Asynchronous personalized federal learning method and system
By employing an asynchronous personalized federated learning method, combining neural networks with shared and personalized layers, and dynamically adjusting weights and latency compensation, the problems of data privacy and heterogeneous data fusion in high-speed train bogie diagnosis are solved, achieving efficient and secure personalized diagnosis.
Patent Information
- Application Number
- CN202510909296.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-11-25
AI Technical Summary
In intelligent fault diagnosis of high-speed train bogies, existing federated learning methods struggle to efficiently integrate diverse heterogeneous data and support asynchronous personalized optimization while ensuring data privacy and security. This results in limited model generalization capabilities, making it unable to adapt to the diagnostic needs of different lines and operating conditions.
An asynchronous personalized federated learning approach is adopted. By training a neural network including a shared layer and a personalized layer on the client side, the aggregation weights are dynamically adjusted using a delay compensation weight mechanism and a meta-learning algorithm. Asynchronous data upload is supported, and model aggregation is performed on the central server to ensure data privacy and personalized optimization.
It improves the personalization capabilities and model convergence speed of federated learning, adapts to asynchronous communication and computation characteristics, enhances the generalization ability and diagnostic accuracy of the model, and ensures data security and efficiency.
Smart Images
Figure CN121010010A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence, and in particular to an asynchronous personalized federated learning method and system. Background Technology
[0002] High-speed trains, as key equipment in the modern railway transportation system, are directly related to the stable operation of the national transportation lifeline and socio-economic development. The bogie, as one of the most crucial components of the high-speed train structure, undertakes multiple tasks including supporting the car body, guiding operation, and buffering vibrations. Due to its complex mechanical structure, numerous components, and long-term operation under different lines and conditions, the bogie is highly susceptible to various external factors such as track unevenness, alternating bridges and tunnels, and climate change, leading to frequent problems such as wheelset wear, suspension system performance degradation, and component loosening. These problems not only exacerbate train vibration and impact but may also trigger cascading failures, even jeopardizing the safe operation of the entire train in severe cases. Therefore, intelligent fault diagnosis technology for high-speed train bogies has become a research hotspot and engineering challenge in the rail transit field.
[0003] In actual operation and maintenance, deep learning-based intelligent diagnostic methods have been widely applied to bogie fault detection and health assessment. Deep learning models rely on a large amount of high-quality labeled data; however, due to limitations in actual data collection conditions and the cost of manual labeling, it is often difficult to obtain sufficient and balanced training samples for a single line or a single piece of equipment. High-speed trains on different lines exhibit significant heterogeneity in data distribution due to differences in operating environment, equipment status, and data collection frequency, and the types of faults and the number of samples contained within each are also extremely uneven. Models trained solely on local data from a single line often have limited generalization ability and are difficult to adapt to the diagnostic needs of other lines or new operating conditions.
[0004] In theory, jointly modeling bogie data from multiple railway lines could significantly improve the model's generalization ability, enabling efficient fault identification across lines and scenarios. However, directly centralizing data from multiple sources not only faces enormous economic and management costs, but more importantly, it presents policy and technical obstacles in terms of data security and privacy protection. Especially in sensitive sectors such as rail transit, healthcare, and finance, data is typically strictly protected and cannot leave local storage devices. Data barriers and competition also exist between enterprises, making data sharing and centralized modeling difficult to implement. Data-driven intelligent diagnostic methods are thus severely limited in real-world industrial environments. How to fully mine the knowledge within distributed data while ensuring data privacy and security has become a major challenge in the field of intelligent modeling.
[0005] Federated learning offers a new approach to solving these problems. By uploading only local model parameters or gradients from each line without sharing the original data, and then aggregating the models on a central server, it achieves a balance between data privacy protection and knowledge sharing. Federated learning can collaboratively model data from multiple lines, devices, and scenarios, theoretically achieving stronger generalization capabilities and higher diagnostic accuracy. However, practical deployment still faces many challenges. First, the data distribution across different lines is extremely uneven; some lines have large amounts of data and rich information, while others have scarce data but may contain key fault characteristics. Traditional federated learning frameworks typically assign equal weights to each client and use simple parameter averaging aggregation, which can easily lead to the model information from small sample lines being "submerged," failing to effectively transfer valuable knowledge and affecting overall model performance.
[0006] Secondly, significant differences in client device performance, network bandwidth, and data acquisition frequency in industrial scenarios lead to highly asynchronous model training and uploading processes. Synchronous aggregation strategies not only reduce system efficiency but also easily affect model convergence and final performance due to "stale gradients." Furthermore, in practical applications, the business needs and data characteristics of each network differ greatly, making it difficult for a single global model to meet the optimal performance requirements of all networks; therefore, the need for personalized modeling is increasingly prominent.
[0007] In summary, federated learning methods still have many shortcomings in handling asynchronous, heterogeneous, and personalized high-speed train bogie data, making it difficult to balance efficiency and personalization. How to achieve a federated learning method that can both protect data privacy and efficiently integrate various heterogeneous data while supporting asynchronous and personalized optimization has become a pressing technical challenge. Summary of the Invention
[0008] The purpose of this application is to provide an asynchronous personalized federated learning method and system that can protect data privacy while efficiently integrating various heterogeneous data and supporting asynchronous personalized optimization. Simultaneously, it can improve the computational efficiency of edge computing devices and servers through asynchronous mechanisms; reduce the storage overhead of models on servers through parameter sharing; and possess good portability and adaptability by supporting multi-line, multi-device collaborative training, making it suitable for large-scale industrial IoT environments.
