial loss, while the domain discriminator aims to increase it. The gradient penalty term is calculated as follows: one sample is randomly selected from the source domain sample and one sample is randomly selected from the target domain sample. Interpolation points are obtained by uniformly sampling along the hidden layer feature connection between the two samples. The gradient norm 2 of the domain discriminator at the interpolation point is calculated, and the square of the difference between the gradient norm 2 and the value 1 is used as the penalty. The gradient penalty term is only used to update the parameters of the domain discriminator. The specific calculation process of the feature scatter regularization term is as follows: For a mini-batch containing multiple source domain samples or target domain samples, extract the hidden layer feature vectors obtained after all samples have passed through the feature extractor, and arrange all the hidden layer feature vectors in rows to form a two-dimensional feature matrix; calculate the average value of each column of the two-dimensional feature matrix to obtain a mean vector; subtract the mean vector element of the corresponding column from each element of the two-dimensional feature matrix to obtain the mean-free feature matrix; calculate the product of the transpose of the mean-free feature matrix and the mean-free feature matrix, and then divide by the number of mini-batch samples minus one to obtain a covariance matrix; calculate the sum of the absolute values ​​of the differences between corresponding elements of the covariance matrix and the identity matrix with the same dimension, and then divide the sum of the absolute values ​​by the total number of dimensions of the covariance matrix to obtain a scalar value, which is used as the feature scatter regularization term. The feature scatter regularization term is also used to update the feature extractor parameters.

6. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 5, characterized in that: The total loss function of the feature extractor is the Wasserstein adversarial loss plus the feature spread regularization term multiplied by the first weight coefficient, and the total loss function of the domain discriminator is the negative Wasserstein adversarial loss plus the gradient penalty term multiplied by the second weight coefficient. The first weight coefficient was determined to be 0.01 within the range of 0.001 to 0.1 through grid search, and the second weight coefficient was set to 10. During training, the RMSprop optimizer was used, with the learning rate of the domain discriminator set to 5 x 10^-5 and the learning rate of the feature extractor set to 1 x 10^-4. In each iteration, the feature extractor was fixed first, and the domain discriminator was updated five times with a mini-batch size of 64 for each update. Then, the domain discriminator was fixed again, and the feature extractor was updated once. The iterations lasted for a total of 500 rounds. After training, all parameters of the domain discriminator were discarded, and only the updated parameters of the feature extractor were retained to obtain the domain-aligned feature extractor.

7. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 6, characterized in that: In the rapid adaptation training module, the source domain historical data is divided into multiple subsets based on the production line identity label, and each subset corresponds to an existing extrusion lamination production line. For each subset corresponding to an existing extrusion lamination production line, five source domain samples are randomly selected as the support set, and fifteen source domain samples mutually exclusive with the support set are selected as the query set. Each virtual task consists of the support set and query set corresponding to an existing extrusion lamination production line. The total number of virtual tasks is equal to the total number of existing extrusion lamination production lines participating in the training, and the total number ranges from five to ten. The complete model is constructed by adding a lightweight regression head to the domain-aligned feature extractor. The regression head is a two-layer fully connected network. The first layer maps the 128-dimensional hidden features to 64 dimensions using ReLU activation function, and the second layer maps the 64-dimensional features to one dimension without activation function. The parameters of the complete model include the feature extractor parameters and the regression head parameters. The feature extractor parameters are initialized to the domain-aligned feature extractor parameters, while the regression head parameters are randomly initialized. The parameter update process of the complete model consists of two stages: an inner loop and an outer loop. In the inner loop, for each virtual task, the prediction loss is calculated using the support set. The prediction loss is in the form of mean squared error. The calculation process is as follows: subtract the corresponding film thickness change from the predicted value of the film thickness change of all samples in the support set, square the difference, sum all the squared differences, and then divide the sum by the number of samples in the support set. The quotient is the prediction loss. Then, a gradient update is performed on the parameters of the complete model. The updated parameters are called temporary parameters. The gradient update calculation process is as follows: subtract the inner loop learning rate multiplied by the gradient of the prediction loss with respect to the parameters of the complete model from the parameters of the complete model. The difference is the temporary parameter. The inner loop learning rate is 0.

01.

8. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 7, characterized in that: The learning rate of the outer loop is set to 0.

