Federal learning-based strip steel mechanical property prediction method

Abstract

The invention discloses a strip steel mechanical property prediction method based on federal learning, and belongs to the technical field of strip steel mechanical property prediction.The method comprises the steps that feature data influencing the strip steel mechanical property are determined; for each production line, characteristic data in the historical production process of the production line are collected, the collected data are preprocessed, and a data set of the corresponding production line is constructed by using the preprocessed data; aiming at each production line, respectively constructing a strip steel mechanical property prediction model, and training the constructed strip steel mechanical property prediction model by adopting the data set of the corresponding production line; each production line is regarded as a client, a federal learning framework is constructed, strip steel mechanical property prediction model parameters of the clients are aggregated, and multi-production-line collaborative prediction is achieved. The strip steel mechanical property prediction method provided by the invention provides a standardized solution for multi-production-line collaborative prediction of the strip steel mechanical property.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of strip steel mechanical property prediction, in particular to a strip steel mechanical property prediction method based on federated learning. BACKGROUND

[0002] Hot-rolled strip steel is a main product of steel production and is widely used in the fields of construction, automobiles, home appliances, ships, etc. In the rolling production, the performance is predicted according to the composition of the steel material and the rolling process, which can not only predict the mechanical properties online, reduce the sampling frequency of detection, shorten the production cycle and improve the efficiency, but also optimize the rolling process and realize online control. The mechanical property prediction of hot-rolled strip steel can be used for online dynamic control of product performance and optimal design of new steel grades. The mechanical properties such as tensile strength and yield strength are important indicators for measuring the quality of strip steel and directly affect the performance and safety of the product.

[0003] Traditional mechanical property detection of strip steel relies on destructive testing, which increases production costs and prolongs the product flow cycle. With the progress of industrial big data and artificial intelligence technology, data-driven mechanical property prediction technology of strip steel has become a research focus. The construction of a prediction model can quickly and accurately predict the mechanical properties of steel without damaging the steel, optimize the production process and improve the product quality. However, the existing mechanical property prediction methods of strip steel are mostly based on single production line data to construct models. Due to the limitation of data sharing between different production lines due to privacy protection, etc., the model learning data is single, the prediction accuracy is limited, and the model of a single production line cannot be generalized to other production lines, which restricts the wide application of the model. SUMMARY

[0004] The present application provides a strip steel mechanical property prediction method based on federated learning to solve the technical problems of the existing strip steel mechanical property prediction method, such as single model learning data, limited prediction accuracy, and difficulty in generalizing the model of a single production line to other production lines, which restricts the wide application of the model.

[0005] To solve the above technical problems, the present application provides the following technical solutions: On the one hand, the present application provides a strip steel mechanical property prediction method based on federated learning, comprising: determining feature data affecting the mechanical properties of strip steel; For each production line, the feature data in its historical production process is collected, and the collected data is preprocessed, and the preprocessed data is used to construct a data set of the corresponding production line; For each production line, a strip steel mechanical property prediction model is constructed, and the data set of the corresponding production line is used to train the constructed strip steel mechanical property prediction model; Each production line is regarded as a client, a federated learning framework is constructed, parameters of the strip steel mechanical property prediction model of each client are aggregated, and collaborative prediction of multiple production lines is realized; wherein, the model of each production line is trained locally, and after the training is completed, only the model parameters are transmitted to the central server, the parameters of the strip steel mechanical property prediction model of each client are combined and updated by the central server, and the updated parameters are distributed to each client.

[0006] Further, the characteristic data affecting the mechanical properties of the strip steel comprises: The actual production process data of the strip steel is collected; wherein, the actual production process data of the strip steel comprises: process parameters and strip steel mechanical property data; wherein, the process parameters comprise: material target thickness, material target width, material target length, material actual thickness, material actual width, material actual length, hot rolling finishing mill outlet temperature, hot rolling coiling temperature, rolling pressure and material chemical composition. Based on the actual production process data, the Pearson correlation coefficients of each process parameter and target variable are calculated, and process parameters with a Pearson correlation coefficient greater than a preset threshold are selected as characteristic data.

[0007] Further, the preprocessing comprises: data cleaning, standardization and normalization.

[0008] Further, the strip steel mechanical property prediction model is an integrated model.

[0009] Further, the integrated model integrates and fuses three kinds of models of CNN, Transformer and MLP.

[0010] Further, the integrated model automatically learns the contribution of different models in prediction through a Softmax layer, adopts a dynamic integration strategy during integrated learning, dynamically adjusts the proportion of different models according to different working conditions, and dynamically weights the prediction results of different models; Wherein, the dynamic integration strategy trains the Softmax layer to output the combined weights of each basic model, so that the integrated model can adaptively select the optimal prediction path under different working conditions, and improve the robustness of the integrated model.

[0011] Further, in the integrated model, CNN adopts a deeper convolutional layer structure, normalizes batch data with a BatchNormalization layer to improve training efficiency and stability, and alleviates the gradient vanishing problem, and also suppresses overfitting through Dropout; Transformer enhances the time sequence feature extraction capability of the process parameter sequence through the position encoding layer and the multi-head attention mechanism; MLP increases the number of hidden layer neurons and uses regularization techniques to improve the modeling capability of the product effect.

[0012] Furthermore, the training strategy for the ensemble model is to use a cosine annealing learning rate schedule with an initial learning rate of 0.001, which dynamically decays to 0.000001 over the training cycle, while using early stopping to avoid overfitting. The loss function uses the Huber loss with a hyperparameter δ=1.0 to balance the impact of outliers and enhance the model's robustness to industrial data noise. A 5-fold cross-validation approach is used, with the training and validation sets repartitioned for each fold to ensure reliable performance evaluation.

[0013] Furthermore, the data set corresponding to each client contains at least 20% of the public steel grade data; wherein the public steel grade data is obtained by selecting several preset steel grades from all the steel grades at the production site as basic steel grades, and using the characteristic data corresponding to the basic steel grades as the public steel grade data; During the parameter aggregation stage of the central server, the central server uses a dynamic weight adjustment mechanism to dynamically adjust the client's weight to achieve weighted average aggregation. At the same time, a differential privacy mechanism is introduced to add Gaussian noise before uploading the parameters to ensure that the gradient information cannot be traced back to the original data characteristics.

[0014] Furthermore, the dynamic weight adjustment mechanism optimizes the federated aggregation process through a dual strategy. The first is data volume weighted aggregation, that is, the aggregation weight is allocated according to the client data scale. Clients with a high data volume contribute more to the global model update to ensure the dominant role of mainstream steel grade data; the second is performance feedback adjustment, that is, an additional 10%-15% weight gain is given to clients with too low MAE to encourage the knowledge contribution of high-quality local models.

