Vehicle data federal dynamic sparse training method, device and equipment and storage medium

Abstract

The invention discloses a vehicle data federated dynamic sparse training method, device and equipment and a storage medium, and the method comprises the steps: obtaining a global model loss function of vehicle data of a current vehicle in a federated learning personalized model, carrying out the minimization of the global model loss function, and obtaining a minimum loss function; dynamically adjusting the sparsity of the federated learning personalized model according to the communication overhead, the computing resource use condition and the minimum loss function, and obtaining an adjusted target personalized model; the target client participating in training is selected according to the client selection mechanism, the target personalized model is tested through the target client, and the final personalized model is generated when the test result reaches the standard, so that the communication overhead can be reduced, the personalized model updating is realized, the adaptability of the model on the specific client is improved, and the user experience is improved. According to the method, computing resources are fully utilized, waste of the computing resources is avoided, the model performance is improved, and the speed and efficiency of vehicle data federation dynamic sparse training are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle control technology, and in particular to a method, apparatus, device, and storage medium for federated dynamic sparse training of vehicle data. Background Technology

[0002] With the development of intelligent and connected vehicles, the amount of data generated by vehicles is increasing dramatically. In order to make full use of this data and improve the level of vehicle intelligence, federated learning (FL), as an emerging distributed machine learning method, has gradually attracted attention. Federated learning allows multiple clients (such as vehicles) to jointly train a global model without sharing the original data, thereby improving the performance of the model while protecting data privacy. However, federated learning faces challenges such as data heterogeneity, communication latency, and computing resource limitations in practical applications.

[0003] Currently, research on federated learning mainly focuses on the following aspects: (1) Reduce communication frequency: Reduce the number of communication cycles by performing multiple local computation iterations between two rounds of communication.

[0004] (2) Use compressed information for communication: Use compressed information instead of the complete parameter vector for communication to reduce communication overhead.

[0005] (3) Sparse regularization technique: By dynamically pruning model parameters, the number of parameters and communication overhead are reduced while maintaining the accuracy of the model.

[0006] (4) Personalized model: Personalized model training is performed based on the data characteristics of each client to improve the adaptability and performance of the model on specific clients.

[0007] (5) Client selection strategy: By intelligently selecting clients to participate in training, the model performance is maximized while the communication overhead is minimized.

[0008] However, these methods have the following drawbacks: (1) High communication overhead: Even with reduced communication frequency and the use of compressed information, communication overhead remains high, especially in large-scale networks.

[0009] (2) Waste of computing resources: On some resource-constrained clients, the utilization rate of computing resources is not high, resulting in low overall efficiency.

[0010] (3) Insufficient personalization: Existing personalization model methods have limited effectiveness when dealing with data heterogeneity and non-independent identically distributed data.

[0011] (4) Inaccurate client selection: Existing client selection strategies may lead to a decrease in model performance when dealing with non-independent and identically distributed data.

[0012] Furthermore, existing federated learning frameworks typically employ fixed sparsity strategies and static client selection methods, which perform poorly when dealing with dynamically changing vehicle data. (1) Lack of dynamic adjustment capability: Fixed sparsification strategies cannot adapt to changes in data distribution, resulting in unstable model performance.

[0013] (2) Inflexible client selection: The static client selection method cannot be adjusted according to real-time data changes, resulting in insufficient resource utilization. Summary of the Invention

[0014] The main objective of this invention is to provide a method, apparatus, device, and storage medium for federated dynamic sparse training of vehicle data. This invention aims to address the technical problems in the prior art where federated learning models employ fixed sparsity strategies, lack dynamic adjustment capabilities, perform poorly when processing dynamically changing vehicle data, and use static client selection methods that cannot be adjusted according to real-time data changes, resulting in insufficient resource utilization, high communication overhead, and a significant waste of computing resources, ultimately leading to a decline in model performance.

[0015] In a first aspect, the present invention provides a method for federated dynamic sparse training of vehicle data, the method comprising the following steps: Obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimize the global model loss function to obtain the minimum loss function; The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data. The sparsity of the federated learning personalized model is dynamically adjusted based on the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model. A client selection mechanism based on deep reinforcement learning is obtained. Target clients are selected to participate in training according to the client selection mechanism. The target personalized model is tested through the target clients. When the test results meet the standards, the final personalized model is generated.

[0016] Optionally, the step of obtaining the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimizing the global model loss function to obtain the minimum loss function, includes: Obtain the vehicle data of the current vehicle. In the federated learning personalized model, the vehicle data is used to participate in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model. The minimum loss function of the global model is obtained by minimizing the loss function.

[0017] Optionally, minimizing the global model loss function to obtain the minimum loss function includes: The minimum loss function of the global model is obtained by minimizing the following formula:

[0018] in, For the minimum loss function, The global model loss function. These are global model parameters. The optimization objective is to minimize the global loss. For federal learning communication rounds, Configure local computing resources. For resource selection decisions.

[0019] Optionally, the step of obtaining the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjusting the sparsity of the federated learning personalized model based on the communication overhead, the computing resource usage, and the minimum loss function to obtain the adjusted target personalized model includes: The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data; Obtain a preset optimized auxiliary dynamic mask, and dynamically adjust the model parameters of the federated learning personalized model according to the preset optimized auxiliary dynamic mask at different training stages of the federated learning personalized model; The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters to obtain the adjusted target personalized model.

[0020] Optionally, the real-time adaptation of communication overhead and computing resource usage, based on the minimum loss function, and ensuring the performance of the federated learning personalized model according to the dynamically adjusted model parameters to obtain the adjusted target personalized model, includes: The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters. The adjusted target personalized model is obtained by the following formula:

[0021] in, The global model loss function. These are global model parameters. For federal learning communication rounds, As constraints, The parameter update rules for the local model after the t-th iteration are as follows: For the local model parameters in round t, For learning rate, Let the gradient of the global loss be the local model parameters in round t. To preset and optimize the auxiliary dynamic mask, This will apply to all local training rounds / participants; For the number of participants, Let be a binary decision variable, representing whether to allocate resources to the k-th participant in the t-th round of communication. Let be the resource gain coefficient of the k-th participant for the i-th type of resource. Let be the resource consumption of the k-th participant in round t. The basic overhead for the i-th type of resource of the k-th participant. For the i-th type of global resource total limit, This means it applies to all resource types; For all participants All communication rounds Resource allocation decisions can only be 0 (no allocation) or 1 (allocation).

