Model training method, state estimation method and device, power utilization equipment group, system and equipment

Abstract

The application provides a model training method, a state estimation method and device, a power consumption equipment group, a system and equipment. The model training method comprises: each power consumption equipment trains a local battery state estimation model by using a local training data set, obtains local model parameters, and sends the local model parameters to an aggregation node; the aggregation node is selected from the power consumption equipment; the aggregation node weights and fuses each local model parameter based on a parameter weight of each power consumption equipment, obtains fusion model parameters, and distributes the fusion model parameters to each power consumption equipment; the parameter weight of each power consumption equipment is determined based on the credibility corresponding to the corresponding local model parameter; and each power consumption equipment updates the local battery state estimation model by using the fusion model parameters. In this way, the influence of low-quality models can be effectively inhibited, the problems of insufficient data quantity of a single node and poor generalization ability can be improved, the robustness and accuracy of the overall model can be improved, and the user privacy of the power consumption equipment can be protected.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a model training method, a state estimation method and apparatus, a group of electrical equipment, a system and equipment. Background Technology

[0002] With the rapid development of new energy electric vehicles, power batteries are widely used due to their advantages such as high energy density and rechargeability. Battery state estimation is one of the key functions in battery management systems, used to assess the battery's state of charge (SOC) and state of health (SOH) to improve user experience and alleviate range anxiety.

[0003] In related technologies, battery state estimation models are optimized by collecting battery operating data and performing feature extraction and dimensionality reduction to achieve battery state estimation. This method relies on a centralized data processing and model training process.

[0004] However, this centralized learning approach has many problems in practical applications, such as a high risk of user privacy leakage, susceptibility to single-node failures, and low tolerance for low-quality data. On the other hand, training using only local data from a single node faces problems such as insufficient data volume and poor generalization ability, making it difficult to meet the accurate modeling requirements in complex scenarios. Summary of the Invention

[0005] This application provides a model training method, a state estimation method and apparatus, a group of electrical devices, a battery management system, a battery system and electrical devices, which can effectively suppress the influence of low-quality models, improve problems such as insufficient data volume of single nodes and poor generalization ability, thereby improving the robustness and accuracy of the overall model, and protecting the user privacy of electrical devices.

[0006] The technical solution of this application embodiment is implemented as follows: This application provides a model training method applied to a group of electrical equipment, which includes multiple electrical devices. The method includes: Each electrical device uses its local training dataset to train its local battery state estimation model, and then sends the local model parameters to the aggregation node. The local training dataset includes at least one set of running data for the device's local battery and the battery state label corresponding to each set of running data. The aggregation node is selected from each electrical device. The aggregation node performs weighted fusion of local model parameters based on the parameter weights of each power device, and then distributes the fused model parameters to each power device. The parameter weights of the power devices are determined based on the credibility of the corresponding local model parameters. Each electrical device uses the fusion model parameters to update its local battery state estimation model.

[0007] Based on the aforementioned technical methods, firstly, each device trains its model using local data and uploads the model parameters to the aggregation node. Secondly, the aggregation node assigns corresponding weights based on the credibility of each device, performing weighted fusion of the local model parameters to generate more accurate global model parameters. Finally, each device updates its own model using these fused model parameters. This approach achieves several advantages. Firstly, by introducing a credibility assessment mechanism, the influence of different nodes in the fusion process can be dynamically adjusted, effectively suppressing the impact of low-quality models and improving issues such as insufficient data volume and poor generalization ability of single nodes, thereby enhancing the robustness and accuracy of the overall model. Secondly, since only model parameters are shared, rather than raw data, distributed collaborative learning can be achieved, protecting user privacy. Furthermore, because the aggregation node is selected from each device and does not rely on a central server, a decentralized model training and update mechanism is implemented, further enhancing the privacy of local training data and the robustness of the model.

[0008] In some embodiments, weighted fusion of the parameters of each local model includes: Obtain the parameter weights for each electrical device in the current round; the parameter weights for the current round are determined based on the first confidence level of the local model parameters obtained from the previous round of training for the corresponding electrical device, and the second confidence level of the fusion model parameters obtained from the previous round of training. Based on the parameter weights of each electrical device in the current round, the local model parameters obtained from the training of each electrical device in the current round are weighted and fused to obtain the fused model parameters for the current round.

[0009] Based on the aforementioned technical methods, the parameter weights in the current round are jointly determined by the credibility of the local model and the fusion model from the previous round, resulting in a more reasonable weight allocation. This allows for a comprehensive consideration of the historical performance of individual nodes and their contribution to the global model, enabling high-credibility nodes to receive higher weights in the new round. Furthermore, by combining the first and second credibility levels, the reliability of the model can be more comprehensively evaluated, improving the fusion effect.

[0010] In some embodiments, the model training method further includes: The electrical equipment uses the validation dataset to validate the local model parameters obtained in the current round of training, and obtains the first confidence level of the electrical equipment in the current round. The aggregation node uses the validation dataset to validate the parameters of the fusion model obtained in the current round of training, and obtains the second confidence level corresponding to the current round. The aggregation node determines the parameter weights for each power device in the next round based on the first credibility and the second credibility of each power device in the current round.

[0011] Based on the aforementioned technical methods, the local model and the fusion model are evaluated separately using a validation dataset to obtain a first confidence level and a second confidence level, which serve as the basis for the next round of weight calculation. This approach improves the objectivity of measuring the model quality of each node, reducing the impact of data bias or noise on model performance. Furthermore, by combining the performance of the local and global models, the weight calculation becomes more valuable, thereby further optimizing the model fusion effect.

[0012] In some embodiments, the parameter weights for each electrical device in the next round are determined based on a first confidence level and a second confidence level corresponding to the current round, including: If the first confidence level of the current wheel of the electrical equipment is greater than the second confidence level of the current wheel, the first count value corresponding to the electrical equipment will be increased by the first step length. Based on the first count value corresponding to the electrical equipment, the parameter weights of the electrical equipment in the next round are determined; the parameter weights of the electrical equipment in the next round are positively correlated with the corresponding first count value.

[0013] Based on the aforementioned technical methods, when a local model outperforms the fusion model, its first count value is increased to reflect its higher reliability, and its weight in the next round is increased accordingly. This encourages high-quality nodes to continuously contribute reliable model parameters, enhancing their influence; furthermore, adjusting weights through count quantization ensures continuous and stable model updates, thereby improving overall training efficiency.

[0014] In some embodiments, determining the parameter weights for each electrical device in the next round based on the first confidence level and the second confidence level corresponding to the current round of each device further includes: If the first confidence level of the current round of the electrical equipment is less than the second confidence level of the current round, the second count value corresponding to the electrical equipment is increased by the second step size; Based on the first count value corresponding to the electrical equipment, determine the parameter weights for the next round of the electrical equipment, including: Based on the first and second count values ​​corresponding to the electrical equipment, the parameter weights for the next round of the electrical equipment are determined; the parameter weights for the next round of the electrical equipment are negatively correlated with the corresponding second count value.

[0015] Based on the aforementioned technical methods, when the local model performs worse than the fusion model, its weight is reduced by increasing its second count value, thereby mitigating its negative impact on the global model. This approach, on the one hand, quickly identifies and weakens the interference of low-quality nodes, preventing them from lowering the overall model accuracy; on the other hand, by introducing a negative feedback mechanism, it makes weight changes more flexible, thus enhancing the model's adaptive capability.

[0016] In some embodiments, the parameter weights of the electrical equipment in the next round are determined based on the first and second count values ​​corresponding to the electrical equipment, including: Based on the first and second count values ​​corresponding to the electrical equipment, a weighted reference value is determined; the weighted reference value is positively correlated with the first count value and negatively correlated with the second count value. The parameter weights for the next round of electrical equipment are determined based on the sum of the weight reference value and the set weight offset value.

[0017] Based on the aforementioned technical methods, by introducing a combination mechanism of weight reference values ​​and offset values, the final weights not only reflect the historical performance of the electrical equipment but also have a certain degree of adjustment space. This approach, on the one hand, avoids excessive weight fluctuations and maintains the stability of model updates; on the other hand, by setting offsets, the weights can be fine-tuned under specific circumstances, thus better addressing the model optimization needs in complex environments.

[0018] In some embodiments, the model training method further includes: If the training termination condition is not met, each power device uses its local training dataset to train its local battery state estimation model for the next round, and obtains and sends the local model parameters for the next round to the aggregation node in each power device. The training termination condition includes at least one of the following: the difference between the second confidence level corresponding to the current round and the second confidence level corresponding to the previous round converges, and the number of training rounds reaches the preset number. The aggregation node performs weighted fusion of local model parameters based on the parameter weights of each power device in the next round, and then distributes the fused model parameters to each power device in the next round. Each electrical device uses the parameters from the next round of fusion model to update its local battery state estimation model.

