Self-adaptive privacy protection federal learning method for material life cycle management
By employing an adaptive privacy-preserving federated learning method, the problem of data privacy protection in the material lifecycle is addressed, achieving a balance between privacy protection and model performance, and improving the optimization effect of green manufacturing processes.
Patent Information
- Application Number
- CN202610056113.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-16
- Publication Date
- 2026-04-28
AI Technical Summary
The lack of effective privacy protection measures during the sharing and use of materials lifecycle data leads to privacy leaks, causing economic losses and reputational damage to enterprises.
An adaptive privacy-preserving federated learning approach is adopted, including a data management module, a virtual computing layer, a privacy collaboration layer, and an application service layer. Through techniques such as dynamic privacy budget allocation, client-side fairness scheduling, gradient pruning, and differential privacy noise addition, data privacy is protected and model performance is improved.
It effectively prevents privacy leaks, ensures model performance and stability, enhances the generalization ability and fairness of the global model, and assists in optimizing green manufacturing processes.
Smart Images
Figure CN121935963A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of materials lifecycle management technology, specifically to an adaptive privacy-preserving federated learning method for materials lifecycle management. Background Technology
[0002] In the field of materials lifecycle management technology, the collection, processing and analysis of data play a crucial role in optimizing green manufacturing processes, improving resource utilization efficiency and reducing environmental impact. With the rapid development of big data and artificial intelligence technologies, materials lifecycle data is showing characteristics of massive growth and multi-source heterogeneity, which provides abundant resources for in-depth data mining and effective utilization. However, materials lifecycle data often involves core secrets and sensitive information of enterprises, such as raw material composition, production process parameters, and product quality standards. If there is a lack of effective privacy protection measures during the sharing and use of this data, it is very easy to lead to privacy leaks, which in turn can cause significant economic losses and reputational damage to enterprises. Therefore, how to effectively protect data privacy while ensuring data availability has become a key issue that urgently needs to be addressed in the field of materials lifecycle management. Therefore, developing a smart neckband for myopia prevention with stronger overall performance and a better user experience is of great practical significance. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides an intelligent anti-myopia neckband and control system based on multimodal data fusion, which solves the problems mentioned in the background section.
[0004] To address the aforementioned technical problems, according to one aspect of the present invention, more specifically, an adaptive privacy-preserving federated learning method for materials lifecycle management, comprising a privacy-preserving federated learning system, the privacy-preserving federated learning system comprising: Data Management Module: Responsible for loading the original material lifecycle dataset, performing data cleaning and normalization, and generating a data index list specific to each client based on the experimental settings, thereby realizing logical data sharding; Virtual computing layer: Contains multiple virtual client instances and a central server controller; virtual client instances encapsulate local training methods and perform forward propagation, backpropagation, gradient clipping, and differential privacy noise addition; the central server controller maintains global model parameters, controls the progression of simulation rounds, and executes the client selection algorithm; Privacy Coordination Layer: Located in the central server, it is responsible for dynamic privacy budget allocation and client fairness scheduling; the dynamic privacy budget allocator is based on an arithmetic sequence dynamic allocation strategy, and the client fairness scheduler is based on a loss-aware fairness selection method. Application service layer: Based on the aggregated global model, it provides material lifecycle prediction services to help optimize green manufacturing processes.
[0005] Furthermore, in the privacy-preserving federated learning system, the dynamic privacy budget allocation method includes the following steps: Initialization phase: Receiving the total privacy budget Total number of communication rounds T and minimum first term threshold As input parameters; Calculate the first term: Based on the total privacy budget and the number of communication rounds, calculate the first term of the arithmetic sequence. The formula is ; First amendment: If Less than the minimum first term threshold Then set the first item to And recalculate the tolerance d, using the formula: ; Dynamic allocation: In each round of communication t, the arithmetic sequence formula is used... Allocate a privacy budget and calculate the corresponding Gaussian noise standard deviation. Noise injection is used in model training, where the privacy budget for round t is determined. Subsequently, based on Gaussian mechanism-based RDP analysis, for a function with sensitivity C, the noise standard deviation required in round t at RDP order α is calculated as follows: .