[0022] Optionally, the step of obtaining a client selection mechanism based on deep reinforcement learning, selecting target clients to participate in training according to the client selection mechanism, testing the target personalized model through the target clients, and generating the final personalized model when the test results meet the standards includes: Obtain a client selection mechanism based on deep reinforcement learning, and select suitable target clients from each client through the optimal selection strategy according to the client selection mechanism; Experimental training of the target personalized model is conducted based on the local data of the target client to obtain experimental results; The closed-loop evaluation determines whether the test results meet the preset performance and efficiency requirements. If the test results meet the preset performance and efficiency requirements, the test results are deemed to be up to standard, and the final personalized model is generated.

[0023] Optionally, the step of obtaining a client selection mechanism based on deep reinforcement learning, and selecting a suitable target client from among the clients using an optimal selection strategy according to the client selection mechanism, includes: Obtain the global model performance of the target personalized model, and obtain the computing power, power consumption, network bandwidth, local data distribution and local data quality of each client; The global model performance, computing power, power consumption, network bandwidth, local data distribution, and local data quality are used as input states for the deep reinforcement learning agent. The weighted result of model training efficiency, performance improvement, and resource consumption is used as a reward signal. The agent continuously iterates and trains with the current environment to learn the optimal selection strategy. The optimal selection strategy is used to select the appropriate target client from among the clients.

[0024] Secondly, to achieve the above objectives, the present invention also proposes a vehicle data federated dynamic sparse training device, the vehicle data federated dynamic sparse training device comprising: The loss function minimization module is used to obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and to minimize the global model loss function to obtain the minimum loss function. A dynamic sparsity adjustment module is used to obtain the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjust the sparsity of the federated learning personalized model according to the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model. The deep reinforcement learning experiment module is used to obtain a client selection mechanism based on deep reinforcement learning, select target clients to participate in training according to the client selection mechanism, conduct experiments on the target personalized model through the target clients, and generate the final personalized model when the experiment results meet the standards.

[0025] Thirdly, to achieve the above objectives, the present invention also proposes a vehicle data federation dynamic sparse training device, the vehicle data federation dynamic sparse training device comprising: a memory, a processor, and a vehicle data federation dynamic sparse training program stored in the memory and executable on the processor, the vehicle data federation dynamic sparse training program being configured to implement the steps of the vehicle data federation dynamic sparse training method as described above.

[0026] Fourthly, to achieve the above objectives, the present invention also proposes a storage medium storing a vehicle data federation dynamic sparse training program, wherein the vehicle data federation dynamic sparse training program, when executed by a processor, implements the steps of the vehicle data federation dynamic sparse training method described above.

[0027] The proposed vehicle data federated dynamic sparse training method involves: obtaining the global model loss function of the current vehicle data in the federated learning personalized model; minimizing the global model loss function to obtain the minimum loss function; obtaining the communication overhead and computing resource usage of the current vehicle from the vehicle data; dynamically adjusting the sparsity of the federated learning personalized model based on the communication overhead, computing resource usage, and the minimum loss function to obtain the adjusted target personalized model; and acquiring a client selection mechanism based on deep reinforcement learning to select target clients for training. The target personalized model is then tested on the target client, and the final personalized model is generated when the test results meet the requirements. This method optimizes communication overhead and computing resource usage by dynamically adjusting the model's sparsity, maintaining model effectiveness, reducing communication overhead, and achieving personalized model updates through fine-tuning strategies. It improves the model's adaptability to specific clients, fully utilizes computing resources, avoids wasting computing resources, improves model performance, and enhances the speed and efficiency of vehicle data federated dynamic sparse training. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the first embodiment of the vehicle data federated dynamic sparse training method of the present invention. Figure 3 This is a flowchart illustrating the second embodiment of the vehicle data federated dynamic sparse training method of the present invention; Figure 4 This is a flowchart illustrating the third embodiment of the vehicle data federated dynamic sparse training method of the present invention. Figure 5 This is a flowchart illustrating the fourth embodiment of the vehicle data federated dynamic sparse training method of the present invention. Figure 6 This is a functional block diagram of the first embodiment of the vehicle data federated dynamic sparse training device of the present invention.

[0029] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0030] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0031] The solution of this invention mainly involves: obtaining the global model loss function of the current vehicle's data in the federated learning personalization model, minimizing the global model loss function to obtain the minimum loss function; obtaining the communication overhead and computing resource usage of the current vehicle from the vehicle data, dynamically adjusting the sparsity of the federated learning personalization model based on the communication overhead, the computing resource usage, and the minimum loss function to obtain the adjusted target personalization model; obtaining a client selection mechanism based on deep reinforcement learning, selecting target clients to participate in training according to the client selection mechanism, testing the target personalization model through the target clients, and generating the final personalization model when the test results meet the standards. This technology optimizes communication overhead and computing resource utilization by dynamically adjusting the sparsity of the model, maintaining model effectiveness and reducing communication overhead. Through fine-tuning strategies, it achieves personalized model updates, improves the model's adaptability to specific clients, fully utilizes computing resources, avoids wasting computing resources, and enhances model performance. It also improves the speed and efficiency of federated dynamic sparse training of vehicle data. This technology solves the technical problems in existing federated learning models that use fixed sparsity strategies, lack dynamic adjustment capabilities, perform poorly when processing dynamically changing vehicle data, and whose static client selection methods cannot be adjusted according to real-time data changes, resulting in insufficient resource utilization, high communication overhead, and a significant waste of computing resources, leading to a decline in model performance.