[0019] Based on the aforementioned technical methods, by setting clear training termination conditions, such as model performance convergence or reaching a preset number of iterations, the training process can be ensured to be controllable and efficient. This avoids unnecessary redundant training, saving computational resources; furthermore, by continuously optimizing model parameters through multiple iterations, the overall model performance is improved, leading to more accurate battery state estimation.

[0020] In some embodiments, the model training method further includes: The device collects at least one set of first operating data from the local battery and generates a first training dataset based on the at least one set of first operating data; The electrical equipment uses a digital twin of the local battery to simulate the operating behavior of the local battery under at least one operating condition, obtains at least one set of second operating data, and generates a second training dataset based on the at least one set of second operating data. The electrical equipment determines its local training dataset based on the first and second training datasets.

[0021] Based on the aforementioned technical methods, constructing a local training dataset by combining real-world data collection with digital twin simulation not only improves data diversity but also enhances the model's generalization ability. This overcomes the limitations of a single data source, enabling the model to adapt to more real-world application scenarios. Furthermore, by introducing digital twin technology, high-quality simulated data can be generated even in the absence of real data, thereby improving training efficiency and model reliability.

[0022] This application provides a state estimation method, including: Obtain operating data of the local battery on the electrical device; The battery state of the local battery is estimated using a battery state estimation model, and the estimation result is obtained. The battery state estimation model is trained using the model training method described above.

[0023] This application provides a group of electrical devices, including multiple electrical devices, which are used to implement the above-described model training method.

[0024] This application provides a state estimation device, including: The acquisition module is used to acquire the operating data of the local battery on the electrical device; The estimation module is used to estimate the battery state of the local battery using a battery state estimation model to obtain an estimation result; wherein the battery state estimation model is trained using the model training method described above.

[0025] This application provides a battery management system for implementing the steps in the above-described state estimation method.

[0026] This application provides a battery system including at least one battery and the battery management system described above.

[0027] This application provides an electrical device including the battery system described above.

[0028] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the above-described state estimation method.

[0029] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described state estimation method. Attached Figure Description

[0030] Figure 1 A schematic diagram illustrating the implementation process of a model training method provided in this application embodiment; Figure 2 A schematic diagram of the implementation process of a state estimation method provided in this application embodiment. Figure 1 ; Figure 3 A schematic diagram illustrating the composition of a group of electrical equipment provided in an embodiment of this application; Figure 4 A schematic diagram of the implementation process of a state estimation method provided in this application embodiment. Figure 2 ; Figure 5 A schematic diagram illustrating the implementation process of model training using a group learning framework, provided in an embodiment of this application; Figure 6 This is a schematic diagram of the composition structure of a state estimation device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the composition structure of a battery system provided in an embodiment of this application; Figure 8 This is a schematic diagram of the composition structure of an electrical device provided in an embodiment of this application.

[0031] It should be noted that the terms "first" and "second" mentioned above are only used to distinguish between different options and do not represent the degree of superiority or inferiority of the options or their priority in the implementation process. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] It should be noted that, unless otherwise specified, the embodiments and technical features in the embodiments of this application can be combined with each other, and the detailed descriptions in the specific implementation should be understood as explanations of the purpose of this application and should not be regarded as undue limitations on this application.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having” and any variations thereof are intended to cover non-exclusive inclusion.

[0035] In the description of the embodiments of this application, technical terms such as "first," "second," and "third" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0036] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects are in an "or" relationship.

[0038] In the following description, the terms local learning, group learning, credibility assessment, and digital twin are important concepts that embody the core technical features of this application and need to be explained first. These terms are closely related to the technical background of this application, and understanding them will help to better grasp the overall concept of this application.

[0039] 1) Local Learning (LL): This refers to a learning method where terminal devices train models using local data without uploading raw data to the cloud or centralized servers. In a group of electrical devices, this manifests as a single device's Battery Management System (BMS) independently updating model parameters based on its own collected battery operating data, thereby learning the battery state estimation model and using this model to estimate the battery state.

[0040] 2) Swarm Learning (SL): A decentralized collaborative learning framework where multiple terminal devices jointly train the system by sharing model parameters instead of raw data. For example, multiple terminal devices can exchange model parameters instead of raw data through a blockchain or peer-to-peer (P2P) network. In a group of electrical devices, each device can dynamically select an aggregation node and achieve knowledge sharing by exchanging model parameters, while ensuring data privacy and system robustness.

[0041] 3) Trustworthiness Evaluation: In the process of group learning, this is a metric used to measure the contribution of the model parameters of each local model to the parameters of the fused model. It can be calculated by the prediction error or accuracy of the model on the validation dataset. The model parameters of local models with high trustworthiness are given higher weights during the fusion process, thereby reducing the interference of low-quality models on the global results.

[0042] 4) Digital Twin: A virtual image of a battery constructed through the fusion of physical modeling and real-time data, capable of dynamically reflecting the physical state and aging process of the physical battery. In this application, the digital twin can be used to generate battery delivery simulation data under various operating conditions to supplement the insufficiency of actual collected data and improve the data diversity and richness of model training.

[0043] 5) Battery State Estimation: This includes estimating SOC, SOH, etc., and is one of the core functions of the battery management system. For example, SOC represents the ratio of the battery's current remaining charge to its maximum usable capacity; SOH characterizes the degree of degradation of the battery's current performance relative to its initial performance, usually expressed as a percentage.

[0044] 6) Deep Neural Network (DNN): A machine learning model containing multiple hidden layers, capable of extracting high-dimensional features through hierarchical nonlinear transformations. For example, in this application, a DNN can be used as the basic structure of a battery state estimation model, with its input being battery operating data or feature information of battery operating data, and its output being the estimated SOC and SOH values.

[0045] 7) Weighted Fusion: In group learning, the parameters of each local model are weighted and averaged according to their credibility to generate the fused global model parameters. This method can effectively suppress the influence of noise or low-quality models, and improve the overall estimation accuracy and robustness.

[0046] 8) Root Mean Square Error Loss Function (RMSE): A commonly used objective function in supervised learning to measure the deviation between predicted and true values. In this application, RMSE can be used as an optimization objective for model training to minimize prediction error.

[0047] 9) Normalization: A data preprocessing method used to scale feature values ​​of different dimensions to a standard range (e.g., [0,1]) to eliminate the impact of dimensional differences on model training. In this application, battery operating data such as voltage and current have been normalized.

[0048] 10) Generalization: The ability of a machine learning model to perform on unseen data. In this application, the generalization ability of the model is improved through group learning and the use of diverse data generated by digital twins, enabling it to maintain high estimation accuracy under different battery packs and operating conditions.

[0049] The model training methods provided in the embodiments of this application can be executed collaboratively by various electrical devices in a group of electrical devices, wherein the electrical devices may include, but are not limited to, electric vehicles, airplanes, electric bicycles, electric motorcycles, electric boats, and / or ships. In some embodiments, the model training method may be executed by an onboard BMS or onboard terminal in a group of electric vehicles.

[0050] This application provides a model training method applied to a group of electrical equipment, which includes multiple electrical devices. Figure 1 This is a schematic diagram illustrating the implementation process of a model training method provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes the following steps S101 to S103: In step S101, each electrical device trains its local battery state estimation model using its local training dataset, obtains and sends the local model parameters to the aggregation node. The local training dataset includes at least one set of running data for the local battery of the electrical device and the battery state label corresponding to each set of running data. The aggregation node is selected from each electrical device.

[0051] Here, each piece of running data in the local training dataset can serve as a sample of running data for the local battery, used to train the local battery state estimation model. Running data can include, but is not limited to, operating information such as voltage, current, and / or temperature of the battery during operation. The running data of the local battery in the local training dataset can be locally collected by the electrical device and / or generated through simulation. For example, the running data in the local training dataset of the electrical device can include real-time operating information such as voltage, current, and temperature of the local battery collected by each electrical device during actual operation, or it can include simulation data generated by digital twin technology.

[0052] The battery state labels corresponding to the running data in the local training dataset can be pre-labeled ground truth values ​​for the battery state corresponding to that running data. The battery state can include, but is not limited to, at least one of SOC, SOH, and battery capacity. The battery state labels correspond to the actual battery state of each running data set, such as SOC labels, SOH labels, and / or battery capacity labels, and are used to supervise the battery state estimated by the model during training. The battery state labels can be obtained through laboratory testing or historical data annotation.

[0053] In some implementations, the running data can be filtered and normalized before being used to train a local battery state estimation model.

[0054] Local model parameters are obtained after each device trains its neural network locally. These parameters represent the weights and / or biases of each layer in the battery state estimation model structure trained by each device. Through local training, each device can independently update its own battery state estimation model parameters without uploading its original battery operating data, thus protecting data privacy. Furthermore, local training can be customized to the specific operating conditions of each device's environment (such as different temperatures and load conditions), improving the model's adaptability.