[0015] On the other hand, the present invention further provides an electronic device, comprising a processor and a memory; wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the above method.

[0016] In yet another aspect, the present invention further provides a computer-readable storage medium, wherein the storage medium stores at least one instruction, and the instruction is loaded and executed by a processor to implement the above method.

[0017] The beneficial effects brought about by the technical solution provided by the present invention include at least: 1. This invention integrates the local feature extraction capability of CNN and the long-term time series modeling advantage of Transformer into a federated learning framework for the first time, solving the nonlinear mapping problem of "process parameters-performance" in strip steel production.

[0018] 2. The present invention proposes a Non-IID data processing strategy of common steel grade anchoring + dynamic weight allocation, which reduces the MAE fluctuation of cross-line prediction by 42%, providing a standardized solution for multi-production line collaboration.

[0019] 3、The application realizes that the R2 of the federal model is reduced by no more than 1.1% under the premise of guaranteeing data security through differential privacy and homomorphic encryption technology, and breaks through the technical bottleneck in industrial application. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the premise of the drawings.

[0021] Figure 1 is a flowchart of the strip mechanical property prediction method based on federal learning provided by the embodiments of the present application; Figure 2 is a Pearson correlation heat map provided by the embodiments of the present application; Figure 3 is an integrated learning framework diagram provided by the embodiments of the present application; Figure 4 is a comparison diagram of the integrated model for the yield strength real value and the predicted value provided by the embodiments of the present application; Figure 5 is a comparison diagram of the integrated model for the tensile strength real value and the predicted value provided by the embodiments of the present application; Figure 6 is a comparison diagram of the integrated model for the elongation at break real value and the predicted value provided by the embodiments of the present application; Figure 7 is a federal learning framework diagram provided by the embodiments of the present application; Figure 8 is a single client independent prediction and federal learning performance comparison column chart provided by the embodiments of the present application; Figure 9 is a system block diagram of an electronic device provided by the embodiments of the present application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present application more clear, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0023] First, it should be noted that in the embodiments of the present application, the words "exemplary", "for example", and the like are used to mean serving as an example, instance, or illustration. Any embodiment or design presented as "exemplary" in the present application is not necessarily to be construed as preferred or advantageous over other embodiments or designs. Rather, use of the word exemplary is intended to present concepts in a concrete manner. In addition, in the embodiments of the present application, the meaning expressed by "and / or" can be both, or can be either one of the two.

[0024] First embodiment

[0025] The present embodiment provides a strip steel mechanical property prediction method based on federated learning. In view of the problems of data isolation, insufficient model generalization ability and privacy protection in traditional strip steel mechanical property prediction, a data-driven prediction method is proposed by introducing a federated learning framework. The key technologies of multi-production line collaborative prediction are explored by comprehensively using data-driven modeling, model optimization and comparative analysis methods. Through multi-model comparison, the convolutional neural network (CNN) combined with the Transformer mechanism is selected. After data preprocessing and feature selection, a high-quality data set is constructed by combining cross-validation. The number of convolution kernels in the model structure is increased and a regularization layer is introduced to strengthen feature extraction. In the training process, dynamic learning rate, early stopping method and robust loss function are used. After optimization, the average absolute error of yield strength, tensile strength and elongation after fracture is reduced by 13%, 6% and 8% respectively, and the determination coefficient is significantly improved, effectively capturing sequence features and complex patterns. In the design of the federated learning framework, the data is divided into three clients according to the steel grade. Fusion is achieved through localized training and parameter aggregation. The FedAvg algorithm is used to aggregate model parameters, and differential privacy and homomorphic encryption technology are introduced to ensure data security. In the non-identically and independently distributed (Non-IID) scenario, federated learning makes the R² of yield strength, tensile strength and elongation after fracture reach 0.9464, 0.9677 and 0.8652 respectively, which is 1.1%, 0.6% and 0.6% higher than that of centralized learning respectively. This verifies the effectiveness of the method in protecting privacy while improving generalization ability. The research aims to build an efficient prediction model to achieve accurate prediction of the mechanical properties of more than 10 steel grades, ensure that the average absolute error of key indicators such as tensile strength and yield strength is ≤40MPa, and provide technical support for strip steel production quality control and process optimization, promoting the collaborative application of federated learning across production lines in the industrial field.

[0026] The method can be implemented by an electronic device, and the execution flow of the method is as shown in Figure 1 , including the following steps: S1, determining the feature data affecting the mechanical properties of the strip steel; It is to be noted that in the data-driven method of strip performance prediction, data preprocessing is a key step. It improves data quality through data cleaning, standardization and normalization, encoding, etc., which will directly affect the effect of model training and the accuracy of prediction. For numerical data, such as strip rolling temperature, the missing values can be filled with the mean value. Data standardization and normalization are very important. It can speed up the model training speed, when using gradient descent to adjust parameters, make the numerical range of different features similar, prevent the numerical size difference of different features from affecting the gradient update direction, help the model find the optimal solution faster. And it can improve the prediction accuracy, avoid the model paying too much attention to the features with larger numerical values, let all features be learned fairly, so as to more reasonably allocate weights.

[0027] The min-max normalization in data normalization maps the data to the [0, 1] interval, making different features in the same scale range, which is convenient for model comparison and learning, and can be expressed as formula (1).

[0028] (1)

[0029] wherein, is the original rolling speed value, and are the minimum and maximum values of the rolling speed in the batch data. This method preserves the original distribution relationship of the data, but is sensitive to outliers.

[0030] Standardization adopts Z-score standardization: makes the data conform to the standard normal distribution with mean 0 and standard deviation 1, which can effectively eliminate the dimension effect. For the carbon content C of the strip, it is expressed as formula (2).

[0031] (2)

[0032] wherein, is the mean value of the carbon content, is the standard deviation of the carbon content. This method is suitable for most data distribution situations, especially when using gradient descent-based optimization algorithms, which helps to speed up model convergence.

