Model training method, electronic device, storage medium, and program product

Abstract

The application provides a model training method, an electronic device, a storage medium and a program product, relates to the technical field of computers, and is used for fusing clustering results in a model, and realizing vehicle battery health degree prediction with adaptive adjustment of multi-task weights. The method comprises the following steps: acquiring a source domain training sample set; performing clustering processing on source domain vehicle operation data to obtain a plurality of clustering clusters; determining battery health degree attenuation modes of the clustering clusters; generating clustering feature vectors corresponding to the clustering clusters; embedding each clustering feature vector into a classification network of a to-be-trained model to obtain an embedded model; inputting the source domain vehicle operation data into the embedded model to obtain source domain predicted battery health degree attenuation data corresponding to the source domain training sample; determining source domain prediction loss based on the source domain predicted battery health degree attenuation data and source domain battery health degree attenuation data; and updating parameters of the embedded model and the clustering feature vectors based on a plurality of source domain prediction losses to obtain a target model.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a model training method, electronic device, storage medium, and program product. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the accurate assessment of the state of health (SOH) of power batteries, as core components, is of great significance for ensuring safe vehicle operation, optimizing battery management strategies, and evaluating battery residual value. Battery health degradation is a complex nonlinear electrochemical process, influenced by various factors such as temperature, charge / discharge rate, and driving habits, exhibiting diverse degradation modes.

[0003] Currently, existing battery health assessment technologies mainly rely on establishing equivalent circuit models or data-driven machine learning models. However, these technologies often suffer from insufficient generalization ability when faced with the complex operating conditions of real vehicles. Traditional single models struggle to simultaneously account for the differentiated degradation characteristics of batteries at different stages or scenarios, leading to significant discrepancies between predicted results and actual operating conditions. Summary of the Invention

[0004] This application provides a model training method, electronic device, storage medium, and program product. By clustering vehicle operation data, it matches dedicated expert networks for different battery degradation modes and performs weighted fusion prediction to solve the problems of huge discrepancies between current evaluation results and actual working conditions and poor model generalization ability.

[0005] Firstly, this application provides a model training method, comprising: acquiring a source domain training sample set; the source domain training sample set includes multiple source domain training samples, each source domain training sample including source domain vehicle operation data and corresponding source domain battery health degradation data; performing clustering processing on the source domain vehicle operation data in the multiple source domain training samples to obtain multiple clusters; determining the battery health degradation mode of each cluster based on the source domain battery health degradation data corresponding to each cluster; generating a clustering feature vector corresponding to each cluster based on the cluster center corresponding to each cluster; embedding the clustering feature vector corresponding to each cluster into the classification network of the model to be trained to obtain an embedded model; the embedded model includes a classification network and multiple expert networks, wherein the classification network is used to perform clustering identification on the input vehicle operation data. The weights of each expert network are determined, and each expert network corresponds to a battery health degradation mode. These networks process the input vehicle operation data and output corresponding battery health degradation prediction results. Source domain vehicle operation data from the source domain training samples are input into the embedding model. Based on the weights of each expert network output by the classification network, the sub-prediction results output by each expert network are weighted and fused to obtain the source domain predicted battery health degradation data corresponding to the source domain training samples. Based on the source domain predicted battery health degradation data and the source domain battery health degradation data corresponding to the source domain training samples, the source domain prediction loss corresponding to the source domain training samples is determined. Based on the source domain prediction losses corresponding to multiple source domain training samples, the parameters and clustering feature vectors of the embedding model are updated to obtain the target model.

[0006] The technical solution provided in this application offers at least the following benefits: By clustering source domain vehicle operation data, various battery health degradation modes are effectively identified, and a unique clustering feature vector is generated for each mode. Based on this, an embedding model comprising a classification network and multiple expert networks is constructed. The clustering feature vector is used to accurately identify the degradation mode of the input sample, and the weights of each expert network are dynamically determined. By weighted fusion of the sub-prediction results of each expert network, the model can adaptively output high-precision prediction results for different degradation modes. Finally, based on the loss feedback between predicted and real data, the model parameters and clustering feature vectors are updated synchronously, enabling the model to continuously optimize its fitting ability to complex degradation modes, thereby significantly improving the accuracy of battery health degradation prediction and the model's generalization ability.