[0055] Each device uses the same model structure for its local battery state estimation model. For example, the battery state estimation model can use any suitable deep neural network model, such as Long Short-Term Memory Network (LSTM), Convolutional Neural Network (CNN), etc.

[0056] In some implementations, the battery state estimation model refers to a deep neural network model deployed on the electrical device to predict the battery's State of Charge (SOC) and / or State of Health (SOH). The battery state estimation model can consist of multiple hidden layers and can process time-series data, such as an LSTM structure. The inputs to the battery state estimation model include, but are not limited to, features such as voltage, current, temperature, open-circuit voltage curve, relaxation voltage curve characteristics, and / or depth of charge / discharge, and the output is the predicted SOC and / or SOH. The relaxation voltage curve refers to the dynamic change curve of the terminal voltage gradually approaching equilibrium over time after the battery's charging and discharging has stopped. Its morphological characteristics are strongly correlated with the battery's internal polarization state and aging degree.

[0057] In some implementations, each device can periodically collect local battery operating data via sensors and combine this data with operating data generated using digital twin technology to construct a training dataset containing various degradation paths and operating conditions. Subsequently, a deep neural network model is used to train the local battery state estimation model locally, employing the root mean square error loss function as the optimization objective to minimize the deviation between the predicted and true values. By using a deep neural network model to train the local battery state estimation model, each device can complete high-quality battery state estimation model training without exposing user privacy.

[0058] The aggregation node can be selected from the various electrical devices using any suitable selection method, and this application embodiment does not limit this. For example, one electrical device can be randomly selected from the various electrical devices as the aggregation node. Alternatively, the electrical devices can be selected in turn as the aggregation node in the current round of training according to a set selection order. Yet another example is that the electrical device with the lowest resource utilization rate can be selected as the aggregation node.

[0059] In some implementations, the aggregation node can be any suitable electrical device dynamically elected by the devices in the group of electrical devices. In practice, the aggregation node can be jointly elected by the devices using any suitable method, and this application does not limit this. For example, the aggregation node can be randomly elected from among the devices. Alternatively, the aggregation node can be the device with the highest credibility among all the devices.

[0060] In some implementations, the aggregation node may be pre-selected by each power-consuming device and remain unchanged in each round of training.

[0061] In some implementations, the latest aggregation node can be selected from each power-consuming device after each round or multiple rounds of training of the local battery state estimation model using the local training dataset.

[0062] In step S102, the aggregation node performs weighted fusion of the parameters of each local model based on the parameter weights of each power device, and then distributes the fused model parameters to each power device; the parameter weights of the power devices are determined based on the credibility of the corresponding local model parameters.

[0063] Here, the aggregation node is responsible for collecting the local model parameters of each power device and performing fusion calculations. The determination of the parameter weights for each power device depends on the credibility evaluation results corresponding to the local model parameters of each device. The credibility of the local model parameters is measured by the prediction error or accuracy on the validation dataset. The more stable the performance of the model corresponding to the local model parameters on the validation dataset, the higher its credibility, and the higher its weight will be assigned during the fusion process. This weighted averaging strategy can effectively suppress the influence of low-quality models, thereby improving the overall performance and robustness of the final battery state estimation model obtained based on the updated fused model parameters.

[0064] A validation dataset is a set of data used to evaluate model performance. It is typically not involved in the model training process but rather used to measure the model's performance on unknown data. In this embodiment, the validation dataset is a public dataset shared by all electrical devices, and may include, but is not limited to, publicly available battery operating data, and / or operating data collected during historical battery testing before the electrical device leaves the factory (excluding privacy data generated during user use of the electrical device). For example, when the electrical device includes an electric vehicle, the validation dataset may include publicly available operating datasets of the power batteries used in the electric vehicle or initial training datasets provided by the automaker.

[0065] In some implementations, the aggregation node may first receive local model parameters from each device and calculate the credibility of each device's local model parameters based on a pre-reserved public validation dataset. Then, a credibility weight prediction algorithm is used to assign corresponding parameter weights to each device. This algorithm dynamically adjusts the parameter weights by comparing the credibility of each device's local model parameters on the validation dataset, ensuring that high-credibility models contribute more to the fused model parameters. Finally, the aggregation node distributes the weighted fused model parameters to each device for use in subsequent training or tasks.

[0066] In step S103, each electrical device updates its local battery state estimation model using the fusion model parameters.

[0067] Here, after receiving the fusion model parameters from the aggregation node, each device can apply these parameters to its local battery state estimation model, replacing the original local model parameters and thus updating the entire model. The updated model maintains high estimation accuracy across a wider range of data distributions and possesses stronger generalization ability. Through continuous iteration of this process, the model can gradually optimize its predictive capabilities while accumulating more data.

[0068] In some implementations, steps S101 to S103 can be used to train the local battery state estimation model of each electrical device in multiple rounds until the training termination condition is met. The training termination condition may include, but is not limited to, the convergence of the difference between the second confidence level corresponding to the fusion model parameters obtained in the current round of training and the second confidence level corresponding to the fusion model parameters obtained in the previous round of training, and / or the number of training rounds reaching a preset number, etc.

[0069] In some implementations, each device can load the current fusion model parameters into its local battery state estimation model and use the currently collected battery operating data to estimate the local battery state, thus obtaining the estimated battery state. Since the updated battery state estimation model integrates the training results from multiple devices, it can provide more stable and accurate battery state estimation results under different operating conditions. Furthermore, the performance of the battery state estimation model will continuously improve as group learning progresses.

[0070] In actual implementation, the steps are closely interconnected: in step S101, each device completes local training and obtains local model parameters, laying the foundation for subsequent model fusion; in step S102, the aggregation node receives the local model parameters from each device and performs weighted fusion to ensure optimal overall model quality; in step S103, each device applies the fused model parameters to improve its own model performance, forming a closed-loop optimization mechanism. The entire process, through distributed collaboration, achieves continuous model evolution and optimization while ensuring data security.

[0071] Based on the aforementioned technical methods, firstly, each device trains its model using local data and uploads the model parameters to the aggregation node. Secondly, the aggregation node assigns corresponding weights based on the credibility of each device, performing weighted fusion of the local model parameters to generate more accurate global model parameters. Finally, each device updates its own model using these fused model parameters. This approach achieves several advantages. Firstly, by introducing a credibility assessment mechanism, the influence of different nodes in the fusion process can be dynamically adjusted, effectively suppressing the impact of low-quality models and addressing issues such as insufficient data volume and poor generalization ability of single nodes, thereby improving the robustness and accuracy of the overall model. Secondly, since only model parameters are shared, rather than raw data, distributed collaborative learning can be achieved, effectively protecting user privacy. Furthermore, because the aggregation node is selected from each device and does not rely on a central server, a decentralized model training and update mechanism is implemented, further enhancing the privacy of local training data and the robustness of the model.

[0072] In some embodiments, the weighted fusion of the local model parameters in step S102 may include the following steps S111 to S112: Step S111: Obtain the parameter weights of each power device in the current round. The parameter weights in the current round are determined based on the first confidence level of the local model parameters obtained from the previous round of training on the corresponding power device and the second confidence level of the fusion model parameters obtained from the previous round of training.

[0073] The parameter weight of an electrical device in the current round refers to a numerical value used in the swarm learning framework to measure the degree of contribution of that device to the parameter fusion process in this round of modeling. A higher parameter weight indicates that a particular electrical device has a greater impact on the fused model parameters in this round. The calculation of the parameter weight of an electrical device in the current round depends on two key indicators: the first confidence level of the local model parameters obtained from the previous round of training and the second confidence level of the fused model parameters obtained from the previous round of training.

[0074] The first confidence level of the local model parameters obtained in each training round of an electrical device can characterize the prediction accuracy or reliability assessment value of the local model parameters obtained in this training round on the validation dataset. The first confidence level reflects the performance of the local model of the electrical device. If the prediction error of the local model is small, the confidence level of the corresponding local model parameters is high, indicating that the local model of the electrical device has high prediction accuracy. For example, in battery health state estimation, if the root mean square error of the local model obtained in the previous training round of an electrical device on the validation dataset is low, then the first confidence level of the local model parameters obtained in the previous training round of the electrical device is high.

[0075] The second confidence level of the fusion model parameters obtained in each round of training reflects the overall performance of the battery state estimation model on the validation dataset when the fusion model parameters obtained in this round are applied to the battery state estimation model. If the battery state estimation model performs well on the validation dataset when the fusion model parameters obtained in this round are applied, then the second confidence level of the fusion model parameters obtained in this round of training is high. In some implementations, the fusion model parameters obtained in this round of training can be applied to the local battery state estimation model of any electrical device in the group of electrical devices, and the prediction accuracy of the battery state estimation model on the validation dataset can be verified to obtain the second confidence level of the fusion model parameters.

