Multi-line rubber calendering quality prediction method based on weighted aggregation federated learning

Abstract

The application discloses a multi-line rubber calendering quality prediction method based on weighted aggregation federal learning, and belongs to the technical field of rubber manufacturing process quality prediction. The technical scheme comprises the following steps: collecting multivariate process parameter time series data by edge nodes corresponding to multiple calendering production lines and constructing local sample sequences; training a time series prediction model combined with a GRU and an attention mechanism by each edge node based on local training samples, and outputting a width accuracy prediction value; evaluating the local model on an independent validation set and obtaining performance indicators; uploading the local model parameters and the performance indicators, normalizing the performance indicators to obtain performance scores in the cloud, and constructing comprehensive aggregation weights in combination with the training data sizes of each node; and updating global model parameters by the cloud according to the weights, and distributing the global model parameters to each edge node for the next round of iteration or for width accuracy prediction. The application has the beneficial effect of providing a multi-line rubber calendering quality prediction method based on weighted aggregation federal learning.

Description

Technical Field

[0001] This invention belongs to the field of rubber manufacturing process quality prediction technology, and specifically relates to a multi-production line rubber calendering quality prediction method based on weighted aggregation federated learning. Background Technology

[0002] Rubber calendering is a crucial step in tire and rubber product manufacturing. It involves the continuous extrusion of rubber materials using a calender to ensure the product meets the requirements for subsequent bonding, cutting, and molding in terms of width, thickness, and surface quality. Among these factors, the width accuracy of the calendered product (e.g., width deviation, width uniformity) directly impacts material utilization, bonding alignment accuracy, and finished product consistency, making it a vital indicator for quality control during production. Unstable width accuracy can lead to increased trimming and defect rates, ultimately increasing production costs.

[0003] In existing technologies, the following two technical approaches are typically adopted to improve the width accuracy control capability of calendered products: Mechanism / experience-based control and compensation: For example, adjusting parameters such as calender speed, gap, and tension based on process experience rules, equipment parameter adjustment strategies, or simple linear / nonlinear compensation models. This type of method is simple to implement, but it is sensitive to factors such as raw material fluctuations, equipment aging, and environmental changes, and it is difficult to adapt to multiple product types and multiple operating conditions, often requiring extensive manual maintenance based on experience. Data-driven predictive modeling: With the widespread use of sensors, PLCs, and MES systems, the production process can collect multivariate time-series process parameters such as extruder speed, extruder pressure, calender speed, and calender current. Existing technologies attempt to use regression models, neural networks, or time-series deep learning models to predict quality indicators, in order to achieve quality prediction and process optimization. However, in practical implementation, data-driven methods still face the following problems: 1) There is a significant time-series correlation and nonlinear coupling between calender width accuracy and multivariate process parameters, and the degree of influence of parameters on the final quality varies at different time periods. Traditional models cannot fully represent complex dynamic processes, resulting in limited prediction accuracy and stability. 2) Rubber companies typically have multiple calendering production lines. The equipment models, sensor configurations, maintenance status, operating strategies, and product specifications of each production line may differ, leading to significant data inconsistencies. For example, older production lines may require adjustments to the calendering gap to ensure width accuracy due to roller wear, and the traction speed ranges corresponding to different skeleton materials also vary. In such cases, models trained on a single production line are difficult to directly transfer to other production lines, limiting their generalization ability.