[0009] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an asynchronous personalized federated learning method, including: Step 1: Initialize the global model. Use the initialized global model as the first global model. Determine the shared layer parameters and version of the first global model. The first global model is a neural network that includes shared layers and personalized layers. Step 2: Using the client's local data, train the client's local model based on the first global model to obtain the client-optimized model, and determine the shared layer parameters and version of the client-optimized model. Step 3: Based on the shared layer parameters of the client-side optimization model, the compensation weights of the client-side optimization model, and the shared layer parameters of the first global model, update the shared layer parameters of the first global model to obtain the updated first global model; the compensation weights of the client-side optimization model are determined by the difference between the version of the first global model and the version of the client-side optimization model. The updated first global model is used as the new first global model. The version of the new first global model is determined, and the process returns to step two until the first global model converges.
[0010] Secondly, this application provides an asynchronous personalized federated system, including a central server and multiple clients connected to the central server; the central server includes: The global variable initialization module is used to initialize the global model, take the initialized global model as the first global model, and determine the shared layer parameters and the version of the first global model; the first global model is a neural network including shared layers and personalized layers. The first global model update module is used to update the shared layer parameters of the first global model based on the shared layer parameters of the client-optimized model, the compensation weights of the client-optimized model, and the shared layer parameters of the first global model, so as to obtain the updated first global model; the compensation weights of the client-optimized model are determined by the difference between the version of the first global model and the version of the client-optimized model. The training termination criterion module is used to send a stop signal to multiple clients based on the convergence of the first global model. The client includes: The local model training module is used to download the first global model and train the client's local model based on the first global model using the client's local data to obtain the client-optimized model. The parameter acquisition module is used to acquire the shared layer parameters and the version of the client optimization model, and upload them to the central server. The training termination module is used to terminate training in the local model training module based on a stop signal.
[0011] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides an asynchronous personalized federated learning method and system. First, a neural network including shared and personalized layers is trained using local data on the client side, balancing global sharing with local differences, thus improving the personalization capability of federated learning. Second, the first global model is updated using the shared layer parameters from the client side, allowing for the fusion of various heterogeneous data to obtain more accurate shared layer parameters for the first global model, accelerating the convergence speed of federated learning. Third, this application supports asynchronous data uploads from the client side; as long as the client-optimized model is available, it can be uploaded at any time, improving the efficiency of federated learning and adapting to the asynchronous communication and computation characteristics of real-world industrial environments. Finally, by introducing a delay compensation weight mechanism based on the version difference between the client-optimized model and the first global model, the outdated client-optimized model caused by asynchronous uploads is addressed through adaptive weighted fusion, improving the convergence speed and final performance of the first global model. Furthermore, this application only uploads the client's shared layer parameters, maintaining the privacy of local data and improving the security of federated learning. In summary, the federated learning method of this application possesses multiple advantages such as asynchronous processing, heterogeneity, personalization, privacy, and high efficiency, significantly improving the performance of federated learning. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0013] Figure 1 An application environment diagram of an asynchronous personalized federated learning method provided in an embodiment of this application; Figure 2 A flowchart illustrating an asynchronous personalized federated learning method provided in an embodiment of this application; Figure 3 A meta-learning framework diagram of an asynchronous personalized federated learning method provided in another embodiment of this application; Figure 4 A first schematic diagram of key components and typical faults of a high-speed train bogie in an asynchronous personalized federated learning method provided in another embodiment of this application; Figure 5 A second schematic diagram illustrating key components and typical faults of a high-speed train bogie in an asynchronous personalized federated learning method provided in another embodiment of this application; Figure 6 A third schematic diagram illustrating key components and typical faults of a high-speed train bogie in an asynchronous personalized federated learning method provided in another embodiment of this application; Figure 7A schematic diagram of the functional modules of an asynchronous personalized federated learning system provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation
[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0015] To make the objectives, features, and advantages of this application more apparent and understandable, the following detailed description of this application is provided in conjunction with the accompanying drawings and specific embodiments.
[0016] The asynchronous personalized federated learning method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the shared layer parameters and version of the client optimization model to server 104. Server 104 updates the shared layer parameters of the first global model based on the shared layer parameters, compensation weights, and shared layer parameters of the first global model, obtaining an updated first global model. Server 104 can then feed back the updated first global model to terminal 102.
[0017] In asynchronous personalized federated learning, server 104 serves as the central server, and terminal 102 acts as the client, with multiple clients. The central server undertakes the core tasks of global model aggregation and dynamic optimization. Unlike traditional synchronous federated learning, the asynchronous federated learning mechanism allows each client to upload its local model parameters to the central server at any time after local model training is complete, without waiting for all clients to synchronize. The central server continuously listens for client upload requests and merges each uploaded local model parameter with the current global model. Since the parameters of the local models uploaded by each client may be based on asynchronous versions of the global model, they possess a certain degree of "staleness" or latency. To address this, the system introduces a latency compensation mechanism, dynamically adjusting aggregation weights to reasonably balance the impact of new and old information. Specifically, the central server calculates the difference between the version of the parameters of the local model uploaded by the client and the version of the current global model, and sets a latency compensation coefficient accordingly. This coefficient decreases as latency increases, ensuring that the impact of stale parameters on the global model is effectively suppressed.