[0076] In some implementations, when the current training round is the k-th round, the parameter weights for the current round are determined based on the first confidence level of the local model parameters obtained by the corresponding device in the (k-1)-th round of training, and the second confidence level of the fusion model parameters obtained in the (k-1)-th round of training; when the current training round is the first round of training, the parameter weights for the current round can be preset initial values. Here, k is an integer greater than 1.

[0077] By combining the first confidence level of the local model parameters obtained from the previous training round on the power-consuming device with the second confidence level of the fused model parameters obtained from the previous training round, the parameter weights of each power-consuming device in the current round can be dynamically adjusted. This enables the suppression of low-quality model parameters and the enhancement of high-quality model parameters, thereby effectively improving the accuracy and stability of the battery state estimation model using fused model parameters and reducing the problem of model performance degradation caused by low-quality data provided by some power-consuming devices.

[0078] Step S112: Based on the parameter weights of each power device in the current round, the local model parameters obtained from the training of each power device in the current round are weighted and fused to obtain the fused model parameters for the current round.

[0079] After obtaining the parameter weights of each power device in the current round, the aggregation node performs a weighted average fusion of the local model parameters uploaded by each power device based on these parameter weights to generate new fused model parameters.

[0080] In some implementations, the weighted fusion process can follow a weighted average strategy, that is, the local model parameters of each power device are multiplied by the parameter weights corresponding to that power device and then added together to obtain the fused model parameters.

[0081] Weighted fusion allows local model parameters of proven high-performing devices to have a greater impact on the fused model parameters during model updates, while the impact of local model parameters of low-confidence devices is weakened. Weighted fusion based on parameter weights can reduce the negative impact of low-quality local model parameters on the fused model parameters, while retaining the effective knowledge learned by high-confidence local model parameters.

[0082] Based on the aforementioned technical methods, the parameter weights in the current round are jointly determined by the credibility of the local model and the fusion model from the previous round, resulting in a more reasonable weight allocation. This approach allows for a comprehensive consideration of the historical performance of individual nodes and their contribution to the global model, granting higher weights to high-credibility nodes in the new round, thereby enhancing the robustness of the global model and reducing the interference of low-quality nodes on model performance. Furthermore, by combining the first and second credibility levels, the reliability of the model can be more comprehensively evaluated, improving the fusion effect.

[0083] In some embodiments, the model training method further includes the following steps S121 to S123: Step S121: The electrical equipment uses the verification dataset to verify the local model parameters obtained in the current round of training, and obtains the first confidence level of the electrical equipment in the current round.

[0084] By applying the local model parameters obtained from the current training iteration of the device to a local battery state estimation model, and validating the model using a validation dataset, the prediction error or other evaluation metrics are calculated. Based on the calculated prediction error or other evaluation metrics, the reliability of the local model parameters obtained from the current training iteration on the validation dataset can be determined, i.e., the first reliability. The first reliability represents the reliability level of the local battery state estimation model trained by the device in the current round. If the local battery state estimation model trained by the device in the current round performs well on the validation dataset, it indicates that the model parameters obtained from the current training iteration have high reliability; otherwise, it may be affected by low-quality data or noise interference, and the reliability of the local model parameters trained by the device in the current round is low.

[0085] For example, in electric vehicle battery state of health estimation, each vehicle-side battery management system trains a local battery state estimation model using locally collected battery operating data and simulation data generated by a digital twin. Then, the local battery state estimation model trained on the current vehicle cycle is applied to a validation dataset. By analyzing the battery operating data in the validation dataset using the local battery state estimation model, the accuracy and stability of the local battery state estimation model in predicting battery capacity and / or state of health (SOH) are evaluated, thereby deriving a first level of confidence.

[0086] In some implementations, the local model parameters obtained from the current training round of the electrical device can be applied to the local battery state estimation model, and the battery state estimation model can be validated using a validation dataset. The prediction error of the battery state estimation model for the battery state can be calculated, and the calculated prediction error can be used as the first confidence level of the local model parameters obtained from the current training round of the electrical device.

[0087] In step S122, the aggregation node uses the validation dataset to validate the parameters of the fusion model obtained in the current round of training, and obtains the second confidence level corresponding to the current round.

[0088] Here, the aggregation node acts as the coordination center in the swarm learning framework. It is responsible for collecting the local model parameters uploaded by each device and performing the fusion operation on these parameters. The aggregation node also uses the same validation dataset to evaluate the performance of the fused model parameters in the current round, thereby obtaining a credibility metric, the second credibility. This second credibility represents the overall predictive ability of the battery state estimation model applying the fused model parameters.

[0089] In some implementations, the fused model parameters of the current round can be applied to the battery state estimation model, and the battery state estimation model can be validated using a validation dataset. The prediction error of the battery state estimation model for the battery state can be calculated, and the calculated prediction error can be used as the second confidence level for the current round.

[0090] In step S123, the aggregation node determines the parameter weights for each power device in the next round based on the first credibility and the second credibility of each power device in the current round.

[0091] Parameter weights determine the degree of influence of each local model parameter on the fused model parameters during the fusion process. In this embodiment, the aggregation node dynamically adjusts the parameter weights of each device in the next round by combining the first credibility of the device in the current round and the second credibility of the fused model parameters in the current round. This allows high-credibility local model parameters to have a higher influence in the next round of fusion, suppresses the interference of low-quality local model parameters, and improves the robustness of the fused model parameters in the next round.

[0092] In some implementations, the ratio between the first confidence level and the second confidence level corresponding to the current round of each electrical device (denoted as the confidence ratio) can be determined separately, and the parameter weights for the next round of each electrical device can be determined based on the confidence ratios corresponding to each electrical device. The parameter weights for the next round of each electrical device are positively correlated with the confidence ratio corresponding to that electrical device.

[0093] In some implementations, the values ​​of the first confidence level and the second confidence level corresponding to the current round for each electrical device can be determined separately (denoted as the confidence level difference). The parameter weights for the next round of each electrical device are then determined based on the confidence level differences corresponding to each device. The parameter weights for the next round of each electrical device are positively correlated with the confidence level difference corresponding to that device.

[0094] Based on the aforementioned technical methods, the local model and the fusion model are evaluated separately using a validation dataset to obtain a first confidence level and a second confidence level, which serve as the basis for the next round of weight calculation. This approach improves the objectivity of measuring the model quality of each node, reducing the impact of data bias or noise on model performance. Furthermore, by combining the performance of the local and global models, the weight calculation becomes more valuable, thereby further optimizing the model fusion effect.

[0095] In some embodiments, determining the parameter weights for each electrical device in the next round based on the first confidence level and the second confidence level corresponding to the current round in step S123 may include the following steps S131 to S132: Step S131: If the first confidence level of the current wheel of the electrical equipment is greater than the second confidence level of the current wheel, increase the first count value of the electrical equipment by the first step length.

[0096] Here, if the first confidence level of the current round of the power device is greater than the second confidence level of the current round, it indicates that the local model parameters of the current round of the power device are better than the fusion model parameters of the current round, and have higher reliability. In this case, a preset step size value (such as 1, 2, etc.) can be added to the first count value corresponding to the power device to record the degree of contribution of the power device.

[0097] The first count value is a record of the number of times the reliability of the local model parameters of the electrical equipment is higher than that of the fused model parameters. Whenever the reliability of the local model parameters of the electrical equipment is detected to be higher than that of the fused model parameters in the same round, the first count value can be incremented according to the set first step length to accumulate the historical better performance of the electrical equipment. By changing the first count value, the long-term stability of the electrical equipment can be dynamically evaluated, and the parameter weights of the local model parameters of the electrical equipment can be adjusted according to the evaluation results.

[0098] For example, suppose that in the current training round of a certain electric vehicle's onboard BMS, if the first confidence level corresponding to the current wheel of the electric vehicle is greater than the second confidence level, it indicates that the battery state estimation model applying the local model parameters of the onboard BMS in this round outperforms the battery state estimation model applying the fused model parameters in this round on the validation dataset. In other words, the electrical device represented by the onboard BMS of the electric vehicle provides high-quality learning results. By incrementally increasing the first count value corresponding to the onboard BMS of the electric vehicle, the weight allocation ratio of the electrical device represented by the BMS node of the electric vehicle in subsequent model parameter fusion can be enhanced, thereby encouraging electrical devices with high confidence to continuously contribute high-quality data and model parameters.

[0099] In this application, the first step length is a positive number, and the value of the first step length can be set by those skilled in the art according to the actual situation. This application does not limit this. For example, the first step length can be 1, 2, 3, or 5, etc.

[0100] Step S132: Based on the first count value corresponding to the electrical equipment, determine the parameter weights of the electrical equipment in the next round; the parameter weights of the electrical equipment in the next round are positively correlated with the corresponding first count value.