[0033] In addition, it should be pointed out that the Pearson correlation coefficient is used to evaluate the importance of the features in this embodiment to determine the feature data affecting the mechanical properties of the strip steel. Among them, the Pearson correlation coefficient is calculated based on the ratio of variable covariance to standard deviation, and generates a value in the interval of-1 to 1. The closer the absolute value of the value is to 1, the more significant the linear relationship is. In practical application, the Pearson correlation coefficient of each feature and the target variable is calculated based on the standardized production data, and then the high correlation features are screened according to the set threshold (such as |p|>0.5). Experiments show that this method can accurately lock the core features such as carbon content and finishing temperature which are highly related to mechanical properties, and effectively screen out low correlation redundant variables such as equipment running time, thereby greatly improving the prediction accuracy of the model while realizing feature dimension reduction, and providing strong support for subsequent mechanical property prediction model construction. Among them, the original production data in the prediction of the mechanical properties of the strip steel usually includes a large number of process parameters such as rolling temperature, pressure and chemical composition. Figure 2 The Pearson correlation coefficient correlation heat map is given, and the correlation degree of the parameters in the map with the target performance indicators such as yield strength and tensile strength is significantly different. The features that have a significant impact on the prediction target are screened through the map.

[0034] The experimental data of this embodiment is derived from the historical production records of a hot rolling production line of a certain steel enterprise, covering 10770 samples, 80 steel grades, including 30 process parameters such as carbon content, finishing temperature and rolling force, and 3 mechanical property indicators such as yield strength, tensile strength and elongation after fracture. In the data preprocessing stage, the missing values are filled with the mean value, the abnormal values are filtered by the IQR method, and the core features with an absolute correlation value greater than or close to 0.5 with each target variable are screened out by the Pearson correlation coefficient.

[0035] S2, for each production line, the feature data in the historical production process is collected, and the collected data is preprocessed, and the preprocessed data is used to construct a data set of the corresponding production line; S3, for each production line, a strip steel mechanical property prediction model is constructed, and the data set of the corresponding production line is used to train the constructed strip steel mechanical property prediction model; Among them, the strip steel mechanical property prediction model in this embodiment is a dynamic weight integrated model. Moreover, in order to select a suitable model, this embodiment selects mean squared error (MSE), mean absolute error (MAE) and coefficient of determination (R²) as the core indicators for evaluating the regression prediction model to comprehensively measure the prediction accuracy, error distribution and goodness of fit of the model: the calculation formulas of the indicators are as follows: (3) (4) (5) wherein, is the predicted value, is the actual value, is the mean of the actual values, is the number of predicted samples.

[0036] Based on the above, the present embodiment aims at the problem of predicting the mechanical properties of strip steel, relying on the historical data of a single production line of a certain steel plant to build a data-driven prediction model. The data contains more than ten thousand samples, covering 30 process parameter characteristics such as rolling temperature, chemical composition, rolling force, and three mechanical property target variables of yield strength, tensile strength, and elongation after fracture. The model realizes high-precision prediction by mining the non-linear relationship between production parameters and performance indicators, and provides decision support for process optimization.

[0037] Combined with the time series characteristics and high-dimensional feature distribution of strip steel data, three types of traditional machine learning models, including Random Forest (RF), Support Vector Machine (SVM), and Radial Basis Function Network (RBFN), and three types of deep learning models, including Multi-layer Perceptron (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and fully connected network with Transformer mechanism, are selected for targeted modeling. The data preprocessing and feature extraction before training of each model are consistent.

[0038] When comparing CNN, LSTM, RF, SVM, RBFN, MLP, and Transformer models, the present embodiment uses a unified data processing method: all models are filled with mean values to fill missing values, IQR method is used to remove outliers, the first 10 features are selected according to the Pearson coefficient, and the data set is divided according to the 8:2 ratio and the same random seed is kept. Next, adjustments are made for different models: SVM and MLP require feature standardization, RF and RBFN are not sensitive to data size and do not require feature scaling, LSTM data needs to be converted to a three-dimensional format of [number of samples, time steps, number of features] and normalized, and Transformer needs to add sequential information by position encoding layer after standardization to change the data to a dimension of [number of samples, number of features, 1]. These adjustments ensure that each model can exert its own advantages under the same benchmark.

[0039] In the single production line prediction of the mechanical properties of steel strips, seven models, including CNN, LSTM, RF, SVM, RBFN, MLP, and Transformer, were compared. The results showed that the performance of CNN, MLP, LSTM, and Transformer was significantly different from that of traditional machine learning models (RF, SVM, and RBFN). From the core evaluation indicators, CNN had a prominent advantage in yield strength (MSE = 581.75, MAE = 14.53 MPa, R² = 0.942) and tensile strength (MSE = 304.58 MPa, MAE = 9.79 MPa, R² = 0.968) prediction. It could effectively capture the local spatial correlation of process parameters (such as the synergistic effect of final rolling temperature and cooling speed) through the hierarchical stacking of convolutional and pooling layers, and had a significant advantage in high-dimensional nonlinear feature interaction scenarios. MLP relied on a fully connected structure and achieved the highest R² (0.945) in yield strength prediction, indicating that it had a stronger modeling capability for global nonlinear product effects such as "carbon content x manganese content" and the best fitting effect. However, the fully connected architecture made it take longer to train and was sensitive to outliers (such as the MSE of 344.21 MPa in tensile strength prediction, which was higher than that of CNN).

[0040] In the time series feature modeling of the Transformer model, the position encoding layer and multi-head attention mechanism played a unique role in tensile strength prediction. It had a clear modeling effect on long-range dependencies in the rolling stage (such as the delayed impact of rough rolling temperature on final performance). Although the MAE was 10.63 MPa, which was higher than that of CNN, its self-attention mechanism could clearly quantify the time series contribution of each process parameter (such as the attention weight of the temperature in the finishing rolling stage reaching 28%). However, the determination coefficient of the Transformer in the elongation prediction was only 0.819, indicating that it was less adaptable to non-strong time series data (such as composition fluctuations) than CNN and MLP, and had a higher computational complexity. As shown in Tables 1, 2, and 3, the MSE, MAE, and R² for yield strength, tensile strength, and elongation, respectively.