[0004] 3) Production data typically involves sensitive information such as enterprise process formulas, equipment parameters, capacity, and quality. Due to data security, compliance requirements, and system isolation, data from multiple production lines is often difficult to centrally aggregate for joint modeling, thus limiting cross-production line knowledge sharing and collaborative optimization. 4) To address the difficulties in collaborative training caused by the lack of centralized data, existing technologies propose using federated learning for industrial quality prediction, allowing each production line to train its model locally and only upload model parameters or gradients for cloud aggregation. However, commonly used aggregation methods such as federated averaging (FedAvg) typically use a weighted average based on the sample size of each participant, failing to fully consider differences in data quality, operational complexity, and model training performance across production lines. However, the actual working conditions in rubber production amplify the shortcomings of the traditional FedAvg algorithm: FedAvg only allocates aggregation weights based on the amount of data from each production line, assuming that "the larger the amount of data, the higher the model's contribution." But in actual rubber production, production lines with large amounts of data may have poor process parameter stability due to aging equipment and batch fluctuations in raw materials, resulting in lower prediction accuracy for the local model. On the other hand, some niche tire production lines, although with smaller amounts of data, have good equipment condition and strict process control, allowing the local model to achieve high-precision predictions. If the FedAvg algorithm is used directly, the low-performance, large-scale model will excessively dominate the update of the global model, leading to slow global model convergence and low prediction accuracy, making it unable to adapt to the actual production needs of multiple rubber production lines. Summary of the Invention

[0005] The purpose of this invention is to provide a quality prediction method that enables cross-production line collaborative training under the condition that the data does not leave the domain, for multi-production line scenarios of rubber calendering.

[0006] This invention is achieved through the following measures: a multi-line rubber calendering quality prediction method based on weighted aggregation federated learning, characterized by comprising: S1, using multiple calendering production lines of a rubber enterprise as multiple edge nodes, acquiring time-series data of multivariate process parameters collected by each edge node during the calendering production process, and constructing a local sample input sequence for training; S2, constructing and training a local quality prediction model at each edge node based on the local training sample set, wherein the local quality prediction model is a time-series prediction model combining GRU and attention mechanism, and outputting a width accuracy prediction value of the calendered product; S3, evaluating the prediction performance of the trained local quality prediction model at each edge node based on an independent validation set, and obtaining the performance index of the corresponding edge node; S4, each edge node uploading the updated local model parameters to a cloud server. The cloud server normalizes the performance metrics of each edge node to obtain a performance score, and constructs a comprehensive aggregation weight based on the training data scale of each edge node and the performance score. S5, the cloud server performs weighted aggregation on the local model parameters uploaded by each edge node according to the comprehensive aggregation weight to obtain updated global model parameters, and distributes the global model parameters to each edge node for the next round of local training iteration or for predicting the width accuracy of the rolled product.

[0007] Optionally, the input for each sample in step S1 is a multivariate time series feature matrix of length T, wherein the multivariate process parameters include extruder speed, extruder current, extruder pressure, calender speed, calender current, and calender temperature.

[0008] The time-series prediction model combining GRU and attention mechanisms described in step S2 includes: a GRU network layer, an attention mechanism layer, and a fully connected layer; wherein: The GRU network layer extracts temporal features from the multivariate time series feature matrix to obtain the hidden state sequence at each time step; the attention mechanism layer calculates attention weights for the hidden states at each time step and performs weighted fusion to obtain a context vector; the fully connected layer outputs the width-accuracy prediction value based on the context vector.

[0009] Optionally, in step S2, the training of the local quality prediction model updates the model parameters by minimizing the loss function between the predicted value and the true quality label; when the width precision is a continuous value, the loss function is the mean squared error loss function; when the width precision is the classification quality level, the loss function is the cross-entropy loss function.

[0010] Optionally, the performance metric in step S3 includes at least the coefficient of determination, and the coefficient of determination is calculated based on the true quality value and the predicted quality value of the validation set.

[0011] Optionally, constructing a comprehensive aggregation weight based on the training data size of each edge node and the performance score specifically includes: Introducing a balance factor Then the first The aggregate weight of each edge node in the current round is defined as follows:

[0012] in, This indicates the weighting ratio based on the data size. This indicates the weighting proportions based on the model's predictive performance. when When the algorithm degenerates into a traditional federated average algorithm based on sample size, it becomes a mere collection of samples. At that time, the aggregation weights are entirely determined by the model's predictive performance.

[0013] Optionally, the updated global model parameters satisfy:

[0014] in Indicates the first Global model parameters after rounds of federated training; Local model parameters.

