Privacy-preserving cold rolling force prediction method based on deep belief network and federated learning

Abstract

The present invention discloses a privacy-preserving cold rolling force prediction method based on deep belief network and federated learning. First, the rolling data is feature selected by the maximum relevant minimum redundancy technology. Then, an improved deep belief network rolling force prediction model is constructed, and unsupervised pre-training is performed using the Leaky‑ReLU RBM and adaptive learning rate weighted contrast divergence algorithm. A PLSR layer is set on top of the deep belief network, and reverse fine-tuning is performed to minimize the loss between the feature set and the label set. Finally, under the framework of federated learning, a client selection mechanism and a federated weighted aggregation mechanism are designed. The model training loss value uploaded by the local client and the cosine distance to the global model are used to select the top u% of local clients to participate in server aggregation. By calculating the p percentile of the historical gradient L2 norm of the local client in each iteration as the clipping threshold of the current iteration, the dynamic gradient clipping localized differential privacy function of the local client is realized.

Description

Technical Field

[0001] The present invention relates to the field of rolling process technology, and specifically to a privacy-preserving cold rolling force prediction method based on deep belief network and federated learning. Background Art

[0002] Metal materials are widely used in industrial production, and cold rolling technology is one of the important links in the field of metal material processing. Cold rolling refers to the industrial production process of pressing metal billets into the required size through multiple rolling passes at room temperature. In the cold rolling process, rolling force is an important physical parameter that directly affects the rolling quality and process efficiency. Accurately predicting rolling force is of vital importance for optimizing process parameters, improving production efficiency, and reducing production costs.

[0003] Traditional rolling force prediction methods are mainly based on empirical models or mathematical modeling, such as multivariate linear regression, support vector machine, etc. These methods usually require a large number of data samples as training sets, and then predict the rolling force by fitting the data and model parameters. However, traditional methods have some limitations:

[0004] 1. Large data demand: Traditional methods usually require a large amount of centralized data for training and fitting, which means that a large amount of rolling process data needs to be collected and stored, increasing the cost of data management and processing.

[0005] 2. High computational complexity: For complex metal processing processes, traditional methods often require complex mathematical models and computational algorithms to handle them, which has high computational complexity and leads to low prediction efficiency.

[0006] 3. Poor data privacy and security: Traditional methods usually require centralized data processing and store data in one place for model training and prediction. This approach has the risk of data privacy leakage. Once the data is illegally obtained or leaked, it will have a serious impact on the company's business interests and competitive advantages. Summary of the invention

[0007] The present invention proposes a privacy-preserving cold rolling force prediction method based on deep belief network and federated learning. The application of this scheme can achieve accurate prediction of rolling force while ensuring data privacy.

[0008] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0009] A privacy-preserving cold rolling force prediction method based on deep belief network and federated learning includes the following steps:

[0010] Step 1: Use the maximum relevance minimum redundancy technique to select the features of the cold rolling data set to select the most representative and predictive feature set;

[0011] Step 2: Perform unsupervised pre-training of the deep belief network on the filtered feature data by stacking Leaky-ReLU RBM and improving the weighted contrast divergence of adaptive learning rate to improve the generalization ability and learning efficiency of the model;

[0012] Step 3: Integrate a partial least squares regression (PLSR) layer on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, so as to effectively capture the complex nonlinear relationship between features and labels;

[0013] Step 4: To address the overfitting problem that may occur during the pre-training and fine-tuning process of the model, the present invention introduces L2 regularization in the error calculation of the contrastive divergence algorithm and the loss function in the fine-tuning stage. This effectively controls the complexity of the model and improves the generalization ability;

[0014] Step 5: Use the reverse fine-tuning algorithm to fine-tune the deep belief network, optimize the model parameters, and minimize the error between the predicted target and the actual output;

[0015] Step 6: Deploy the improved deep belief network to each cold rolling mill (as a client) in the federated learning environment;

[0016] Step 7: Introduce dynamic gradient clipping local differential privacy for each client. By calculating the L2 norm of the gradient in each iteration and selecting the p percentile of the historical gradient L2 norm as the clipping threshold for the current iteration, the privacy protection strength is adjusted according to the sensitivity of the data and the privacy requirements of the client to ensure the privacy security of the data during the federated learning process.