[0041] Table 1 Comparison of evaluation indicators for each model (yield strength)

[0042] Table 2 Comparison of evaluation indicators for each model (tensile strength)

[0043] Table 3 Comparison of evaluation indicators for each model (elongation)

[0044] To improve the generalization ability and prediction accuracy of the strip steel mechanical property prediction model, the embodiment introduces an ensemble learning framework to systematically optimize the base model structure and training strategy. By fusing the prediction results of CNN, Transformer and MLP, a dynamic weight ensemble model (Ensemble Model) is constructed. The ensemble learning framework used in the embodiment is shown in Figure 3 The core of the model is to automatically learn the contribution of different models in prediction through the Softmax layer, achieving complementary advantages: CNN is good at capturing local spatial correlation of process parameters (such as temperature and rolling force synergy effect), Transformer focuses on modeling long-range dependence of time series features (such as cumulative influence of historical parameters in the rolling stage), and MLP excavates global nonlinear interaction between features through fully connected structure. The attention mechanism enhances the key feature extraction capability through adaptive weight allocation. The dynamic integration strategy outputs the combined weights of each base model (value range 0-1 and sum 1) through the training weight network. During ensemble learning, the proportion of different prediction methods is dynamically adjusted according to different working conditions, so that the model can adaptively select the optimal prediction path in different working conditions (such as CNN weight ratio up to 45% in high-dimensional nonlinear scenarios, and Transformer weight can be increased to 50% in time series feature dominant working conditions), significantly improving the robustness of the model in complex industrial data.

[0045] In the optimization of the base model, CNN uses deeper convolution layer structure (i.e. more layers of convolution), normalizes batch data with BatchNormalization layer to improve training efficiency and stability, and alleviates the problem of gradient disappearance. It also uses Dropout to suppress overfitting; Transformer uses position encoding layer and multi-head attention mechanism, i.e. adding sequential information after the position encoding layer to change the data dimension to [number of samples, number of features, 1], enhancing the time series feature extraction capability of process parameter sequence, such as capturing the delayed influence of temperature fluctuation in the finishing rolling stage on the final yield strength when processing rolling temperature sequence; MLP increases the number of hidden layer neurons (512→256) and uses regularization techniques to improve the modeling capability of product effects such as "carbon content x manganese content". Specifically, in the embodiment, the CNN model internal structure sequence is: input data → convolution layer (32 3x3 convolution kernels) → BatchNorm layer → ReLU activation function → Dropout layer → max pooling layer → convolution layer (64 3x3 convolution kernels) → BatchNorm layer → ReLU activation function → Dropout layer → max pooling layer → flattening layer → fully connected layer → ReLU activation function → Dropout layer → fully connected layer (output target variable).

[0046] The training strategy of the integrated model is to use an initial learning rate of 0.001, and dynamically decay to 0.000001 with cosine annealing learning rate scheduling, while cooperating with early stopping method to avoid overfitting. The loss function selects Huber loss with hyperparameter δ = 1.0 to balance the influence of outliers and enhance the robustness of the model to industrial data noise. The experiment adopts 5-fold cross-validation, and each fold re-divides the training set and validation set to ensure the reliability of performance evaluation. The integrated learning and model optimization strategy improves the precision and stability of the prediction of the mechanical properties of the strip steel through the three mechanisms of structural complementation, dynamic weighting and training enhancement, providing a more reliable technical solution for real-time quality control in industrial scenarios. Table 4 is a comparison table of the integrated learning model and other models.

[0047] Table 4 Comparison of integrated learning model and other models

[0048] The test performance of the integrated model is shown in Figure 4 , Figure 5 , Figure 6 It can be seen from Figure 4 , Figure 5 , Figure 6 that the average absolute error (MAE) of the yield strength, tensile strength and elongation of the integrated model is 13.00%, 5.93% and 8.51% lower than that of the single optimal base model CNN, respectively, and the determination coefficient (R²) is increased by 1.97%, 0.76% and 2.1%, respectively. The prediction bias is significantly lower than that of the single model in extreme working conditions, verifying the adaptability of integrated learning to complex industrial scenarios. The weight visualization results show that CNN contributes an average weight of 42% in tensile strength prediction, Transformer contributes 38% weight in yield strength prediction, and MLP contributes 35% in elongation prediction, reflecting the differentiated advantages of different model feature modeling.

[0049] S4, each production line is regarded as a client, a federated learning framework is constructed, the parameters of the strip steel mechanical property prediction model of each client are aggregated, and collaborative prediction of multiple production lines is realized; wherein the model of each production line is trained locally, after training, only the model parameters are transmitted to the central server, the model parameters of each client are combined and updated by the central server, and the updated parameters are distributed to each client.

[0050] It is worth noting that in recent years, artificial intelligence technology has been continuously developing, and emerging technologies such as deep learning and federated learning have provided new paths for the prediction of the mechanical properties of steel strips. Federated learning is a main method of distributed machine learning. This method allows multiple clients to train models locally and then transmit the training results to the cloud for aggregation. This way, without sharing raw data, different clients can optimize and update the model together. This not only protects data privacy and security, but also combines data resources from different production lines, so the adaptability and prediction accuracy of the model in complex industrial scenarios are significantly improved.

[0051] In the context of steel strip production, the industrial adaptability of federated learning is reflected in its ability to handle multi-source heterogeneous data. Due to differences in equipment models, process parameters, and steel types, the data from multiple production lines of hot-rolled steel strips exhibits non-independent and identically distributed characteristics (for example, a production line focuses on high-carbon steel production, and its chemical composition and rolling temperature distribution are significantly different from those of other production lines). Traditional centralized models are difficult to generalize in such data, while federated learning allows each client to learn the local characteristics of a specific production line (such as the unique "final rolling temperature-carbon content" coupling mode of a production line), and then aggregates global parameters to capture common rules across production lines (such as the universal influence of rolling force on yield strength), achieving the dual goals of "local feature capture + global knowledge fusion".

[0052] The real-time requirements of steel strip production and the iterative mechanism of federated learning are well matched. In industrial sites, second-level mechanical property prediction is required. Federated learning uses miniaturized model design (such as simplified CNN structure) and optimizes communication methods (such as sparse gradients) to reduce the data aggregation time to minutes, thus meeting the requirements of real-time adjustment of production lines. In view of the time series characteristics of steel strip production data, a suitable time series model is selected. Each client individually establishes a time series relationship model for the rolling stage (such as using rough rolling temperature changes to predict the final elongation), and then shares model parameters to improve the model's adaptability to long-term process fluctuations.

[0053] Based on the above, the present embodiment uses a federated learning method to predict the mechanical properties of steel strips. The federated learning framework is used to integrate data from multiple production lines. Each production line first trains the model on the local device, and then transmits the training results to the cloud for aggregation. This way, the data privacy of each factory is protected, and the model learns more diverse data features. After such training, the model's ability and accuracy in predicting mechanical properties are improved. This method breaks down the data isolation between production lines. Different production lines can share experiences without revealing privacy and improve the prediction model together. This not only improves the technical level of the entire steel strip industry, but also has reference value for other fields that require data privacy protection. Below, the construction process of the federated learning framework is described.