[0015] Optionally, the global model parameters are sent to each edge node for the next round of local training iteration or for predicting the width accuracy of the rolled product, including: the cloud server sends the global model parameters to the edge computing nodes of each production line through a secure communication channel to replace or update the local model parameters at each edge computing node; After receiving the global model parameters, each edge computing node uses the global model parameters as the initialization parameters for the next round of local training, and continues to perform local training based on the newly generated production data. During each round of training, each edge computing node only uploads local model parameters or gradient information to the cloud server and does not upload the original process data. Training terminates based on a preset convergence criterion or the maximum number of iterations. The convergence criterion includes: when the performance improvement of the global model on the validation set is lower than a set threshold for multiple consecutive iterations, the model is determined to have reached convergence and training is terminated. Once the model reaches convergence, each edge computing node retains the final global model, which is used to input the corresponding time series process parameters after the complete process cycle of the new production sample ends, and to perform inference through the time series prediction model to output the prediction results.

[0016] Optionally, the quality prediction method also includes periodically triggering a new round of federated training as new labeled data is continuously generated during the production process.

[0017] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows: The present invention adopts a federated learning architecture, in which each production line only trains the model locally and uploads the model parameters, without sharing the original production data, effectively avoiding the risk of centralized storage of core industrial data such as process parameters, and meeting the requirements of industrial data security and privacy protection.

[0018] For industrial scenarios where the distribution of process parameters varies greatly across multiple production lines and exhibits non-independent and identically distributed characteristics, this invention achieves the collaborative fusion of feature information from different production lines through multiple rounds of federated iterative training. This enables the model to adapt to different equipment conditions and process fluctuation environments, thereby improving the model's generalization ability in multiple scenarios.

[0019] Compared to the traditional federated averaging method that only assigns weights based on the number of samples, this invention introduces model prediction performance indicators into the weight calculation, making the aggregation process more consistent with the objective situation of the differences in the contribution of each production line model in actual industrial scenarios, thereby improving the applicability of federated learning in industrial quality prediction tasks. Attached Figure Description

[0020] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings listed below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the multi-production-line rubber calendering quality prediction method based on weighted aggregation federated learning in an embodiment of the present invention; Figure 2 This is a comparison chart showing the effects of applying this application and existing models on production line A; Figure 3 This is a comparison chart showing the effects of applying this application and existing models on production line B; Figure 4 This is a comparison chart showing the effects of applying this application and existing models on the C production line. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] See Figures 1-4 The embodiment itself provides a method for predicting the quality of multi-production line rubber calendering based on weighted aggregate federated learning. In the following description, This serves as an index for edge nodes, which are deployed on their respective rolling production lines. Since each edge node is configured in a one-to-one correspondence with a rolling production line, Simultaneously characterize the corresponding first Strip calendering production line.

[0024] A method for predicting the quality of multi-production-line rubber calendering based on weighted aggregation federated learning includes: S1, taking multiple calendering production lines of a rubber enterprise as multiple edge nodes, acquiring time series data of multivariate process parameters collected by each edge node during the calendering production process, and constructing a local sample input sequence for training; the multivariate process parameters include extruder speed, extruder current, extruder pressure, calender speed, calender current, and calender temperature.

[0025] Specifically, in a multi-production-line tire manufacturing scenario, the different production lines create a heterogeneous data environment with varying statistical distributions due to differences in equipment models, operating conditions, and process parameter settings. To construct a distributed quality prediction model under a federated learning framework, it is necessary to acquire the time-series data of multivariate process parameters collected by each edge node during the calendering production process.

[0026] Assume the system has a total of The first production line, the The original data samples collected from each edge node are represented as follows:

[0027] in, Indicates the first The number of samples at each edge node; To indicate the first The edge node of the first Multivariate time series input features corresponding to each training sample The sampling time step length within the production cycle. For the dimensions of the collected process parameters; This corresponds to the quality label.

[0028] Furthermore, due to the complex environment of industrial production sites, the collected data is inevitably affected by factors such as equipment vibration, electromagnetic interference, sensor accuracy errors, and network transmission fluctuations. The original process parameter data may contain abnormal fluctuations, missing sampling points, and noise interference. Therefore, each production line needs to independently complete data preprocessing operations on its corresponding edge computing node and build its own local training and validation sample sets.