[0018] This application ensures that models with higher freshness receive greater weight in the aggregation, while models with higher latency are appropriately suppressed to prevent outdated information from affecting global convergence and to mitigate global model drift caused by asynchronous processing. For updates with excessive latency, federated learning automatically discards them, ensuring the timeliness and robustness of the global model.
[0019] In one exemplary embodiment, such as Figure 2 As shown, an asynchronous personalized federated learning method is provided, which includes the following steps one through three. Wherein: Step 1: Initialize the global model. The initialized global model is used as the first global model. The shared layer parameters and version of the first global model are determined. The first global model is a neural network including shared layers and individualized layers. Specifically, the first global model contains multiple shared feature extraction layers and independent individualized output layers. During the initialization of the global model, this application allocates resources such as global model parameters, global model version, asynchronous update queue, global lock, and training statistics to ensure the concurrency and consistency of federated learning. Each client node loads its local dataset and establishes an asynchronous communication mechanism with the central server.
[0020] Step 2: Using the client's local data, train the client's local model based on the first global model to obtain the client-optimized model, and determine the shared layer parameters and version of the client-optimized model. Asynchronously uploading the shared layer parameters of each client-optimized model to the central server avoids leakage of raw data. In addition, the client supports breakpoint resume training and multiple rounds of local iteration, improving communication efficiency and local model convergence speed.
[0021] Step 3: Based on the shared layer parameters of the client-side optimization model, the compensation weights of the client-side optimization model, and the shared layer parameters of the first global model, update the shared layer parameters of the first global model to obtain the updated first global model; the compensation weights of the client-side optimization model are determined by the difference between the version of the first global model and the version of the client-side optimization model.
[0022] The updated first global model is used as the new first global model. The version of the new first global model is determined, and the process returns to step two until the first global model converges.
[0023] In another exemplary embodiment of this application, there are two clients: Client 1 and Client 2. At the start of federated learning, the global model is initialized, and the initialized global model is used as the first global model. The version of the first global model is determined to be the global first version, and this first global model is used as the local model for both clients. Each local model is trained using local data from each client to obtain the optimized model for Client 1, and the version of the optimized model for Client 1 is determined to be the local first version. The local model for Client 2 has not yet been trained due to network speed or other issues.
[0024] The parameters of the optimized model from client 1 are uploaded to the central server to update the first global model. This updated version of the first global model is then used as the second global version. Client 1 downloads the updated first global model as its local model and trains it, obtaining the local second version of the optimized model from client 1. At this point, client 2 has completed its training, and its optimized model version is determined to be the local first version. Using the parameters of the local second version of the optimized model from client 1 and the parameters of the local first version of the optimized model from client 2, the updated first global model (the second global version) is updated again.
[0025] At this point, during the update, the local version of the parameters of client-side optimized model 1 is consistent with the global version of the updated first global model, both being version 2. However, the version of the parameters of client-side optimized model 2 is different from the version of the updated first global model. Based on the version difference, different weights are assigned to the parameters of client-side optimized model 1 and client-side optimized model 2, respectively.
[0026] In this way, federated learning can dynamically fuse information from multiple local models in an asynchronous environment, ensuring stable convergence of the global model and improving its adaptability to asynchronous and heterogeneous data. Furthermore, the central server records real-time statistics such as latency distribution, client contribution, and accuracy for each aggregation during the aggregation process, providing data support for subsequent processes.
[0027] By implementing steps one through three above, a neural network including shared and personalized layers is used to balance global sharing with local differences, thus improving the personalization capabilities of federated learning. By fusing multiple heterogeneous data sets, more accurate shared layer parameters for the first global model are obtained, accelerating the convergence speed of federated learning. Furthermore, a delay compensation weight mechanism based on the version difference between the client-optimized model and the first global model is introduced, enabling adaptive weighted fusion in federated learning and improving the convergence speed and final performance of the first global model. Asynchronous data upload by the client improves the efficiency of federated learning and adapts to the asynchronous communication and computation characteristics of real-world industrial environments. Uploading only the shared layer parameters from the client enhances the security of federated learning.
[0028] In another exemplary embodiment of this application, meta-learning algorithms are fully integrated during the local model training phase on the client side, achieving an organic unity of personalization and generalization capabilities. Step two specifically includes: Based on the first global model, the client's local model is trained using the client's local data and a meta-learning algorithm and a multilayer perceptron (MLP) to obtain the client-optimized model.
[0029] In terms of local model structure design, to simultaneously support general feature modeling across lines and line-specific local optimization, this application adopts a dual-branch neural network architecture of "general feature branch + line-specific branch". Specifically, the shared layer, or general branch, consists of multiple fully connected neural networks and is used to extract common vibration features between different lines; these parameters are aggregated and updated by the server during federated learning. The specific layer, or special branch, is unique to each client and typically consists of single-layer or multi-layer classifiers. It is updated only on local data and does not participate in global synchronization, preserving the line-specific diagnostic feature extraction capability. This structure, while ensuring collaborative training among different clients, also takes into account local feature differences and heterogeneity, improving the model's generalization ability and personalized diagnostic capabilities.