[0032] Reference Figure 1 , Figure 1 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of the present invention.

[0033] like Figure 1 As shown, the device may include: a processor 1001, such as a CPU; a communication bus 1002; a user interface 1003; a network interface 1004; and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0034] Those skilled in the art will understand that Figure 1The device structure shown does not constitute a limitation on the device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0035] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a network communication module, a user interface module, and a vehicle data federation dynamic sparse training program.

[0036] The device of the present invention calls the vehicle data federated dynamic sparse training program stored in the memory 1005 through the processor 1001 and performs the following operations: Obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimize the global model loss function to obtain the minimum loss function; The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data. The sparsity of the federated learning personalized model is dynamically adjusted based on the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model. A client selection mechanism based on deep reinforcement learning is obtained. Target clients are selected to participate in training according to the client selection mechanism. The target personalized model is tested through the target clients. When the test results meet the standards, the final personalized model is generated.

[0037] The device of the present invention, through processor 1001 calling the vehicle data federated dynamic sparse training program stored in memory 1005, also performs the following operations: Obtain the vehicle data of the current vehicle. In the federated learning personalized model, the vehicle data is used to participate in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model. The minimum loss function of the global model is obtained by minimizing the loss function.

[0038] The device of the present invention, through processor 1001 calling the vehicle data federated dynamic sparse training program stored in memory 1005, also performs the following operations: The minimum loss function of the global model is obtained by minimizing the following formula:

[0039] in, For the minimum loss function, The global model loss function. These are global model parameters. The optimization objective is to minimize the global loss. For federal learning communication rounds, Configure local computing resources. For resource selection decisions.

[0040] The device of the present invention, through processor 1001 calling the vehicle data federated dynamic sparse training program stored in memory 1005, also performs the following operations: The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data; Obtain a preset optimized auxiliary dynamic mask, and dynamically adjust the model parameters of the federated learning personalized model according to the preset optimized auxiliary dynamic mask at different training stages of the federated learning personalized model; The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters to obtain the adjusted target personalized model.

[0041] The device of the present invention, through processor 1001 calling the vehicle data federated dynamic sparse training program stored in memory 1005, also performs the following operations: The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters. The adjusted target personalized model is obtained by the following formula:

[0042] in, The global model loss function. These are global model parameters. For federal learning communication rounds, As constraints, The parameter update rules for the local model after the t-th iteration are as follows: For the local model parameters in round t, For learning rate, Let the gradient of the global loss be the local model parameters in round t. To preset and optimize the auxiliary dynamic mask, This will apply to all local training rounds / participants; For the number of participants, Let be a binary decision variable, representing whether to allocate resources to the k-th participant in the t-th round of communication. Let be the resource gain coefficient of the k-th participant for the i-th type of resource. Let be the resource consumption of the k-th participant in round t. The basic overhead for the i-th type of resource of the k-th participant. For the i-th type of global resource total limit, This means it applies to all resource types; For all participants All communication rounds Resource allocation decisions can only be 0 (no allocation) or 1 (allocation).

[0043] The device of the present invention, through processor 1001 calling the vehicle data federated dynamic sparse training program stored in memory 1005, also performs the following operations: Obtain a client selection mechanism based on deep reinforcement learning, and select suitable target clients from each client through the optimal selection strategy according to the client selection mechanism; Experimental training of the target personalized model is conducted based on the local data of the target client to obtain experimental results; The closed-loop evaluation determines whether the test results meet the preset performance and efficiency requirements. If the test results meet the preset performance and efficiency requirements, the test results are deemed to be up to standard, and the final personalized model is generated.

[0044] The device of the present invention, through processor 1001 calling the vehicle data federated dynamic sparse training program stored in memory 1005, also performs the following operations: Obtain the global model performance of the target personalized model, and obtain the computing power, power consumption, network bandwidth, local data distribution and local data quality of each client; The global model performance, computing power, power consumption, network bandwidth, local data distribution, and local data quality are used as input states for the deep reinforcement learning agent. The weighted result of model training efficiency, performance improvement, and resource consumption is used as a reward signal. The agent continuously iterates and trains with the current environment to learn the optimal selection strategy. The optimal selection strategy is used to select the appropriate target client from among the clients.

[0045] This embodiment, through the above-described scheme, obtains the global model loss function of the current vehicle data in the federated learning personalized model, minimizes the global model loss function to obtain the minimum loss function; obtains the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjusts the sparsity of the federated learning personalized model based on the communication overhead, computing resource usage, and the minimum loss function to obtain the adjusted target personalized model; obtains a client selection mechanism based on deep reinforcement learning, selects target clients to participate in training according to the client selection mechanism, and conducts experiments on the target personalized model through the target clients. When the experimental results meet the standards, the final personalized model is generated. By dynamically adjusting the sparsity of the model, the communication overhead and computing resource usage are optimized, the model's effectiveness is maintained, communication overhead is reduced, and personalized model updates are achieved through fine-tuning strategies. This improves the model's adaptability on specific clients, makes full use of computing resources, avoids wasting computing resources, improves model performance, and increases the speed and efficiency of federated dynamic sparse training of vehicle data.

[0046] Based on the above hardware structure, an embodiment of the vehicle data federated dynamic sparse training method of the present invention is proposed.

[0047] Reference Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the vehicle data federated dynamic sparse training method of the present invention.

[0048] In the first embodiment, the vehicle data federated dynamic sparse training method includes the following steps: Step S10: Obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimize the global model loss function to obtain the minimum loss function.

[0049] It should be noted that by defining a global model loss function for the current vehicle's data in the federated learning personalization model, and by minimizing the global model loss function to obtain the minimum loss function, the performance of the federated learning personalization model can be optimized.

[0050] Step S20: Obtain the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjust the sparsity of the federated learning personalized model according to the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model.