[0006] Furthermore, the dynamic privacy budget allocation method also includes an adaptive boundary protection mechanism, specifically: Tolerance check: Check the calculated tolerance d before each round of communication; Policy switching: If d is less than or equal to 0, the policy automatically switches to a uniform allocation policy, meaning the privacy budget allocated in each round is... This is to ensure the robustness of the algorithm and the effectiveness of privacy protection.
[0007] Furthermore, in the privacy-preserving federated learning system, the client-side fairness scheduler employs a loss-aware fairness selection method, including the following steps: Define metrics: Define fairness metrics, including accuracy variance and the average accuracy of the worst 10% of clients; Dynamically adjust the probability: based on the client's training loss from the previous round. and historical participation Dynamically adjust the probability of it being selected. ; Non-uniform sampling: By using non-uniform sampling, the probability of selecting poorly performing clients is increased, thereby enhancing the generalization ability and fairness of the global model.
[0008] Furthermore, in the privacy-preserving federated learning system, the virtual client instance performs gradient clipping and differential privacy noise addition methods during local training, including the following steps: Gradient clipping: Clip the calculated gradient to ensure that the gradient norm does not exceed a preset threshold C; Noise injection: based on the privacy budget of the current round. Calculate the corresponding Gaussian noise standard deviation. And Gaussian noise is injected into the clipped gradient; Model update: The model parameters are updated using the noisy gradient to protect data privacy.
[0009] Furthermore, the privacy-preserving federated learning system also includes a dynamic regularization mechanism, specifically: Define the loss function: Define the total loss function. Includes mean squared error loss and dynamic regularization terms; Adaptive update coefficients: The dynamic regularization coefficient µt is adaptively updated based on the loss change trend of the previous training round. The update rule is: if the loss increased in the previous round, µt is increased to quickly suppress model divergence; if the loss decreased in the previous round, µt is decreased to promote fine-grained convergence of parameters. Optimize the training process: Through dynamic regularization, optimize the model training process to improve the model's stability and prediction accuracy.
[0010] Furthermore, in the privacy-preserving federated learning system, the method for providing material lifecycle prediction services at the application service layer includes the following steps: Input reception: Receives material lifecycle-related parameters input by the user; Model prediction: Predictions are made using the aggregated global model to obtain prediction results for the material life cycle; Results Feedback: The prediction results are fed back to the user to help optimize green manufacturing processes.
[0011] Furthermore, the deployment method of the privacy-preserving federated learning system includes the following steps: Environment configuration: Configure the system operating environment on the industrial edge device, including installing the necessary software and libraries; Parameter settings: Configure system parameters, including total privacy budget, number of communication rounds, number of clients, minimum first-order threshold, etc. System startup: The privacy-preserving federated learning system is launched, and multiple rounds of federated learning training are conducted; Service provision: Provides material lifecycle prediction services based on the aggregated global model.
[0012] Furthermore, the deployment method of the privacy-preserving federated learning system includes the following steps in its multi-round training method: Initialize global model: Initialize global model parameters on the central server; Client selection: Based on the selection results of the client fairness scheduler, the clients participating in this round of training are determined; Local training: The selected client uses local data for local training, performing gradient clipping and differential privacy noise addition methods; Parameter aggregation: The central server collects the updated parameters from each client, executes a weighted average aggregation algorithm, and updates the global model; Iterative training: Repeat the above steps for multiple rounds of training until the stopping condition is met.
[0013] Furthermore, in the multi-round training method of the privacy-preserving federated learning system, the stopping condition determination method includes the following steps: Define stopping conditions: Define stopping conditions, such as reaching the maximum number of communication rounds, model convergence, or privacy budget exhaustion; Conditional judgment: After each round of training, determine whether the stopping condition is met; Training termination: If the stopping conditions are met, training will be terminated; otherwise, the next round of training will continue.
[0014] The beneficial effects of this invention's adaptive privacy-preserving federated learning method for materials lifecycle management are as follows: This invention introduces a dynamic privacy budget allocation mechanism based on an arithmetic sequence dynamic allocation strategy to ensure that the privacy budget allocation in each round of communication is both scientific and reasonable, effectively preventing the risk of privacy leakage. In specific implementation, the privacy budget is allocated according to the arithmetic sequence formula in each round of communication, and the corresponding Gaussian noise standard deviation is calculated for noise injection in model training, thereby maximizing the privacy protection effect while ensuring model performance.