[0101] The parameter weights of electrical equipment are positively correlated with the first count value of the electrical equipment. That is, the higher the first count value, the greater the weight of the electrical equipment in the next round of weighted fusion of local model parameters.

[0102] For example, in a parameter fusion process, if the first count value corresponding to device A is 5, the first count value corresponding to device B is 2, and the first count value corresponding to device C is 3, then the local model parameters of device A will participate in the fusion with a higher weight, while the local model parameters of devices B and C will have a relatively smaller impact. That is, the parameter weight of device A in the next round is greater than the parameter weight of devices B and C in the next round.

[0103] Based on the aforementioned technical methods, when a local model outperforms the fusion model, its first count value is increased to reflect its higher reliability, and its weight in the next round is increased accordingly. This encourages high-quality nodes to continuously contribute reliable model parameters, enhancing their influence; furthermore, adjusting weights through count quantization ensures continuous and stable model updates, thereby improving overall training efficiency.

[0104] In some embodiments, the step S123, which determines the parameter weights for each electrical device in the next round based on the first confidence level and the second confidence level corresponding to the current round, may further include the following step S141: Step S141: If the first confidence level of the current round of the electrical equipment is less than the second confidence level of the current round, increase the second count value of the electrical equipment by a second step.

[0105] Here, if the first confidence level of the current cycle of the electrical device is less than the second confidence level of the current cycle, it indicates that the local model parameters of the current cycle of the electrical device are not as good as the fusion model parameters of the current cycle, and have lower reliability. In this case, a preset step size value (such as 1, 2, etc.) can be added to the second count value corresponding to the electrical device to record the historical poor performance of the electrical device.

[0106] The second count value records the number of times the reliability of the local model parameters of the electrical equipment is lower than that of the fused model parameters. Whenever the reliability of the local model parameters of the electrical equipment is detected to be lower than that of the fused model parameters in the same round, the second count value can be incremented according to a set second step size to accumulate the historical poor performance of the electrical equipment. By changing this second count value, the long-term stability of the electrical equipment can be dynamically evaluated, and the parameter weights of the local model parameters of the electrical equipment can be adjusted according to the evaluation results, thereby suppressing the continuous impact of low-quality local model parameters on the accuracy of the global model.

[0107] For example, suppose that in the current training round of a certain electric vehicle's onboard BMS, if the second confidence level corresponding to the current wheel of the electric vehicle is lower than the second confidence level of the current wheel, it indicates that the battery state estimation model applying the local model parameters of the onboard BMS in this round performs worse on the validation dataset than the battery state estimation model applying the fused model parameters of this round. In other words, the electrical device represented by the onboard BMS of the electric vehicle provides lower-quality learning results. By incrementing the second count value corresponding to the onboard BMS of the electric vehicle, the weight allocation ratio of the electrical device represented by the BMS node of the electric vehicle in subsequent model parameter fusion can be suppressed, thereby preventing the electrical device with low confidence from continuously affecting high-quality data and models.

[0108] In this embodiment, the second step length is a positive number, and its value can be set by those skilled in the art based on actual circumstances; this embodiment does not limit this. For example, the second step length can be 1, 2, 3, or 5, etc. The second step length can be the same as or different from the first step length; this embodiment does not limit this.

[0109] The step S132, which determines the parameter weights of the electrical equipment in the next round based on the first count value corresponding to the electrical equipment, may include the following step S142: Step S142: Based on the first and second count values ​​corresponding to the electrical equipment, determine the parameter weights for the next round of the electrical equipment; the parameter weights for the next round of the electrical equipment are negatively correlated with the corresponding second count value.

[0110] The parameter weights of electrical equipment are negatively correlated with the second count value of the electrical equipment. That is, the higher the second count value, the smaller the weight of the electrical equipment in the next round of local model parameter weighting fusion process.

[0111] For example, in a parameter fusion process, if the second count value corresponding to device A is 1, the second count value corresponding to device B is 4, and the second count value corresponding to device C is 3, then the local model parameters of device B will participate in the fusion with a lower weight, while the local model parameters of devices A and C will have a relatively larger impact. That is, the parameter weight of device B in the next round is less than the parameter weight of devices A and C in the next round.

[0112] Based on the aforementioned technical methods, when the local model performs worse than the fusion model, its weight is reduced by increasing its second count value, thereby mitigating its negative impact on the global model. This approach, on the one hand, quickly identifies and weakens the interference of low-quality nodes, preventing them from lowering the overall model accuracy; on the other hand, by introducing a negative feedback mechanism, it makes weight changes more flexible, thus enhancing the model's adaptive capability.

[0113] In some implementations, a first counter and a second counter can be set for each electrical device, with the first counter counting a first count value and the second counter counting a second count value.

[0114] In some embodiments, determining the parameter weights of the electrical equipment in the next round based on the first and second count values ​​corresponding to the electrical equipment in step S142 includes the following steps S151 to S152: Step S151: Based on the first count value and the second count value corresponding to the electrical equipment, determine the weight reference value; the weight reference value is positively correlated with the first count value and negatively correlated with the second count value.

[0115] Step S152: Based on the sum of the weight reference value and the set weight offset value, determine the parameter weights of the electrical equipment in the next round.

[0116] The weight reference value is an intermediate variable used to calculate the final parameter weights. Its magnitude reflects the overall reliability level of the electrical equipment. Since the weight reference value is positively correlated with the first count value, this indicates that when the reliability of the local model parameters of an electrical equipment is high, the weight reference value corresponding to that equipment will increase, thus giving it a greater influence on the next round of fusion. Simultaneously, the weight reference value is negatively correlated with the second count value, which reflects the unreliability of the local model parameters of the electrical equipment. Since the weight reference value is negatively correlated with the second count value, this indicates that when the reliability of the local model parameters of an electrical equipment is low, the weight reference value corresponding to that equipment will decrease. Therefore, through this positive and negative correlation, the influence of each electrical equipment in the next round can be dynamically adjusted, enhancing the system's robustness.

[0117] In some implementations, the difference between the first count value and the second count value can be determined as the weight reference value.

[0118] In some implementations, the ratio of the first count value to the second count value can be determined as a weighted reference value.

[0119] In some implementations, the sum of a first count value and a second count value can be calculated, and the ratio between the first count value and the sum can be determined as a weighted reference value.

[0120] The weight offset is a preset constant used to adjust the distribution range of the weight reference values ​​to prevent drastic weight fluctuations caused by large initial parameter weight deviations in the electrical equipment. The weight offset controls the aggression of parameter weight changes. The weight reference values ​​and the weight offset are added together to form the final parameter weights, which participate in the next round of weighted fusion of local model parameters.

[0121] For example, when the weight reference value is in the range of [0,1], if the weight offset value is 3, then the parameter weights of all electrical equipment are in the range of [3,4], the parameter weights fluctuate less, so that the influence of low-quality electrical equipment is controlled within a reasonable range, while high-quality electrical equipment can still play a dominant role; if the weight offset value is 0, then the parameter weights of all electrical equipment are in the range of [0,1], and the parameter weights change more aggressively.

[0122] Understandably, the setting of the weight reference value ensures that the local model parameters of high-quality electrical equipment can continue to play a positive role, while the introduction of the weight offset value further balances the weight differences between different electrical equipment, preventing system instability caused by excessively large or small weights for individual electrical equipment. By adding the weight reference value and the weight offset value, an effective model can be quickly converged in the early stages, while maintaining small weight changes in the later stages to preserve stability.

[0123] Based on the aforementioned technical methods, by introducing a combination mechanism of weight reference values ​​and offset values, the final weights not only reflect the historical performance of the electrical equipment but also have a certain degree of adjustment space. This approach, on the one hand, avoids excessive weight fluctuations and maintains the stability of model updates; on the other hand, by setting offsets, the weights can be fine-tuned under specific circumstances, thus better addressing the model optimization needs in complex environments.

[0124] In some embodiments, the model training method further includes the following steps S161 to S163: Step S161: If the training termination condition is not met, each power device uses its local training dataset to train its local battery state estimation model for the next round, and obtains and sends the next round of local model parameters to the aggregation node in each power device. The training termination condition includes at least one of the following: the difference between the second confidence level corresponding to the current round and the second confidence level corresponding to the previous round converges, and the training rounds reach the preset number of rounds.

[0125] In some implementations, the training termination condition may include the convergence of the difference between the second confidence level of the current round and the second confidence level of the previous round. If the difference between the second confidence level of the current round and the second confidence level of the previous round has converged, that is, the change in the difference is minimal, it indicates that the performance of the display model tends to stabilize.

[0126] In some implementations, the training termination condition may include reaching a preset number of training rounds.

[0127] The technical advantage of this step is that by using local training and dynamically determining the training termination condition, unnecessary computational resource consumption can be reduced while maintaining model accuracy, thus improving system efficiency. Furthermore, continuing training even when the termination condition is not met helps improve the model's robustness and stability.