[0051] It should be understood that the communication overhead and computing resource usage of the current vehicle can be obtained from the vehicle data. The sparsity of the federated learning personalized model can be dynamically adjusted based on the communication overhead, the computing resource usage, and the minimum loss function, that is, the communication overhead and computing resource usage can be optimized to obtain the adjusted target personalized model.

[0052] Step S30: Obtain the client selection mechanism based on deep reinforcement learning, select the target client to participate in the training according to the client selection mechanism, conduct experiments on the target personalized model through the target client, and generate the final personalized model when the experiment results meet the standards.

[0053] Understandably, the client selection mechanism based on deep reinforcement learning can intelligently select clients to participate in training, and then conduct experiments on the target personalized model through the target client. When the experimental results meet the preset standards, the final personalized model is generated.

[0054] This embodiment, through the above-described scheme, obtains the global model loss function of the current vehicle data in the federated learning personalized model, minimizes the global model loss function to obtain the minimum loss function; obtains the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjusts the sparsity of the federated learning personalized model based on the communication overhead, computing resource usage, and the minimum loss function to obtain the adjusted target personalized model; obtains a client selection mechanism based on deep reinforcement learning, selects target clients to participate in training according to the client selection mechanism, and conducts experiments on the target personalized model through the target clients. When the experimental results meet the standards, the final personalized model is generated. By dynamically adjusting the sparsity of the model, the communication overhead and computing resource usage are optimized, the model's effectiveness is maintained, communication overhead is reduced, and personalized model updates are achieved through fine-tuning strategies. This improves the model's adaptability on specific clients, makes full use of computing resources, avoids wasting computing resources, improves model performance, and increases the speed and efficiency of federated dynamic sparse training of vehicle data.

[0055] Furthermore, Figure 3 This is a flowchart illustrating the second embodiment of the vehicle data federated dynamic sparse training method of the present invention, as shown below. Figure 3 As shown, based on the first embodiment, a second embodiment of the vehicle data federated dynamic sparse training method of the present invention is proposed. In this embodiment, step S10 specifically includes the following steps: Step S11: Obtain the vehicle data of the current vehicle. In the federated learning personalized model, the vehicle data is used to participate in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model.

[0056] It should be noted that after obtaining the vehicle data of the current vehicle, in the federated learning personalized model, the vehicle data of the current vehicle can participate in the global model aggregation through local training, thereby affecting the calculation of the global model loss function, that is, obtaining the global model loss function of the federated learning personalized model.

[0057] Step S12: Minimize the global model loss function to obtain the minimum loss function.

[0058] It is understandable that minimizing the global model loss function yields the minimum loss function.

[0059] Furthermore, step S12 specifically includes the following steps: The minimum loss function of the global model is obtained by minimizing the following formula:

[0060] in, For the minimum loss function, The global model loss function. These are global model parameters. The optimization objective is to minimize the global loss. For federal learning communication rounds, Configure local computing resources. For resource selection decisions.

[0061] It should be understood that by minimizing the global model loss function This can optimize the performance of personalized federated learning models.

[0062] This embodiment, through the above scheme, obtains the vehicle data of the current vehicle, and in the federated learning personalized model, the vehicle data participates in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model; minimizing the global model loss function to obtain the minimum loss function can optimize the performance of the federated learning personalized model and improve the speed and efficiency of federated dynamic sparse training of vehicle data.

[0063] Furthermore, Figure 4 This is a flowchart illustrating the third embodiment of the vehicle data federated dynamic sparse training method of the present invention, as shown below. Figure 4 As shown, based on the first embodiment, a third embodiment of the vehicle data federated dynamic sparse training method of the present invention is proposed. In this embodiment, step S20 specifically includes the following steps: Step S21: Obtain the current vehicle's communication overhead and computing resource usage from the vehicle data.

[0064] It should be noted that communication overhead refers to the network resources consumed during data transmission (such as bandwidth, traffic, latency, etc.), while computing resource usage refers to the device's usage status of computing power (such as the utilization rate of the central processing unit (CPU) / graphics processing unit (GPU)), memory, storage, and other resources in tasks such as data processing and model training. The communication overhead and computing resource usage of the current vehicle can be obtained from the vehicle data.

[0065] Step S22: Obtain a preset optimized auxiliary dynamic mask, and dynamically adjust the model parameters of the federated learning personalized model according to the preset optimized auxiliary dynamic mask at different training stages of the federated learning personalized model.

[0066] It is understandable that the model parameters are updated through a dynamic mask. This dynamic mask can selectively prune and reparameterize the model parameters at different training stages. After obtaining the preset optimization auxiliary dynamic mask, the model parameters of the federated learning personalized model can be dynamically adjusted according to the preset optimization auxiliary dynamic mask at different training stages of the federated learning personalized model, thereby reducing computational and communication overhead while maintaining the effectiveness of the model.

[0067] Step S23: Adapt the communication overhead and computing resource usage in real time, and based on the minimum loss function, ensure the performance of the federated learning personalized model according to the dynamically adjusted model parameters to obtain the adjusted target personalized model.

[0068] It should be understood that by adapting the communication overhead and computing resource usage in real time, the performance of the federated learning personalized model can be guaranteed based on the minimum loss function and the dynamically adjusted model parameters, thereby obtaining the adjusted target personalized model.

[0069] Furthermore, step S23 specifically includes the following steps: The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters. The adjusted target personalized model is obtained by the following formula:

[0070] in, The global model loss function. These are global model parameters. For federal learning communication rounds, As constraints, The parameter update rules for the local model after the t-th iteration are as follows: For the local model parameters in round t, For learning rate, Let the gradient of the global loss be the local model parameters in round t. To preset and optimize the auxiliary dynamic mask, This will apply to all local training rounds / participants; For the number of participants, Let be a binary decision variable, representing whether to allocate resources to the k-th participant in the t-th round of communication. Let be the resource gain coefficient of the k-th participant for the i-th type of resource. Let be the resource consumption of the k-th participant in round t. The basic overhead for the i-th type of resource of the k-th participant. For the i-th type of global resource total limit, This means it applies to all resource types; For all participants All communication rounds Resource allocation decisions can only be 0 (no allocation) or 1 (allocation).