[0007] In one possible implementation, the weights of each expert network are determined as follows: for any source domain training sample, the feature distance between the source domain vehicle operation data and each cluster feature vector in the source domain training sample is determined, and the feature distance is normalized to obtain the weights of each expert network corresponding to the source domain training sample.

[0008] In one possible implementation, after obtaining the target model, the method further includes: acquiring a target domain training sample set; performing transfer learning on the target model based on the source domain training sample set and the target domain training sample set to obtain a transfer model; wherein the source domain training sample set is obtained based on a vehicle model and / or battery, and the target domain training sample set is obtained based on a different vehicle model and / or battery than the source domain training sample set, and the target domain training sample set includes multiple target domain training samples, which include target domain vehicle operation data and corresponding target domain battery health degradation data.

[0009] In one possible implementation, the source domain training sample set further includes source domain battery electrochemical features, the target domain training sample set further includes target domain battery electrochemical features, and the source domain battery health degradation data includes source domain battery health degradation curves. Transfer learning is performed on the target model based on the source domain training sample set and the target domain training sample set to obtain a transfer model, including: determining the source domain battery aging rate based on the source domain battery health degradation curves; inputting target domain vehicle operation data into the target model to obtain target domain predicted battery health degradation data, and determining the target domain battery aging rate based on the target domain predicted battery health degradation data; determining physical loss based on the source domain battery aging rate and the target domain battery aging rate; and characterizing the source domain battery health. The study examines whether the monotonicity of the decay curve is consistent with that of the battery health decay curve predicted in the target domain; determines the physical weights based on the electrochemical feature similarity between the source domain battery electrochemical features and the target domain battery electrochemical features; the physical weights are positively correlated with the electrochemical feature similarity between the source domain battery electrochemical features and the target domain battery electrochemical features; determines the transfer loss based on the task loss, physical loss, and physical weights, and updates the parameters and clustering feature vectors of the target model based on multiple transfer losses to obtain the transfer model; the task loss characterizes the difference between the source domain prediction accuracy of the target model when predicting battery health decay data based on source domain vehicle operation data and the target domain prediction accuracy when predicting battery health decay data based on target domain vehicle operation data.

[0010] In one possible implementation, before updating the parameters and clustering feature vectors of the target model based on multiple transfer losses to obtain the transfer model, the method further includes: determining the parameter adjustment range when performing transfer learning on the target model based on electrochemical feature similarity; the parameter adjustment range is positively correlated with the electrochemical feature similarity; updating the parameters and clustering feature vectors of the target model based on multiple transfer losses to obtain the transfer model includes: updating the parameters and clustering feature vectors of the target model within the parameter adjustment range based on multiple transfer losses to obtain the transfer model.

[0011] In one possible implementation, the source domain training sample set also includes source domain battery temperature distribution features, and the target domain training sample set also includes target domain battery temperature distribution features. When the source domain training sample set and the target domain training sample set are obtained based on the same battery but different vehicle models, the transfer loss is determined based on task loss, physical loss, and physical weights. This includes: determining temperature loss based on the source domain battery temperature distribution features and the target domain battery temperature distribution features; the temperature loss characterizes the temperature distribution difference between the source domain battery temperature distribution features and the target domain battery temperature distribution features; determining the transfer loss based on task loss, physical loss, physical weights, and temperature loss; and the temperature loss is used to instruct the transfer model to compensate for the battery health degradation caused by the temperature distribution difference.