[0029] First, the collected raw data undergoes local cleaning. For each process parameter channel, outliers are identified and removed based on statistical characteristics, and sampling points that significantly deviate from the normal fluctuation range are deleted or interpolated for repair. At the same time, missing data in the time series is completed to ensure the continuity and integrity of the production cycle data.

[0030] Secondly, after data cleaning, the data from each production line is standardized to ensure that each feature channel is distributed within a uniform scale, thereby improving model convergence stability and training efficiency. After preprocessing, each production line forms a time-series dataset with a unified structure and consistent scale, which is stored at local edge nodes and used only for local model training without cross-production line data transmission, providing a reliable data foundation for subsequent federated learning collaborative modeling. S2. At each edge node, a local quality prediction model is constructed and trained based on the local training sample set. The local quality prediction model is a time-series prediction model combining GRU and attention mechanisms, outputting the width accuracy prediction value of the rolled product. After data preprocessing, each production line independently builds and trains a quality prediction model on edge computing nodes based on locally generated standardized time-series data (local training sample set) to learn the nonlinear mapping relationship between manufacturing process parameters and the final tire quality indicators. To capture the temporal dependency characteristics of manufacturing process data, this invention adopts a prediction model structure that integrates gated recurrent units (GRUs) and attention mechanisms.

[0031] For the The edge node of the first There are samples, and their input is a length of... The multivariate time series feature matrix is ​​first input into the GRU network layer. Through update and reset gates, historical states are dynamically remembered and forgotten, thereby extracting the temporal features of the production process and obtaining the hidden state sequence at each time step. The GRU hidden state update process can be represented as:

[0032] in For time step process parameter vector, This is the output of the hidden state at the corresponding time.

[0033] To further enhance the model's ability to focus on key process stages, an attention mechanism is introduced based on the GRU output sequence. Attention weights are calculated for the hidden states at each time step and then weighted and fused. The attention weights are automatically calculated based on the current sequence features, enabling the model to focus on key time segments that significantly impact quality results. The weighted context vector can be represented as:

[0034] in, For the first The attention weights at each time step satisfy the weight normalization constraint.

[0035] Finally, the fused context feature vector is input into the fully connected layer, which outputs the quality prediction value for the corresponding sample.

[0036] in, This is the predicted value for width accuracy.

[0037] During model training, each production line optimizes its parameters using only local data and does not share raw data with other production lines. Model parameters are updated by minimizing the loss function between the predicted value and the true quality label. If the width accuracy is a continuous value, mean squared error is used as the loss function; if the width accuracy is a classification quality level, cross-entropy loss is used. The general form of the loss function can be expressed as:

[0038] in, For the first The average loss value used in local training for each edge node sample. For the first The number of samples for each edge node. For the first The edge node The single-sample loss function for each sample. If the quality index is a continuous value: The mean squared error loss can be used; if it is a classification quality level, Cross-entropy loss is preferable.

[0039] Through multiple rounds of local iterative training, each production line obtains a set of optimized local model parameters, providing a foundation for subsequent federated weighted aggregation.

[0040] This step takes each calendering production line of a rubber enterprise as an independent edge node. Combining the temporal coupling characteristics of key parameters such as extruder speed, extruder pressure, calender speed, and calender current during the tire tread winding process, a deep temporal prediction model combining GRU and attention mechanisms is constructed locally to establish a dynamic mapping relationship between multivariate process parameters and calender width accuracy. The temporal prediction model uses the GRU structure to extract the dependence of process parameters over time and uses the attention mechanism to identify key time intervals and key process parameters that significantly affect width deviation and width uniformity, thereby enhancing the model's ability to distinguish between stable and fluctuating calendering stages. S3: At each edge node, the predicted performance of the trained local quality prediction model is evaluated based on an independent validation set to obtain the performance indicators of the corresponding edge nodes. After the local quality prediction model is trained, each production line needs to evaluate the performance of its trained model to quantify its predictive ability on local data, providing a reference for subsequent federated weighted aggregation. The evaluation process is conducted independently on the edge computing nodes of each production line to ensure data privacy. Each production line delineates an independent validation set from its historical production data. This validation set does not overlap with the training set and includes the time series of process parameters for the complete production cycle and the corresponding width and accuracy prediction values. The time series data from the validation set is input into the locally trained time series prediction model to obtain the predicted quality value for each sample. .