[0071] It should be noted that the federated learning framework in this embodiment combines dynamic sparse training with a precise client selection strategy, enabling the targeted training of personalized models in a federated learning environment. This design not only considers the performance of the global model but also takes into account the specificity of each client, thereby improving the model's adaptability and efficiency on specific clients. In automotive applications, this means that more accurate driver assistance systems or autonomous driving models can be trained for different vehicle models, driving habits, and road conditions.

[0072] In practice, to meet the needs of different clients, each client performs personalized optimization on its local data to generate the optimal local model parameters. This embodiment optimizes communication overhead and computational resource utilization to suit the characteristics of federated learning training. By dynamically adjusting the sparsity of the model, it can maintain model effectiveness while reducing the number of parameters. Model parameter updates are achieved through a dynamic mask M, which selectively prunes and reparameterizes model parameters at different training stages, thereby reducing computational and communication overhead while maintaining model effectiveness. This dynamic adjustment mechanism allows the system to flexibly optimize the model according to the specific needs of the client. Simultaneously, a fine-tuning strategy is introduced to adapt to the data characteristics of individual clients, further enabling personalized model updates within the federated learning framework. In automotive scenarios, this method can significantly reduce the burden on onboard computing units, extend battery life, and improve the efficiency of real-time data processing.

[0073] It should be understood that in the training and deployment of federated learning personalized models, the sparsity of the model (i.e., the proportion of zero or near-zero parameters in the model parameters) is a key regulatory variable for balancing performance and resource consumption. To achieve efficient and accurate model optimization, this sparsity can be dynamically adjusted based on real-time monitoring of communication overhead, computing resource usage, and the model's minimum loss function: when the communication bandwidth between the terminal device and the server is limited and the transmission latency is high (high communication overhead), the model sparsity can be increased to reduce the amount of parameter transmission and reduce communication pressure; when the terminal device has insufficient computing power, excessive memory usage, or limited energy consumption (tight computing resources), the model structure can be simplified by increasing sparsity, reducing the computational load and resource consumption of local training.

[0074] Understandably, using the minimum loss function of the model as a constraint can prevent excessive sparsity from causing a significant drop in model performance. If the loss function value increases significantly after increasing sparsity, then the sparsity needs to be appropriately reduced to ensure the model's prediction accuracy. Conversely, if the loss function is stable within an acceptable range, then the sparsity can be maintained or further increased to optimize resource utilization. This dynamic adjustment mechanism can achieve a dynamic balance between resource constraints and model performance, enabling federated learning personalized models to meet the needs of efficient deployment in complex real-world scenarios while maintaining good personalized learning effects.

[0075] This embodiment, through the above-described scheme, obtains the current vehicle's communication overhead and computing resource usage from the vehicle data; obtains a preset optimized auxiliary dynamic mask; dynamically adjusts the model parameters of the federated learning personalized model at different training stages based on the preset optimized auxiliary dynamic mask; adapts to the communication overhead and computing resource usage in real time; uses the minimum loss function as a benchmark; and ensures the performance of the federated learning personalized model based on the dynamically adjusted model parameters to obtain the adjusted target personalized model. By dynamically adjusting the sparsity of the model, it optimizes the use of communication overhead and computing resources, maintains the model's effectiveness, reduces communication overhead, and achieves personalized model updates through fine-tuning strategies, thereby improving the model's adaptability on specific clients.

[0076] Furthermore, Figure 5 This is a flowchart illustrating the fourth embodiment of the vehicle data federated dynamic sparse training method of the present invention, as shown below. Figure 5 As shown, based on the first embodiment, a fourth embodiment of the vehicle data federated dynamic sparse training method of the present invention is proposed. In this embodiment, step S30 specifically includes the following steps: Step S31: Obtain the client selection mechanism based on deep reinforcement learning, and select the appropriate target client from each client through the optimal selection strategy according to the client selection mechanism.

[0077] It should be noted that the client selection mechanism based on deep reinforcement learning can intelligently select clients to participate in training, thereby maximizing model performance. That is, the optimal selection strategy is used to select suitable target clients from among the clients according to the client selection mechanism.

[0078] Furthermore, step S31 specifically includes the following steps: Obtain the global model performance of the target personalized model, and obtain the computing power, power consumption, network bandwidth, local data distribution and local data quality of each client; The global model performance, computing power, power consumption, network bandwidth, local data distribution, and local data quality are used as input states for the deep reinforcement learning agent. The weighted result of model training efficiency, performance improvement, and resource consumption is used as a reward signal. The agent continuously iterates and trains with the current environment to learn the optimal selection strategy. The optimal selection strategy is used to select the appropriate target client from among the clients.

[0079] Understandably, a client selection mechanism based on deep reinforcement learning can maximize model performance and minimize communication overhead by intelligently selecting clients to participate in training. It typically considers various resource constraints, including time, computational resources, and network bandwidth, while ensuring that each client's resource consumption in each training round does not exceed its available resource limit. This method in this embodiment is particularly suitable for handling non-independent and identically distributed data problems, effectively mitigating the negative impact of client data distribution differences and preventing client drift.

[0080] In practical implementation, the real-time state of the federated learning system (e.g., computing power, power consumption, network bandwidth, local data distribution and quality, current global model performance, etc.) can be used as the input state of the deep reinforcement learning (DRL) agent. The action space can be "selecting a subset of clients", and the reward signal can be the weighted result of "model training efficiency (e.g., convergence speed), performance improvement (e.g., accuracy, generalization ability), and resource consumption (e.g., communication cost, computational load)". The optimal selection strategy can be learned through continuous interaction with the environment (iterative training).