[0012] In one possible implementation, the battery health degradation data includes battery aging rate, remaining lifespan, drop point location, and drop probability. The drop point represents the turning point where battery health undergoes a sudden change. The battery aging rate and remaining lifespan output by the target model satisfy the condition that the difference between the theoretical remaining lifespan determined based on the battery aging rate and the actual remaining lifespan is less than a preset difference. After obtaining the target model, the method further includes: acquiring current vehicle operation data; inputting the vehicle operation data into the target model to obtain current battery health degradation data; when the current drop probability is greater than a preset probability threshold, using the current drop point location as a dividing point, dividing the degradation process corresponding to the remaining lifespan into a stable segment before the drop and an acceleration segment after the drop, and predicting the remaining lifespan for each segment; wherein, the weights of the expert network in the target model are different when predicting the remaining lifespan for different segments.

[0013] Secondly, this application provides a model training apparatus, comprising: The acquisition module is used to acquire the source domain training sample set; the source domain training sample set includes multiple source domain training samples, which include source domain vehicle operation data and corresponding source domain battery health degradation data. The processing module is used to perform clustering processing on the source domain vehicle operation data in multiple source domain training samples to obtain multiple clusters; The processing module is also used to determine the battery health degradation mode of each cluster based on the source domain battery health degradation data corresponding to each cluster. The processing module is also used to generate cluster feature vectors corresponding to each cluster based on the cluster center corresponding to each cluster. The processing module is also used to embed the clustering feature vector corresponding to each cluster into the classification network of the model to be trained to obtain the embedding model. The embedding model includes a classification network and multiple expert networks. The classification network is used to identify clusters of the input vehicle operation data and determine the weights of each expert network. Each expert network corresponds to a battery health degradation mode and is used to process the input vehicle operation data and output the corresponding battery health degradation prediction results. The training module is used to input the source domain vehicle operation data in the source domain training samples into the embedding model. Based on the weights of each expert network output by the classification network, the sub-prediction results output by each expert network are weighted and fused to obtain the source domain predicted battery health degradation data corresponding to the source domain training samples. The training module is also used to determine the source domain prediction loss corresponding to the source domain training samples based on the source domain predicted battery health decay data and the source domain battery health decay data corresponding to the source domain training samples. The training module is also used to update the parameters and clustering feature vectors of the embedded model based on the source domain prediction loss corresponding to multiple source domain training samples, so as to obtain the target model.

[0014] Thirdly, this application provides an electronic device, including: a processor and a memory, the processor being coupled to the memory; the memory being used to store computer instructions, which are loaded and executed by the processor to enable the computer device to implement the method described in the first aspect.

[0015] Fourthly, this application provides a computer-readable storage medium comprising: computer software instructions; when the computer software instructions are executed in an electronic device, they cause the electronic device to implement the method described in the first aspect.

[0016] Fifthly, this application provides a computer program product comprising a computer program; when the computer program is run in an electronic device, it causes the electronic device to implement the method described in the first aspect.

[0017] The beneficial effects of the second to fifth aspects mentioned above are described in the corresponding description of the first aspect and will not be repeated here. Attached Figure Description

[0018] Figure 1 This application provides a schematic diagram of the architecture of a vehicle battery health assessment system. Figure 2 A schematic flowchart illustrating a model training method provided in an embodiment of this application; Figure 3 This is a schematic diagram of a clustering result provided in an embodiment of this application; Figure 4This is a schematic diagram of a network structure for an embedded model provided in an embodiment of this application; Figure 5 A schematic diagram of a transfer learning process provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0020] The terms "first" and "second," etc., used in the specification and drawings of this application are used to distinguish different objects or to distinguish different treatments of the same object, rather than to describe a specific order of objects.

[0021] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0022] It should be noted that in the embodiments of this application, the words "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0023] To facilitate a clear description of the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with essentially the same function and effect. Those skilled in the art can understand that the terms "first" and "second" are not intended to limit the quantity or execution order.