[0054] 1. Client processing

[0055] Non-identically distributed data refers to data that does not satisfy the characteristics of independence and identical distribution, or at least violates one of the conditions. It is a common form of real data in machine learning and statistics, in contrast to the traditional independent and identically distributed (IID) assumption. The non-identically distributed characteristics of strip production data under the federated learning framework pose a significant challenge to the generalization ability of the model. The research adopts a comprehensive strategy combining stratified data alignment, adaptive regularization, and meta-learning initialization to ensure the stability and accuracy of cross-production line model training.

[0056] In view of the heterogeneity of the distribution of steel types in each client when the data is distributed to three clients, the system aligns the data distribution through common steel types, defines basic steel types such as SPHC, Q235B, and LXXXX as the common data set, and requires each client to contain at least 20% of the common steel data. Through the Pearson correlation coefficient, shared key features such as carbon content, finish rolling temperature, and rolling force in the common steel are selected to construct a consistent feature space across production lines and ensure the parameter consistency of the model on the common features. For steel types specific to the client, the client is allowed to add trace elements content, special rolling stage parameters, and other exclusive features locally, and through the local-global feature interaction mechanism of federated learning, the weight of the common features and local features is dynamically adjusted through the attention mechanism during server aggregation.

[0057] When dealing with non-identically distributed strip production data in federated learning, the adaptive regularization and gradient calibration mechanism suppresses model gradient drift through a double constraint strategy. This mechanism adds a global parameter alignment term and a branch structure calibration term to the client training loss function. The former constrains the deviation of the client's global parameters from the server's parameters to ensure the consistency of the basic feature mapping, and the latter implements differential constraints for CNN, Transformer, and MLP sub-model branches to prevent local overfitting. The regularization coefficient is dynamically adjusted according to the client data distribution difference and branch complexity, and the gradient calibration forces the local gradient direction to be consistent with the global direction.

[0058] The meta-learning initialization and dynamic weight adjustment strategy under the federated learning framework improves the cross-production line adaptability of the model to non-identically distributed data through pre-training of the global model and dynamic aggregation mechanism. The system generates initial parameters robust to heterogeneous data based on model-independent meta-learning algorithms during the pre-training phase by traversing representative data sets from each client.

[0059] The dynamic weight adjustment mechanism optimizes the federated aggregation process through a double strategy. First, data volume weighted aggregation, according to the client data size, the aggregation weight is allocated, the client with high data volume proportion contributes more to the global model update to ensure the dominant role of the mainstream steel grade data. Second, performance feedback adjustment, the client with lower MAE is additionally given a weight gain of 10-15% to encourage the knowledge contribution of high-quality local model. The synergy of meta-learning initialization and dynamic weight makes the system balance global universality and local specificity in cross-production line data collaboration, providing a robust federated learning solution for strip steel mechanical property prediction.

[0060] 2. Federated learning system framework

[0061] The federated learning system used in this paper operates in a client-server mode, mainly including three collaboration levels, and the framework diagram is shown in Figure 7 The specific workflow is: the data of each production line is trained locally, and after training, only the parameters are transmitted to the central server, and the server updates the parameters. This has two benefits: first, the original data is always kept within the factory, which meets the safety regulations; second, different production lines can share experience through parameter exchange. In the whole process, the system will encrypt the transmitted parameters and update the model version of all nodes in time. This mechanism not only guarantees the security of data flow, but also effectively integrates the machine learning achievements of each link.

[0062] The data layer uses a combination of physical isolation and logical association design. The system divides more than 100 steel grades into 3 client clusters (30-40 steel grades per cluster) according to production processes, while sharing basic steel grades such as SPHC and Q235B as public data. This not only maintains the uniqueness of each client data, but also improves the model alignment effect through public steel grades. Differential privacy (ε=0.5, δ=1e-5) is added during data preprocessing, and Gaussian noise is added during mean filling and IQR outlier processing to ensure the safety of the original data.

[0063] The client layer uses a heterogeneous model collaboration framework, dynamically scheduling CNN, Transformer and MLP models according to different steel data characteristics: CNN captures local feature interactions through convolutional layers, Transformer handles time series dependencies in the rolling process through position encoding and multi-head attention mechanism, and MLP focuses on non-linear mapping of global features such as chemical composition. Client training uses an adaptive learning rate strategy that combines cosine annealing scheduling and gradient clipping (clip norm=1.0) to ensure convergence speed while suppressing gradient drift caused by non-IID data. To reduce the computational burden of edge devices, model parameters are quantized to 8 bits, reducing communication volume to less than 10% of the original.

[0064] The server layer realizes three-layer aggregation optimization as a global knowledge hub: first, it aggregates client model parameters through a weighted federated average algorithm (weights are dynamically adjusted according to client data volume and training loss), second, it introduces a model ensemble mechanism to dynamically weight and fuse the prediction results of CNN, Transformer, and MLP, and deploys a homomorphic encryption module on the server side to encrypt gradients with Paillier encryption before parameter aggregation to ensure the privacy of intermediate results. The system also designs a federated learning monitoring subsystem to monitor the training progress, gradient distribution, and model performance of each client, and automatically triggers FedProx regularization when abnormal fluctuations are detected to constrain the extent to which the client model deviates from the global model through a proximal term. This architecture effectively solves the contradiction between industrial data privacy protection and model generalization through the collaborative mechanism of parameter encryption transmission, model dynamic adaptation, and knowledge hierarchical aggregation.

[0065] Based on the above modules, a strip steel mechanical property prediction system based on federated learning is constructed, which consists of: a data preprocessing module that realizes data cleaning, standardization, and feature selection, and outputs a data set that meets the modeling requirements; a single production line modeling module that integrates CNN-Transformer neural networks, supports model structure optimization and training strategy configuration; a federated learning module composed of a client (edge device deployment) and a server (cloud aggregation), the client is responsible for local training and parameter upload, the server performs dynamic weighted aggregation and privacy protection; a prediction and feedback module that outputs real-time mechanical property prediction results and interfaces with the production system to realize dynamic adjustment of process parameters.

[0066] 3. Federated improvement of data-driven models

[0067] The federated improvement of data-driven models is the core link to realize cross-production line collaborative prediction, and its core goal is to embed the integrated model (CNN-Transformer-MLP) trained on a single production line into the federated learning framework, and improve the model generalization ability under the premise of protecting data privacy through the "local training-parameter aggregation-global update" mechanism.