[0030] The core objective of this application is to enable the global model to quickly adapt to new tasks or environments using a small number of samples. In this application, each client's local model consists of two parts: a shared layer and a personalized layer. The shared layer extracts general features, while the personalized layer is specifically designed for adaptive optimization of local data, such as... Figure 3 As shown.
[0031] During local model training, the client first divides the local dataset into a support set and a query set to support the inner and outer loop optimization mechanism of the meta-learning algorithm. Specifically, training is divided into two phases: First, in the inner loop phase, the local model undergoes several rapid adaptation steps on the support set, primarily updating the parameters of the personalization layer to fully leverage local data features and achieve personalized model customization. This process can be described by the following formula: .
[0032] in, Here, i represents the iteration number of the local model, and i is the client-side optimized model index. For personalized layer parameters, The meta-learning rate, To support the loss function on the set, To optimize the personalization layer parameters of the model for the i-th client in the (a+1)th iteration, To optimize the personalization layer parameters of the model for the i-th client in the a-th iteration, To The gradient.
[0033] Subsequently, in the outer loop phase, the local model evaluates its performance on the query set after rapid adaptation and optimizes the meta-loss accordingly. The goal of the outer loop is to minimize the loss on the query set, specifically expressed as: .
[0034] in, For shared layer parameters, Optimize the parameters of the model after the inner loop update for the i-th client. The loss function on the query set, This represents the expected value of the loss function.
[0035] Through the aforementioned meta-learning algorithm, the local model can quickly adapt to the local data distribution on each client. After each round of training of the local model, the client uploads the optimized shared layer parameters and personalized layer performance to the central server to participate in the asynchronous aggregation of the first global model.
[0036] In another exemplary embodiment of this application, in order to effectively prevent low-quality models caused by client-side local anomalies or data drift from being mistakenly included in global aggregation and to enhance the robustness of federated learning, the historical performance of the accuracy of client-side optimized models is used to screen them. Client-side optimized models whose accuracy is lower than the historical performance anomalies are updated, and federated learning automatically rejects aggregation, thereby improving the robustness of the first global model. Step three specifically includes steps 31 to 33.
[0037] Step 31: Determine the accuracy of the client-side optimization model and the historical average accuracy of the client-side optimization model.
[0038] Step 32: Based on the accuracy of the client-side optimization model and its historical average accuracy, the client-side optimization model is initially screened to obtain the initially screened client-side optimization model.
[0039] This application continuously records the average accuracy of each client-side optimization model based on its historical uploads, and uses this as a benchmark to filter client-side optimization models, setting historical performance criteria: .in, To optimize the accuracy of the model for the i-th client during the t-th update of the first global model, Tolerance coefficient for historical performance The historical average accuracy of the optimized model for the i-th client.
[0040] Step 33: Based on the shared layer parameters of the client-side optimization model after initial screening, the compensation weights of the client-side optimization model after initial screening, and the shared layer parameters of the first global model, update the shared layer parameters of the first global model to obtain the updated first global model.
[0041] In another exemplary embodiment of this application, a multi-level, dynamically adjusted local model selection strategy is proposed to address the complexity of heterogeneous data and the heterogeneity of local models in an asynchronous environment. The central server dynamically adjusts the accuracy threshold of the client-optimized model based on the update frequency of the first global model, achieving adaptive selection of the client-optimized model. Step 34 specifically includes steps 341 to 344.
[0042] Step 341, determine the dynamic accuracy threshold; the dynamic accuracy threshold is determined by the initial threshold, the threshold growth rate, the upper limit of the threshold and the number of updates of the first global model.
[0043] Step 342: Based on the dynamic accuracy threshold and the accuracy of the client-side optimization model after initial screening, the client-side optimization model after initial screening is screened again to obtain the screened client-side optimization model.
[0044] Step 343: Determine the compensation weight of the filtered client-side optimization model based on the difference between the version of the first global model and the version of the filtered client-side optimization model.
[0045] Step 344: Based on the shared layer parameters of the filtered client-side optimization model, the compensation weights of the filtered client-side optimization model, and the shared layer parameters of the first global model, update the shared layer parameters of the first global model to obtain the updated first global model.
[0046] In another exemplary embodiment of this application, a dynamic accuracy threshold is set for each round of client-side optimization model updates. This dynamic accuracy threshold starts from an initial threshold and gradually increases as the first global model is continuously updated. The dynamic accuracy threshold is determined using the following formula: .
[0047] Where t is the number of times the first global model is updated. This is the dynamic accuracy threshold for the t-th update of the first global model. Initial threshold, The threshold growth rate, This represents the upper limit of the threshold.
[0048] This dynamic accuracy threshold ensures the inclusiveness of federated learning towards diversity and exploration in the early stages, while gradually converging to a high-quality first global model in the later stages, thus improving overall performance. This application employs a dual-screening mechanism, combining the accuracy threshold of the current client-optimized model with the historical average accuracy to further filter out anomalous or low-quality data, thereby enhancing the robustness of the first global model.
[0049] In another exemplary embodiment of this application, the filtered first... Compensation weights for each client-side optimization model: .
[0050] in, For the filtered first Compensation weights for each client-side optimization model. The sequence number of the optimized client model after screening. As the attenuation base, To compensate for the lower limit of the weight, For the filtered first The difference between the version of the client-side optimized model and the version of the first global model. This is the base value for compensating for the weights.