[0081] It should be understood that the target clients selected based on this mechanism must meet two conditions: on the one hand, their local data must be representative (e.g., covering multiple scene distributions) and of high quality (e.g., high label accuracy) to effectively support the training of personalized models; on the other hand, their resource status must be adapted to training needs (e.g., sufficient computing power and stable network) to reduce training latency and failure risks.

[0082] Step S32: Conduct experimental training of the target personalized model based on the local data of the target client, and obtain experimental results.

[0083] It is understandable that experimental training of the target personalized model can be carried out based on the local data of the target client, thereby obtaining the corresponding experimental results.

[0084] In the specific implementation, the target client can conduct experimental training on the federated learning personalized model (e.g., a client-specific sub-model fine-tuned based on the global model, or a hybrid model containing a personalized layer) based on its own local data: the client loads the initial personalized model parameters, performs multiple rounds of iterative optimization using local data, generates experimental results including local gradients, parameter updates, or performance feedback (such as loss values ​​and accuracy), and uploads them to the server.

[0085] Step S33: Closed-loop evaluation of whether the test results meet the preset performance requirements and preset efficiency requirements. When the test results meet the preset performance requirements and preset efficiency requirements, the test results are deemed to be up to standard, and the final personalized model is generated.

[0086] It should be understood that by evaluating the test results in a closed loop, it can be determined whether the current test results meet the preset performance requirements and preset efficiency requirements. When the test results meet the preset performance requirements and preset efficiency requirements, the test results are deemed to be up to standard, thereby generating the final personalized model.

[0087] In practice, the server can aggregate experimental data from various target clients and evaluate the experimental results using preset metrics (such as the model's accuracy achievement rate on local client tasks, loss function convergence threshold, and personalization fit score). If the experimental results meet the preset standards (such as core metrics exceeding thresholds and stable fluctuations), a final personalized model is generated based on the current experimental data (such as fixing optimal parameters for each client or forming a dynamically adaptable personalized strategy). If the standards are not met, the server sends experimental feedback (such as performance shortcomings or resource bottlenecks) back to the DRL mechanism, triggering the next round of client selection strategy adjustments, repeating the "selection-experimentation-evaluation" process until a final personalized model that meets the requirements is generated.

[0088] It should be noted that a series of experiments verified the effectiveness of the framework. The experimental results show that the framework has achieved significant results in improving communication efficiency, saving computing resources, and accelerating model convergence. For example, the average communication overhead was reduced by about 6.71%, and the average computing overhead was reduced by about 24.04%. Within a time span of 6000 milliseconds, the client was selected by the server to participate in training less than 6 times on average, effectively reducing resource overhead. In resource-constrained environments, the framework demonstrated good practicality and scalability, showing its great potential in practical applications. At the same time, the effectiveness of the framework in handling the problems of non-independent and identically distributed data and client heterogeneity was also verified, which further improved the model performance and convergence speed in federated learning scenarios.

[0089] This embodiment, through the above-described scheme, obtains a client selection mechanism based on deep reinforcement learning, and selects suitable target clients from various clients using an optimal selection strategy according to the client selection mechanism; conducts experimental training of the target personalized model based on the local data of the target client, and obtains experimental results; performs closed-loop evaluation on whether the experimental results meet preset performance and efficiency requirements. When the experimental results meet the preset performance and efficiency requirements, the experimental results are deemed satisfactory, and the final personalized model is generated. By dynamically adjusting the sparsity of the model, communication overhead and the use of computing resources are optimized, the effectiveness of the model is maintained, communication overhead is reduced, personalized model updates are achieved through fine-tuning strategies, the adaptability of the model on specific clients is improved, computing resources are fully utilized, computing resource waste is avoided, model performance is improved, and the speed and efficiency of vehicle data federated dynamic sparse training are increased.

[0090] Accordingly, the present invention further provides a vehicle data federated dynamic sparse training device.

[0091] Reference Figure 6 , Figure 6 This is a functional block diagram of the first embodiment of the vehicle data federated dynamic sparse training device of the present invention.

[0092] In a first embodiment of the vehicle data federation dynamic sparse training device of the present invention, the vehicle data federation dynamic sparse training device includes: The loss function minimization module 10 is used to obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and to minimize the global model loss function to obtain the minimum loss function.

[0093] The dynamic sparsity adjustment module 20 is used to obtain the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjust the sparsity of the federated learning personalized model according to the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model.

[0094] The deep reinforcement learning experiment module 30 is used to obtain a client selection mechanism based on deep reinforcement learning, select target clients to participate in training according to the client selection mechanism, conduct experiments on the target personalized model through the target clients, and generate the final personalized model when the experiment results meet the standards.

[0095] The loss function minimization module 10 is also used to acquire the vehicle data of the current vehicle, and in the federated learning personalized model, the vehicle data participates in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model; and the global model loss function is minimized to obtain the minimum loss function.

[0096] The loss function minimization module 10 is further configured to minimize the global model loss function, obtaining the minimum loss function using the following formula:

[0097] in, For the minimum loss function, The global model loss function. These are global model parameters. The optimization objective is to minimize the global loss. For federal learning communication rounds, Configure local computing resources. For resource selection decisions.

[0098] The dynamic sparsity adjustment module 20 is further configured to obtain the communication overhead and computing resource usage of the current vehicle from the vehicle data; obtain a preset optimized auxiliary dynamic mask; dynamically adjust the model parameters of the federated learning personalized model according to the preset optimized auxiliary dynamic mask at different training stages of the federated learning personalized model; adapt the communication overhead and computing resource usage in real time; and, based on the minimum loss function, ensure the performance of the federated learning personalized model according to the dynamically adjusted model parameters to obtain the adjusted target personalized model.