[0024] In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0025] State of Health (SOH) assessment has become a core component in ensuring safe vehicle operation and optimizing battery management strategies. However, the real-world vehicle operating environment is highly dynamic and random, with current rate, ambient temperature, and depth of discharge constantly undergoing non-steady-state changes, fundamentally different from ideal laboratory conditions. While current battery health assessment technologies have made some progress, they often exhibit limitations when faced with complex and variable real-world vehicle data, such as insufficient generalization ability, strong data dependence, and inadequate feature mining, making it difficult to meet the demand for accurate predictions across vehicle models and the entire battery lifecycle. Specifically, current technologies mainly suffer from the following shortcomings: The contradiction between operational adaptability and data dependence: Constant current and constant temperature tests under laboratory conditions cannot truly reflect the randomly changing current and environmental stress during actual vehicle operation. Current methods such as curve analysis heavily rely on complete charge and discharge segments, while actual vehicle charging behavior is often fragmented and irregular, resulting in a large amount of incomplete real-vehicle data that cannot be effectively utilized, leading to significant discrepancies between evaluation results and actual operating conditions.

[0026] Limitations of Model Generalization and Cross-Domain Transfer: Models calibrated for different battery cells or specific vehicle models exhibit significant data distribution shifts when facing different vehicle models or new usage scenarios due to differences in battery system design, thermal management strategies, and control logic. Current transfer learning methods are mostly simple porting of general frameworks, lacking in-depth consideration of the physical laws of battery aging (such as differences in electrochemical systems and the impact of thermal management), resulting in weak model generalization ability across vehicle models and operating conditions, and high adaptation costs.

[0027] Limited feature mining dimensions and lack of degradation pattern identification: Traditional solutions (such as SOH estimation based on random forests) mainly focus on the battery's own electrothermal parameters, neglecting the coupled influence of vehicle status (driving conditions, energy consumption) and user behavior habits (driving style, charging habits) on aging. Furthermore, battery aging trajectories exhibit diversity (e.g., linear degradation, precipitous degradation). Existing technologies typically use a uniform regression model to process all samples, lacking clustering identification and differentiated modeling for specific degradation patterns. This results in insufficient early warning capabilities for nonlinear, drastic degradation such as "precipitous" degradation, and an inability to provide joint prediction information on aging rate, remaining lifespan, and other multi-dimensional dimensions.

[0028] To address the aforementioned issues, this application provides a model training method, electronic device, storage medium, and program product. By clustering vehicle operation data, it matches dedicated expert networks to different battery degradation modes and performs weighted fusion prediction, thereby at least solving the problems of significant discrepancies between current evaluation results and actual operating conditions, and poor model generalization ability.

[0029] The embodiments provided in this application will now be described in detail with reference to the accompanying drawings.

[0030] Figure 1 This is a schematic diagram of the architecture of a vehicle battery health assessment system provided in an embodiment of this application, such as... Figure 1 The vehicle battery health assessment system 100 includes an in-vehicle terminal 110, an edge node 120, a cloud server 130, and a user terminal 140.

[0031] The vehicle terminal 110 acquires vehicle operation data from sources such as the controller area network (CAN) bus, vehicle controller, charging records, driving logs, battery management system (BMS), and battery historical data. The vehicle terminal then sends the vehicle operation data acquired by 110 to the edge node 120.

[0032] In one implementation, vehicle operating data may include any one or more of the following: average vehicle speed, acceleration distribution, mileage, energy consumption rate, air conditioning usage intensity, state of charge (SOC) distribution at the start of charging, charging frequency, depth of discharge distribution, frequency of rapid acceleration / deceleration, characteristics of charging / discharging voltage curves, estimated internal resistance, temperature distribution, and characteristics of capacity increment curves.

[0033] Edge node 120 is used to clean, segment, and perform feature statistics on the acquired vehicle operation data, which will then be used as input for subsequent models.

[0034] In one implementation, data cleaning / preprocessing includes operations such as removing outliers and filling in missing values; segmentation can be the division of time-series data into segments, such as by charging events, driving events, and stationary events; feature statistics refer to calculating the statistical features within each event, such as mean, variance, extreme values, slope, and integral area; finally, a fixed-length feature vector is constructed as the input for subsequent model training.