[0051] In another exemplary embodiment of this application, the shared layer parameters of the first global model are finally updated using a weighted aggregation strategy for all client-optimized models that pass the screening, specifically expressed as follows: .
[0052] in, These are the shared layer parameters updated for the t-th time in the first global model. These are the shared layer parameters updated in the (t+1)th iteration of the first global model. For the filtered first The client-side optimization model optimizes the shared layer parameters during the t-th update of the first global model.
[0053] This weighted aggregation not only integrates knowledge from multi-source heterogeneous data, but also achieves adaptive compensation for asynchrony and latency through dynamic weight allocation.
[0054] This application also provides an application scenario in which the above-described asynchronous personalized federated learning method is applied. Specifically, the asynchronous personalized federated learning method provided in this embodiment can be applied to multi-track intelligent fault diagnosis of high-speed train bogies. Specifically, in high-speed train bogies, bogie data from multiple tracks can be jointly modeled, which can greatly improve the generalization ability of the model and achieve efficient fault identification across tracks and scenarios.
[0055] In this embodiment, the local data collected by the client consists of bogie fault data from multiple high-speed train lines. To obtain high-quality and representative high-speed train bogie fault data, the SIMPACK multibody dynamics simulation platform was used to systematically simulate the train operation status on different lines. Since conducting bogie fault experiments directly in a real industrial environment is not only costly but also poses significant safety risks, the simulation platform can efficiently collect abundant fault and normal operating condition data under controllable conditions. The high-speed train dynamics model used in the experiment was provided by the State Key Laboratory of Traction Power at Southwest Jiaotong University. Based on the nonlinear dynamic characteristics of the CRH380A high-speed train, the overall system has 62 degrees of freedom. By deploying multiple sets of sensors at key parts of the bogie, multidimensional vibration signals such as acceleration and displacement can be simultaneously collected in the vertical, lateral, and longitudinal directions, ultimately obtaining 58-dimensional high-dimensional data features.
[0056] In the bogie structure of high-speed trains, air springs (ASs) are primarily responsible for bearing the weight of the car body and providing vertical vibration damping, lateral dampers (LDs) are used to suppress lateral vibrations of the car body and improve the smoothness of train operation, while anti-hunting dampers (ADs) are used to prevent the train from hunting at high speeds, ensuring the safety and comfort of the train. These critical components are prone to failures such as air leaks and oil leaks during long-term service. Figure 4 , Figure 5 and Figure 6 The images show air leakage in AS, oil leakage in LD, and oil leakage in AD. These faults can lead to bogie performance degradation, which in turn can exacerbate train vibration, affect the riding experience, and even threaten operational safety.
[0057] To comprehensively reflect actual operating conditions, the experiment was designed with seven typical states, including normal, lateral damper oil leakage (LD), anti-hunting damper oil leakage (AD), air spring leakage (AS), and multi-component combined failures (such as LD+AD, AS+LD, AS+AD). Each operating condition corresponds to a unique label, as shown in Table 1.
[0058] Table 1 Common Operating Conditions of Bogies
[0059] During data acquisition, the sampling frequency of the vibration signal was set to 243Hz, and the train speed was 200 km / h to ensure the timeliness and representativeness of the data. A large number of time-series vibration signals were collected under each operating condition, fully covering various typical fault scenarios. Specifically, this application selected three representative high-speed railway lines in China: the Wuhan-Guangzhou line, the Zhengzhou-Xi'an line, and the Beijing-Tianjin line. Simulation excitation was performed on each line based on its actual track spectrum, comprehensively reproducing the differences in the operating environment and conditions of the bogies under different lines.
[0060] In terms of data volume design, 5,000 vibration signals were collected for each of the three lines under each operating condition, totaling 35,000 data points per line. Each data point contains 58 feature dimensions. This not only ensures data balance but also provides a solid data foundation for subsequent asynchronous personalized federated learning modeling.
[0061] Using this application for multi-track fault diagnosis of high-speed train bogies allows clients (with data sources from different tracks) to upload only model parameters without sharing raw data, thus avoiding sensitive data leakage. It effectively integrates data uploaded from different clients, improving the generalization ability of the global model. It allows clients to train independently based on their own computing resources and data acquisition progress, while retaining the personalized characteristics of their local models (such as fault modes specific to a particular track).
[0062] This application constructs a highly robust federated learning method for practical engineering scenarios through multiple mechanisms, including dynamic accuracy thresholds, historical performance criteria, and compensation weights. Theoretically, this method guarantees the effectiveness and security of the first global model update. In engineering implementation, it achieves efficient and scalable deployment through global variable management, queue mechanisms, and thread synchronization, significantly improving the intelligent diagnostic capabilities for high-speed train bogies in multi-track data scenarios.
[0063] Based on the same inventive concept, this application also provides an asynchronous personalized federated system for implementing the asynchronous personalized federated learning method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of the one or more asynchronous personalized federated system embodiments provided below can be found in the limitations of the asynchronous personalized federated learning method described above, and will not be repeated here.
[0064] In one exemplary embodiment, such as Figure 7 As shown, an asynchronous personalized federated system is provided, including a central server and multiple clients connected to the central server. The central server includes: The global variable initialization module is used to initialize the global model, take the initialized global model as the first global model, and determine the shared layer parameters and the version of the first global model; the first global model is a neural network including a shared layer and a personalized layer.
