A cutting film quality tracing system based on digital twinning

Abstract

The application relates to a film cutting quality tracing system based on digital twinning, in particular to the field of film cutting quality tracing of a film cutting machine, rapid retrieval and migration of historical process knowledge are realized through a cross-task process element knowledge base, the new task cold start problem is effectively solved, virtual simulation and constraint verification of candidate parameters are carried out by using a digital twinning model, physical trial and error risks are avoided in advance, production safety is ensured, a closed-loop feedback mechanism based on online learning and Bayesian optimization realizes self-adaptive optimization of process parameters, and the optimization efficiency and product quality are significantly improved, finally, the trajectory and model formed in the whole optimization process are deposited as new element knowledge, self-learning and continuous evolution ability are realized, the dependence on expert experience is reduced, the process debugging cycle is shortened, and material loss is reduced.

Description

Technical Field

[0001] This invention relates to the field of film cutting quality traceability, and more specifically, to a film cutting quality traceability system based on digital twins. Background Technology

[0002] As the manufacturing industry undergoes a profound transformation towards intelligence and flexibility, the film cutting processing field faces an increasing demand for multi-variety, small-batch customized production. In actual production scenarios, the introduction of a new film material, complex and intricate cutting patterns, or the replacement and adjustment of key components of the film cutting machine all represent a completely new process exploration. Traditional solutions heavily rely on the personal experience of operators and scattered, static historical process records. Operators need to manually set initial core parameters such as cutting speed, blade pressure, and tension based on experience, and approximate qualified products through a trial-and-error cycle of "production-inspection-adjustment." This process is not only lengthy in terms of debugging and significant waste of raw materials, but also makes it difficult to accumulate and standardize process knowledge for reuse. Although the introduction of digital twin technology provides a new approach for process simulation and virtual debugging, enabling certain predictions of parameter effects based on physical models, its prediction accuracy has inherent limitations in scenarios lacking sufficient actual production data for model calibration and training, especially when facing entirely new combinations of production conditions. It is difficult to directly use it for high-confidence parameter decisions.

[0003] Currently, existing technologies for determining parameters under the aforementioned new process conditions have the following shortcomings: First, methods relying on static process parameter databases for matching and querying rely on the complete coverage of historical data with the current scenario, lacking flexibility and unable to handle unseen combinations of production conditions. Second, methods entirely dependent on human experience are inefficient, yielding results that vary from person to person, and failing to systematically balance the core contradiction between "using existing knowledge to ensure a basic production pass rate" and "exploring the unknown parameter space to seek a better process window" during the exploration process. Third, while some data-driven optimization methods based on machine learning, such as Bayesian optimization or reinforcement learning, provide a framework for automated parameter optimization, their efficient operation typically depends on... A certain number of initial samples or online interactive data are needed to build accurate proxy models or strategies. In the "cold start" phase of a new process, historical data available for model training is extremely sparse or even missing. Direct application of such data may lead to excessively high quality risks in the initial recommended parameters, resulting in a large number of defective products. Conversely, if the system is overly conservative in order to avoid risks, it may become trapped in local optima, failing to fully utilize the potential of the equipment and the characteristics of the materials, and losing the management value of achieving process breakthroughs and continuously improving production quality and efficiency through intelligent means. Therefore, there is an urgent need for an intelligent decision-making method that can effectively integrate physical laws, limited experience, and online learning capabilities under conditions of data scarcity, and can proactively manage and explore risks to support robust and adaptive optimization of film cutting process parameters. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a film cutting quality traceability system based on digital twins, 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 cross-task process meta-knowledge base construction module, a proxy model initialization module, a parameter pre-screening module, and an online optimization and update module connected in sequence, wherein; Cross-task process meta-knowledge base construction module: When receiving a new film cutting task, it constructs a cross-task process meta-knowledge base based on the historical task feature vectors, historical optimal parameter sets and historical optimized trajectory data stored in the system, and calculates the generalized process similarity between the task feature vector of the new film cutting task and the feature vectors of each historical task in the cross-task process meta-knowledge base, so as to retrieve the subset of similar historical task feature vectors and their associated optimized trajectory data from the cross-task process meta-knowledge base; Proxy model initialization module: Using a small amount of historical optimization trajectory data corresponding to each historical task in the feature vector subset of similar historical tasks as a support set, the module performs gradient update operation on the shared base model pre-trained through the meta-learning framework to generate an initial proxy model for the new membrane cutting task. Parameter pre-screening module: Generates an initial acquisition function based on the initialization proxy model, and samples multiple candidate parameter combinations within the preset parameter space; calls the system's built-in digital twin model to perform virtual cutting simulation on each candidate parameter combination, and obtains the predicted quality index corresponding to each candidate parameter combination; based on whether each predicted quality index violates the system's preset process constraints, dynamically trims and reorders all candidate parameter combinations, and outputs the recommended parameter combinations that are ranked first and verified by virtual simulation to meet the process constraints to the physical film cutting machine for execution; Online optimization and update module: Collect real quality feedback data after the physical film cutting machine executes the recommended parameter combination, use the real quality feedback data to iteratively update the initial proxy model, obtain the updated proxy model, and send the updated proxy model, the final optimized trajectory data of the new film cutting task and the task feature vector as the new task package back to the cross-task process meta-knowledge base construction module to incrementally update the cross-task process meta-knowledge base; In a preferred embodiment, in the cross-task process meta-knowledge base construction module, the task feature vector of the new film cutting task is composed of physical property parameters for characterizing the processed film material and geometric feature parameters for characterizing the cutting path.

[0006] In a preferred embodiment, the process of calculating the generalized process similarity between the task feature vector of the new film cutting task and the feature vectors of each historical task in the cross-task process meta-knowledge base specifically includes: Extract the task feature vector of the new film cutting task, and extract a historical record from the cross-task process meta-knowledge base. This historical record contains a historical task feature vector, a historical optimal parameter set, and a historical optimization trajectory. The task feature vector of the new film cutting task and the historical task feature vector in the historical record are input into the generalized process similarity calculation model; the generalized process similarity calculation model calculates the first distance between the task feature vector of the new film cutting task and the historical task feature vector in the original space; The generalized process similarity calculation model also includes a pre-trained encoder neural network. The encoder neural network takes the historical task feature vector and the historical optimization trajectory as input to the historical record and outputs the task embedding vector of the historical record. For a new membrane cutting task, its task embedding vector is generated by the encoder neural network based on the task feature vector of the new membrane cutting task and combined with the response pattern represented by the historical optimization trajectory in at least one historical record that has the minimum first distance from the new membrane cutting task. Calculate the second distance between the task embedding vector of the new membrane cutting task and the task embedding vector of the historical record; add the product of the first distance and a trainable balancing weight to the product of the second distance and the difference between the numerical value and the trainable balancing weight to obtain a composite distance value; calculate the exponential function value with the natural constant e as the base and the negative of the composite distance value as the exponent, and use the exponential function value as the generalized process similarity between the new membrane cutting task and the historical task represented by the historical record.