[0035] In one implementation, the edge device can maintain a small-scale sample cache pool locally (containing data from the most recent 1000 charging cycles). Before the model version update is distributed from the cloud, the newly added data in the local cache pool is used to perform online incremental learning on the lightweight edge model. For example, the elastic weightconsolidation (EWC) method is used to add additional regularization constraints to the parameters in the model that are related to existing knowledge to prevent catastrophic forgetting. The incremental learning process is executed in a low-power mode during vehicle charging periods (non-driving state) to avoid affecting the normal use of the vehicle. The model parameters after incremental learning are uploaded to the cloud periodically (e.g., once a week) via the vehicle network for aggregation and reference for version iteration of the main model in the cloud.

[0036] The cloud server 130 acquires the feature vectors preprocessed by the edge node 120 as the source domain training sample set, which includes multiple source domain training samples. The source domain training samples include source domain vehicle operation data and corresponding source domain battery health degradation data. The cloud server can perform unsupervised clustering on the source domain vehicle operation data, embed the clustering results into the model to be trained, and train the model.

[0037] In one implementation, due to varying environmental temperature distributions, road conditions, and charging infrastructure across different regions, data distribution exhibits regional differences. Therefore, a federated learning framework can be deployed in the cloud, with edge models from each region (divided by climate zone or geographical area) serving as participating nodes in the federated learning process. Each regional node independently trains or fine-tunes its model on local data; each node only uploads model parameter updates (gradients or weight increments), without needing to upload the original data, thus protecting user privacy and data security; the cloud server performs weighted aggregation of the parameter updates uploaded by each node (weights are determined by the data volume and model performance of each region), generating a global model version; the global model version is then distributed to each regional node, achieving cross-regional knowledge sharing while maintaining the local adaptability of each regional model.

[0038] In one implementation, the cloud server can also perform transfer learning on the trained target model based on the target domain training sample set to obtain a transfer model. The cloud server can store multiple versions of the model and distribute the model to the edge nodes 120 according to actual needs.

[0039] In one implementation, the model trained on the cloud server (containing all four task branches) has approximately 15M-25M parameters, making it difficult to run in real-time on in-vehicle edge computing units (typically with a computing power of 1-5 TOPS). Therefore, a layered deployment strategy can be adopted: for example, a lightweight inference model is deployed at the edge. This lightweight model contains only a lightweight network with a SOH estimation branch and an aging rate prediction branch, with parameters compressed to 2M-4M and inference latency <50ms. Knowledge distillation technology is used, with the complete cloud model as the teacher network and the lightweight edge model as the student network, training the lightweight model through distillation loss. The complete model is retained in the cloud: the complete model containing all task branches is deployed in the cloud for offline batch analysis and in-depth evaluation of complex scenarios. Condition-triggered cloud inference: when the edge detects the following trigger conditions, the current data is packaged and uploaded to the cloud for complete model inference: the SOH estimate decreases by more than 3% compared to the previous period (abnormal decay may occur); the aging rate prediction exceeds the upper limit of the normal range by 1.5 times; and the vehicle's cumulative mileage increases by 10,000 kilometers.

[0040] In one implementation, both the cloud server and the edge node 120 can perform State of Health (SOH) assessments based on the acquired vehicle operating data and send the assessment results to the user terminal 140. Simultaneously, the cloud server can also generate a battery health report or battery health warning based on the SOH assessment results and send it to the user terminal 140.

[0041] The client 140 is used to receive the SOH assessment results sent by the cloud server or edge node 120 and to display them visually.

[0042] In one implementation, the user terminal can also generate lifespan warnings, maintenance recommendations, vehicle residual value assessments, and vehicle battery health reports based on the SOH assessment results.

[0043] The aforementioned vehicle terminal 110, edge node 120, cloud server 130, and user terminal 140 can be electronic devices, and they can communicate with each other via wired or wireless means.