[0065] The first global model update module is used to update the shared layer parameters of the first global model based on the shared layer parameters of the client-optimized model, the compensation weights of the client-optimized model, and the shared layer parameters of the first global model, to obtain the updated first global model; the compensation weights of the client-optimized model are determined by the difference between the version of the first global model and the version of the client-optimized model.
[0066] The training termination criterion module is used to send a stop signal to multiple clients based on the convergence of the first global model.
[0067] The client includes: The local model training module is used to download the first global model and train the client's local model based on the first global model using the client's local data to obtain the client-optimized model.
[0068] The parameter acquisition module is used to acquire the shared layer parameters and the version of the client optimization model, and upload them to the central server.
[0069] The training termination module is used to terminate training in the local model training module based on a stop signal.
[0070] As an optional implementation, the parameter acquisition module is also used to acquire the accuracy and version of the client-side optimization model and upload them to the central server.
[0071] The first global model update module includes: a client-side optimized model filtering module, a latency compensation module, and a global model aggregation module.
[0072] The client-side optimization model filtering module is used to determine the historical average accuracy and dynamic accuracy threshold of the client-side optimization model; the dynamic accuracy threshold is determined by the initial threshold, the threshold growth rate, the threshold upper limit, and the number of updates of the first global model.
[0073] The client-side optimization model filtering module is also used to filter client-side optimization models based on their accuracy, historical average accuracy, and dynamic accuracy threshold, to obtain the filtered client-side optimization models.
[0074] The delay compensation module is used to determine the compensation weight of the filtered client-optimized model based on the difference between the version of the first global model and the version of the filtered client-optimized model.
[0075] The global model aggregation module is used to update the shared layer parameters of the first global model based on the shared layer parameters of the filtered client-side optimization model, the compensation weights of the filtered client-side optimization model, and the shared layer parameters of the first global model, so as to obtain the updated first global model.
[0076] As an optional implementation, the client optimization model filtering module includes: a dynamic filtering module and a client update processing module.
[0077] The client update processing module is used to determine the historical average accuracy of the client optimization model, and based on the accuracy of the client optimization model and the historical average accuracy of the client optimization model, to perform an initial screening of the client optimization model and obtain the initial screening client optimization model.
[0078] The dynamic filtering module is used to determine a dynamic accuracy threshold, and based on the dynamic accuracy threshold and the accuracy of the client-side optimized model after the initial screening, to further filter the client-side optimized model after the initial screening, so as to obtain the filtered client-side optimized model.
[0079] As an optional implementation, the client also includes a data acquisition module for acquiring local data from the client using a multi-source simulation platform (such as SIMPACK) and actual sensor deployment.
[0080] In the application environment of high-speed train bogie fault detection, it is necessary to collect multi-dimensional time-series data of high-speed train bogies under different lines and operating conditions to ensure the diversity and representativeness of data distribution. Data acquisition channels are established for each line (such as the Beijing-Tianjin line, Wuhan-Guangzhou line, and Zhengzhou-Xi'an line) to collect signals from multiple types of sensors, including acceleration, displacement, and temperature, and then perform unified data structuring processing.
[0081] Furthermore, the collected data is labeled with operating conditions to form a high-quality, labeled dataset, supporting subsequent personalized modeling and federated training. Data from each line is distributed and stored locally on nodes to ensure data privacy and security, avoiding centralized transmission of raw data.
[0082] As an optional implementation method, such as Figure 7 As shown, the system also includes a main control module; the main control module is used to start the client training thread according to the first manual instruction; the client training thread includes the local model training module and the parameter acquisition module.
[0083] The main control module is also used to start an update processing thread according to a second manual instruction; the update processing thread includes the global variable initialization module, the first global model update module, and the training termination criterion module. The first global model update module includes a client update processing module, a dynamic filtering module, a latency compensation module, and a global model aggregation module. Figure 7 The delay compensation and dynamic filtering module includes a delay compensation module and a dynamic filtering module.
[0084] The main control module is also used to wait for the local model training module to complete training according to a third manual instruction.
[0085] As an optional implementation, the training termination module sets multiple termination criteria globally, comprehensively considering multiple dimensions such as the number of training rounds of the client's local model, the number of aggregation rounds of the first global model, the performance improvement of the first global model, client activity, and the runtime of the federated learning system. Specifically, the federated learning system monitors the training progress of each client in real time through global variables, including the cumulative number of updates, the number of completed local model training rounds, historical aggregation accuracy, and the client's activity status. To ensure that the first global model fully learns from the heterogeneous data distribution, the federated learning system requires each client to complete at least a preset minimum number of local model training rounds. and the minimum number of aggregations in the first global model ,Right now .
[0086] in, Optimize the number of local training rounds completed by the model for the i-th client. To optimize the number of times the i-th client's model participates in the first global model aggregation, Optimize the model set for the client.
[0087] To prevent overtraining by some clients due to data quality issues or network fluctuations, the federated learning system sets a maximum allowed number of updates for each client. The number of aggregations is consistent with that of the first global model; and a maximum training duration is set globally. Once any client reaches Or the global training time exceeds The federated learning system automatically sends a stop signal: .
[0088] in, This represents the cumulative runtime since the Federated Learning System was launched.