[0041] Comparison with predicted values Compared with the true quality value The performance metrics of the calculation model are as follows. The coefficient of determination is used for evaluation, and its expressions are as follows:

[0042] in, For the sample size, To verify the average value of the true quality values ​​in the set.

[0043] After completing the above calculations, a performance evaluation result of the local model is generated. This evaluation result is used as one of the important parameters in the subsequent federated weighted aggregation process to dynamically adjust the contribution weight of each production line model in the global model update. This ensures that the local model with better prediction performance occupies a higher proportion in the aggregation process, thereby improving the overall accuracy and generalization ability of multi-production line quality prediction. S4. Each edge node uploads the updated local model parameters to the cloud server. The cloud server normalizes the performance metrics of each edge node to obtain a performance score, and constructs a comprehensive aggregation weight based on the training data scale of each edge node and the performance score. After each edge node completes local model training and prediction performance evaluation, each edge node will update the local model parameters for this round. The corresponding performance metrics are uploaded to the cloud server. Raw production data remains locally and is not transmitted across nodes. The cloud server is responsible for weighted fusion of the model parameters uploaded from each edge node to form new global model parameters.

[0044] Suppose the system includes The first production line, the The local training data size corresponding to each edge node is The coefficient of determination calculated on the validation set The cloud server normalizes the determination coefficients of each edge node to obtain a performance score.

[0045] First, the performance metrics of each node are normalized to obtain the performance score:

[0046] : This indicates that the decision coefficients of all edge nodes participating in the current round of federated training / aggregation are summed, and j is the summation index; For the first Performance scores for each edge node.

[0047] Next, a comprehensive aggregation weight is constructed based on the training data size of each edge node and the performance score. To balance the impact of data size and model prediction performance, a balancing factor is introduced. Then the first The aggregate weight of each edge node in the current round is defined as follows:

[0048] in, Indicates the weighting ratio based on the data size; This indicates the weighting proportions based on the model's predictive performance. Indicates the first The size of the local dataset for each edge node, i.e., the number of samples used for local training at that node; This represents the local dataset of the j-th edge node. Its sample size.

[0049] By adjusting the balance factor This allows control over the relative proportions of data volume and performance factors during the aggregation process. When When the algorithm degenerates into a traditional federated average algorithm based on sample size, it becomes inefficient. At that time, the aggregation weights are entirely determined by the model's predictive performance.

[0050] S5. The cloud server performs weighted aggregation on the local model parameters uploaded by each edge node according to the comprehensive aggregation weight, obtains the updated global model parameters, and sends the global model parameters to each edge node for the next round of local training iteration or for predicting the width accuracy of the rolled product.

[0051] The updated global model parameters satisfy the following: the aggregate weights of each node are calculated. Then, the cloud updates the global model parameters using a weighted average method:

[0052] in Indicates the first The global model parameters are updated after each round of federated training. Once updated, the new global model will be used in the next training iteration or deployed to each production line for real-time quality prediction. Through this performance-based weighted aggregation mechanism, production lines with stronger predictive capabilities and reasonable data scales receive higher contribution weights during model fusion, thereby improving the convergence efficiency and generalization performance of the global model in a multi-production-line data heterogeneous environment.

[0053] The global model parameters are distributed to each edge node for the next round of local training iteration or for predicting the width accuracy of rolled products. This includes: the cloud server distributes the global model parameters to the edge computing nodes of each production line through a secure communication channel to replace or update the local model parameters at each edge computing node. After receiving the global model parameters, each edge computing node uses the global model parameters as the initialization parameters for the next round of local training, and continues to perform local training based on the newly generated production data; thus forming a cyclical update mechanism of "local training - cloud aggregation - model distribution - retraining".