[0007] In a preferred embodiment, the process of retrieving a subset of similar historical task feature vectors and their associated optimized trajectory data from the cross-task process meta-knowledge base specifically involves: After calculating the generalized process similarity of the new film cutting task, all historical records are sorted in descending order according to the numerical value of the generalized process similarity. Based on a preset positive integer threshold, multiple historical records with the number equal to the threshold are selected from the first few records in the descending order of historical records. The historical optimization trajectories contained in these selected historical records are extracted to form a set of similar historical optimization trajectories. The historical task feature vectors contained in these selected historical records are extracted to form a subset of similar historical task feature vectors. The cross-task process meta-knowledge base construction module outputs the subset of similar historical task feature vectors and the set of similar historical optimization trajectories to the proxy model initialization module.

[0008] In a preferred embodiment, the specific operation of the proxy model initialization module in using a small amount of historical optimized trajectory data corresponding to each historical task in the subset of similar historical tasks as a support set is as follows: The system receives the set of similar historical optimization trajectories and the subset of similar historical task feature vectors output by the cross-task process meta-knowledge base construction module; extracts a preset number of parameter and performance data pairs located at the beginning of each historical optimization trajectory from each historical optimization trajectory in the set of similar historical optimization trajectories; assigns a weight coefficient to each parameter and performance data pair extracted from the corresponding historical optimization trajectory based on the value of the generalized process similarity between the historical task corresponding to each historical task feature vector in the subset of similar historical task feature vectors and the new film cutting task; and aggregates the parameter and performance data pairs from all historical tasks with attached weight coefficients to form a weighted support set.

[0009] In a preferred embodiment, the process of performing gradient update operation on the shared base model pre-trained through the meta-learning framework specifically includes: Calculate the overall loss function value of the shared base model on the weighted support set. The overall loss function value is the sum of the weighted losses of each parameter and performance data pair in the weighted support set. The loss of each data pair is composed of the product of its weight coefficient and the difference function value between the predicted performance of the parameter in the data pair and the true performance in the data pair by the shared base model. Calculate the gradient vector of the overall loss function value with respect to the trainable parameters of the shared base model. Obtain a meta-gradient modulation factor vector, which is generated during the meta-learning pre-training phase of the shared base model and has the same dimension as the trainable parameters of the shared base model. The gradient vector is multiplied element-wise with the meta-gradient modulation factor vector to obtain the modulated gradient vector; a preset learning rate is used to multiply the modulated gradient vector to obtain the parameter update amount; the parameter update amount is subtracted from the current trainable parameters of the shared base model to obtain the updated model parameters. The model defined by the updated model parameters is the initialized surrogate model. The surrogate model initialization module outputs the initialized surrogate model to the parameter pre-screening module.

[0010] In a preferred embodiment, the parameter pre-screening module generates an initial acquisition function based on an initialization proxy model and samples multiple candidate parameter combinations within a preset parameter space as follows: The module receives an initialized surrogate model from the surrogate model initialization module. This initialized surrogate model is configured to output a prediction performance score and an uncertainty value characterizing the prediction confidence for the input parameter combination. The parameter pre-selection module constructs an initial acquisition function. The calculation process of the initial acquisition function for a parameter combination is as follows: Add a non-negative exploration coefficient to the predictive performance score corresponding to this parameter combination and the product of the uncertainty value corresponding to this parameter combination, and then subtract a non-negative constraint avoidance coefficient and the estimated confidence level of this parameter combination violating process constraints. The result is the acquisition function value corresponding to this parameter combination. Within a preset parameter space, the parameter pre-screening module generates multiple parameter combinations of a first preset number based on the distribution of the acquisition function values ​​of the constructed initial acquisition function using the Monte Carlo sampling method, forming an initial candidate parameter combination set.

[0011] In a preferred embodiment, the specific operation of dynamically trimming and reordering all candidate parameter combinations based on whether each predicted quality index violates the system's preset process constraints is as follows: A digital twin model is invoked to perform virtual cutting simulation on each parameter combination in the initial candidate parameter combination set, obtaining a set of predicted quality indicators. The predicted quality indicators of each parameter combination are compared with the system's preset process constraints, generating a constraint violation flag vector for each parameter combination. Parameter combinations with at least one flag value of one in all constraint violation flag vectors are deleted, forming a safe candidate parameter subset. For each remaining parameter combination in the safe candidate parameter subset, a comprehensive ranking score is calculated. This comprehensive ranking score is obtained by weighted summation of three parts: the first part is the model guidance score, which is the product of a model guidance weight coefficient and the normalized initial acquisition function value; the second part is the simulation performance score. The first part is the product of a simulation performance weight coefficient and a comprehensive simulation performance score, which is obtained by normalizing and weighting the various prediction quality indicators output by the digital twin model. The second part is the diversity score, which is the product of a diversity reward weight coefficient and a diversity reward item, which is calculated based on the minimum distance between the current parameter combination and all parameter combinations in a set of parameter combinations selected into a temporary recommendation set in the parameter space. According to the calculated comprehensive ranking score, all parameter combinations in the safe candidate parameter subset are sorted in descending order, and the top-ranked parameter combinations with a quantity of the second preset number are selected to form a recommended parameter combination set, which is then output to the physical film cutting machine.

[0012] In a preferred embodiment, the online optimization and update module uses real quality feedback data to iteratively update the initial agent model to obtain the updated agent model. The specific process is as follows: The system receives the set of recommended parameter combinations and the predicted quality index of the corresponding parameter combinations in the digital twin model from the parameter pre-screening module, and receives the actual quality feedback data returned after execution by the physical film cutting machine. For each recommended parameter combination in the set of recommended parameter combinations and its corresponding actual quality feedback data, a data confidence score is calculated. The value of the data confidence score is equal to the product of two factors: the first factor is an exponential function with the natural constant e as the base, where the exponent of the exponential function is a negative real number, which is obtained by dividing the absolute value of the difference between the actual quality index and the predicted quality index by a normalized constant; the second factor is the absolute value of a difference, which is obtained by subtracting a preset uncertainty threshold value from the value of the prediction uncertainty of the recommended parameter combination by the initial surrogate model. Then, using all recommended parameter combinations and their corresponding data confidence scores, the system calculates a data confidence score. The quality feedback data is used to perform a weighted incremental update on the initial surrogate model. The update process is achieved by minimizing a composite loss function, which is the sum of three terms: the first term is the weighted sum of the losses of all data points in the original dataset of the initial surrogate model, where the weight of each original data point is a coefficient that decays over time; the second term is the weighted sum of the losses of all data points in the newly added real quality feedback dataset, where the weight of each new data point is the calculated data confidence score; the third term is a regularization term, which is obtained by multiplying the KL divergence between the probability distribution of the trainable parameters of the updated surrogate model and the probability distribution of the trainable parameters of the surrogate model before the update by a regularization strength coefficient; the trainable parameters of the updated surrogate model are obtained by minimizing the composite loss function, and the model defined by the updated trainable parameters is the updated surrogate model.