[0128] In step S162, the aggregation node performs weighted fusion of the local model parameters based on the parameter weights of each power device in the next round, and then distributes the fused model parameters to each power device.

[0129] In step S163, each electrical device updates its local battery state estimation model using the parameters from the next round of fusion model.

[0130] Based on the aforementioned technical methods, by setting clear training termination conditions, such as model performance convergence or reaching a preset number of iterations, the training process can be ensured to be controllable and efficient. This avoids unnecessary redundant training, saving computational resources; furthermore, by continuously optimizing model parameters through multiple iterations, the overall model performance is improved, leading to more accurate battery state estimation.

[0131] In some embodiments, the model training method further includes the following steps S171 to S173: Step S171: The electrical device collects at least one set of first running data from the local battery and generates a first training dataset based on the at least one set of first running data.

[0132] First-line operational data refers to real-world operational data directly collected from the battery management system (BMS). This data is highly representative and reliable, reflecting the battery's behavior in real-world environments. By collecting multi-dimensional, multi-time-period first-line operational data, researchers can capture trends in battery performance over time and load, thereby improving the generalization ability of trained models.

[0133] In some implementations, the first operational data refers to real-time or periodically recorded data generated by the battery during actual use, including but not limited to parameters such as voltage, current, temperature, and charge / discharge time. The real-time or periodically recorded data generated by the battery during actual use can reflect the battery's performance and aging trends under different operating conditions, and this data serves as the basis for constructing the first training dataset.

[0134] In some implementations, the process of generating the first training dataset may include data cleaning, normalization, and feature extraction. For example, after filtering and denoising the collected raw voltage and current signals, key features such as open-circuit voltage curves, relaxation voltage curves, and cumulative charge-discharge capacity are extracted as input features for subsequent model training. This improves the quality and consistency of the first training dataset, which helps to enhance the model's learning efficiency and prediction accuracy.

[0135] Step S172: The electrical equipment uses the digital twin of the local battery to simulate the operating behavior of the local battery under at least one operating condition, obtains at least one set of second operating data, and generates a second training dataset based on the at least one set of second operating data.

[0136] A digital twin is a virtual mirror system built through high-fidelity modeling and real-time data fusion, capable of simulating the behavior of a physical battery under various operating conditions. A digital twin module typically integrates multiple sub-models, including physical, electrochemical, and thermodynamic models, accurately reflecting the battery's response characteristics under different temperature and load conditions. Digital twins can generate a large amount of high-quality simulation data without increasing additional hardware costs, thus compensating for the lack of real-world operational data.

[0137] The second operational data, generated by digital twin simulation, can include battery behavior information under various operating conditions, such as voltage, internal resistance, and capacity decay parameters under high temperature, low temperature, and high current charging and discharging conditions. The second operational data generated by digital twin simulation not only covers extreme operating conditions that are difficult to obtain in the real environment, but also provides rich data dimensions. Using this type of data can improve the robustness and generalization ability of the model.

[0138] In some implementations, the process of creating a second training dataset may include data preprocessing and feature extraction. By standardizing the simulation data and extracting features related to battery degradation, a supplementary dataset that complements the first training dataset can be constructed. This combination of virtual and real-world methods effectively expands the training sample space and enhances the model's adaptability to different degradation paths and complex operating conditions.

[0139] Step S173: The electrical equipment determines a local training dataset based on the first training dataset and the second training dataset.

[0140] The local training dataset is the final training data set formed by combining the first and second training datasets. It not only includes real-world operational data but also incorporates simulation data generated by digital twins, covering various operating conditions and degradation paths, thus exhibiting greater diversity and representativeness. By appropriately combining and weighting the first and second training datasets, the model's training performance can be further optimized, making the model's training results more closely resemble the needs of real-world application scenarios.

[0141] The second training dataset generated by the digital twin maintains consistency with the first training dataset collected in real-world conditions in terms of feature dimensions, facilitating fusion processing between the two. Furthermore, when generating the second training dataset, a noise perturbation mechanism can be introduced to simulate uncertainties in the real-world environment, thereby enhancing the model's robustness against interference.

[0142] Based on the aforementioned technical methods, constructing a local training dataset by combining real-world data collection with digital twin simulation not only improves data diversity but also enhances the model's generalization ability. This overcomes the limitations of a single data source, enabling the model to adapt to more real-world application scenarios. Furthermore, by introducing digital twin technology, high-quality simulated data can be generated even in the absence of real data, thereby improving training efficiency and model reliability.

[0143] Figure 2 A schematic diagram of the implementation process of a state estimation method provided in this application embodiment. Figure 1 ,like Figure 2 As shown, the method includes the following steps S201 to S202: Step S201: Obtain the operating data of the local battery on the electrical device.

[0144] Step S202: The battery state of the local battery is estimated using the battery state estimation model to obtain the estimation result; wherein, the battery state estimation model is trained using the model training method described above.

[0145] Based on the aforementioned technical means, during model training, on the one hand, by introducing a credibility evaluation mechanism, the influence of different nodes in the fusion process can be dynamically adjusted, effectively suppressing the influence of low-quality models and improving problems such as insufficient data volume and poor generalization ability of single nodes, thereby improving the robustness and accuracy of the overall model. On the other hand, since only model parameters are shared rather than raw data, distributed collaborative learning can be achieved, effectively protecting user privacy. Furthermore, because the aggregation nodes are selected from various electrical devices and do not rely on a central server, a decentralized model training and update mechanism is implemented, further enhancing the privacy of local training data and the robustness of the model. Thus, the accuracy, robustness, and generalization ability of the battery state estimation model can be effectively improved, thereby increasing the accuracy of battery state estimation.

[0146] This application provides a group of electrical equipment, such as Figure 3 As shown, the group of electrical equipment 30 includes multiple electrical equipment 31, which are used to implement the model training method described above.

[0147] In some implementations, multiple electrical devices refer to a network node set composed of multiple terminal devices with local computing and communication capabilities, such as the vehicle-mounted BMS system in electric vehicles or the control unit of an energy storage system. This network node set, composed of multiple terminal devices with local computing and communication capabilities, can complete data acquisition, feature extraction, and model training tasks locally. It also shares model parameters of a deep neural network-based battery state estimation method through a swarm learning mechanism, rather than raw data, thereby ensuring user privacy and security. Furthermore, it effectively addresses real-world problems such as uneven data distribution and inconsistent data quality among multiple electrical devices.

[0148] The following describes the application of the model training method and state estimation method provided in the embodiments of this application in real-world scenarios. The battery state estimation scenario for electric vehicles will be used as an example.

[0149] In the field of new energy electric vehicles, power batteries are widely used due to their advantages such as high energy density, rechargeability, safety, and environmental friendliness. Battery state estimation, as a crucial function of onboard battery management systems, includes SOC estimation and SOH estimation. To meet users' needs for accurate control over battery life and driving range, improving the accuracy of battery state estimation has become a key focus for both industry and academia.

[0150] In related technologies, data-driven methods for estimating the health status of lithium batteries involve collecting operational data, extracting features and preprocessing them, using nearest neighbor component analysis for dimensionality reduction, and applying an improved sparrow search algorithm to optimize the parameters of a deep neural network model, thereby achieving lithium battery health status estimation. However, in practical applications, single-node local learning faces problems of insufficient or low-quality data, while centralized learning suffers from privacy leaks and poor system robustness. Therefore, a decentralized learning framework is needed that can both protect data privacy and enhance model robustness.

[0151] The core of this application's embodiments lies in introducing a dynamic credibility evaluation mechanism to address the impact of low-quality data nodes on the performance of the battery state estimation model, thereby enhancing the model's robustness. By employing a swarm learning approach, each node shares only model parameters rather than the original data, thus improving the model's generalization ability while ensuring privacy. Furthermore, digital twin technology is used to generate training data under various operating conditions, further enriching the local training dataset. This enriched local training dataset improves the model's adaptability and prediction accuracy.

[0152] Figure 4 A schematic diagram of the implementation process of a state estimation method provided in this application embodiment. Figure 2 ,like Figure 4 As shown, the method includes the following steps S401 to S405: Step S401, Data Acquisition and Processing.

[0153] Each vehicle-side terminal acts as an independent single node (corresponding to the electrical equipment in the aforementioned embodiments). The BMS system at this vehicle-side terminal periodically collects raw data such as the terminal voltage, current, and temperature of the local battery, and performs preprocessing operations such as filtering and normalization. Simultaneously, digital twins are used to simulate the battery's behavior under different temperature and load conditions, generating additional training samples. The real data (corresponding to the first training dataset in the aforementioned embodiments) is combined with the simulated data (corresponding to the second training dataset in the aforementioned embodiments) to expand the sample space, covering various degradation paths and operating conditions.