[0068] To adapt to the distributed training mode of federated learning, the structure of the single production line integrated model needs to be optimized. The input interfaces of CNN, Transformer, and MLP are unified into a three-dimensional tensor format [number of samples, number of features, 1] to ensure consistency in data format across different clients. CNN captures local spatial features through a one-dimensional convolutional layer, Transformer injects timing information through a positional encoding layer, and MLP receives a flattened two-dimensional feature vector to retain the ability to model global interactions. When deployed on the client side, the model is parameterized (e.g., represented by 8-bit fixed-point numbers) and structurally pruned to compress the model size to 65% of the original size, reducing the computational load on edge devices. At the same time, the SoftMax weight layer of the integrated model is retained to adaptively adjust the contribution of different base models, improving the cross-scenario generalization ability and prediction accuracy of data-driven models under the federated learning framework. The improved federated learning process consists of four stages: initialization, local training, parameter aggregation, and global update.

[0069] In the initialization stage, the server builds a global integrated model and distributes the initial model parameters to each client. Clients load local data based on steel type division and divide the data into three client clusters, each containing 30-40 types of steel, to ensure that the data distribution has both differences and common intersections (e.g., sharing basic steel types such as SPHC and Q235B).

[0070] In the local training stage, each client independently trains the model based on local data. The cosine annealing decay adaptive learning rate strategy and early stopping method are used to avoid overfitting. For example, when client 1 trains on high-carbon steel data, it increases the number of Transformer layer attention heads to 6 to strengthen the modeling of long-range dependencies on rolling temperature sequences.

[0071] In the parameter aggregation stage, the server performs weighted average aggregation based on the data volume of each client, with larger data clients having a higher weight proportion. The differential privacy mechanism is introduced, and Gaussian noise is added before parameter transmission to ensure that gradient information cannot be traced back to the original data features.

[0072] In the global update stage, the server distributes the aggregated model parameters to each client to trigger the next round of training. Through 20 iterations, the model gradually converges to demonstrate the effectiveness of cross-production line knowledge fusion.

[0073] 4. Single client independent prediction and federated learning performance comparison analysis

[0074] In the prediction of strip mechanical properties, single client models are limited by local data distribution and cannot capture cross-production line common features. Federated learning achieves global knowledge fusion through parameter aggregation. This paper compares the performance differences between single client independent prediction and global model after federated learning to verify the effectiveness of federated learning.

[0075] When predicting single clients independently, the performance of each client is significantly affected by data heterogeneity. Taking yield strength as an example, the MAE of client 2 is 14.26 MPa and the R² reaches 0.9605 (the performance is the best but depends on the distribution balance of the client data). The R²s of clients 1 and 3 are 0.9470 and 0.9503, respectively, indicating the sensitivity of the single-client model to data fluctuations. In the prediction of tensile strength, the MAE of each client ranges from 10.16 to 11.07 MPa, and the R² exceeds 0.96 (indicating a strong linear relationship between this indicator and the process parameters). However, the R² of client 2 (0.9720) is still significantly higher than that of other clients, reflecting the influence of data distribution differences on the model. The prediction accuracy of elongation is relatively low (R²=0.85-0.86), with the MAE of each client ranging from 2.23 to 2.35 MPa. This may be because this indicator is coupled with complex factors such as rolling temperature gradient and cooling rate, making it difficult to fully model the local data.

[0076] By aggregating the model parameters of three clients, federated learning significantly improves the overall performance of the global model. The overall R² of yield strength reaches 0.9527 (higher than most single clients, indicating that cross-production line data fusion effectively enhances the model generalization ability), the overall MAE of tensile strength is 10.64MPa, R²=0.9695 (maintaining high accuracy while balancing the prediction deviations of each client), and the overall R² of elongation after fracture is 0.8559 (slightly lower than the optimal value of client 2), and the MAE is reduced to 2.297MPa (reflecting the robustness of federated learning to data noise).

[0077] The comparative results show that federated learning retains the single-client modeling capability of local data through the "local training-global aggregation" mechanism, while improving the global adaptability of the model through cross-production line knowledge fusion. The R² of the single-client model fluctuates greatly (for example, the yield strength R² difference reaches 0.0135), while federated learning effectively suppresses this fluctuation through dynamic weight allocation, making the overall R² stable. Experiments show that federated learning can significantly improve the consistency and reliability of model predictions in non-independent and identically distributed scenarios, and provide an effective solution for collaborative prediction of multiple strip steel production lines. Figure 8 The figure is a bar chart comparing the performance of single-client independent prediction and federated learning.

[0078] 5. Verification of Federated Learning Collaborative Performance

[0079] The experiment simulates a federated learning scenario and divides the data into three clients. Client 1 contains 26 steel grades, client 2 contains 26 steel grades, and client 3 contains 28 steel grades. Each client shares data enhancement model alignment for three basic steel grades, such as SPHC and Q235B. The Z-score method is used to standardize the data so that its features are scaled to a distribution with mean 0 and standard deviation 1. The data is then reshaped into a three-dimensional tensor format [number of samples, number of features, 1] suitable for one-dimensional CNN and Transformer input.

[0080] The experiment used 5-fold cross-validation to evaluate model performance and compared the prediction accuracy of centralized learning, federated learning, and single-client models. Core evaluation metrics included mean absolute error (MAE), mean squared error (MSE), and coefficient of determination (R²). The federated learning process was set to 20 iterations, and the client-side local training used a batch size of 128 and 30 epochs. The server-side aggregated parameters using the FedAvg algorithm and introduced a data volume weighting strategy in which the aggregation weight was determined by the client's data volume.

[0081] The collaborative performance of federated learning was verified by comparing the performance differences between single-client independent prediction and the global federated learning model to verify its effectiveness in cross-production line data collaboration. The experiment divided the data into three clients to simulate the heterogeneous data scenario of multiple production lines. The single-client model was limited by the local data distribution and the performance fluctuations were significant: in the yield strength prediction, client 2 (MAE=14.26MPa, R²=0.9605) performed best due to the better data distribution balance. The R² of clients 1 and 3 were 0.9470 and 0.9503 respectively; in the tensile strength prediction, the MAE of each client was between 10.16-11.07MPa, and the R² exceeded 0.96, but client 2 (R²=0.9720) was significantly better than other clients; in the elongation prediction, the MAE of a single client was 2.23-2.35MPa, and the R² was 0.85-0.86. The local data modeling capability was limited due to complex coupling factors.