[0013] In a preferred embodiment, the specific process of transmitting the updated proxy model, the final optimized trajectory data of the new film cutting task, and the task feature vector as a new task package back to the cross-task process meta-knowledge base construction module is as follows: After completing the iterative update of the initial agent model, the current optimization process is evaluated to determine whether it should be terminated. The evaluation method involves calculating a continued optimization benefit index, which is equal to the product of three factors: the first factor is the difference between the average expected improvement value after the model update and before the model update on the subset of safe candidate parameters generated by the parameter pre-screening module, divided by the moving average of the average expected improvement values ​​of the most recent iterations before the model update; the second factor is the average prediction uncertainty value before the model update divided by the average prediction uncertainty value after the model update; the third factor is the remaining physical experiment budget value divided by the total physical experiment budget value. If the continued optimization benefit index is lower than a preset threshold, or the model's predicted performance... If the score reaches the preset target, the optimization process is terminated. After the optimization process terminates, the online optimization update module extracts the complete final optimization trajectory data of this new film cutting task, along with multiple dynamic meta-information that records key features of the optimization process. The final optimization trajectory data, task feature vector, updated proxy model, and dynamic meta-information are packaged together into a new task package, which is then sent to the cross-task process meta-knowledge base construction module. After receiving the new task package, the cross-task process meta-knowledge base construction module updates the encoder neural network parameters and trainable balanced weights in its internal generalized process similarity calculation model based on the final optimization trajectory data and dynamic meta-information contained in the package, thereby achieving incremental updates to the cross-task process meta-knowledge base.

[0014] The beneficial effects of this invention are as follows: By enabling rapid retrieval and transfer of historical process knowledge through a cross-task process meta-knowledge base, the cold start problem of new tasks is effectively solved; by using a digital twin model to perform virtual simulation and constraint verification of candidate parameters, the risk of physical trial and error is avoided in advance, ensuring production safety; based on the closed-loop feedback mechanism of online learning and Bayesian optimization, adaptive optimization of process parameters is achieved, significantly improving optimization efficiency and product quality; finally, the trajectory and model formed by the entire optimization process are precipitated as new meta-knowledge, realizing self-learning and continuous evolution capabilities, reducing dependence on expert experience, shortening the process debugging cycle and reducing material loss. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] 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.

[0017] 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.

[0018] 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.

[0019] Example 1 This embodiment provides, for example Figure 1-2 The illustrated film cutting quality traceability system includes, in detail: a cross-task process meta-knowledge base construction module, a proxy model initialization module, a parameter pre-screening module, and an online optimization and update module, which are connected sequentially. Cross-task process meta-knowledge base construction module: When receiving a new film cutting task, it constructs a cross-task process meta-knowledge base based on the historical task feature vectors, historical optimal parameter sets and historical optimized trajectory data stored in the system, and calculates the generalized process similarity between the task feature vector of the new film cutting task and the feature vectors of each historical task in the cross-task process meta-knowledge base, so as to retrieve the subset of similar historical task feature vectors and their associated optimized trajectory data from the cross-task process meta-knowledge base; Proxy model initialization module: Using a small amount of historical optimization trajectory data corresponding to each historical task in the feature vector subset of similar historical tasks as a support set, the module performs gradient update operation on the shared base model pre-trained through the meta-learning framework to generate an initial proxy model for the new membrane cutting task. Parameter pre-screening module: Generates an initial acquisition function based on the initialization proxy model, and samples multiple candidate parameter combinations within the preset parameter space; calls the system's built-in digital twin model to perform virtual cutting simulation on each candidate parameter combination, and obtains the predicted quality index corresponding to each candidate parameter combination; based on whether each predicted quality index violates the system's preset process constraints, dynamically trims and reorders all candidate parameter combinations, and outputs the recommended parameter combinations that are ranked first and verified by virtual simulation to meet the process constraints to the physical film cutting machine for execution; Online optimization and update module: Collect real quality feedback data after the physical film cutting machine executes the recommended parameter combination, use the real quality feedback data to iteratively update the initial proxy model to obtain the updated proxy model, and send the updated proxy model, the final optimized trajectory data of the new film cutting task and the task feature vector as the new task package back to the cross-task process meta-knowledge base construction module to incrementally update the cross-task process meta-knowledge base.