[0017] Step 8: Design a client selection mechanism and federated weighted aggregation mechanism to select the top u% of local clients to participate in server aggregation by using the model training loss value uploaded by the local client and the cosine distance with the global model;

[0018] As a further optimization of this solution, feature selection uses the maximum relevance and minimum redundancy technology to select features related to cold rolling force prediction, and ensures the maximum relevance and minimum redundancy of the selected feature set. The techniques used in feature selection and dimensionality reduction processing include but are not limited to mutual information analysis and maximum relevance and minimum redundancy (mRMR) algorithm to ensure that the selected features can effectively improve the prediction accuracy of the model.

[0019] As a further optimization of this solution, the unsupervised pre-training process of the deep belief network includes stacking Leaky-ReLU RBM and introducing weighted contrast divergence with adaptive learning rate to improve the robustness and generalization ability of the model. The stacked Leaky-ReLU RBM achieves adaptive learning for different data distributions by adjusting the number of hidden layer nodes, learning rate and other hyperparameters, and enhances the model's ability to extract complex data features.

[0020] As a further optimization of this scheme, a partial least squares regression (PLSR) layer is integrated on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, so as to effectively capture the complex nonlinear relationship between features and labels. The introduction of the PLSR layer aims to optimize the mapping between features and labels, and improve the accuracy and generalization ability of the prediction model by minimizing the mean square error between the feature set and the label set.

[0021] As a further optimization of this solution, dynamic gradient clipping local differential privacy achieves privacy protection by calculating the L2 norm of the gradient and selecting the p percentile of the historical gradient L2 norm as the clipping threshold, protecting user sensitive information while improving the robustness and privacy protection level of the model.

[0022] As a further optimization of this scheme, the client selection mechanism and federated aggregation mechanism under the federated learning framework select local clients participating in server aggregation by combining loss value and cosine distance to improve the convergence speed and prediction performance of the model. The client selection mechanism selects the top u% clients with the smallest loss and the smallest difference from the global model based on the client model loss and the similarity with the global model (measured by cosine distance). The weighted aggregation algorithm takes into account the client's data volume, model quality, and training progress, and adopts a dynamic weight allocation mechanism to ensure that the global model update can fairly reflect the learning results of all selected clients.

[0023] The method first selects the rolling data features by the maximum correlation minimum redundancy technique, and then uses the stacked Leaky-ReLU RBM and the weighted contrastive divergence with improved adaptive learning rate to perform unsupervised pre-training of the deep belief network. Then, a PLSR layer is set on the top of the deep belief network to realize the mapping between features and labels, and L2 regularization is introduced in the error calculation and loss function of the fine-tuning stage of the contrastive divergence algorithm. Finally, the loss between the feature set and the label set is minimized by the reverse fine-tuning technique. Under the federated learning framework, the improved deep belief network is deployed as a local model, and a client selection mechanism and a federated weighted aggregation mechanism are designed. The model training loss value uploaded by the local client and the cosine distance with the global model are used to select the top u% of local clients to participate in server aggregation. At the same time, dynamic gradient clipping local differential privacy is introduced for each client. The privacy protection function is realized by calculating the L2 norm of the gradient in each iteration and selecting the p percentile of the historical gradient L2 norm as the clipping threshold of the current iteration.

[0024] In view of the limitations of traditional methods, in recent years, with the development of deep learning and data privacy protection technology, people have begun to try to apply deep learning models to the field of rolling force prediction, and explore the use of distributed computing and privacy protection technology to solve data privacy issues. Deep learning models have powerful representation and generalization capabilities, and can automatically learn and extract high-level features of data to better predict rolling force. Distributed computing and privacy protection technology can ensure the security and privacy of data and avoid the risk of data leakage.

[0025] Therefore, the present invention combines deep belief network and federated learning technology to propose a new rolling force prediction method. This method realizes the efficient use of data from multiple cold rolling mills through distributed computing and model aggregation, and protects the privacy security of the data through differential privacy technology. This method not only improves the accuracy and efficiency of rolling force prediction, but also effectively solves the problem of data privacy leakage, and has broad application prospects and market value. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0027] Figure 1 This is a structural diagram of the improved deep belief network of the present invention;

[0028] Figure 2 This is a diagram of the improved privacy protection network framework based on deep belief network and federated learning of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] Example 1, a privacy-preserving cold rolling force prediction method based on deep belief network and federated learning, screens feature sets for original data, and uses maximum relevance minimum redundancy technology to perform feature selection on cold rolling data sets to screen out the most representative and predictive feature sets;