[0082] Federated learning aggregates the model parameters of three clients through the FedAvg algorithm. The global model has significant advantages in non-independent and identically distributed scenarios. The overall R² of yield strength reaches 0.9527 (higher than most single clients, indicating that cross-production line data fusion effectively enhances the model's generalization ability), the overall MAE of tensile strength is 10.64MPa, and R²=0.9695 (balancing the prediction deviations of each client while maintaining high accuracy), the overall R² of elongation after fracture is 0.8559, and the MAE is reduced to 2.297MPa (reflecting robustness to data noise). The maximum R² difference of single-client models reaches 0.0135. Federated learning stabilizes the overall performance through dynamic weight allocation and parameter aggregation mechanisms, verifying its effectiveness in suppressing the impact of data heterogeneity.

[0083] The experimental results show that the federated learning significantly improves the global prediction consistency and reliability of the model by the "local feature capture + global knowledge fusion" mechanism under the premise of protecting the data privacy of each production line. Compared with the single-client model, the global model improves the R² of the yield strength, tensile strength, and elongation after fracture by 0.32%-1.35%, and reduces the MAE by 3.86%-5.68%, which highlights the advantages of balancing privacy protection and prediction accuracy in the cross-production line collaborative scenario, and provides a feasible technical path for the intelligent quality control of the strip steel multi-production line.

[0084] In summary, the embodiment aims to solve the problems of data isolation, insufficient model generalization ability, and privacy protection in traditional strip steel mechanical property prediction. A data-driven prediction framework based on federated learning is constructed, and the prediction accuracy and privacy security are improved through multi-dimensional modeling and cross-production line collaboration mechanism.

[0085] At the data-driven modeling level, the research systematically preprocesses the hot-rolled strip steel production data, selects core features such as carbon content and final rolling temperature through the Pearson correlation coefficient, and analyzes the data set containing 10770 samples. Compared with 7 kinds of models such as RF, SVM, CNN, etc., it is found that CNN performs best in capturing sequence features and complex nonlinear relationships. After optimization, the average absolute error (MAE) of yield strength, tensile strength, and elongation after fracture is reduced by 13%, 6%, and 8%, respectively, and the determination coefficient (R²) is significantly improved. Further, through the integration of CNN, Transformer, and MLP, and the dynamic weighting strategy to enhance the robustness of the model, the adaptability of multi-model collaboration to complex industrial data is verified.

[0086] In the design of the federated learning framework, a three-layer client-server architecture is constructed, the data is divided into three clients, the FedAvg algorithm is used for parameter aggregation through localized training and data weight, the differential privacy and homomorphic encryption technology are introduced to ensure data security, and the public steel anchor, adaptive regularization, and other strategies are used to optimize the differences between client models. The experiment shows that the federated learning in cross-production line prediction has R² of yield strength, tensile strength, and elongation after fracture of 0.9464, 0.9677, and 0.8652, respectively, which is 0.6%-1.1% higher than that of centralized learning, verifying the effectiveness of improving the generalization ability under the premise of privacy protection.

[0087] The experimental results show that the CNN has a significant advantage in yield strength and tensile strength prediction in a single production line scenario, the random forest performs well in elongation prediction, the federated learning improves the model generalization ability while protecting privacy and has a higher determination coefficient of each index, and the model average absolute error is ≤40 MPa through cross-production line knowledge fusion and communication overhead reduction. The data-driven prediction method of federated learning constructed in this study provides technical support for strip steel production, and future research can further explore lightweight model design and fusion with physical metallurgical models.

[0088] Second embodiment

[0089] The embodiment provides an electronic device, such as Figure 9 As shown in the figure, the electronic device comprises a processor and a memory; wherein the processor and the memory can be connected through a communication bus; the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment. In addition, the electronic device can also include a transceiver, and the processor and the transceiver can be connected through a communication bus, and the transceiver is used for communication with other devices.

[0090] Next, combined with Figure 9 The various components of the electronic device will be described in detail: The processor is the control center of the electronic device, and the electronic device can include multiple processors, each of which can be a single-CPU or a multi-CPU. The processor here can be a processor or a general term for multiple processing elements. For example, the processor is one or more central processing units (CPU), which can also be other general-purpose processors, application specific integrated circuits (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application, such as one or more microprocessors (digital signal processors, DSP), or one or more field programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor can execute various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory.

[0091] In a specific implementation, as an embodiment, the processor may include one or more CPUs, such as Figure 9 The CPU0 and CPU1 shown in FIG are, of course, only exemplary.

[0092] The memory is used to store the software program for executing the solution of the present invention, and the execution is controlled by the processor. The specific implementation method can refer to the above method embodiment and will not be repeated here.

[0093] Optionally, the memory may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compact disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory may be integrated with the processor or exist independently and accessed through the interface circuit ( Figure 9 (not shown) is coupled to the processor, which is not specifically limited in this embodiment of the present invention.

[0094] The transceiver may include a receiver and a transmitter ( Figure 9 The receiver is used to implement the receiving function, and the transmitter is used to implement the sending function. The transceiver can be integrated with the processor or exist independently and communicate with the electronic device through the interface circuit ( Figure 9 (not shown) is coupled to the processor, which is not specifically limited in this embodiment of the present invention.

[0095] In addition, it should be noted that Figure 9 The structure of the electronic device shown in the figure does not constitute a limitation on the device. The actual device may include more or fewer components than shown, or may combine certain components, or arrange the components differently. In addition, the technical effects achieved by the electronic device when executing the method of the first embodiment can refer to the technical effects described in the first embodiment above, and therefore will not be repeated here.

[0096] Third embodiment

[0097] The embodiment provides a computer readable storage medium, at least one instruction is stored in the storage medium, the instruction is loaded and executed by a processor to realize the method of the first embodiment. Wherein, the computer readable storage medium can be a ROM, a random access memory, a CD-ROM, a magnetic tape, a floppy disk and an optical data storage device, etc. The instruction stored therein can be loaded and executed by the processor in the terminal to execute the above method.

[0098] In addition, it should be noted that the present application can be provided as a method, an apparatus or a computer program product. Therefore, the embodiments of the present application can be in the form of entirely or partially hardware embodiments, entirely or partially software embodiments or embodiments combining software and hardware aspects. Moreover, when implemented by software, the embodiments of the present application can be in the form of a computer program product implemented on one or more computer usable storage media including computer usable program code. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded and executed on a computer, the flow or function described in the embodiments of the present application is generated entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (such as infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium sets. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid state disk.