[0020] In this embodiment, it is specifically noted that in the cross-task process meta-knowledge base construction module, the task feature vector of the new film cutting task is composed of physical property parameters used to characterize the processed film material and geometric feature parameters used to characterize the cutting path. The physical property parameters may include at least one of the film material's thickness, elastic modulus, elongation at break, and coefficient of thermal expansion. The geometric feature parameters may include at least one of the total length of the cutting path, the number of turning points, the minimum radius of curvature, and the graphic complexity index. By quantifying the specific physical and geometric properties into feature vectors, a measurable and unified digital representation basis is provided for subsequent similarity calculations, making different tasks comparable. The specific process for calculating the generalized process similarity between the task feature vector of the new membrane cutting task and the feature vectors of each historical task in the cross-task process meta-knowledge base is as follows: Extract the task feature vector of the new film cutting task, and extract a historical record from the cross-task process meta-knowledge base. The historical record contains a historical task feature vector, a historical optimal parameter set, and a historical optimization trajectory consisting of multiple parameter and performance data pairs arranged in chronological order. The parameters include cutting speed and blade pressure, and the performance includes cutting accuracy score. The task feature vector of the new film cutting task and the historical task feature vector in this historical record are input into the generalized process similarity calculation model. The generalized process similarity calculation model first calculates the first distance between the task feature vector of the new film cutting task and the historical task feature vector in the original space spanned by all feature parameters. The first distance is the normalized Euclidean distance. The normalization process adopts the minimum-maximum scaling method to linearly transform the value of each dimension in the task feature vector to the interval [0,1] to eliminate the bias caused by different physical dimensions and numerical magnitudes in the distance calculation, and ensure that each dimension of the feature has an equally important contribution to the similarity measurement. The generalized process similarity calculation model also includes a pre-trained encoder neural network. The encoder neural network can employ a hybrid structure containing long short-term memory (LSM) layers and fully connected layers. The LSM layers process historical optimization trajectories in sequence form, capturing the dynamic dependencies between parameters and performance over time. The fully connected layers fuse the processed trajectory features with historical task feature vectors, ultimately outputting a fixed-dimensional task embedding vector. The training objective of this network is to learn, through a large amount of historical task data, to map tasks with similar optimization behaviors to positions close to each other in the embedding space. For this historical record, the encoder neural network takes as input the historical task feature vector and the historical optimization trajectory, and outputs a fixed-dimensional task embedding vector representing the position of this historical record in the implicit embedding space. For the new membrane cutting task, its task embedding vector is generated by the encoder... The encoder neural network generates a response pattern based on the task feature vector of the new membrane cutting task and combined with the response pattern jointly represented by the historical optimization trajectory in at least one historical record with the minimum first distance to the new membrane cutting task. A specific generation method is as follows: First, select the top K historical records most similar to the new membrane cutting task based on the first distance; then, extract the historical optimization trajectory of these K historical records, and calculate the average trend or principal component of these trajectories with respect to parameters on key performance indicators (such as cutting accuracy and burr height) as an aggregated response pattern; finally, input the task feature vector of the new membrane cutting task and this aggregated response pattern into the trained encoder neural network, and the network outputs the task embedding vector of the new membrane cutting task. This method enables the new task to obtain a meaningful embedding representation through the group behavior of similar tasks even if it does not have its own historical data. Calculate the second distance between the task embedding vector of the new membrane cutting task and the task embedding vector of this historical record. The second distance is the value obtained by subtracting the cosine similarity between the two task embedding vectors. The product of the first distance and a trainable balancing weight is added to the product of the second distance and the difference between the first distance and the trainable balancing weight, to obtain a composite distance value. An exponential function value is calculated with the natural constant e as the base and the negative of the composite distance value as the exponent. This exponential function value is used as the generalized process similarity between the new membrane cutting task and the historical task represented by this historical record. The trainable balancing weight ranges from zero to one. The trainable balancing weight is obtained through cross-validation training on the historical task dataset. The training objective is to minimize the difference between the ranking order of the generalized process similarity prediction and the transferability ranking of optimization difficulty between real tasks. When its value is close to 1, the generalized process similarity mainly depends on static feature similarity; when its value is close to 0, it mainly depends on dynamic optimization behavior similarity. By automatically learning this weight, the model can adaptively adjust the similarity metric according to different types of tasks, improving retrieval accuracy. The specific process of retrieving a subset of similar historical task feature vectors and their associated optimized trajectory data from the cross-task process meta-knowledge base is as follows: After calculating the generalized process similarity between the task feature vector of the new membrane cutting task and the historical task feature vector contained in each historical record in the cross-task process meta-knowledge base, all historical records are sorted in descending order according to the value of the generalized process similarity from largest to smallest. Based on a preset positive integer threshold, select multiple historical records that are at the top of the descending sequence of historical records and whose number equals the threshold. The threshold can be set according to the overall size of the historical task library and computing resources. For example, it can usually be set to an integer between 5 and 20. If the threshold is too small, it may lead to insufficient supporting information, and if it is too large, it may introduce noise from irrelevant tasks. By setting a reasonable threshold, a balance can be achieved between the sufficiency of knowledge transfer and computational efficiency. Extract the historical optimization trajectories contained in these selected historical records to form a set of similar historical optimization trajectories; Extract the historical task feature vectors contained in these selected historical records to form a subset of similar historical task feature vectors; The cross-task process meta-knowledge base construction module outputs a subset of similar historical task feature vectors and a set of similar historical optimization trajectories to the proxy model initialization module. The elements in the subset of similar historical task feature vectors and the set of similar historical optimization trajectories are arranged in descending order of generalized process similarity. The cross-task process meta-knowledge base construction module packages these two sets and the corresponding generalized process similarity ranking information together and outputs them to the proxy model initialization module. The proxy model initialization module then uses this ranking information to prioritize the use of data corresponding to historical tasks with higher similarity to initialize the proxy model, thereby ensuring the effectiveness and priority of knowledge transfer.

[0021] In this embodiment, it is specifically necessary to explain the operation of using a small amount of historical optimized trajectory data corresponding to each historical task in the similar historical task subset as the support set in the proxy model initialization module as follows: Receives a set of similar historical optimization trajectories and a subset of similar historical task feature vectors output by the cross-task process meta-knowledge base construction module; extracts a preset number of parameter-performance data pairs located at the beginning of each historical optimization trajectory from each historical optimization trajectory in the set of similar historical optimization trajectories; the preset number is usually set to three to ten data points, depending on the total length of the historical optimization trajectory. Extracting data points at the beginning aims to capture the parameter-performance relationship of the historical task in the early stage of exploration. These relationships have key guiding value for the initial model construction of the new task, avoiding the use of data that may have been overfitted to a specific historical task in the later stages; Based on the generalized process similarity value between the historical task and the new film cutting task, calculated and provided by the cross-task process meta-knowledge base construction module, for each historical task feature vector in the similar historical task feature vector subset, a weight coefficient is assigned to each parameter and performance data pair extracted from the corresponding historical optimization trajectory. The weight coefficients are assigned in the following way: First, the generalized process similarity value between the current historical task and the new film cutting task is divided by the number of parameter and performance data pairs extracted from that historical task to obtain the basic weight of each parameter and performance data pair under that historical task. Then, all basic weights are normalized so that the sum of the weight coefficients of all data pairs is one. This allocation method ensures that the data of high-similarity historical tasks receive higher importance, while maintaining the stability of the weights in terms of value. The higher the value of the generalized process similarity of the historical task, the larger the weight coefficient assigned to its corresponding data pair. By aggregating the historical tasks corresponding to all similar historical task feature vector subsets, along with parameter and performance data pairs with weighted coefficients, a weighted support set is formed. Constructing the weighted support set essentially involves fusing initial exploration experiences with different confidence levels from multiple similar historical tasks to form a prior knowledge dataset with confidence weights for the new membrane cutting task. This provides training samples of uneven quality but with clear direction for subsequent gradient update operations, which is the key data preparation for achieving fast and accurate model initialization. The specific process of performing gradient update operations on a shared base model pre-trained through a meta-learning framework is as follows: Calculate the overall loss function value of the shared base model on the weighted support set. The overall loss function value is the sum of the weighted losses of each parameter and performance data pair in the weighted support set. The loss of each data pair is composed of its weight coefficient and the product of the difference function value between the predicted performance of the parameter in the data pair and the actual performance of the data pair by the shared base model. The difference function is usually the mean squared error function or the mean absolute error function, which is used to quantify the deviation between the model's predicted value and the actual performance value. Calculate the gradient vector of the overall loss function value relative to the trainable parameters of the shared base model. A meta-gradient modulation factor vector is obtained. This vector is generated during the meta-learning pre-training phase of the shared base model, and its dimension is the same as that of the trainable parameters of the shared base model. During the meta-learning pre-training phase, the shared base model is trained on different historical tasks, and the gradient vector of each trainable parameter on the loss function of multiple tasks is recorded. For each trainable parameter, the average cosine similarity between its gradient directions on different tasks is calculated. This average cosine similarity value is transformed by a Sigmoid function and scaled to the [0,1] interval to obtain the value of the corresponding dimension in the meta-gradient modulation factor vector. The higher the cosine similarity, the more consistent the optimization direction of the parameter is across different tasks, and the closer its modulation factor is to 1. The larger the adjustment range allowed when initializing a new task model, the higher the value of each dimension in the meta-gradient modulation factor vector is. The value of each dimension is a real number between zero and one, which is used to characterize the consistency of the gradient direction of the trainable parameter at the corresponding position when it adapts to different historical tasks during the meta-learning training phase. The higher the consistency, the closer the value of this dimension is to one, and vice versa. The gradient vector is multiplied element-wise with the meta-gradient modulation factor vector to obtain the modulated gradient vector. A preset learning rate is used to multiply the modulated gradient vector to obtain the parameter update amount. The learning rate is usually set to a small positive number, such as between 0.001 and 0.1, to ensure the stability of parameter updates and avoid model performance oscillation or degradation due to excessively large single update steps. The parameter update amount is subtracted from the current trainable parameters of the shared base model to obtain the updated model parameters. The model defined by the updated model parameters is the initialized surrogate model. The surrogate model initialization module outputs the initialized surrogate model to the parameter pre-screening module. By introducing a meta-gradient modulation factor to modulate the original gradient element-wise, the gradient update operation implemented in this step can differentiate between different parameters in the shared base model: parameters that exhibit general and stable optimization directions in historical tasks are adjusted rapidly and significantly to adapt them quickly to new tasks; while parameters that have variable optimization directions in different tasks and are more task-specific are adjusted more gradually or suppressively to preserve their generalization properties. This method achieves a fine balance between "rapid adaptation" and "prevention of overfitting" during model initialization with limited prior data for new tasks, significantly improving the starting performance and robustness of the initial surrogate model.