[0031] Training feature data: The unsupervised pre-training of the deep belief network is performed on the filtered feature data by stacking Leaky-ReLU RBM and improving the weighted contrast divergence of the adaptive learning rate; the feature selection selects the features related to the cold rolling force prediction through the maximum relevance minimum redundancy technology, and ensures the maximum relevance and minimum redundancy of the selected feature set. The techniques used in feature selection and dimensionality reduction processing include but are not limited to mutual information analysis and maximum relevance minimum redundancy algorithm to ensure that the selected features can effectively improve the prediction accuracy of the model. The unsupervised pre-training process of the deep belief network includes stacking Leaky-ReLU RBM and introducing weighted contrast divergence with adaptive learning rate. The stacked Leaky-ReLU RBM adjusts the number of hidden layer nodes, learning rate and other hyperparameters to achieve adaptive learning of different data distributions and enhance the model's ability to extract complex data features.

[0032] Constructing mapping relationships: Integrating a partial least squares regression layer on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, so as to capture the complex nonlinear relationship between features and labels; Integrating a partial least squares regression layer on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, so as to effectively capture the complex nonlinear relationship between features and labels. The introduction of the PLSR layer aims to optimize the mapping between features and labels, and improve the accuracy and generalization ability of the prediction model by minimizing the mean square error between the feature set and the label set.

[0033] Regularization: L2 regularization is introduced in the error calculation of the contrastive divergence algorithm and the loss function in the fine-tuning stage;

[0034] Model optimization: Use the reverse fine-tuning algorithm to fine-tune the deep belief network, optimize the model parameters, and minimize the error between the predicted target and the actual output; dynamic gradient clipping local differential privacy achieves privacy protection by calculating the L2 norm of the gradient and selecting the p percentile of the historical gradient L2 norm as the clipping threshold, protecting user sensitive information while improving the robustness and privacy protection level of the model.

[0035] Network deployment: In the federated learning environment, the improved deep belief network is deployed in each cold rolling mill as a client; the client selection mechanism and federated aggregation mechanism under the federated learning framework select local clients participating in server aggregation by combining loss value and cosine distance to improve the convergence speed and prediction performance of the model. The client selection mechanism selects the top u% clients with the smallest loss and the smallest difference from the global model based on the client model loss and similarity with the global model. The weighted aggregation algorithm takes into account the client's data volume, model quality, and training progress, and adopts a dynamic weight allocation mechanism.

[0036] Enhanced privacy and security: Dynamic gradient clipping and local differential privacy are introduced for each client. By calculating the L2 norm of the gradient in each iteration, the privacy protection strength is adjusted according to the privacy requirements to ensure the privacy and security of data in the process of federated learning.

[0037] Aggregation mechanism setting: Set the client selection mechanism and the federated weighted aggregation mechanism to select the top u% of local clients to participate in server aggregation by utilizing the model training loss value uploaded by the local client and the cosine distance with the global model.

[0038] Specific embodiment 2, step 1: data processing

[0039] (1) Outlier processing: Through data cleaning and the 3σ principle, invalid and missing records in the data set are removed to ensure the quality of data for subsequent processing.

[0040] (2) Feature selection: The maximum relevance minimum redundancy technique is used to select features from the cold rolling data set. This technique can select the most representative and valuable feature set for predicting rolling force from a huge data set.

[0041] Step 2: Build an improved deep belief network

[0042] (1) Improved restricted Boltzmann machine model structure and training method: layer-by-layer unsupervised pre-training of deep belief networks is performed by stacking Leaky-ReLU RBMs and improving weighted contrast divergence with adaptive learning rates.

[0043]

[0044]

[0045] Where D and d represent the learning rate increase coefficient and decrease coefficient respectively, 0<d<1<D. The principle of adaptive change of learning rate is that when the parameter update direction (positive or negative change amount) after two consecutive iterations is the same, the learning rate will increase, otherwise it will decrease.

[0046] (2) Introducing the least squares regression layer: The partial least squares regression (PLSR) layer is integrated on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, which helps to effectively capture the complex nonlinear relationship between features and labels and improve the accuracy and generalization ability of the prediction model.

[0047] (3) Model fine-tuning: Use the back-propagation algorithm to fine-tune the entire DBN-PLSR model. By adjusting the weights of all layers, the error between the predicted target and the actual output is further reduced, and the model performance is optimized.

[0048] Step 3: Deployment in a Federated Learning Environment

[0049] The improved deep belief network is deployed in various cold rolling mills as clients to achieve model update and parameter optimization in a federated learning environment.