[0054] During each round of training, each edge computing node only uploads local model parameters or gradient information to the cloud server without uploading the original process data; this ensures that production data is always kept locally, thereby meeting the requirements for industrial data privacy and security.

[0055] As the number of federated training rounds increases, the global model parameters gradually integrate feature information from multiple production lines, enabling cross-production line knowledge sharing and collaborative optimization. In practical applications, training terminates based on a preset convergence criterion or the maximum number of iteration rounds. The convergence criterion includes: when the performance improvement of the global model on the validation set is lower than a set threshold for several consecutive rounds, the model is considered to have stabilized, the model is determined to have reached convergence, and training is terminated. Once the model reaches convergence, each edge computing node retains the final global model, which is used to input the corresponding time series process parameters after the complete process cycle of the new production sample ends, and to perform inference through the time series prediction model to output the prediction results.

[0056] It also includes periodically triggering a new round of federated training when new labeled data is continuously generated during the production process, in order to adapt to the data distribution drift caused by changes in equipment status or adjustments in process parameters, thereby ensuring the long-term stable operation of the model.

[0057] This step quantifies the predictive performance of each production line model by introducing a performance metric based on the coefficient of determination, and determines weighting coefficients accordingly to achieve performance-aware weighted fusion of parameters from multiple production line models. This aggregation method allows production line models with superior breadth prediction performance to play a greater role in global model updates, while ensuring that data remains local.

[0058] Example 2: Based on Example 1, and within a federated learning framework, the proposed multi-production-line rubber calendering quality prediction method based on weighted aggregation federated learning was implemented on three production lines (A, B, and C). Figures 2-4The MAE (Magnitude of Effect) on all three production lines was significantly lower than that of the independent local training model and the federated model based on the traditional FedAvg method. This result indicates that the multi-production-line rubber calendering quality prediction method proposed in this invention can more effectively mine complementary information between different production lines, while mitigating the adverse effects of data heterogeneity on model performance to some extent. Especially on production line C, where operating conditions change frequently and process non-stationarity is strong, the MAE is reduced by approximately 22.9% compared to the FedAvg method, demonstrating that the two-factor weighting mechanism can improve the robustness and generalization ability of the model in complex, time-varying industrial environments.

[0059] While centralized training demonstrates the best overall performance across various evaluation metrics, it relies on access to raw process data from multiple production lines, including sensitive operating parameters such as extruder load, pressure, and calendering temperature. In real-world industrial scenarios, this centralized data-sharing approach is often difficult to implement due to factors such as data privacy protection, information security strategies, and communication bandwidth limitations. In contrast, the DFW-FedAgg strategy proposed in this paper achieves predictive performance close to that of centralized training while maintaining data locality and without transmitting raw data, making it more suitable for the deployment needs of real-world manufacturing systems.