[0022] In this embodiment, it is specifically necessary to explain that in the parameter pre-screening module, the process of generating an initial acquisition function based on the initialization proxy model and sampling multiple candidate parameter combinations within a preset parameter space is as follows: The module receives an initialized surrogate model from the surrogate model initialization module. This initialized surrogate model is configured to output a prediction performance score and an uncertainty value characterizing the prediction confidence for the input parameter combination. The parameter pre-selection module constructs an initial acquisition function. The calculation process of the initial acquisition function for a parameter combination is as follows: Add a non-negative exploration coefficient to the predictive performance score corresponding to this parameter combination and the product of the uncertainty value corresponding to this parameter combination, and then subtract a non-negative constraint avoidance coefficient and the estimated confidence level of this parameter combination violating process constraints. The result is the acquisition function value corresponding to this parameter combination. The exploration coefficient adjusts the degree of attention paid to the uncertainty of model predictions. A higher value indicates a greater tendency to sample regions with high model uncertainty, thus promoting exploration. It typically ranges from 1.0 to 3.0. The constraint avoidance coefficient adjusts the strength of avoidance of regions where the prediction may violate constraints. A higher value indicates a greater tendency to sample regions with high model confidence levels. It typically ranges from 0.5 to 2.0. The specific values ​​of these two coefficients can be determined through cross-validation on historical data. The estimated confidence level of process constraint violations can be output by an auxiliary classifier that initializes the surrogate model. This classifier is trained alongside the surrogate model during training to learn the relationship between parameter combinations and the probability of violating various process constraints. Alternatively, it can be estimated by analyzing the constraint violation frequency of similar parameter combinations in historical data. Within a preset parameter space, the parameter pre-screening module generates multiple parameter combinations of a first preset number based on the distribution of the acquisition function values ​​of the constructed initial acquisition function using the Monte Carlo sampling method, forming an initial candidate parameter combination set. The initial preset number determines the breadth of the initial sampling, which is usually set to a large value, such as between 100 and 1000, to ensure that the potential regions in the parameter space are fully covered and to provide enough candidate samples for subsequent virtual simulation verification. The Monte Carlo sampling method can perform random but sampled data guided by the acquisition function in the parameter space, which is more efficient than grid search. It is especially suitable for scenarios with high dimensionality or irregular shape of the parameter space and can obtain a batch of promising candidate points at a controllable computational cost. Based on whether each predicted quality indicator violates the system's preset process constraints, the specific operation of dynamically trimming and reordering all candidate parameter combinations is as follows: The digital twin model is invoked to perform virtual cutting simulation on each parameter combination in the initial candidate parameter combination set to obtain a set of predicted quality indicators based on physical calculations. The predicted quality indicators of each parameter combination are compared with the system's preset process constraints. Each process constraint is defined in the form of an inequality. A constraint violation flag vector is generated for each parameter combination. Each dimension of the constraint violation flag vector corresponds to a process constraint. If the predicted quality indicator of this parameter combination violates the corresponding process constraint, the flag value of the corresponding dimension is one; otherwise, it is zero. All parameter combinations in the constraint violation flag vector with at least one flag value of one are deleted to form a safe candidate parameter subset. For each remaining parameter combination in the subset of safety candidate parameters, a comprehensive ranking score is calculated. This comprehensive ranking score is obtained by a weighted sum of three parts: the first part is the model guidance score, which is the product of a model guidance weight coefficient and the normalized initial acquisition function value; the second part is the simulation performance score, which is the product of a simulation performance weight coefficient and a comprehensive simulation performance score. The comprehensive simulation performance score is calculated by weighted aggregation of the various predicted quality indicators output by the digital twin model after normalization. First, for each predicted quality indicator, such as cutting accuracy, burr height, and cutting efficiency, minimum-maximum normalization is performed to transform it to the [0,1] interval, with larger values ​​indicating better performance; then, an importance weight is assigned to each normalized indicator; finally, all weighted indicator values ​​are summed to obtain the comprehensive simulation performance score. The indicator weights can be set according to the priority of the actual production process; the third part is... The diversity score is the product of a diversity reward weight coefficient and a diversity reward item. The diversity reward item is calculated based on the minimum distance in the parameter space between the current parameter combination and all parameter combinations in a set of parameter combinations selected into a temporary recommendation set. The larger the distance, the larger the value of the diversity reward item. The sum of the model guidance weight coefficient and the simulation performance weight coefficient is one, and both are real numbers between zero and one. The diversity reward weight coefficient is a non-negative real number. The model guidance weight coefficient and the simulation performance weight coefficient reflect the degree of trust in the surrogate model's predictive information and the direct results of the digital twin simulation, respectively. In the early stages of optimization, a higher model guidance weight can be set to encourage exploration. In the later stages or when the safety requirements are extremely high, a higher simulation performance weight can be set. The diversity reward weight coefficient is used to control the degree of encouragement for exploring diversity, and its value is usually small (e.g., 0.05 to 0.2) to prevent excessive pursuit of diversity at the expense of performance. Based on the calculated comprehensive ranking score, all parameter combinations in the safe candidate parameter subset are sorted in descending order. The top-ranked parameter combinations, which are a second preset number, are selected to form a recommended parameter combination set, which is then output to the physical film cutting machine. The second preset number determines the number of parameter combinations actually sent to the physical film cutting machine for execution in each iteration. It is usually set to a small positive integer, such as 3 to 5. This allows for multiple sets of feedback to be obtained through parallel or fast serial execution to accelerate optimization, while also controlling the trial and error cost and time of a single iteration. By first using a digital twin model to conduct rigorous hard verification of constraint compliance to eliminate unsafe options, and then combining model guidance, simulation performance, and exploration diversity for refined reordering, this step ensures that the final output set of recommended parameter combinations not only absolutely meets process safety constraints, but also represents the optimal combination of potential, performance, and exploration value under current understanding. This maximizes the efficiency and value of each execution of the physical film cutting machine, achieving synergy between full optimization in virtual space and efficient verification in physical space.