[0050] Step 4: Dynamic Gradient Clipping for Local Differential Privacy

[0051] Introduces dynamic gradient clipping local differential privacy for each client by calculating the L2 norm of the gradient in each iteration And select the p percentile of the historical gradient L2 norm as the clipping threshold for the current iteration Gradient clipping is performed to obtain Finally, differential privacy technology with Gaussian noise is added to ensure the privacy security of data during federated learning. The privacy protection strength is dynamically adjusted according to the sensitivity of the data and the privacy requirements of the client, while improving the robustness and privacy protection level of the model.

[0052]

[0053] Step 5: Client Selection and Federated Weighted Aggregation

[0054] (1) Client selection: The client calculates the loss function of the local model participating in this round of training and the cosine distance with the global model, and then uploads the obtained parameters to the server. The server multiplies the parameters uploaded by the clients participating in the training and sorts them. The top u% of clients are selected to issue local model upload permissions, and then the uploaded local models are aggregated into the global model of this round of training to accelerate the convergence speed of the model and improve the generalization ability.

[0055]

[0056] CDL m =Cosine Distance(A,B)×L m

[0057] Among them, A and B are two vectors, A·B represents the dot product of the vectors, ||A|| and ||B|| represent the Euclidean norm of the vectors (i.e., the length of the vector), L m represents the loss value of the local model,

[0058] (2) Federated weighted aggregation: Perform weighted aggregation on the models of the selected clients, dynamically assign weights based on the client’s data volume, model quality, and training progress, fairly reflect the learning outcomes of all participating clients, and update the global deep belief network model.

[0059] First, calculate the relative weight of the training loss of client m in round t. The calculation formula is as follows:

[0060]

[0061] When determining the aggregate weight of the client model in each round of training, a multiplication factor is used to combine the relative weight of the training loss and the amount of data as the aggregate weight of the model of client m in round t. The specific formula is as follows:

[0062]

[0063] Where: s is a hyperparameter used to adjust In calculation The impact of time

[0064] In general, the federated aggregation algorithm proposed in this section is based on the FedProx algorithm, replacing the data items whose weights are calculated based on the amount of client data with The final model aggregation formula is as follows:

[0065]

[0066] This invention significantly improves the accuracy of rolling force prediction by innovating the key technologies of traditional deep belief network (DBN) and federated learning framework, and strengthens the protection of data privacy in a distributed computing environment. The original algorithm is improved mainly in the following aspects:

[0067] Maximum relevance and minimum redundancy (mRMR): The mRMR algorithm is used to further optimize the feature selection and select the feature set that is highly correlated with the rolling force and has low redundancy with each other.

[0068] Traditional deep belief networks use restricted Boltzmann machines (RBMs) as building blocks. This paper improves it into Leaky-ReLU RBMs. This new type of RBM introduces tiny gradient leakage when processing activation functions, allowing information to be transmitted at a lower probability, thereby enhancing the model's ability to capture subtle features and improving the efficiency of data feature extraction.

[0069] In order to solve the problem that the fixed learning rate in the traditional contrast divergence algorithm may cause the training process to end prematurely or be difficult to converge, the present invention designs and implements an adaptive learning rate mechanism. On the basis of the weighted contrast divergence algorithm, by setting the maximum learning rate and the minimum learning rate limits, the learning rate is dynamically adjusted to ensure the stability and efficiency of the model in the fine-tuning stage.

[0070] In order to solve the overfitting problem that may occur during the pre-training and fine-tuning process of the model, the present invention introduces L2 regularization in the error calculation of the contrastive divergence algorithm and the loss function in the fine-tuning stage. By adding a penalty term proportional to the square of the weight to the loss function, the model is promoted to lean towards a smoother or more concise expression, effectively controlling the complexity of the model and improving the generalization ability.

[0071] Considering that the traditional single-layer BP network has problems such as local minimum traps and slow convergence speed when the deep belief network is being supervised, the present invention uses a partial least squares regression (PLSR) layer to replace the top-level BP network. The PLSR layer accelerates the convergence speed of the model and significantly improves the tuning accuracy by establishing an efficient mapping relationship between features and rolling force.

[0072] In order to solve the problem of lack of sufficient prior knowledge to reasonably set the clipping threshold in most learning tasks, this paper designs a dynamic gradient clipping strategy (DCDP). This strategy dynamically adjusts the gradient clipping threshold by real-time monitoring the L2 norm of the gradient during training. Compared with the traditional fixed gradient clipping method, DCDP can adaptively adjust the clipping boundary and add noise as needed, which not only ensures the rationality of noise addition, but also effectively improves the overall performance of the model.