[0089] In addition, the federated learning system continuously monitors the aggregate accuracy of the first global model. If the accuracy improvement falls below a threshold for several consecutive rounds, or if no effective client updates are received within a set time window, the training is determined to have converged or stalled, and training is automatically terminated. See the formula for details: .
[0090] in, This indicates the time interval since the last update of the first global model. The maximum allowed idle time, Let g be the aggregate accuracy of the first global model in round g. The aggregate accuracy of the first global model in round g-1. This is the upper limit of the time window. This represents the size of the time window.
[0091] Upon satisfying any termination criterion, the federated learning system broadcasts a stop signal to all clients, ensuring the orderly convergence of all asynchronous training and aggregation processes. To guarantee the robustness and data consistency of the federated learning system, a buffer period is reserved to handle residual client updates in the queue, ultimately completing the synchronization of the first global model and the safe release of resources. The federated learning system also outputs the final states of the first global model and each client's local model, generating detailed training statistics reports to provide a basis for subsequent model deployment and evaluation.
[0092] This application, through the aforementioned engineering process, realizes an efficient, robust, and scalable federated learning system for asynchronous, heterogeneous, personalized, and dynamic environments. Applying this federated learning system to fault diagnosis of high-speed train bogies, the overall accuracy rate of fault diagnosis for bogies on various lines exceeds 92%, significantly enhancing the practical application value of distributed intelligent modeling.
[0093] This application adopts an asynchronous federated learning architecture, supporting multiple clients to asynchronously upload optimized model parameters in parallel. The central server automatically identifies and rejects low-quality or abnormal updates through a dual screening mechanism of dynamic accuracy threshold and historical performance, improving the robustness of the first global model. A meta-learning-driven neural network structure is introduced to achieve adaptive optimization of client models, balancing global sharing and local differences. To address model latency issues in asynchronous environments, a latency compensation aggregation method based on model version differences is proposed. By dynamically adjusting the compensation coefficient, the negative impact of outdated updates on the first global model is suppressed. Furthermore, a dynamic threshold screening and robust aggregation mechanism are designed, combining client historical performance and real-time accuracy to dynamically adjust aggregation criteria, improving the adaptability and security of the aggregation process. The training termination criterion employs multi-condition fusion to ensure efficient and controllable training. The synergistic effect of these technical solutions effectively enhances the personalization capabilities, aggregation robustness, and system convergence efficiency of federated learning in heterogeneous, asynchronous, and dynamic environments, while ensuring the privacy and security of multi-party data. This provides a feasible solution for implementing federated learning among users with imbalanced data in industrial scenarios.
[0094] The beneficial technical effects of this application are as follows: 1. The asynchronous personalized federated learning system proposed in this application supports multi-client concurrency, asynchronous uploading and aggregation, significantly improving system throughput and fault tolerance, and adapting to the asynchronous communication and computing characteristics in actual industrial environments.
[0095] 2. By introducing a delay compensation mechanism based on model version differences, adaptive weighted fusion is achieved to address the model obsolescence problem caused by asynchronous uploads, thereby improving the convergence speed and final performance of the first global model.
[0096] 3. Design a dual screening mechanism of dynamic accuracy threshold and historical performance to automatically reject low-quality updates and improve the generalization ability and robustness of the first global model.
[0097] 4. Supports a model structure that separates the personalized layer and the shared layer. Combined with meta-learning optimization strategies, it enables efficient personalized modeling in multi-client heterogeneous data scenarios, taking into account both local adaptability and global collaboration.
[0098] 5. The entire process is engineered and implemented, supporting multi-threading, exception retries, early stopping, statistical monitoring, etc. It is suitable for actual industrial-grade federated learning deployment needs and has good scalability and practicality.
[0099] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores the shared layer parameters of the first global model and the version of the first global model. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an asynchronous personalized federated learning method.
[0100] Those skilled in the art will understand that Figure 8 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0101] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0102] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. An asynchronous personalized federated learning method, characterized in that, The method includes: Step 1: Initialize the global model. Use the initialized global model as the first global model. Determine the shared layer parameters and version of the first global model. The first global model is a neural network that includes shared layers and personalized layers. Step 2: Using the client's local data, train the client's local model based on the first global model to obtain the client-optimized model, and determine the shared layer parameters and version of the client-optimized model. Step 3: Based on the shared layer parameters of the client-side optimization model, the compensation weights of the client-side optimization model, and the shared layer parameters of the first global model, update the shared layer parameters of the first global model to obtain the updated first global model; the compensation weights of the client-side optimization model are determined by the difference between the version of the first global model and the version of the client-side optimization model. The updated first global model is used as the new first global model. The version of the new first global model is determined, and the process returns to step two until the first global model converges.
2. The asynchronous personalized federated learning method according to claim 1, characterized in that, Using the client's local data and based on the first global model, a local model is trained on the client to obtain a client-optimized model, specifically including: Based on the first global model, using the client's local data, the client's local model is trained using a meta-learning algorithm and a multilayer perceptron to obtain the client-optimized model.