[0015] This invention introduces a dynamic regularization mechanism. By defining a total loss function that includes mean squared error loss and a dynamic regularization term, and adaptively updating the regularization coefficient, the model training process is adjusted according to the loss change trend of the previous training round. When the loss increases, the regularization coefficient is increased to quickly suppress model divergence, and when the loss decreases, the coefficient is decreased to promote fine-grained convergence of parameters, thereby improving the stability and prediction accuracy of the model.
[0016] The client fairness scheduler in this application adopts a loss-aware fairness selection method. By defining fairness indicators such as accuracy variance and the average accuracy of the worst 10% of clients, it dynamically adjusts the probability of a client being selected. Through non-uniform sampling, it increases the participation of poorly performing clients, effectively enhancing the generalization ability and fairness of the global model. Attached Figure Description
[0017] The present invention will now be described in further detail with reference to the accompanying drawings and specific implementation methods.
[0018] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation
[0019] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.
[0020] like Figure 1 As shown, according to one aspect of the present invention, an adaptive privacy-preserving federated learning method for materials lifecycle management is provided, including a privacy-preserving federated learning system, the privacy-preserving federated learning system comprising: Data Management Module: Responsible for loading the original material lifecycle dataset, performing data cleaning and normalization, and generating a data index list specific to each client based on the experimental settings, thereby realizing logical data sharding; Virtual computing layer: Contains multiple virtual client instances and a central server controller; virtual client instances encapsulate local training methods and perform forward propagation, backpropagation, gradient clipping, and differential privacy noise addition; the central server controller maintains global model parameters, controls the progression of simulation rounds, and executes the client selection algorithm; Privacy Coordination Layer: Located in the central server, it is responsible for dynamic privacy budget allocation and client fairness scheduling; the dynamic privacy budget allocator is based on an arithmetic sequence dynamic allocation strategy, and the client fairness scheduler is based on a loss-aware fairness selection method. Application service layer: Based on the aggregated global model, it provides material lifecycle prediction services to help optimize green manufacturing processes.
[0021] In this embodiment, the dynamic privacy budget allocation method in the privacy-preserving federated learning system includes the following steps: Initialization phase: Receiving the total privacy budget Total number of communication rounds T and minimum first term threshold As input parameters; Calculate the first term: Based on the total privacy budget and the number of communication rounds, calculate the first term of the arithmetic sequence. The formula is ; First amendment: If Less than the minimum first term threshold Then set the first item to And recalculate the tolerance d, using the formula: ; Dynamic allocation: In each round of communication t, the arithmetic sequence formula is used... Allocate a privacy budget and calculate the corresponding Gaussian noise standard deviation. Noise injection is used in model training, where the privacy budget for round t is determined. Subsequently, based on Gaussian mechanism-based RDP analysis, for a function with sensitivity C, the noise standard deviation required in round t at RDP order α is calculated as follows: .
[0022] In this embodiment, the dynamic privacy budget allocation method further includes an adaptive boundary protection mechanism, specifically: Tolerance check: Check the calculated tolerance d before each round of communication; Policy switching: If d is less than or equal to 0, the policy automatically switches to a uniform allocation policy, meaning the privacy budget allocated in each round is... This is to ensure the robustness of the algorithm and the effectiveness of privacy protection.
[0023] In this embodiment, the privacy-preserving federated learning system employs a loss-aware fairness selection method in its client-side fairness scheduler, comprising the following steps: Define metrics: Define fairness metrics, including accuracy variance and the average accuracy of the worst 10% of clients; Dynamically adjust the probability: based on the client's training loss from the previous round. and historical participation Dynamically adjust the probability of it being selected. ; Non-uniform sampling: By using non-uniform sampling, the probability of selecting poorly performing clients is increased, thereby enhancing the generalization ability and fairness of the global model.
[0024] In this embodiment, in the privacy-preserving federated learning system, the virtual client instance performs gradient clipping and differential privacy noise addition methods during local training, including the following steps: Gradient clipping: Clip the calculated gradient to ensure that the gradient norm does not exceed a preset threshold C; Noise injection: based on the privacy budget of the current round. Calculate the corresponding Gaussian noise standard deviation. And Gaussian noise is injected into the clipped gradient; Model update: The model parameters are updated using the noisy gradient to protect data privacy.