[0073] In the federated learning framework, the present invention designs a new client selection mechanism and weighted clustering method to address the problem of non-independent and identically distributed data and the model deviation problem that may be caused by localized updates. This method can effectively improve the clustering effect of the global model, ensure that the global model is closer to the global optimal solution, and enhance the robustness and generalization ability of the model.

[0074] As shown in Tables 1 to 3, the present invention shows significant advantages compared with several prior arts. δ represents the relative error.

[0075] Table 1 Comparison of cold rolling experiments

[0076]

[0077] Table 2 Comparison of federated learning experiments

[0078]

[0079] Table 3 Comparison of dynamic threshold differential privacy and fixed threshold differential privacy experiments

[0080]

[0081] In various embodiments, the hardware implementation of the technology can directly adopt existing intelligent devices, including but not limited to industrial computers, PCs, smart phones, handheld stand-alone computers, floor-standing stand-alone computers, etc. Its input device preferably adopts a screen keyboard, its data storage and calculation module adopts existing memory, calculator, controller, its internal communication module adopts existing communication ports and protocols, and its remote communication adopts existing GPRS network, World Wide Internet, etc.

[0082] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0083] In the embodiments provided by the present invention, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0084] Each functional unit in each embodiment of the present invention may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of a software functional unit. If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and may also instruct the relevant hardware to complete it through a computer program. The computer program may be stored in a computer-readable storage medium, and the computer program may be executed by the processor to implement the steps of the above-mentioned method embodiments. Among them, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media does not include electrical carrier signals and telecommunications signals.

[0085] The above are only preferred embodiments of the present invention and are 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 in the protection scope of the present invention.

Claims

1. A privacy-preserving cold rolling force prediction method based on deep belief network and federated learning, characterized in that: The method is as follows: Screening feature sets for raw data: For the cold rolling data set, the maximum relevance minimum redundancy technology is used to select features of the cold rolling data to screen out the most representative and predictive feature sets; Training feature data: Perform unsupervised pre-training of the deep belief network on the filtered feature data by stacking Leaky-ReLU RBM and improving weighted contrast divergence with adaptive learning rate; Constructing mapping relationships: Integrating a partial least squares regression layer on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, capturing the complex nonlinear relationship between features and labels; Regularization: L2 regularization is introduced in the error calculation of the contrastive divergence algorithm and the loss function in the fine-tuning stage; Model optimization: Use the reverse fine-tuning algorithm to fine-tune the deep belief network, optimize the model parameters, and minimize the error between the predicted target and the actual output; Network deployment: In a federated learning environment, the improved deep belief network is deployed in various cold rolling mills as a client model; Enhanced privacy and security: Dynamic gradient clipping and local differential privacy are introduced for each client. By calculating the L2 norm of the gradient in each iteration, the privacy protection strength is adjusted according to the privacy budget and relaxation term to ensure the privacy and security of data in the process of federated learning. Aggregation mechanism settings: Set the client selection mechanism and the federated weighted aggregation mechanism to select the top u% of local clients to participate in server aggregation by utilizing the model training loss value uploaded by the local client and the cosine distance with the global model.

2. The privacy-preserving cold rolling force prediction method based on deep belief network and federated learning according to claim 1 is characterized in that: In the screening feature set, feature selection selects features related to the cold rolling force prediction through the maximum relevance and minimum redundancy technology, and ensures the maximum relevance and minimum redundancy of the selected feature set. The technologies used in feature selection and dimensionality reduction processing include but are not limited to mutual information analysis and maximum relevance and minimum redundancy algorithm to ensure that the selected features can effectively improve the prediction accuracy of the model.

3. The privacy-preserving cold rolling force prediction method based on deep belief network and federated learning according to claim 1 is characterized in that: In the training feature data, the unsupervised pre-training process of the deep belief network includes stacking Leaky-ReLU RBM and introducing weighted contrast divergence with adaptive learning rate. The stacked Leaky-ReLU RBM achieves adaptive learning of different data distributions by adjusting the number of hidden layer nodes, learning rate and other hyperparameters, thereby enhancing the model's ability to extract complex data features.

4. The privacy-preserving cold rolling force prediction method based on deep belief network and federated learning according to claim 1 is characterized in that: In constructing the mapping relationship, a partial least squares regression layer is integrated on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, so as to effectively capture the complex nonlinear relationship between features and labels. The introduction of the PLSR layer aims to optimize the mapping between features and labels, and improve the accuracy and generalization ability of the prediction model by minimizing the mean square error between the feature set and the label set.