[0023] In this embodiment, it is specifically necessary to explain that in the online optimization and update module, the process of iteratively updating the initial agent model using real quality feedback data to obtain the updated agent model is as follows: The online optimization and update module receives a set of recommended parameter combinations and their predicted quality indices in the digital twin model from the parameter pre-screening module. It also receives real quality feedback data returned by the physical film cutting machine after executing these recommended parameter combinations. For each recommended parameter combination and its corresponding real quality feedback data, a data confidence score is calculated. The data confidence score is equal to the product of two factors: the first factor is an exponential function with the natural constant e as its base, where the exponent is a negative real number. This negative real number is obtained by dividing the absolute value of the difference between the real quality index in the real quality feedback data corresponding to this recommended parameter combination and the predicted quality index in the digital twin model by a normalized constant. The second factor is a product of the difference between the two factors. The difference is obtained by subtracting a preset uncertainty threshold from the predicted uncertainty of the initial surrogate model for this combination of recommended parameters. This ensures that the closer the predicted quality indicator is to the actual quality indicator and the closer the predicted uncertainty is to the preset uncertainty threshold, the higher the calculated data confidence score. The normals used for normalization are usually set as the average or median of the absolute difference between the predicted and actual quality indicators in historical data to ensure that the input value of the exponential function is within a reasonable range. The uncertainty threshold can be set according to the statistical distribution of the predicted uncertainty of the initial surrogate model in the historical safe parameter region (such as the 70th percentile). Its purpose is to distinguish between the "relatively certain" and "relatively uncertain" regions of the model, thereby giving higher information value weights to data points that fall near the cognitive boundary. Then, using all recommended parameter combinations and their associated real quality feedback data with data confidence scores, the initialized surrogate model is updated incrementally with weights. The update process is achieved by minimizing a composite loss function, which is the sum of three terms: the first term is the weighted sum of the losses of all data points in the dataset consisting of the weighted support set used in the generation process of the initialized surrogate model from the surrogate model initialization module and the historical data associated from the cross-task process meta-knowledge base. The weight of each original data point is a coefficient that decays over time, monotonically decreasing as the data point's storage time increases; the second term is the weight of the data point itself. The time decay coefficient can be calculated using the exponential decay rule, i.e., the coefficient is equal to a power of the "data point age" with a base less than one (e.g., 0.9), where the "data point age" is in units of optimization iterations. This design gradually weakens the influence of earlier historical data, preventing outdated knowledge from hindering the model's adaptation to recent patterns, and achieving smooth updating and forgetting of model knowledge. The second term is the weighted sum of the losses of all data points in the newly added real quality feedback dataset, where the weight of each new data point is the calculated data confidence score. The third term is a regularization term, which is applied to the updated surrogate model. The trainable parameters of the updated surrogate model are obtained by multiplying the KL divergence between the probability distribution of the trainable parameters of the new surrogate model and the probability distribution of the trainable parameters of the surrogate model before the update by a regularization strength coefficient. The regularization strength coefficient controls the degree of similarity that the parameter distributions of the updated and new models should maintain. Its value is usually set to a small positive number (e.g., 0.01 to 0.1). The larger the coefficient, the closer the updated model is to the original model, which helps stabilize the training process, prevents overfitting on a small amount of new data, and achieves a balance between adapting to new evidence and retaining existing knowledge. The trainable parameters of the updated surrogate model are obtained by minimizing the composite loss function. The updated surrogate model is defined by the updated trainable parameters. By designing data confidence scores that simultaneously consider "simulation-reality consistency" and "cognitive boundary proximity," and by adopting different weighting strategies for new and old data (new data by confidence, old data by time decay), supplemented by KL divergence regularization, the weighted incremental update process guided by the composite loss function can intelligently filter high-value physical experiment data, gradually reduce outdated knowledge, and stably integrate new information. This significantly improves the robustness and efficiency of model updates, ensuring that the model can quickly converge to a better state after a small number of physical experiments. The specific process of sending the updated proxy model, the final optimized trajectory data of the new film cutting task, and the task feature vector as a new task package back to the cross-task process meta-knowledge base construction module is as follows: After completing the iterative update of the initial agent model, the current optimization process is evaluated to determine whether it should be terminated. The evaluation method involves calculating a continued optimization benefit index, which is equal to the product of three factors: the first factor is the difference between the average expected improvement value after the model update and before the model update on the safe candidate parameter subset generated by the parameter pre-screening module, divided by the moving average of the average expected improvement values ​​from the most recent iterations before the model update. The average expected improvement value is calculated by averaging the initial acquisition function constructed in the parameter pre-screening module on the corresponding parameter set. The second factor is the average prediction uncertainty value before the model update divided by the average prediction uncertainty value after the model update. The predicted uncertainty is calculated from the uncertainty value output by the initial surrogate model on the corresponding parameter set. The third factor is the remaining physical test budget divided by the total physical test budget. The physical test budget limits the total number of recommended parameter combinations that can be sent to the physical film cutting machine for execution. The total physical test budget can be determined based on the production batch size of the new film cutting task, the cost per test, and the allowable debugging time. For example, for small-batch, high-value tasks, the budget may be set lower (e.g., 10-20 tests); for large-scale production or process research, the budget may be set higher (e.g., 50-100 tests). This budget serves as a hard constraint, ensuring that the total cost of the optimization process is controllable. If the optimization yield index falls below a preset threshold, typically a near-zero positive decimal (e.g., 0.05), it indicates that further iterations will result in very limited performance improvement, and the optimization process may have plateaued. Terminating optimization at this point conserves computational and physical experimental resources, or if the model's predicted performance score has reached the preset target, the optimization process is terminated. After termination, the online optimization update module extracts the complete final optimization trajectory data for the new film-cutting task. This data includes all tried parameter combinations, the digital twin model's predicted quality indices for each combination, the actual quality feedback data returned by the physical film-cutting machine, and the number of optimization iterations, along with multiple dynamic metadata recording key features of the optimization process. The dynamic metadata includes the shape of the learning curve describing the model performance changes with iteration, the types of process constraints repeatedly violated during the optimization process, and the direction of trainable parameters that have been significantly adjusted during the update process of the surrogate model. Recording the shape of the learning curve helps to judge the optimization difficulty of similar tasks in the future; recording the types of process constraints repeatedly violated can reveal the common process bottlenecks of specific materials or patterns; recording the direction of model parameters that have been significantly adjusted can reflect the core model features on which the optimization of this task depends. This dynamic metadata provides the meta-knowledge base with process knowledge about "how to optimize" that goes beyond static input-output relationships. The final optimization trajectory data, task feature vector, updated surrogate model and dynamic metadata are packaged together into a new task package and sent to the cross-task process meta-knowledge base construction module. After receiving a new task package, the cross-task process meta-knowledge base construction module updates the encoder neural network parameters and trainable balancing weights in its internal generalized process similarity calculation model based on the final optimized trajectory data and dynamic meta-information contained in the package. This allows the generalized process similarity calculation model to learn from the deep patterns revealed in the current optimization process, thereby enhancing its ability to judge the similarity of future new film cutting tasks and achieving incremental enhancement and evolution of the cross-task process meta-knowledge base. By packaging and transmitting back the final optimized trajectory and dynamic meta-information containing rich process information and using them to update the generalized process similarity calculation model, the system extracts transferable process cognition and optimization strategies from the optimization experience of a single task. This not only increases the amount of data in the meta-knowledge base, but more importantly, improves its quality and intelligence, enabling the system to make more accurate similarity judgments and knowledge retrievals when dealing with future new tasks, thereby achieving continuous self-enhancement and performance evolution.