[0044] The electronic device can be a server, such as a server cluster consisting of multiple servers, a single server, a computer, or a processor or processing chip within a server or computer. For example, the electronic device can also be a terminal, such as a mobile phone, tablet computer, wearable device, in-vehicle device, laptop computer, or ultra-mobile personal computer. Mobile personal computers (UMPCs), netbooks, personal digital assistants (PDAs), etc. It should be noted that the embodiments of this application do not limit the specific device form of the electronic equipment.

[0045] Figure 2 This is a flowchart illustrating a model training method provided in an embodiment of this application. The model training method provided in this application can be implemented via a cloud server.

[0046] The model trained on the aforementioned cloud server can be integrated into automakers' cloud-based big data platforms (Internet of Vehicles) to provide real-time battery health monitoring and early warning for vehicles. It can also be deployed as a core algorithm on third-party battery health assessment service platforms to provide analysis services. Furthermore, this model can provide battery life and residual value assessment data for used new energy vehicle battery residual value assessment systems, and provide data support for new energy vehicle insurance actuarial and risk assessment systems to develop more reasonable battery-related insurance products.

[0047] The services offered by this platform can ultimately be implemented across various hardware terminals and operational scenarios. In terms of vehicle type, it covers battery health assessments for various models, including battery electric vehicles (BEVs), plug-in hybrid electric vehicles (PHEVs), and extended-range electric vehicles (EREVs). Regarding operational scenarios, it can be applied to battery operation management and early warning for ride-sharing or car-hailing platforms, battery health grading, scheduling and retirement decisions for battery swapping stations, battery screening and lifespan prediction in secondary utilization scenarios, and lifespan prediction and operation and maintenance decisions for energy storage systems.

[0048] like Figure 2 As shown, the method specifically includes S201~S208: S201. Obtain the source domain training sample set.

[0049] Model training requires sufficient labeled data. The source domain refers to the data area that already has complete annotations (i.e., known battery health degradation results), typically derived from a large amount of historically accumulated vehicle test data. Obtaining a training sample set in the source domain provides learning material for subsequent model training, enabling the model to learn the mapping pattern between vehicle operating characteristics and battery degradation from existing data.

[0050] The source domain training sample set includes multiple source domain training samples, which include source domain vehicle operation data and corresponding source domain battery health degradation data. Here, "source domain" refers to an existing domain with sufficient labeled data, as opposed to a target domain lacking labeled data. In this embodiment, the source domain is the set of vehicle operation and battery degradation data; the source domain vehicle operation data consists of actual operating parameters from the source domain vehicles, such as time-series characteristic data like charging / discharging current, voltage, temperature, SOC, and mileage; the source domain battery health degradation data is labeled data showing the change in battery health (SOH) over time or cycle count corresponding to the operation data, which is the label the model needs to learn.

[0051] Specifically, electronic devices can extract a large amount of operating logs (such as CAN bus data) and their corresponding battery SOH calibration data from retired or long-term operating vehicles from vehicle networking platforms or historical test databases. After data cleaning (removing outliers and filling in missing values), time alignment, and standardization, a structured training sample set is formed. Each sample consists of a set of operating feature sequences and the corresponding SOH decay value.

[0052] S202. Perform clustering processing on the source domain vehicle operation data in multiple source domain training samples to obtain multiple clusters.

[0053] The usage scenarios of different vehicles (such as urban commuting, long-distance freight, and operation in high / low temperature regions) vary greatly, resulting in different battery degradation patterns. By using clustering, vehicles with similar operating characteristics can be grouped together, thereby identifying different usage patterns.

[0054] Specifically, feature extraction (such as average charge / discharge rate, average temperature, average daily mileage, and discharge depth distribution) is performed on the source domain vehicle operation data. Then, clustering algorithms such as K-Means, DBSCAN, or Gaussian mixture model are used to perform unsupervised clustering on the feature vectors, ultimately obtaining K clusters, each containing several training samples with similar operating modes.