001. The query loss of the temporary parameters is calculated using the query set of each virtual task in the outer loop. The query loss is also in the form of mean squared error. The calculation process is as follows: subtract the corresponding membrane thickness change from the predicted value of the membrane thickness change of all samples in the query set, square the difference, sum all the squared differences, and divide the sum by the number of samples in the query set. The quotient is the query loss. The query losses of all virtual tasks are summed, and the sum is used as the meta-target. The gradient of the meta-target with respect to the parameters of the complete model is calculated using the backpropagation algorithm, and this gradient is used as the meta-gradient. The parameters of the complete model are then updated using the meta-gradient. The calculation process of meta-gradient update is as follows: subtract the outer loop learning rate multiplied by the meta-gradient from the parameters of the complete model, and the difference is the updated parameters; the inner loop and outer loop are executed alternately for a total of 300 rounds, and source domain samples in the support set and query set are randomly re-sampled in each round to enhance task diversity; After training, the parameters of the complete model are saved as meta-initialization parameters.

9. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 8, characterized in that: In the target production line one-step fine-tuning and deployment module, the process of obtaining the fine-tuned complete model is as follows: The normal operation data was randomly divided into a support set and a validation set in a ratio of 7:

3. The number of samples in the support set was 70% of the total data in half an hour. The total data in half an hour was the total number of samples corresponding to a collection time of 1800 seconds and a sampling frequency of 1 Hz. During the fine-tuning process, a regularization term based on elastic weight consolidation is introduced. The calculation process of the regularization term is as follows: for each parameter of the complete model, the gradient of the log likelihood of the parameter with respect to each sample in the historical data of the source domain is calculated on the historical data of the source domain. The gradient is squared and then averaged over all samples in the source domain. The result is used as the importance measure of the parameter. Subsequently, the importance metric of each parameter is multiplied by the square of the difference between that parameter and the corresponding parameter in the meta-initialization parameters, and then multiplied by the regularization strength coefficient, which is set to 0.

1. Finally, the product of all parameters is summed to obtain the regularization term. The total loss function for fine-tuning consists of the mean squared error loss on the support set and the regularization term. The mean squared error loss is calculated as follows: subtract the corresponding film thickness change from the predicted value of the film thickness change of all samples in the support set, square the difference, sum all the squared differences, and then divide the sum by the number of samples in the support set. The quotient is the mean squared error loss. During fine-tuning, the initial values ​​of the meta-initialization parameters are used as the initial values. The gradient of the total loss function with respect to the parameters of the complete model is calculated on the support set, and a gradient descent update is performed: the parameters of the complete model are subtracted from the learning rate multiplied by the gradient. The learning rate is set to 0.001, and the fine-tuned complete model is obtained.

10. The adaptive control system for plastic extrusion lamination process based on machine learning according to claim 9, characterized in that: The learning rate for the step gradient descent update is set to 0.001, and the optimizer uses stochastic gradient descent with momentum, with the momentum coefficient set to 0.

9. The finely tuned complete model is deployed in the adaptive control of the new production line. The complete model receives the five-channel sensor sequence from 60 seconds prior to the current moment in real time. The five-channel sensor sequence contains time series data of melt temperature, die pressure, screw speed, traction speed, and film thickness measurement values, forming an input tensor and outputting a predicted value of the film thickness change for the next 10 seconds. The predicted value of the film thickness change is converted into screw speed compensation and temperature setpoint correction through a predefined process mapping table and sent to the underlying PID controller for execution. Meanwhile, the prediction error of the fine-tuned complete model is continuously monitored using the validation set. The prediction error is expressed as root mean square error. The calculation process is as follows: subtract the corresponding film thickness change from the predicted value of the film thickness change of all samples in the validation set, square the difference, sum all the squared differences, divide the sum by the number of samples in the validation set, and finally square root the quotient to obtain the root mean square error. If the root mean square error exceeds a preset threshold (0.8%), a re-fine-tuning process is triggered: collect half an hour of normal operating data again, randomly divide the normal operating data into a support set and a validation set in a 7:3 ratio, use the meta-initialization parameters as the initial values, and repeat the fine-tuning steps. The fine-tuning steps include calculating the regularization term, the total loss function, and performing one step of gradient descent update until the root mean square error does not exceed 0.8% or the preset maximum number of re-fine-tuning times is reached.

Images