5. The privacy-preserving cold rolling force prediction method based on deep belief network and federated learning according to claim 1 is characterized in that: In model optimization, dynamic gradient clipping local differential privacy achieves privacy protection by calculating the L2 norm of the gradient and selecting the p percentile of the historical gradient L2 norm as the clipping threshold, protecting user sensitive information while improving the robustness and privacy protection level of the model.

6. The privacy-preserving cold rolling force prediction method based on deep belief network and federated learning according to claim 1 is characterized in that: In network deployment, the client selection mechanism and federated aggregation mechanism under the federated learning framework select local clients participating in server aggregation by combining loss value and cosine distance to improve the convergence speed and prediction performance of the model. The client selection mechanism is based on the client model loss and the similarity with the global model, and selects the top u% clients with the smallest loss and the smallest difference with the global model. The weighted aggregation algorithm takes into account the client's data volume, model quality and training progress, and adopts a dynamic weight allocation mechanism.

7. The privacy-preserving cold rolling force prediction method based on deep belief network and federated learning according to claim 1 is characterized in that: Step 1: Data Processing (1) Outlier processing: Through data cleaning and the 3σ principle, invalid and missing records in the data set are removed to ensure the quality of data for subsequent processing. (2) Feature selection: The maximum relevance minimum redundancy technique is used to select features from the cold rolling data set. This technique can select the most representative and predictive feature set for predicting rolling force from a huge data set. Step 2: Build an improved deep belief network (1) Improved restricted Boltzmann machine model structure and training method: layer-by-layer unsupervised pre-training of deep belief networks is performed by stacking Leaky-ReLU RBMs and improving weighted contrast divergence with adaptive learning rates. Where D and d represent the learning rate increase coefficient and decrease coefficient respectively, 0<d<1<D. The principle of adaptive change of learning rate is that when the parameter update direction after two consecutive iterations is the same, the learning rate will increase, otherwise it will decrease. (2) Introducing the partial least squares regression layer: The partial least squares regression layer is integrated on top of the deep belief network to construct an efficient mapping relationship between features and rolling force labels, which helps to effectively capture the complex nonlinear relationship between features and labels and improve the accuracy and generalization ability of the prediction model. (3) Model fine-tuning: Use the back-propagation algorithm to fine-tune the entire DBN-PLSR model. By adjusting the weights of all layers, the error between the predicted target and the actual output is further reduced, and the model performance is optimized. Step 3: Deployment in a Federated Learning Environment The improved deep belief network is deployed in various cold rolling mills as clients to achieve model update and parameter optimization in a federated learning environment. Step 4: Dynamic Gradient Clipping for Local Differential Privacy Introduces dynamic gradient clipping local differential privacy for each client by calculating the L2 norm of the gradient in each iteration And select the p percentile of the historical gradient L2 norm as the clipping threshold for the current iteration Gradient clipping is performed to obtain Finally, differential privacy technology with Gaussian noise is added to ensure the privacy security of data during federated learning. The privacy protection strength is dynamically adjusted according to the sensitivity of the data and the privacy requirements of the client, while improving the robustness and privacy protection level of the model. Step 5: Client Selection and Federated Weighted Aggregation (1) Client selection: The client calculates the loss function of the local model participating in this round of training and the cosine distance with the global model, and then uploads the obtained parameters to the server. The server multiplies the parameters uploaded by the clients participating in the training and sorts them. The top u% of clients are selected to issue local model upload permissions, and then the local models are uploaded to the global model aggregation of this round of training to accelerate the convergence speed of the model and improve the generalization ability. CDL m =Cosine Distance(A,B)×L m Among them, A and B are two vectors, A·B represents the dot product of the vectors, ||A|| and ||B|| represent the Euclidean norm of the vectors (i.e., the length of the vector), L m represents the loss value of the local model, (2) Federated weighted aggregation: Perform weighted aggregation on the models of the selected clients, dynamically assign weights based on the client’s data volume, model quality, and training progress, fairly reflect the learning outcomes of all participating clients, and update the global deep belief network model. First, calculate the relative weight of the training loss of client m in round t. The calculation formula is as follows: When determining the aggregate weight of the client model in each round of training, a multiplication factor is used to combine the relative weight of the training loss and the amount of data as the aggregate weight of the model of client m in round t. The specific formula is as follows: Where: s is a hyperparameter used to adjust In calculation The impact of time In general, the federated aggregation algorithm proposed in this section is based on the FedProx algorithm, replacing the data items whose weights are calculated based on the amount of client data with The final model aggregation formula is as follows:

Images