[0099] The dynamic sparsity adjustment module 20 is also used to adapt the communication overhead and computing resource usage in real time, and based on the minimum loss function, ensure the performance of the federated learning personalized model according to the dynamically adjusted model parameters, and obtain the adjusted target personalized model through the following formula:

[0100] in, The global model loss function. These are global model parameters. For federal learning communication rounds, As constraints, The parameter update rules for the local model after the t-th iteration are as follows: For the local model parameters in round t, For learning rate, Let the gradient of the global loss be the local model parameters in round t. To preset and optimize the auxiliary dynamic mask, This will apply to all local training rounds / participants; For the number of participants, Let be a binary decision variable, representing whether to allocate resources to the k-th participant in the t-th round of communication. Let be the resource gain coefficient of the k-th participant for the i-th type of resource. Let be the resource consumption of the k-th participant in round t. The basic overhead for the i-th type of resource of the k-th participant. For the i-th type of global resource total limit, This means it applies to all resource types; For all participants All communication rounds Resource allocation decisions can only be 0 (no allocation) or 1 (allocation).

[0101] The deep reinforcement learning-based experimental module 30 is further configured to acquire a client selection mechanism based on deep reinforcement learning, select a suitable target client from among the clients using the optimal selection strategy according to the client selection mechanism, conduct experimental training of the target personalized model based on the local data of the target client, and obtain experimental results; conduct closed-loop evaluation of whether the experimental results meet the preset performance requirements and preset efficiency requirements, and determine that the experimental results meet the standards when the experimental results meet the preset performance requirements and preset efficiency requirements, and generate the final personalized model.

[0102] The deep reinforcement learning experiment module 30 is further used to obtain the global model performance of the target personalized model, and to obtain the computing power, power consumption, network bandwidth, local data distribution, and local data quality of each client; the global model performance, computing power, power consumption, network bandwidth, local data distribution, and local data quality are used as the input state of the deep reinforcement learning agent, and the weighted result of model training efficiency, performance improvement, and resource consumption is used as the reward signal to continuously iterate and train with the current environment to learn the optimal selection strategy; and the appropriate target client is selected from each client according to the optimal selection strategy.

[0103] The steps for implementing each functional module of the vehicle data federated dynamic sparse training device can be referred to in the various embodiments of the vehicle data federated dynamic sparse training method of the present invention, and will not be repeated here.

[0104] Furthermore, this embodiment of the invention also proposes a storage medium storing a vehicle data federated dynamic sparse training program, which, when executed by a processor, performs the following operations: Obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimize the global model loss function to obtain the minimum loss function; The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data. The sparsity of the federated learning personalized model is dynamically adjusted based on the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model. A client selection mechanism based on deep reinforcement learning is obtained. Target clients are selected to participate in training according to the client selection mechanism. The target personalized model is tested through the target clients. When the test results meet the standards, the final personalized model is generated.

[0105] Furthermore, when the vehicle data federated dynamic sparse training program is executed by the processor, it also performs the following operations: Obtain the vehicle data of the current vehicle. In the federated learning personalized model, the vehicle data is used to participate in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model. The minimum loss function of the global model is obtained by minimizing the loss function.

[0106] Furthermore, when the vehicle data federated dynamic sparse training program is executed by the processor, it also performs the following operations: The minimum loss function of the global model is obtained by minimizing the following formula:

[0107] in, For the minimum loss function, The global model loss function. These are global model parameters. The optimization objective is to minimize the global loss. For federal learning communication rounds, Configure local computing resources. For resource selection decisions.

[0108] Furthermore, when the vehicle data federated dynamic sparse training program is executed by the processor, it also performs the following operations: The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data; Obtain a preset optimized auxiliary dynamic mask, and dynamically adjust the model parameters of the federated learning personalized model according to the preset optimized auxiliary dynamic mask at different training stages of the federated learning personalized model; The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters to obtain the adjusted target personalized model.

[0109] Furthermore, when the vehicle data federated dynamic sparse training program is executed by the processor, it also performs the following operations: The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters. The adjusted target personalized model is obtained by the following formula:

[0110] in, The global model loss function. These are global model parameters. For federal learning communication rounds, As constraints, The parameter update rules for the local model after the t-th iteration are as follows: For the local model parameters in round t, For learning rate, Let the gradient of the global loss be the local model parameters in round t. To preset and optimize the auxiliary dynamic mask, This will apply to all local training rounds / participants; For the number of participants, Let be a binary decision variable, representing whether to allocate resources to the k-th participant in the t-th round of communication. Let be the resource gain coefficient of the k-th participant for the i-th type of resource. Let be the resource consumption of the k-th participant in round t. The basic overhead for the i-th type of resource of the k-th participant. For the i-th type of global resource total limit, This means it applies to all resource types; For all participants All communication rounds Resource allocation decisions can only be 0 (no allocation) or 1 (allocation).

[0111] Furthermore, when the vehicle data federated dynamic sparse training program is executed by the processor, it also performs the following operations: Obtain a client selection mechanism based on deep reinforcement learning, and select suitable target clients from each client through the optimal selection strategy according to the client selection mechanism; Experimental training of the target personalized model is conducted based on the local data of the target client to obtain experimental results; The closed-loop evaluation determines whether the test results meet the preset performance and efficiency requirements. If the test results meet the preset performance and efficiency requirements, the test results are deemed to be up to standard, and the final personalized model is generated.

[0112] Furthermore, when the vehicle data federated dynamic sparse training program is executed by the processor, it also performs the following operations: Obtain the global model performance of the target personalized model, and obtain the computing power, power consumption, network bandwidth, local data distribution and local data quality of each client; The global model performance, computing power, power consumption, network bandwidth, local data distribution, and local data quality are used as input states for the deep reinforcement learning agent. The weighted result of model training efficiency, performance improvement, and resource consumption is used as a reward signal. The agent continuously iterates and trains with the current environment to learn the optimal selection strategy. The optimal selection strategy is used to select the appropriate target client from among the clients.

[0113] Those skilled in the art will understand that all or part of the steps in the methods described above can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium is a computer-readable storage medium, including: USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and other media that can store program code.