[0024] 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.

[0025] 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.

[0026] 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.

[0027] 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.

[0028] 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.

[0029] 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.

[0030] 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. A film cutting quality traceability system based on digital twins, characterized in that, Specifically, it includes: The cross-task process meta-knowledge base construction module, the proxy model initialization module, the parameter pre-screening module, and the online optimization and update module are connected sequentially, among which; Cross-task process meta-knowledge base construction module: When receiving a new film cutting task, it constructs a cross-task process meta-knowledge base based on the historical task feature vectors, historical optimal parameter sets and historical optimized trajectory data stored in the system, and calculates the generalized process similarity between the task feature vector of the new film cutting task and the feature vectors of each historical task in the cross-task process meta-knowledge base, so as to retrieve the subset of similar historical task feature vectors and their associated optimized trajectory data from the cross-task process meta-knowledge base; Proxy model initialization module: Using a small amount of historical optimization trajectory data corresponding to each historical task in the feature vector subset of similar historical tasks as a support set, the module performs gradient update operation on the shared base model pre-trained through the meta-learning framework to generate an initial proxy model for the new membrane cutting task. Parameter pre-screening module: Generates an initial acquisition function based on the initialization proxy model, and samples multiple candidate parameter combinations within the preset parameter space; calls the system's built-in digital twin model to perform virtual cutting simulation on each candidate parameter combination, and obtains the predicted quality index corresponding to each candidate parameter combination; based on whether each predicted quality index violates the system's preset process constraints, dynamically trims and reorders all candidate parameter combinations, and outputs the recommended parameter combinations that are ranked first and verified by virtual simulation to meet the process constraints to the physical film cutting machine for execution; Online optimization and update module: Collect real quality feedback data after the physical film cutting machine executes the recommended parameter combination, use the real quality feedback data to iteratively update the initial proxy model to obtain the updated proxy model, and send the updated proxy model, the final optimized trajectory data of the new film cutting task and the task feature vector as the new task package back to the cross-task process meta-knowledge base construction module to incrementally update the cross-task process meta-knowledge base.

2. The film cutting quality traceability system based on digital twin according to claim 1, characterized in that: In the cross-task process meta-knowledge base construction module, the task feature vector of a new film cutting task is composed of physical property parameters used to characterize the processed film material and geometric feature parameters used to characterize the cutting path.

3. The film cutting quality traceability system based on digital twin according to claim 2, characterized in that: The process of calculating the generalized process similarity between the task feature vector of the new membrane cutting task and the feature vectors of each historical task in the cross-task process meta-knowledge base is as follows: Extract the task feature vector of the new film cutting task, and extract a historical record from the cross-task process meta-knowledge base. This historical record contains a historical task feature vector, a historical optimal parameter set, and a historical optimization trajectory. The task feature vector of the new film cutting task and the historical task feature vector in the history are input into the generalized process similarity calculation model; the generalized process similarity calculation model calculates the first distance between the task feature vector of the new film cutting task and the historical task feature vector in the original space; The generalized process similarity calculation model also includes a pre-trained encoder neural network. The encoder neural network takes the historical task feature vector and the historical optimization trajectory as input to the historical record and outputs the task embedding vector of the historical record. For a new membrane cutting task, its task embedding vector is generated by the encoder neural network based on the task feature vector of the new membrane cutting task and combined with the response pattern represented by the historical optimization trajectory in at least one historical record that has the minimum first distance from the new membrane cutting task. Calculate the second distance between the task embedding vector of the new membrane cutting task and the task embedding vector of the historical record; add the product of the first distance and a trainable balancing weight to the product of the second distance and the difference between the numerical value and the trainable balancing weight to obtain a composite distance value; calculate the exponential function value with the natural constant e as the base and the negative of the composite distance value as the exponent, and use the exponential function value as the generalized process similarity between the new membrane cutting task and the historical task represented by the historical record.

4. The film cutting quality traceability system based on digital twin according to claim 3, characterized in that: The process of retrieving a subset of similar historical task feature vectors and their associated optimized trajectory data from the cross-task process meta-knowledge base is as follows: After calculating the generalized process similarity of the new membrane cutting task, all historical records are sorted in descending order according to the numerical value of the generalized process similarity. Based on a preset positive integer threshold, multiple historical records that are at the beginning of the descending historical record sequence and whose number is equal to the threshold are selected. The historical optimization trajectories contained in these selected historical records are extracted to form a set of similar historical optimization trajectories. The historical task feature vectors contained in these selected historical records are extracted to form a subset of similar historical task feature vectors. The cross-task process meta-knowledge base construction module outputs a subset of similar historical task feature vectors and a set of similar historical optimization trajectories to the surrogate model initialization module.

5. A film cutting quality traceability system based on digital twins according to claim 4, characterized in that: In the proxy model initialization module, the specific operation of using a small amount of historical optimized trajectory data corresponding to each historical task in the similar historical task subset as a support set is as follows: The system receives a set of similar historical optimization trajectories and a subset of similar historical task feature vectors output by the cross-task process meta-knowledge base construction module. From each historical optimization trajectory in the set of similar historical optimization trajectories, a preset number of parameter and performance data pairs located at the beginning of the historical optimization trajectory are extracted. Based on the value of the generalized process similarity between the historical task and the new film cutting task corresponding to each historical task feature vector in the subset of similar historical task feature vectors, a weight coefficient is assigned to each parameter and performance data pair extracted from the corresponding historical optimization trajectory. The parameter and performance data pairs from all historical tasks, with attached weight coefficients, are aggregated to form a weighted support set.