[0055] S203. Based on the source domain battery health degradation data corresponding to each cluster, determine the battery health degradation mode of each cluster.

[0056] Different clusters represent different use cases, and their battery degradation patterns (such as degradation rate, degradation curve shape, and the presence of accelerated degradation inflection points) often differ significantly. Clearly defining the degradation pattern of each cluster allows for increasing the weight of the corresponding expert network when using the model for prediction, thereby improving the model's prediction accuracy.

[0057] For example, dynamic time warping is used to calculate the similarity between SOH sequences, and then K-means clustering algorithm is used for grouping. The clustering results are as follows: Type A (linear decay): Capacity decreases at a constant rate over time; the most common aging mode. Type B (early rapid decline / early plunge): The decline is relatively rapid in the initial stage, and then tends to level off; Type C (late-stage plunge): Slow decline in the early stage, followed by a sudden acceleration in the later stage; Type D (Oscillating): Capacity fluctuates irregularly, possibly due to temperature changes or irregular usage. The final output is a decay pattern label for each sample, used for task grouping or weighting in subsequent multi-task learning.

[0058] Figure 3 This is a schematic diagram of a clustering result provided in an embodiment of this application, such as... Figure 3 As shown in the figure, the banded shading filled with various grids represents the corresponding distribution range / confidence interval of battery health degradation data; the curves in the figure represent the corresponding battery degradation types; the larger the shading bandwidth, the stronger the health dispersion among individual batteries of the same type; the curve in the middle of the shading represents the average degradation trend of this type.

[0059] The importance of samples with different decay modes varies across different learning tasks. For example, for type A (linear decay) samples, SOH estimation and aging rate prediction tasks have higher reliability and are assigned higher task weights; for type C (late-stage drop) samples, the weights of drop point detection and remaining lifetime prediction tasks are significantly increased. Specifically, the model defines a set of task weight vectors for each decay mode type k. These correspond to the weights of four tasks: SOH estimation, aging rate prediction, bottoming point detection, and remaining lifetime prediction. That is, the learning weights of the decay mode corresponding to each expert network differ for different prediction tasks. During training, the total loss function of a sample is calculated by weighting the weight vectors of the cluster category to which the sample belongs, achieving differentiated loss weighting at the cluster category level.

[0060] S204. Generate cluster feature vectors corresponding to each cluster based on the cluster center corresponding to each cluster.

[0061] Cluster centers are the most representative prototypes of each cluster, containing the core feature information of that cluster. After being transformed into cluster feature vectors, they can be embedded as prior knowledge into the classification network of the model, so that the classification network has a clear reference benchmark when identifying which cluster new data belongs to, rather than learning the cluster boundaries from scratch.

[0062] Specifically, the cluster center coordinates output by each clustering algorithm are directly taken as the initial feature vector, or further encoded (e.g., mapped to a fixed-dimensional embedding vector through a trainable fully connected layer) to obtain the cluster feature vector corresponding to each cluster. This vector can be updated and optimized along with the model parameters during subsequent training.

[0063] S205. Embed the clustering feature vector corresponding to each cluster into the classification network of the model to be trained to obtain the embedding model.

[0064] The core of this step is to construct a hybrid expert model architecture. Clustering feature vectors are embedded into a classification network (i.e., a gating network), enabling the classification network to utilize predefined clustering prior knowledge to identify and assign weights to the input data. Each expert network corresponds to a different decay mode. This architecture is more expressive than a single model and can handle multiple decay modes simultaneously.

[0065] The embedded model includes a classification network and multiple expert networks. The classification network is used to cluster the input vehicle operation data and determine the weights of each expert network. Each expert network corresponds to a battery health degradation mode and is used to process the input vehicle operation data and output the corresponding battery health degradation prediction results.