[0025] In this embodiment, the privacy-preserving federated learning system also includes a dynamic regularization mechanism, specifically: Define the loss function: Define the total loss function. Includes mean squared error loss and dynamic regularization terms; Adaptive update coefficients: The dynamic regularization coefficient µt is adaptively updated based on the loss change trend of the previous training round. The update rule is: if the loss increased in the previous round, µt is increased to quickly suppress model divergence; if the loss decreased in the previous round, µt is decreased to promote fine-grained convergence of parameters. Optimize the training process: Through dynamic regularization, optimize the model training process to improve the model's stability and prediction accuracy.
[0026] In this embodiment, the method for providing material lifecycle prediction services at the application service layer in the privacy-preserving federated learning system includes the following steps: Input reception: Receives material lifecycle-related parameters input by the user; Model prediction: Predictions are made using the aggregated global model to obtain prediction results for the material life cycle; Results Feedback: The prediction results are fed back to the user to help optimize green manufacturing processes.
[0027] In this embodiment, the deployment method of the privacy-preserving federated learning system includes the following steps: Environment configuration: Configure the system operating environment on the industrial edge device, including installing the necessary software and libraries; Parameter settings: Configure system parameters, including total privacy budget, number of communication rounds, number of clients, minimum first-order threshold, etc. System startup: The privacy-preserving federated learning system is launched, and multiple rounds of federated learning training are conducted; Service provision: Provides material lifecycle prediction services based on the aggregated global model.
[0028] In this embodiment, the multi-round training method in the deployment method of the privacy-preserving federated learning system includes the following steps: Initialize global model: Initialize global model parameters on the central server; Client selection: Based on the selection results of the client fairness scheduler, the clients participating in this round of training are determined; Local training: The selected client uses local data for local training, performing gradient clipping and differential privacy noise addition methods; Parameter aggregation: The central server collects the updated parameters from each client, executes a weighted average aggregation algorithm, and updates the global model; Iterative training: Repeat the above steps for multiple rounds of training until the stopping condition is met.
[0029] In this embodiment, the stopping condition determination method in the multi-round training method of the privacy-preserving federated learning system includes the following steps: Define stopping conditions: Define stopping conditions, such as reaching the maximum number of communication rounds, model convergence, or privacy budget exhaustion; Conditional judgment: After each round of training, determine whether the stopping condition is met; Training termination: If the stopping conditions are met, training will be terminated; otherwise, the next round of training will continue.
[0030] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention are also within the protection scope of the present invention.
Claims
1. An adaptive privacy-preserving federated learning method for materials lifecycle management, characterized in that: This includes a privacy-preserving federated learning system, which includes: Data Management Module: Responsible for loading the original material lifecycle dataset, performing data cleaning and normalization, and generating a data index list specific to each client based on the experimental settings, thereby realizing logical data sharding; Virtual computing layer: Contains multiple virtual client instances and a central server controller; virtual client instances encapsulate local training methods and perform forward propagation, backpropagation, gradient clipping, and differential privacy noise addition; the central server controller maintains global model parameters, controls the progression of simulation rounds, and executes the client selection algorithm; Privacy Coordination Layer: Located in the central server, it is responsible for dynamic privacy budget allocation and client fairness scheduling; the dynamic privacy budget allocator is based on an arithmetic sequence dynamic allocation strategy, and the client fairness scheduler is based on a loss-aware fairness selection method. Application service layer: Based on the aggregated global model, it provides material lifecycle prediction services to help optimize green manufacturing processes.
2. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 1, characterized in that: The dynamic privacy budget allocation method in the privacy-preserving federated learning system includes the following steps: Initialization phase: Receiving the total privacy budget Total number of communication rounds T and minimum first term threshold As input parameters; Calculate the first term: Based on the total privacy budget and the number of communication rounds, calculate the first term of the arithmetic sequence. The formula is ; First amendment: If Less than the minimum first term threshold Then set the first item to And recalculate the tolerance d, using the formula: ; Dynamic allocation: In each round of communication t, the arithmetic sequence formula is used... Allocate a privacy budget and calculate the corresponding Gaussian noise standard deviation. Noise injection is used in model training, where the privacy budget for round t is determined. Subsequently, based on Gaussian mechanism-based RDP analysis, for a function with sensitivity C, the noise standard deviation required in round t at RDP order α is calculated as follows: .
3. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 2, characterized in that: The dynamic privacy budget allocation method also includes an adaptive boundary protection mechanism, specifically: Tolerance check: Check the calculated tolerance d before each round of communication; Policy switching: If d is less than or equal to 0, the policy automatically switches to a uniform allocation policy, meaning the privacy budget allocated in each round is... This is to ensure the robustness of the algorithm and the effectiveness of privacy protection.
4. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 1, characterized in that: In the privacy-preserving federated learning system, the client-side fairness scheduler employs a loss-aware fairness selection method, including the following steps: Define metrics: Define fairness metrics, including accuracy variance and the average accuracy of the worst 10% of clients; Dynamically adjust the probability: based on the client's training loss from the previous round. and historical participation Dynamically adjust the probability of it being selected. ; Non-uniform sampling: By using non-uniform sampling, the probability of selecting poorly performing clients is increased, thereby enhancing the generalization ability and fairness of the global model.
5. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 1, characterized in that: In the privacy-preserving federated learning system, the virtual client instance performs gradient clipping and differential privacy noise addition methods during local training, including the following steps: Gradient clipping: Clip the calculated gradient to ensure that the gradient norm does not exceed a preset threshold C; Noise injection: based on the privacy budget of the current round. Calculate the corresponding Gaussian noise standard deviation. And Gaussian noise is injected into the clipped gradient; Model update: The model parameters are updated using the noisy gradient to protect data privacy.
6. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 1, characterized in that: The privacy-preserving federated learning system also includes a dynamic regularization mechanism, specifically: Define the loss function: Define the total loss function. Includes mean squared error loss and dynamic regularization terms; Adaptive update coefficients: The dynamic regularization coefficient µt is adaptively updated based on the loss change trend of the previous training round. The update rule is: if the loss increased in the previous round, µt is increased to quickly suppress model divergence; if the loss decreased in the previous round, µt is decreased to promote fine-grained convergence of parameters. Optimize the training process: Through dynamic regularization, optimize the model training process to improve the model's stability and prediction accuracy.
7. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 1, characterized in that: The privacy-preserving federated learning system provides a method for the application service layer to offer material lifecycle prediction services, which includes the following steps: Input reception: Receives material lifecycle-related parameters input by the user; Model prediction: Predictions are made using the aggregated global model to obtain prediction results for the material life cycle; Results Feedback: The prediction results are fed back to the user to help optimize green manufacturing processes.
8. The adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 1, characterized in that: The deployment method of the privacy-preserving federated learning system includes the following steps: Environment configuration: Configure the system operating environment on the industrial edge device, including installing the necessary software and libraries; Parameter settings: Configure system parameters, including total privacy budget, number of communication rounds, number of clients, minimum first-order threshold, etc. System startup: The privacy-preserving federated learning system is launched, and multiple rounds of federated learning training are conducted; Service provision: Provides material lifecycle prediction services based on the aggregated global model.
9. An adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 8, characterized in that: The deployment method of the privacy-preserving federated learning system includes the following steps in its multi-round training method: Initialize global model: Initialize global model parameters on the central server; Client selection: Based on the selection results of the client fairness scheduler, the clients participating in this round of training are determined; Local training: The selected client uses local data for local training, performing gradient clipping and differential privacy noise addition methods; Parameter aggregation: The central server collects the updated parameters from each client, executes a weighted average aggregation algorithm, and updates the global model; Iterative training: Repeat the above steps for multiple rounds of training until the stopping condition is met.
10. An adaptive privacy-preserving federated learning method for materials lifecycle management according to claim 9, characterized in that: The stopping condition determination method in the multi-round training method of the privacy-preserving federated learning system includes the following steps: Define stopping conditions: Define stopping conditions, such as reaching the maximum number of communication rounds, model convergence, or privacy budget exhaustion; Conditional judgment: After each round of training, determine whether the stopping condition is met; Training termination: If the stopping conditions are met, training will be terminated; otherwise, the next round of training will continue.