Differential privacy industrial equipment anomaly detection method based on personalized federation
By employing a personalized federated learning framework and differential privacy technology, the problems of model instability and privacy leakage in industrial equipment anomaly detection are solved, achieving stable detection and privacy protection in dynamic environments.
Patent Information
- Application Number
- CN202511182733.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-22
- Publication Date
- 2025-11-11
AI Technical Summary
In industrial equipment anomaly detection, existing federated learning algorithms perform poorly in dynamic environments and pose a risk of data privacy leakage, especially due to the instability of the model caused by non-independent and identically distributed client data, while lacking effective privacy protection measures.
A personalized federated learning framework is adopted, which combines hypernetworks and differential privacy technology. By adding Laplace noise to the parameter changes and using EWC regularization terms to constrain model updates, a stable personalized model is generated, and parameter aggregation and updates are performed on the server side.
It achieves stable anomaly detection of industrial equipment in dynamic environments, while enhancing data privacy protection, mitigating the risk of differential attacks, and ensuring the stability and privacy of the model.
Smart Images
Figure CN120930500A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to federated learning, continuous learning, hypernetworks, and differential privacy; more specifically, it relates to a differential privacy method for detecting anomalies in industrial equipment based on personalized federation. Background Technology
[0002] The safety risks posed by abnormal operation during the manufacturing process of industrial equipment have attracted widespread attention. Meanwhile, due to the sensitivity of the data, centralized training on a central server poses a risk of privacy breaches.
[0003] To address the data privacy issues arising from centralized training, federated learning emerged. Federated learning is a distributed machine learning algorithm that enables collaborative training while protecting data privacy. However, due to variations in environmental conditions and equipment wear and tear, the operational data of distributed industrial equipment is dynamic and not independently and identically distributed. This leads to poor performance of federated learning in practical applications of anomaly detection in industrial equipment.
[0004] To address the challenges of federated learning algorithms in dynamic environments and further enhance data privacy, this method utilizes a hypernetwork constrained by malleable weights to improve the stability of targeted updates to client models. This enables the generation of stable, personalized models for clients with non-independent and identically distributed data in dynamic industrial environments. Furthermore, this method leverages the principle of differential privacy by adding Laplace noise to parameter variations to mitigate differential attacks, further strengthening the protection of industrial data privacy.
[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention
[0006] To address the problems in related technologies, this invention proposes a differential privacy-based industrial equipment anomaly detection method based on personalized federation, characterized by overcoming the aforementioned technical problems existing in existing related technologies. Therefore, the specific technical solution adopted by this invention is as follows:
[0007] A differential privacy-based anomaly detection method for industrial equipment based on personalized federation is characterized by the following steps:
[0008] S1. Design a federated learning architecture consisting of one server and multiple client nodes. Modules such as the noise generator, noise additive, hypernetwork updater, Fisher matrix aggregator, and EWC regularization term calculator are deployed on the server side. Modules such as the identity vector generator, feature extraction network, and Fisher matrix calculator are deployed on each client.
[0009] S2. In each round of training, the client model is updated using the client's local industrial equipment data to obtain information such as client model parameters, changes in difference parameters, model prediction results, and identity vectors, and then sent to the server.
[0010] S3. After obtaining the identity vector and the change in differential parameters, the server updates the hypernetwork model parameters through the hypernetwork updater.
[0011] S4. The hypernetwork deployed on the server side generates personalized model parameters for the feature extraction network of each client through the parameter generator and sends them to the corresponding client node.
[0012] S5. The server calculates a regularization term based on the client model parameters and model prediction results sent by each client. The client adds the regularization term to the loss function to constrain the update of the client model parameters.
[0013] Furthermore, the client model is updated in S2, which includes the following steps:
[0014] S21. The client uses a feature extraction network to extract features from industrial equipment data to obtain data features;
[0015] S22. The data features are used as input to the personalization layer to obtain the model prediction results;
[0016] S23. Through backpropagation, update the parameters of the client feature extraction network based on the model prediction results to obtain the updated client model parameters θ. n and parameter change Δθ n ;
[0017] S24. Add Laplace noise to the parameter variation. This generates the variation in the differential parameter Δθ', where ε represents the privacy budget and s represents the sensitivity.
[0018] Furthermore, the calculation of the regularization term in S5 includes the following steps:
[0019] S51. Based on the model prediction results and client model parameters, the client uses a Fisher matrix calculator to calculate the client's Fisher matrix F. n And send it to the server;
[0020] S52. The server obtains the Fisher matrix from each client, averages and sums them to obtain the global Fisher matrix F;
[0021] S53. The EWC regularization term calculator deployed on the server side calculates the regularization term based on the global Fisher matrix and sends it to each client.
[0022] The beneficial effects of this invention are as follows:
[0023] This invention proposes a differential privacy-based anomaly detection method for industrial equipment based on personalized federation, and designs a novel trusted personalized federated learning framework based on hypernetworks and differential privacy. This paper uses hypernetworks to implement targeted updates of the client model, and uses EWC to constrain the parameter updates of the client model in a dynamic environment, thereby generating stable personalized models for clients with non-independent and identically distributed data. Furthermore, a differential privacy mechanism is used to mitigate differential attacks and protect the client model during the federated learning process. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 A flowchart illustrating the present invention's method for differential privacy industrial equipment anomaly detection based on personalized federation;
[0026] Figure 2 A flowchart illustrating client-side local training in a federated architecture for this invention;
[0027] Figure 3 This is a flowchart illustrating the process of generating personalized model parameters for each client using a hypernetwork, as described in this invention.