[0114] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0115] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0116] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for federated dynamic sparse training of vehicle data, characterized in that, The vehicle data federated dynamic sparse training method includes: Obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimize the global model loss function to obtain the minimum loss function; The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data. The sparsity of the federated learning personalized model is dynamically adjusted based on the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model. A client selection mechanism based on deep reinforcement learning is obtained. Target clients are selected to participate in training according to the client selection mechanism. The target personalized model is tested through the target clients. When the test results meet the standards, the final personalized model is generated.

2. The vehicle data federated dynamic sparse training method as described in claim 1, characterized in that, The step of obtaining the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and minimizing the global model loss function to obtain the minimum loss function, includes: Obtain the vehicle data of the current vehicle. In the federated learning personalized model, the vehicle data is used to participate in the global model aggregation through local training to obtain the global model loss function of the federated learning personalized model. The minimum loss function is obtained by minimizing the global model loss function.

3. The vehicle data federated dynamic sparse training method as described in claim 2, characterized in that, Minimizing the global model loss function to obtain the minimum loss function includes: The minimum loss function of the global model is obtained by minimizing the following formula: in, For the minimum loss function, The global model loss function. These are global model parameters. The optimization objective is to minimize the global loss. For federal learning communication rounds, Configure local computing resources. For resource selection decisions.

4. The vehicle data federated dynamic sparse training method as described in claim 1, characterized in that, The step of obtaining the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjusting the sparsity of the federated learning personalized model based on the communication overhead, the computing resource usage, and the minimum loss function to obtain the adjusted target personalized model includes: The communication overhead and computing resource usage of the current vehicle are obtained from the vehicle data; Obtain a preset optimized auxiliary dynamic mask, and dynamically adjust the model parameters of the federated learning personalized model according to the preset optimized auxiliary dynamic mask at different training stages of the federated learning personalized model; The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters to obtain the adjusted target personalized model.

5. The vehicle data federated dynamic sparse training method as described in claim 4, characterized in that, The real-time adaptation of communication overhead and computing resource usage, based on the minimum loss function, and ensuring the performance of the federated learning personalized model according to the dynamically adjusted model parameters, to obtain the adjusted target personalized model includes: The communication overhead and computing resource usage are adapted in real time. Based on the minimum loss function, the performance of the federated learning personalized model is guaranteed according to the dynamically adjusted model parameters. The adjusted target personalized model is obtained by the following formula: in, The global model loss function is... These are global model parameters. For federal learning communication rounds, As constraints, The parameter update rules for the local model after the t-th iteration are as follows: For the local model parameters in round t, For learning rate, Let the gradient of the global loss be the local model parameters in round t. To pre-optimize the auxiliary dynamic mask, This will apply to all local training rounds / participants; For the number of participants, Let be a binary decision variable, representing whether to allocate resources to the k-th participant in the t-th round of communication. Let be the resource gain coefficient of the k-th participant for the i-th type of resource. Let be the resource consumption of the k-th participant in round t. The basic overhead for the i-th type of resource of the k-th participant. For the i-th type of global resource total limit, This means it applies to all resource types; For all participants All communication rounds Resource allocation decisions can only be 0 (no allocation) or 1 (allocation).

6. The vehicle data federated dynamic sparse training method as described in claim 1, characterized in that, The process of acquiring a client selection mechanism based on deep reinforcement learning, selecting target clients for training according to the client selection mechanism, testing the target personalized model through the target clients, and generating the final personalized model when the test results meet the standards includes: Obtain a client selection mechanism based on deep reinforcement learning, and select suitable target clients from each client through the optimal selection strategy according to the client selection mechanism; Experimental training of the target personalized model is conducted based on the local data of the target client to obtain experimental results; The closed-loop evaluation determines whether the test results meet the preset performance and efficiency requirements. If the test results meet the preset performance and efficiency requirements, the test results are deemed to be up to standard, and the final personalized model is generated.

7. The vehicle data federated dynamic sparse training method as described in claim 6, characterized in that, The acquisition of a client selection mechanism based on deep reinforcement learning, and the selection of suitable target clients from among the clients using an optimal selection strategy according to the client selection mechanism, includes: Obtain the global model performance of the target personalized model, and obtain the computing power, power consumption, network bandwidth, local data distribution and local data quality of each client; The global model performance, computing power, power consumption, network bandwidth, local data distribution, and local data quality are used as input states for the deep reinforcement learning agent. The weighted result of model training efficiency, performance improvement, and resource consumption is used as a reward signal. The agent continuously iterates and trains with the current environment to learn the optimal selection strategy. The optimal selection strategy is used to select the appropriate target client from among the clients.

8. A vehicle data federated dynamic sparse training device, characterized in that, The vehicle data federated dynamic sparse training device includes: The loss function minimization module is used to obtain the global model loss function of the current vehicle's vehicle data in the federated learning personalized model, and to minimize the global model loss function to obtain the minimum loss function. A dynamic sparsity adjustment module is used to obtain the communication overhead and computing resource usage of the current vehicle from the vehicle data, and dynamically adjust the sparsity of the federated learning personalized model according to the communication overhead, the computing resource usage and the minimum loss function to obtain the adjusted target personalized model. The deep reinforcement learning experiment module is used to obtain a client selection mechanism based on deep reinforcement learning, select target clients to participate in training according to the client selection mechanism, conduct experiments on the target personalized model through the target clients, and generate the final personalized model when the experiment results meet the standards.

9. A vehicle data federated dynamic sparse training device, characterized in that, The vehicle data federated dynamic sparse training device includes: a memory, a processor, and a vehicle data federated dynamic sparse training program stored in the memory and executable on the processor, wherein the vehicle data federated dynamic sparse training program is configured to implement the steps of the vehicle data federated dynamic sparse training method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a vehicle data federation dynamic sparse training program, which, when executed by a processor, implements the steps of the vehicle data federation dynamic sparse training method as described in any one of claims 1 to 7.

Images