[0154] Step S402, Feature extraction.

[0155] Feature extraction is performed on the preprocessed battery operating data, selecting discharge plateau voltage (i.e., open-circuit voltage curve), relaxation voltage curve features, cumulative charge / discharge capacity (such as real-time SOC, depth of discharge (DOD), etc.), and state of charge change rate as input features. Discharge plateau voltage, relaxation voltage curve features, cumulative charge / discharge capacity, and state of charge change rate can reflect the changes in battery internal resistance and the degree of degradation, providing a basis for health status and capacity prediction.

[0156] It should be noted that, in some implementations, the real-time SOC can be used as input to the battery state estimation model, and is used together with other battery operating data to input into the battery state estimation model. The battery state estimation model can estimate the battery health status and / or battery capacity based on the input operating data. In some implementations, the SOC can be used as output to the battery state estimation model, and the battery's open-circuit voltage, current, and other operating data are input into the battery state estimation model. The battery state estimation model can estimate the battery's SOH and / or SOC based on the input operating data.

[0157] Step S403: Local neural network model construction and training.

[0158] At each vehicle-side node, a deep neural network can be constructed based on extracted features to serve as a local battery state estimation model. All vehicles can use the same type of features and the same network structure, but due to the different data sources for each vehicle-side node, the parameters of the locally trained local models will differ. The local battery state estimation model can be trained using historical operational data collected locally at each vehicle (such as operational data collected during the historical charge-discharge cycles of the local battery) and operational data generated by digital twins, yielding the local model parameters for each vehicle. During training, a mean squared error loss function can be used to enable the battery state estimation model to map input features to output battery capacity or SOH value.

[0159] Step S404, Model parameter coordination and fusion.

[0160] Each vehicle in the vehicle group periodically or triggered by a specific event uploads its trained local model parameters to the aggregation node. The aggregation node collects all uploaded local model parameters and calculates the prediction error or confidence index (corresponding to the first confidence index in the aforementioned embodiment) for each local model based on a pre-reserved validation dataset. Based on the confidence index, the parameter weights of each local model parameter are calculated; a higher parameter weight indicates greater confidence in the local model parameter. Then, a weighted average or other fusion strategy is used to fuse the local model parameters into global model parameters (corresponding to the fused model parameters in the aforementioned embodiment). After completing the fusion operation, the aggregation node distributes the latest global model parameters to other vehicle terminals, and the vehicle group continuously iteratively updates the local battery state estimation model of each vehicle based on the received global model parameters.

[0161] Step S405, global estimation output.

[0162] The aggregation node publishes the fused global model parameters to each vehicle terminal for real-time estimation of the local battery health status and / or battery capacity. Each vehicle terminal can apply the fused global model parameters to its local battery state estimation model to obtain a fused model. The fused model inputs current features extracted from the local battery's current operating data and outputs the current SOH and / or capacity prediction results for the local battery. Based on these SOH and / or capacity prediction results, the vehicle's BMS can optimize charging and discharging strategies or predict remaining battery life.

[0163] In this embodiment, a group learning framework is used to integrate and distribute the training results among the various vehicles, thereby ensuring that the trained battery state estimation model can effectively utilize the data provided by each vehicle and has strong generalization ability and high estimation accuracy.

[0164] In the swarm learning framework, a vehicle swarm consists of multiple vehicle terminals. The process of training a model using the swarm learning framework involves initialization steps, multiple synchronization (parameter fusion) operations, and reaching training standards to complete the training.

[0165] Initialization involves each vehicle initializing its own battery state estimation model and beginning group learning.

[0166] Multiple fusions refer to the process where each vehicle uses the results of its local model training (i.e., local model parameters) to refine the global model. The group learning process ends only after a given number of fusion rounds or when certain set criteria are met. The global model refers to the battery state estimation model that applies the global model parameters.

[0167] Figure 5 This application provides a schematic diagram of an implementation process for model training using a group learning framework, as shown in the embodiments below. Figure 5 As shown, the training process includes the following steps S501 to S510: Step S501, node model initialization; Each vehicle initializes its local battery state estimation model and begins group learning.

[0168] Step S502: Verify the second confidence level of the global model parameters from the previous round; The validation dataset is input into the battery state estimation model obtained after the previous round of fusion (i.e., the global model of the previous round) on any vehicle end. The prediction accuracy of the battery state estimation model on the validation dataset is obtained and used as the second confidence of the parameters of the global model of the previous round.

[0169] Step S503: Train a local battery state estimation model; Each vehicle independently trains its local battery state estimation model using its own local training dataset for a finite number of epochs.

[0170] For example, each vehicle can train a local battery state estimation model at synchronization intervals.

[0171] Step S504: Verify the first confidence level of the local model parameters in the current round; The vehicle inputs the validation dataset into the local battery state estimation model trained for the current round, and obtains the prediction accuracy of the local battery state estimation model on the validation dataset, which is used as the first confidence level of the local model parameters for the current round.

[0172] Step S505: Share the results and update the local model parameters; The aggregation node extracts the local model parameters of the current wheel from each vehicle and the corresponding weights for each local model parameter. It then merges the local model parameters of the current wheel from all vehicles using a weighted average to obtain the global model parameters for the current wheel, and sends these global model parameters to each vehicle. Each vehicle uses the global model parameters of the current wheel as its updated local model parameters to obtain the updated global model. At this point, the local battery state estimation models of all vehicles are identical, all being global models.

[0173] Step S506: Verify the second confidence level of the global model parameters in the current round; The validation dataset is input into the battery state estimation model (i.e., the global model of the current wheel) obtained by fusing the current wheel of any vehicle. The prediction accuracy of the battery state estimation model on the validation dataset is obtained and used as the second confidence level of the global model parameters of the current wheel.

[0174] Step S507: Determine whether the training termination condition is met; If yes, proceed to step S508; if no, proceed to step S509.

[0175] Step S508: Test the estimation accuracy of the global model; Step S509: Update parameter weights; Step S510: Output the result.

[0176] Each vehicle determines whether the training termination condition has been met. If it has, proceed to step S508 to end group learning, input the test dataset into the global model, obtain the latest prediction accuracy of the global model, and output the result; if it has not been met, update the parameter weights to obtain the parameter weights of each local model in the next round of parameter fusion, and proceed to the next round of training.

[0177] In some implementations, the parameter weights can be updated using a confidence weight prediction algorithm.

[0178] The verification dataset and test dataset used by each vehicle can be shared among all vehicles.

[0179] In some implementations, the output of step S510 may include, but is not limited to, the prediction accuracy of the latest global model, which can reflect the results of this group learning.

[0180] In the process of sharing learning outcomes in group learning, the fusion method of local model parameters can be at least one of the following: weighted average, arithmetic average, taking the maximum value, taking the minimum value, etc.

[0181] In some implementations, a weighted average algorithm can be used to fuse the local model parameters of each vehicle. The parameter weights corresponding to the local model parameters of each vehicle can be determined using a confidence weight prediction algorithm. In this confidence weight prediction algorithm, the parameter weights corresponding to the local model parameters of vehicle i can be calculated as shown in Formula 1 below: (1); Where i is a positive integer. and These are the count values ​​of the positive counter on vehicle i (corresponding to the first count value in the aforementioned embodiment) and the count values ​​of the negative counter (corresponding to the second count value in the aforementioned embodiment). This represents the weight offset value. After each round of local model parameter fusion, the first confidence level corresponding to vehicle i in the current round is... The second confidence level of the global model parameters after the current round of fusion is First credibility Second credibility This can be used as input to the credibility weight prediction algorithm. Through the credibility weight prediction algorithm, the first credibility corresponding to the current wheel of vehicle i can be determined. The second confidence level of the global model parameters after fusion with the current round is Comparison, if the first confidence level of vehicle i corresponding to the current wheel... If the value is higher, then the count value of the positive counter at vehicle terminal i will be increased. Increment by one, otherwise increment the count value of the negative counter at vehicle i. Increment by one. After each round of local model parameter fusion, divide the positive counter value of each vehicle by the sum of the positive and negative counter values, and then add a constant. As a weight offset, the confidence prediction weight of each vehicle in the current round can be obtained, which serves as the parameter weight of the local model parameters for the corresponding vehicle in the next round.