[0028] Figure 4 A flowchart illustrating the use of plastic weights to consolidate and enhance the stability of the client-side model in this invention;
[0029] Figure 5 A flowchart for adding Laplace noise to parameter variations is provided for this invention. Detailed Implementation
[0030] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.
[0031] According to embodiments of the present invention, a method for detecting anomalies in industrial equipment based on personalized federation is provided.
[0032] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a differential privacy industrial equipment anomaly detection method based on personalized federation includes the following steps:
[0033] Step S1: Design a federated learning architecture comprising one server and multiple client nodes. Modules such as the noise generator, noise additive, supernetwork updater, Fisher matrix aggregator, and EWC regularization term calculator are deployed on the server side. Modules such as the identity vector generator, feature extraction network, and Fisher matrix calculator are deployed on each client.
[0034] Step S2: In each round of training, update the client model using the client's local industrial equipment data. The specific calculation formula is as follows: Where η represents the learning rate for updating model parameters. Finally, information such as client-side model parameters, changes in difference parameters, model prediction results, and identity vectors are obtained and sent to the server.
[0035] Updating the client model includes the following steps:
[0036] [1] Based on the formula feature=f(D;θ), the client uses the feature extraction network to extract features from the industrial equipment data to obtain data features, where f represents the feature extraction network, D represents the industrial equipment data of the client, and θ represents the parameters of the feature extraction network;
[0037] [2] The client model can be defined as Model(D) = g(feature, γ), where the data features are used as input to the personalization layer to obtain the model prediction results, and γ represents the personalization layer parameters;
[0038] [3] Through backpropagation, the client feature extraction network is updated with parameters based on the model prediction results to obtain the updated client model parameters θ. n and parameter change Δθ n The specific formula is as follows
[0039] [4] Regarding the parameter change Δθ n Add Laplace noise To generate the difference parameter change Δθ n The specific formula is: Where ε represents the privacy budget and s represents sensitivity.
[0040] Step S3: After obtaining the identity vector and the changes in the difference parameters, the server updates the hypernetwork model parameters through the hypernetwork updater. The calculation formula is as follows: Where α represents the learning rate of the hypernetwork training, based on the formula Implement the update of the identity vector.
[0041] Step S4: The hypernetwork deployed on the server generates personalized model parameters for the feature extraction network of each client through a parameter generator, and sends them to the corresponding client nodes. The calculation formula is as follows:
[0042] Step S5: The server calculates a regularization term based on the client model parameters and model prediction results sent by each client. The client adds the regularization term to the loss function to constrain the update of the client model parameters. The formula for calculating the loss function is as follows: Where λ is a hyperparameter representing the correlation degree of the global Fisher matrix; the formula for calculating the client model parameter update is: Where η represents the learning rate for updating the model parameters.
[0043] Calculating the regularization term involves the following steps:
[0044] [1] The client calculates the Fisher matrix F using the Fisher matrix calculator based on the model prediction results and the client model parameters. n And send it to the server. The specific formula for calculating the Fisher matrix is as follows: in Represents θ n The optimal solution;
[0045] [2] The server obtains the Fisher matrix from each client, according to the formula... By averaging and summing these values, we obtain the global Fisher matrix F.
[0046] [3] The EWC regularization calculator deployed on the server side calculates the regularization terms based on the global Fisher matrix and sends them to each client.
Claims
1. A differential privacy-based industrial equipment anomaly detection method based on personalized federation, characterized in that, The method includes the following steps: S1. Design a federated learning architecture consisting of one server and multiple client nodes. Modules such as the noise generator, noise additive, hypernetwork updater, Fisher matrix aggregator, and EWC regularization term calculator are deployed on the server side. Modules such as the identity vector generator, feature extraction network, and Fisher matrix calculator are deployed on each client. S2. In each round of training, the client model is updated using the client's local industrial equipment data to obtain information such as client model parameters, changes in difference parameters, model prediction results, and identity vectors, and then sent to the server. S3. After obtaining the identity vector and the change in differential parameters, the server updates the hypernetwork model parameters through the hypernetwork updater. S4. The hypernetwork deployed on the server side generates personalized model parameters for the feature extraction network of each client through the parameter generator, and sends them to the corresponding client nodes. S5. The server calculates a regularization term based on the client model parameters and model prediction results sent by each client. The client adds the regularization term to the loss function to constrain the update of the client model parameters.
2. The method for anomaly detection of industrial equipment based on personalized federation according to claim 1, characterized in that, The client model update in S2 includes the following steps: S21. The client uses a feature extraction network to extract features from industrial equipment data to obtain data features; S22. The data features are used as input to the personalization layer to obtain the model prediction results; S23. Through backpropagation, update the parameters of the client feature extraction network based on the model prediction results to obtain the updated client model parameters θ. n and parameter change Δθ n ; S24. Add Laplace noise to the parameter variation. This generates the difference parameter variation Δθ', where ε represents the privacy budget and s represents the sensitivity.
3. The method for anomaly detection of industrial equipment based on personalized federation according to claim 1, characterized in that, The calculation of the regularization term in S5 includes the following steps: S51. Based on the model prediction results and client model parameters, the client uses a Fisher matrix calculator to calculate the client's Fisher matrix F. n And send it to the server; S52. The server obtains the Fisher matrix from each client, averages and sums them to obtain the global Fisher matrix F; S53. The EWC regularization term calculator deployed on the server side calculates the regularization term based on the global Fisher matrix and sends it to each client.