[0099] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams according to the method, terminal device (system) and computer program product of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, an embedded processor or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device produce a machine that implements the flow specified in the flowcharts and / or block diagrams. Figure 1 The apparatus for implementing one or more flows and / or blocks Figure 1 The apparatus for implementing one or more flows and / or blocks

[0100] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks of the block diagrams. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow or flows and / or blocks of the block diagrams. Figure 1 function specified in the flow or flows and / or blocks of the block diagrams. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow or flows and / or blocks of the block diagrams. Figure 1 function specified in the flow or flows and / or blocks of the block diagrams. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow or flows and / or blocks of the block diagrams. Figure 1 Figure 1 function specified in the flow or flows and / or blocks of the block diagrams. These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow or flows and / or blocks of the block diagrams.

[0101] It should also be noted that, in the specification, the terms "first" and "second" and the like are used merely to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between such entities or actions. The terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... " does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element. In addition, the term "and / or" is merely an associative relationship of associated objects, and means that there can be three relationships, for example, A and / or B, which means that A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. In addition, the character " / " in this specification generally represents an "or" relationship between the associated objects, but can also represent an "and / or" relationship, which can be understood in the context of the preceding and following text. "One or more" means one or more, and "multiple" means two or more. "At least one" or similar expressions mean any combination of these items, including single or multiple combinations. For example, at least one of a, b, or c can mean a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, and c can be singular or plural.

[0102] In addition, it should be understood that the sequence of the above processes does not mean the order of execution in various embodiments of the present application, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0103] Those skilled in the art can appreciate that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. A skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0104] In several embodiments provided by the present application, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, for example, the division of functional modules / units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another device, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms. The units described as separate components can be or can not be physically separated, and the components displayed as units can be or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the present embodiment. In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0105] If the method is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0106] Finally, it should be pointed out that the above description is only the preferred embodiments of the present application. It should be noted that although the preferred embodiments of the present application have been described, for those skilled in the art, once the basic creative concept of the present application is known, a number of improvements and refinements can be made without departing from the principles of the present application. These improvements and refinements should also be considered as the protection scope of the present application. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all modifications and variations falling within the scope of the embodiments of the present application.

Claims

1. A method for predicting the mechanical properties of steel strips based on federated learning, characterized in that: include: Determine the characteristic data that affects the mechanical properties of the strip; For each production line, we collect characteristic data from its historical production process, pre-process the collected data, and use the pre-processed data to construct a data set for the corresponding production line; For each production line, a strip steel mechanical properties prediction model is constructed and trained using the dataset of the corresponding production line. Each production line is regarded as a client, and a federated learning framework is constructed to aggregate the strip mechanical properties prediction model parameters of each client to achieve collaborative prediction of multiple production lines. The model of each production line is trained locally. After the training is completed, only the model parameters are transmitted to the central server. The central server merges and updates the strip mechanical properties prediction model parameters of each client and sends the updated parameters to each client.

2. The method for predicting mechanical properties of steel strips based on federated learning according to claim 1, wherein: The determining of characteristic data affecting the mechanical properties of the strip steel includes: Collecting actual production process data of the steel strip; wherein the actual production process data of the steel strip includes: process parameters and mechanical property data of the steel strip; wherein the process parameters include: target material thickness, target material width, target material length, actual material thickness, actual material width, actual material length, hot rolling finish rolling outlet temperature, hot rolling coiling temperature, rolling pressure, and material chemical composition; Based on the actual production process data, the Pearson correlation coefficient between each process parameter and the target variable is calculated, and the process parameters with a Pearson correlation coefficient greater than the preset threshold are screened out as feature data.

3. The method for predicting mechanical properties of steel strips based on federated learning according to claim 1, wherein: The preprocessing includes: data cleaning, standardization and normalization.

4. The method for predicting mechanical properties of steel strips based on federated learning according to claim 1, wherein: The strip steel mechanical property prediction model is an integrated model.

5. The method for predicting mechanical properties of steel strips based on federated learning according to claim 4, characterized in that: The integrated model integrates three models: CNN, Transformer and MLP.

6. The method for predicting mechanical properties of steel strips based on federated learning according to claim 4, characterized in that: The ensemble model automatically learns the contribution of different models in prediction through the Softmax layer. During ensemble learning, a dynamic ensemble strategy is adopted to dynamically adjust the proportions of different models according to different working conditions, and to fuse the prediction results of different models with dynamic weights. Among them, the dynamic integration strategy trains the Softmax layer to output the combined weights of each basic model, so that the integrated model can adaptively select the optimal prediction path under different working conditions, thereby improving the robustness of the integrated model.

7. The method for predicting strip mechanical properties based on federated learning according to claim 5, characterized in that: In the integrated model, CNN adopts a deeper convolutional layer structure and uses the Batch Normalization layer to normalize batch data to improve training efficiency and stability, alleviate the gradient vanishing problem, and suppress overfitting through Dropout; Transformer enhances the ability to extract temporal features of process parameter sequences through the position encoding layer and multi-head attention mechanism; MLP increases the number of hidden layer neurons and uses regularization technology to improve the ability to model the product effect.

8. The method for predicting mechanical properties of steel strips based on federated learning according to claim 5, wherein: The training strategy for the ensemble model is to use a cosine annealing learning rate schedule with an initial learning rate of 0.001, which dynamically decays to 0.000001 over the training cycle, while also using early stopping to avoid overfitting. The loss function uses the Huber loss with a hyperparameter δ=1.0 to balance the impact of outliers and enhance the model's robustness to industrial data noise. A 5-fold cross-validation approach is used, with the training and validation sets repartitioned for each fold to ensure reliable performance evaluation.

9. The method for predicting mechanical properties of steel strips based on federated learning according to claim 1, wherein: The data set corresponding to each client contains at least 20% of the public steel grade data; wherein, the public steel grade data is obtained by selecting several preset steel grades from all the steel grades in the production site as basic steel grades, and using the characteristic data corresponding to the basic steel grades as the public steel grade data; During the parameter aggregation stage of the central server, the central server uses a dynamic weight adjustment mechanism to dynamically adjust the client's weight to achieve weighted average aggregation. At the same time, a differential privacy mechanism is introduced to add Gaussian noise before uploading the parameters to ensure that the gradient information cannot be traced back to the original data characteristics.

10. The method for predicting mechanical properties of steel strips based on federated learning according to claim 9, characterized in that: The dynamic weight adjustment mechanism optimizes the federated aggregation process through a dual strategy. The first is data volume weighted aggregation, that is, the aggregation weight is allocated according to the client data scale. Clients with a high data volume contribute more to the global model update to ensure the dominant role of mainstream steel grade data; the second is performance feedback adjustment, that is, an additional 10%-15% weight gain is given to clients with too low MAE to encourage the knowledge contribution of high-quality local models.

Images