3. The asynchronous personalized federated learning method according to claim 1, characterized in that, Based on the shared layer parameters of the client-side optimization model, the compensation weights of the client-side optimization model, and the shared layer parameters of the first global model, the shared layer parameters of the first global model are updated to obtain the updated first global model, specifically including: Determine the accuracy of the client-side optimization model and its historical average accuracy. Based on the accuracy of the client-side optimization model and its historical average accuracy, the client-side optimization model is initially screened to obtain the initial screened client-side optimization model. Based on the shared layer parameters of the client-side optimization model after initial screening, the compensation weights of the client-side optimization model after initial screening, and the shared layer parameters of the first global model, the shared layer parameters of the first global model are updated to obtain the updated first global model.
4. The asynchronous personalized federated learning method according to claim 3, characterized in that, Based on the shared layer parameters of the initially screened client-side optimization model, the compensation weights of the initially screened client-side optimization model, and the shared layer parameters of the first global model, the shared layer parameters of the first global model are updated to obtain the updated first global model, specifically including: A dynamic accuracy threshold is determined; the dynamic accuracy threshold is determined by an initial threshold, a threshold growth rate, an upper limit of the threshold, and the number of updates to the first global model; Based on the dynamic accuracy threshold and the accuracy of the client-side optimized model after initial screening, the client-side optimized model after initial screening is screened again to obtain the screened client-side optimized model. The compensation weight of the filtered client-optimized model is determined based on the difference between the version of the first global model and the version of the filtered client-optimized model. Based on the shared layer parameters of the filtered client-side optimization model, the compensation weights of the filtered client-side optimization model, and the shared layer parameters of the first global model, the shared layer parameters of the first global model are updated to obtain the updated first global model.
5. The asynchronous personalized federated learning method according to claim 4, characterized in that, The dynamic accuracy threshold is determined using the following formula: ； Where t is the number of times the first global model is updated. This is the dynamic accuracy threshold for the t-th update of the first global model. Initial threshold, The threshold growth rate, This represents the upper limit of the threshold.
6. The asynchronous personalized federated learning method according to claim 4, characterized in that, The following formula is used to determine the filtered first... Compensation weights for each client-side optimization model: ； in, For the filtered first Compensation weights for each client-side optimization model. The sequence number of the optimized client model after screening. As the attenuation base, To compensate for the lower limit of the weight, For the filtered first The difference between the version of the client-side optimized model and the version of the first global model. This is the base value for compensating for the weights.
7. The asynchronous personalized federated learning method according to claim 6, characterized in that, The shared layer parameters of the first global model in the (t+1)th update are determined using the following formula: ； in, These are the shared layer parameters updated for the t-th time in the first global model. These are the shared layer parameters updated in the (t+1)th iteration of the first global model. For the filtered first The client-side optimization model optimizes the shared layer parameters of the first global model during the t-th update, where t is the number of updates to the first global model.
8. An asynchronous personalized federated system, comprising a central server and multiple clients connected to the central server, characterized in that, The central server, applied to the asynchronous personalized federated learning method according to any one of claims 1-7, comprises: The global variable initialization module is used to initialize the global model, take the initialized global model as the first global model, and determine the shared layer parameters and the version of the first global model; the first global model is a neural network including shared layers and personalized layers. The first global model update module is used to update the shared layer parameters of the first global model based on the shared layer parameters of the client-optimized model, the compensation weights of the client-optimized model, and the shared layer parameters of the first global model, so as to obtain the updated first global model; the compensation weights of the client-optimized model are determined by the difference between the version of the first global model and the version of the client-optimized model. The training termination criterion module is used to send a stop signal to multiple clients based on the convergence of the first global model. The client includes: The local model training module is used to download the first global model and train the client's local model based on the first global model using the client's local data to obtain the client-optimized model. The parameter acquisition module is used to acquire the shared layer parameters and the version of the client optimization model, and upload them to the central server. The training termination module is used to terminate training in the local model training module based on a stop signal.
9. The asynchronous personalized federated learning system according to claim 8, characterized in that, The parameter acquisition module is also used to acquire the accuracy and version of the client-side optimized model and upload them to the central server. The first global model update module includes: a client-side optimized model filtering module, a latency compensation module, and a global model aggregation module; The client-side optimization model filtering module is used to determine the historical average accuracy and dynamic accuracy threshold of the client-side optimization model; the dynamic accuracy threshold is determined by the initial threshold, the threshold growth rate, the threshold upper limit, and the number of updates of the first global model; The client-side optimization model filtering module is also used to filter client-side optimization models based on their accuracy, historical average accuracy, and dynamic accuracy threshold, to obtain filtered client-side optimization models. The delay compensation module is used to determine the compensation weight of the filtered client-optimized model based on the difference between the version of the first global model and the version of the filtered client-optimized model. The global model aggregation module is used to update the shared layer parameters of the first global model based on the shared layer parameters of the filtered client-side optimization model, the compensation weights of the filtered client-side optimization model, and the shared layer parameters of the first global model, so as to obtain the updated first global model.
10. The asynchronous personalized federated learning system according to claim 9, characterized in that, The client-side optimization model filtering module includes: a dynamic filtering module and a client-side update processing module; The client update processing module is used to determine the historical average accuracy of the client optimization model, and based on the accuracy of the client optimization model and the historical average accuracy of the client optimization model, to perform an initial screening of the client optimization model and obtain the initial screening client optimization model. The dynamic filtering module is used to determine a dynamic accuracy threshold, and based on the dynamic accuracy threshold and the accuracy of the client-side optimized model after the initial screening, to further filter the client-side optimized model after the initial screening, so as to obtain the filtered client-side optimized model.