[0182] According to the credibility weight prediction algorithm provided in this application, significant weight changes can be achieved in the early stages when there are fewer fusion iterations, enabling the model to quickly converge to a compliant range. In the later stages with more fusion iterations, the weight changes are smaller and more precise, allowing for minor adjustments to the parameter weights of each vehicle to accurately determine its credibility. Simultaneously, the aggressiveness of the weight prediction can be well controlled under the influence of the hyperparameter weight offset value. For example, when the weight offset value is 0, the weight values ​​of each vehicle range from 0 to 1. Vehicles with higher first credibility have parameter weights closer to 1, and vice versa. Therefore, in parameter weighted averaging, the local model parameters of vehicles with lower first credibility have a small impact on the global model parameters. Furthermore, when the weight offset value is large, such as 3, the weight values ​​of each vehicle range from 3 to 4. In this case, the difference in the impact of the local model parameters of each vehicle on the global model parameters in parameter weighted averaging will be much smaller than when the weight offset value is 0. The credibility weight prediction algorithm is suitable for situations where some vehicles participating in group learning have low-quality local training datasets. The local model parameters generated during training by these low-quality vehicles often perform poorly on the validation dataset, resulting in lower initial credibility. This credibility weight prediction algorithm reduces the impact of these low-credibility local model parameters on the global model parameters during parameter fusion. Conversely, high-quality local model parameters from vehicles have a stronger influence during the fusion process.

[0183] In summary, the embodiments of this application achieve collaborative battery state estimation under the premise of data privacy protection by introducing a dynamic credibility assessment mechanism and a group learning framework. Furthermore, digital twins can be used to aggregate large amounts of battery data throughout its entire lifecycle, and deep neural networks can be used to predict battery aging trends. The credibility-weighted fusion mechanism suppresses low-quality models and emphasizes high-credibility models, thereby significantly improving the accuracy and robustness of battery state estimation.

[0184] This application provides a state estimation device, such as... Figure 6 As shown, the state estimation device 60 includes: an acquisition module 61 and an estimation module 62, wherein: The acquisition module 61 is used to acquire the operating data of the local battery on the electrical device; The estimation module 62 is used to estimate the battery state of the local battery using a battery state estimation model to obtain an estimation result; wherein the battery state estimation model is trained using the model training method described above.

[0185] This application provides a battery management system for implementing the steps in the above-described state estimation method.

[0186] This application provides a battery system, such as... Figure 7 As shown, the battery system 70 includes at least one battery 71 and the aforementioned battery management system 72.

[0187] This application provides an embodiment of an electrical device, such as... Figure 8 As shown, the electrical device 31 includes the aforementioned battery system 70.

[0188] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the state estimation method described above.

[0189] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the state estimation method described above.

[0190] It should be noted that the descriptions of the various embodiments above tend to emphasize the differences between them, while their similarities or commonalities can be referred to interchangeably. The descriptions of the system, device, apparatus, storage medium, and program product embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the system, device, apparatus, storage medium, and program product embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0191] In this application embodiment, if the above-described method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application embodiment, in essence or the part that contributes to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a control device to execute all or part of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, this application embodiment is not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0192] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, the functional units in the embodiments of this application may all be integrated into one processing unit, or each unit may be a separate unit, or two or more units may be integrated into one unit; the integrated unit may be implemented in hardware or in a combination of hardware and software functional units.

[0193] It should be understood that in the description of this application, the reference to terms such as "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," "in some implementations," "in other implementations," or "exemplary," etc., refers to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the embodiments of this application. In this application, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine the different embodiments or examples described in this application, as well as the features of the different embodiments or examples.

[0194] 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.

[0195] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and devices can be implemented in other ways. The apparatus and device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0196] The above are merely exemplary embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A model training method, characterized in that, Applied to a group of electrical equipment, wherein the group of electrical equipment includes multiple electrical devices, the method includes: Each of the aforementioned electrical devices uses a local training dataset to train a local battery state estimation model, obtains and sends the local model parameters to the aggregation node; the local training dataset includes at least one set of running data of the local battery of the electrical device and a battery state label corresponding to each set of running data, and the aggregation node is selected from each of the aforementioned electrical devices; The aggregation node performs weighted fusion of the local model parameters based on the parameter weights of each of the electrical devices, and then distributes the fused model parameters to each of the electrical devices. The parameter weights of the electrical devices are determined based on the credibility of the corresponding local model parameters. The credibility of the local model parameters includes the credibility of the local model parameters on the validation dataset. The higher the credibility of the local model parameters, the greater the parameter weight of the corresponding electrical device. The validation dataset is shared by all the electrical devices, and the parameter weights in the current round are determined by comparing the first credibility of the local model parameters obtained from the previous round of training on the validation dataset and the second credibility of the fused model parameters obtained from the previous round of training on the validation dataset. Each of the aforementioned electrical devices uses the fusion model parameters to update its local battery state estimation model.

2. The model training method according to claim 1, characterized in that, The weighted fusion of the local model parameters includes: Obtain the parameter weights of each of the aforementioned electrical devices in the current round; Based on the parameter weights of each of the electrical devices in the current round, the local model parameters obtained from the training of each of the electrical devices in the current round are weighted and fused to obtain the fused model parameters for the current round.

3. The model training method according to claim 2, characterized in that, Also includes: The electrical equipment uses a validation dataset to validate the local model parameters obtained from the current round of training, thereby obtaining the first confidence level of the electrical equipment in the current round. The aggregation node uses the verification dataset to verify the parameters of the fusion model obtained in the current round of training, and obtains the second confidence level corresponding to the current round. The aggregation node determines the parameter weights for the next round of each of the electrical devices based on the first credibility and the second credibility of each of the current rounds.

4. The model training method according to claim 3, characterized in that, The step of determining the parameter weights for the next round of each of the electrical devices based on the first confidence level and the second confidence level corresponding to the current round includes: If the first confidence level of the current wheel of the electrical equipment is greater than the second confidence level of the current wheel, the first count value corresponding to the electrical equipment is increased by the first step length; Based on the first count value corresponding to the electrical equipment, the parameter weights of the electrical equipment in the next round are determined; the parameter weights of the electrical equipment in the next round are positively correlated with the corresponding first count value.

5. The model training method according to claim 4, characterized in that, The step of determining the parameter weights for the next round of each of the electrical devices based on the first confidence level and the second confidence level corresponding to the current round of each device further includes: If the first confidence level of the current wheel of the electrical equipment is less than the second confidence level of the current wheel, the second count value corresponding to the electrical equipment is increased by a second step. Based on the first count value corresponding to the electrical equipment, determine the parameter weights of the electrical equipment in the next round, including: Based on the first and second count values ​​corresponding to the electrical equipment, the parameter weights of the electrical equipment in the next round are determined; the parameter weights of the electrical equipment in the next round are negatively correlated with the corresponding second count value.

6. The model training method according to claim 5, characterized in that, The step of determining the parameter weights of the electrical equipment in the next round based on the first and second count values ​​corresponding to the electrical equipment includes: A weighted reference value is determined based on the first and second count values ​​corresponding to the electrical equipment; the weighted reference value is positively correlated with the first count value and negatively correlated with the second count value. The parameter weights for the next round of the electrical equipment are determined based on the sum of the weight reference value and the set weight offset value.

7. The model training method according to any one of claims 1 to 6, characterized in that, Also includes: If the training termination condition is not met, each of the electrical devices uses its local training dataset to train its local battery state estimation model for the next round, and obtains and sends the local model parameters for the next round to the aggregation node in each of the electrical devices; the training termination condition includes at least one of the following: the difference between the second confidence level corresponding to the current round and the second confidence level corresponding to the previous round converges, and the training rounds reach the preset number of times. The aggregation node performs a weighted fusion of the local model parameters of each of the electrical devices in the next round based on the parameter weights of each device in the next round, and then sends the fused model parameters of the next round to each of the electrical devices. Each of the aforementioned electrical devices uses the parameters from the next round of fusion model to update its local battery state estimation model.

8. The model training method according to any one of claims 1 to 6, characterized in that, Also includes: The electrical device collects at least one set of first operating data from the local battery and generates a first training dataset based on the at least one set of first operating data; The electrical equipment uses a digital twin of the local battery to simulate the operating behavior of the local battery under at least one operating condition, obtains at least one set of second operating data, and generates a second training dataset based on the at least one set of second operating data. The electrical equipment determines its local training dataset based on the first training dataset and the second training dataset.

9. A state estimation method, characterized in that, include: Obtain operating data of the local battery on the electrical device; The battery state of the local battery is estimated using a battery state estimation model to obtain an estimation result; wherein the battery state estimation model is trained using the model training method as described in any one of claims 1 to 8.

10. A group of electrical equipment, characterized in that, It includes multiple electrical devices, which are used to implement the model training method as described in any one of claims 1 to 8.

11. A state estimation device, characterized in that, include: The acquisition module is used to acquire the operating data of the local battery on the electrical device; An estimation module is used to estimate the battery state of the local battery using a battery state estimation model to obtain an estimation result; wherein the battery state estimation model is trained using the model training method as described in any one of claims 1 to 8.

12. A battery management system, characterized in that, Used to implement the steps in the method of claim 9.

13. A battery system, characterized in that, It includes at least one battery and the battery management system as described in claim 12.

14. An electrical appliance, characterized in that, Includes the battery system as described in claim 13.

Images