[0060] The time-series prediction curves reveal distinct dynamic characteristics across the three production lines. Line A exhibits significant high-frequency width fluctuations, typically related to rapid adjustments in extruder speed and transient disturbances in rubber compound flow. Line B shows relatively smooth overall fluctuations but exhibits persistent periodic oscillations, reflecting relatively stable but periodically changing operating conditions during the calendering stage. Line C experiences pronounced abrupt changes in operating status at several time points, potentially caused by product switching, process parameter reconfiguration, or human intervention. Compared to benchmark methods, the multi-line rubber calendering quality prediction method proposed in this paper accurately tracks the overall trend of width changes across the three production lines while maintaining a timely response to sudden changes, without significant prediction lag or over-smoothing. This demonstrates that the method can simultaneously capture short-term process disturbances and long-term operating status evolution characteristics in complex industrial environments.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the quality of multi-production line rubber calendering based on weighted aggregate federated learning, characterized in that, include: S1. Using multiple calendering production lines of a rubber enterprise as multiple edge nodes, acquire time series data of multivariate process parameters collected by each edge node during the calendering production process, and construct a local sample input sequence for training. S2. At each edge node, a local quality prediction model is constructed and trained based on the local training sample set. The local quality prediction model is a time-series prediction model combining GRU and attention mechanism, which outputs the width accuracy prediction value of the rolled product. S3. At each edge node, the prediction performance of the trained local quality prediction model is evaluated based on an independent validation set to obtain the performance index of the corresponding edge node. S4. Each edge node uploads the updated local model parameters to the cloud server. The cloud server normalizes the performance metrics of each edge node to obtain a performance score, and constructs a comprehensive aggregation weight based on the training data scale of each edge node and the performance score. S5. The cloud server performs weighted aggregation on the local model parameters uploaded by each edge node according to the comprehensive aggregation weight, obtains the updated global model parameters, and sends the global model parameters to each edge node for the next round of local training iteration or for predicting the width accuracy of the rolled product. A comprehensive aggregation weight is constructed based on the training data scale of each edge node and the performance score, specifically including: Introducing a balance factor Then the first The aggregate weight of each edge node in the current round is defined as follows: in, Indicates the weighting ratio based on the data size; This indicates the weighting proportions based on the model's predictive performance. Indicates the first The size of the local dataset for each edge node, i.e., the number of samples used for local training at that node; This represents the local dataset of the j-th edge node. Its sample size, For N production lines, For the first Raw data samples collected from each edge node; when When the algorithm degenerates into a traditional federated average algorithm based on sample size, it becomes inefficient. At that time, the aggregation weights are entirely determined by the model's predictive performance.

2. The prediction method according to claim 1, characterized in that, In step S1, the input for each sample is a multivariate time series feature matrix of length T. The multivariate process parameters include extruder speed, extruder current, extruder pressure, calender speed, calender current, and calender temperature.

3. The prediction method according to claim 2, characterized in that, The time-series prediction model combining GRU and attention mechanisms described in step S2 includes: a GRU network layer, an attention mechanism layer, and a fully connected layer; wherein: The GRU network layer extracts temporal features from the multivariate time series feature matrix to obtain the hidden state sequence at each time step; the attention mechanism layer calculates attention weights for the hidden states at each time step and performs weighted fusion to obtain a context vector; the fully connected layer outputs the width-accuracy prediction value based on the context vector.

4. The prediction method according to claim 3, characterized in that, In step S2, the training of the local quality prediction model updates the model parameters by minimizing the loss function between the predicted value and the true quality label; when the width precision is a continuous value, the loss function is the mean squared error loss function; when the width precision is the classification quality level, the loss function is the cross-entropy loss function.

5. The prediction method according to claim 1, characterized in that, In step S3, the performance metric includes at least the coefficient of determination, and the coefficient of determination is calculated based on the true quality value and the predicted quality value of the validation set.

6. The prediction method according to claim 1, characterized in that, The updated global model parameters satisfy: in Indicates the first Global model parameters after rounds of federated training.

7. The prediction method according to claim 6, characterized in that, The global model parameters are distributed to each edge node for the next round of local training iteration or for predicting the width accuracy of rolled products. This includes: the cloud server distributes the global model parameters to the edge computing nodes of each production line through a secure communication channel to replace or update the local model parameters at each edge computing node. After receiving the global model parameters, each edge computing node uses the global model parameters as the initialization parameters for the next round of local training, and continues to perform local training based on the newly generated production data. During each round of training, each edge computing node only uploads local model parameters or gradient information to the cloud server and does not upload the original process data. Training terminates based on a preset convergence criterion or the maximum number of iterations. The convergence criterion includes: when the performance improvement of the global model on the validation set is lower than a set threshold for multiple consecutive iterations, the model is determined to have reached convergence and training is terminated. Once the model reaches convergence, each edge computing node retains the final global model, which is used to input the corresponding time series process parameters after the complete process cycle of the new production sample ends, and to perform inference through the time series prediction model to output the prediction results.

8. The prediction method according to claim 7, characterized in that, It also includes periodically triggering a new round of federated training when new labeled data is continuously generated during the production process.

Images