6. A film cutting quality traceability system based on digital twins according to claim 5, characterized in that: The specific process of performing gradient update operation on the shared base model pre-trained through the meta-learning framework is as follows: Calculate the overall loss function value of the shared base model on the weighted support set. The overall loss function value is the sum of the weighted losses of each parameter and performance data pair in the weighted support set. The loss of each data pair is composed of the product of its weight coefficient and the difference function value between the predicted performance of the parameter in the data pair and the true performance in the data pair by the shared base model. Calculate the gradient vector of the overall loss function value with respect to the trainable parameters of the shared base model. Obtain a meta-gradient modulation factor vector, which is generated during the meta-learning pre-training phase of the shared base model and has the same dimension as the trainable parameters of the shared base model. The gradient vector is multiplied element-wise with the meta-gradient modulation factor vector to obtain the modulated gradient vector; a preset learning rate is used to multiply the modulated gradient vector to obtain the parameter update amount; the parameter update amount is subtracted from the current trainable parameters of the shared base model to obtain the updated model parameters. The model defined by the updated model parameters is the initialized surrogate model. The surrogate model initialization module outputs the initialized surrogate model to the parameter pre-screening module.

7. A film cutting quality traceability system based on digital twins according to claim 6, characterized in that: In the parameter pre-screening module, the process of generating an initial acquisition function based on the initialization proxy model and sampling multiple candidate parameter combinations within a preset parameter space is as follows: The module receives an initialized surrogate model from the surrogate model initialization module. This initialized surrogate model is configured to output a prediction performance score and an uncertainty value characterizing the prediction confidence for the input parameter combination. The parameter pre-selection module constructs an initial acquisition function. The calculation process of the initial acquisition function for a parameter combination is as follows: Add a non-negative exploration coefficient to the predictive performance score corresponding to this parameter combination and the product of the uncertainty value corresponding to this parameter combination, and then subtract a non-negative constraint avoidance coefficient and the estimated confidence level of this parameter combination violating process constraints. The result is the acquisition function value corresponding to this parameter combination. Within a preset parameter space, the parameter pre-screening module generates multiple parameter combinations of a first preset number based on the distribution of the acquisition function values ​​of the constructed initial acquisition function using the Monte Carlo sampling method, forming an initial candidate parameter combination set.

8. A film cutting quality traceability system based on digital twins according to claim 7, characterized in that: The specific operation of dynamically trimming and reordering all candidate parameter combinations based on whether each predicted quality index violates the system's preset process constraints is as follows: The digital twin model is invoked to perform virtual cutting simulation on each parameter combination in the initial candidate parameter combination set to obtain a set of predicted quality indicators. The predicted quality indicators of each parameter combination are compared with the system's preset process constraints, and a constraint violation flag vector is generated for each parameter combination. All parameter combinations with at least one flag value of one in the constraint violation flag vectors are deleted to form a safe candidate parameter subset. For each remaining parameter combination in the safe candidate parameter subset, a comprehensive ranking score is calculated. The comprehensive ranking score is obtained by weighted summation of three parts: the first part is the model guidance score, which is the product of a model guidance weight coefficient and the normalized initial acquisition function value. The second part is the simulation performance score, which is the product of a simulation performance weight coefficient and a comprehensive simulation performance score. The comprehensive simulation performance score is calculated by normalizing and weighting the various prediction quality indicators output by the digital twin model. The third part is the diversity score, which is the product of a diversity reward weight coefficient and a diversity reward item. The diversity reward item is calculated based on the minimum distance in the parameter space between the current parameter combination and all parameter combinations in a set of parameter combinations selected into a temporary recommendation set. Based on the calculated comprehensive ranking score, all parameter combinations in the safe candidate parameter subset are sorted in descending order. The top-ranked parameter combinations, with a quantity equal to the second preset number, are selected to form a recommended parameter combination set, which is then output to the physical film cutting machine.

9. A film cutting quality traceability system based on digital twins according to claim 8, characterized in that: In the online optimization and update module, the process of iteratively updating the initial agent model using real quality feedback data to obtain the updated agent model is as follows: The system receives a set of recommended parameter combinations and their corresponding predicted quality indices in the digital twin model from the parameter pre-screening module, and also receives real quality feedback data returned after execution by the physical film cutting machine. For each recommended parameter combination and its corresponding real quality feedback data, a data confidence score is calculated. This data confidence score is equal to the product of two factors: the first factor is an exponential function with the natural constant e as its base, where the exponent is a negative real number, obtained by dividing the absolute value of the difference between the real quality index and the predicted quality index by a normalized constant; the second factor is the absolute value of a difference, obtained by subtracting a preset uncertainty threshold value from the initial surrogate model's prediction uncertainty for the recommended parameter combination. Then, using all recommended parameter combinations and their associated real quality feedback data with data confidence scores, the initialized proxy model is updated incrementally with weights. The update process is achieved by minimizing a composite loss function, which is the sum of three terms: the first term is the weighted sum of the losses of all data points in the original dataset of the initialized proxy model, where the weight of each original data point is a coefficient that decays over time; the second term is the weighted sum of the losses of all data points in the newly added real quality feedback dataset, where the weight of each new data point is the calculated data confidence score. The third term is a regularization term, which is obtained by multiplying the KL divergence between the probability distribution of the trainable parameters of the updated surrogate model and the probability distribution of the trainable parameters of the surrogate model before the update by a regularization strength coefficient. The trainable parameters of the updated surrogate model are obtained by minimizing the composite loss function. The model defined by the updated trainable parameters is the updated surrogate model.

10. A film cutting quality traceability system based on digital twins according to claim 9, characterized in that: The specific process of sending the updated proxy model, the final optimized trajectory data of the new film cutting task, and the task feature vector as a new task package back to the cross-task process meta-knowledge base construction module is as follows: After completing the iterative update of the initial surrogate model, the current optimization process is evaluated to determine whether it should be terminated. The evaluation method is to calculate a continued optimization benefit index, which is equal to the product of three factors: the first factor is the difference between the average expected improvement value of the model after the update and the model before the update on the safe candidate parameter subset generated by the parameter pre-screening module, divided by the moving average of the average expected improvement value of the most recent iterations before the update; the second factor is the average prediction uncertainty value before the update divided by the average prediction uncertainty value after the update; the third factor is the remaining physical experiment budget value divided by the total physical experiment budget value. If the continued optimization benefit index is lower than a preset threshold, or the model prediction performance score has reached the preset target, the optimization process is determined to be terminated. After the optimization process is terminated, the online optimization update module extracts the complete final optimization trajectory data of this new film cutting task, along with multiple dynamic meta-information recording the key features of the optimization process, and packages the final optimization trajectory data, task feature vector, updated surrogate model and dynamic meta-information into a new task package, which is then sent to the cross-task process meta-knowledge base construction module. After receiving a new task package, the cross-task process meta-knowledge base construction module updates the encoder neural network parameters and trainable balanced weights in its internal generalized process similarity calculation model based on the final optimized trajectory data and dynamic meta-information contained in the package, so as to achieve incremental updates to the cross-task process meta-knowledge base.

Images