[0066] Classification networks are typically implemented using fully connected networks or attention mechanisms. Their input is vehicle operation data features, while clustering feature vectors serve as learnable parameters or auxiliary inputs. Activation functions such as softmax output weight vectors for each expert network (the sum of all weights is 1). Expert networks can employ independent fully connected networks, long short-term memory (LSTM) networks, or sequence models such as Transformers. Each expert network has the same structure but independent parameters, each corresponding to a different decay mode.

[0067] S206. Input the source domain vehicle operation data in the source domain training samples into the embedding model. Based on the weights of each expert network output by the classification network, perform weighted fusion of the sub-prediction results output by each expert network to obtain the source domain predicted battery health degradation data corresponding to the source domain training samples.

[0068] In practical applications, a vehicle's operational data may not belong entirely to a single cluster, but rather exhibit mixed characteristics of multiple clusters. Through weighted fusion, the model can integrate the prediction results of multiple expert networks, achieving smoother and more robust predictions.

[0069] In one implementation, the cluster centers of each decay mode category (i.e., the representative vectors of that category in the feature space) are used as learnable prototype vectors (i.e., cluster feature vectors) and embedded into the feature fusion layer of a multi-expert / task network. For each input sample, the network calculates the similarity (using cosine similarity or Euclidean distance) between its fused features and each prototype vector, obtaining the membership score of that sample to each decay mode. This membership score... As a gating signal, the outputs of different expert networks are weighted and modulated:

[0070] in, As a feature of fusion, Let k be the cluster feature vector of the k-th cluster category. For distance measurement function, For temperature coefficient, This is the predicted value for the j-th task branch corresponding to the k-th category.

[0071] In another implementation, the weights of each expert network can be determined as follows: for any source domain training sample, determine the feature distance between the source domain vehicle operation data and each cluster feature vector in the source domain training sample, and normalize the feature distance to obtain the weights of each expert network corresponding to the source domain training sample.

[0072] In practical applications, each time vehicle operation data is input into the model for prediction, the model will identify it and determine the weights of each expert network. Therefore, the weights of each expert network may be different in each prediction. Thus, the expert network weights are specifically determined for each input of vehicle operation data.

[0073] For example, the distance between the vehicle operation data input to the model and each cluster feature vector can be determined by Euclidean distance or cosine distance, and the distance can be further normalized to map it to a specific range, such as 0-1, as the weight of the corresponding expert network.

[0074] This method of determining expert network weights retains the advantages of flexible integration of hybrid expert models while avoiding the training overhead of complex classification networks, making it a computationally efficient and highly interpretable implementation.

[0075] For example, Table 1 shows a network structure of an embedded model provided in an embodiment of this application, wherein the vehicle operation data input to the embedded model may include three-dimensional data such as the vehicle state feature vector, the user behavior feature vector, and the battery aging time series features.

[0076] Table 1

[0077] Correspondingly, Figure 4 This is a schematic diagram of a network structure for an embedded model provided in an embodiment of this application, such as... Figure 4 As shown, the embedding model 400 includes an input layer 410, a feature extraction layer 420, a feature fusion layer 430, a shared hidden layer 440, a task branching layer 450, and an output layer 460.

[0078] Input layer 410 receives 32-dimensional vehicle state features, 24-dimensional user behavior features, and 50-dimensional data. 16-dimensional battery time-series features. The battery time-series features are two-dimensional time series with 50 time steps and 16 battery parameters per step.

[0079] The feature extraction layer has 420 layers, employing adapted extraction networks for three types of features: Vehicle state features: Fully connected layer (FC) (32→128) + rectified linear unit (ReLU); the fully connected layer upscales the 32 dimensions to 128, ReLU is activated, and nonlinear features are extracted. User behavior features: FC (24→128) + ReLU; the fully connected layer maps 24 dimensions to 128 dimensions. Battery time-series features: LSTM (16→256) + Dropout (0.3); LSTM is used specifically to process time-series data, capturing the time-series aging patterns of battery voltage / current changes over time; Dropout (0.3) suppresses overfitting.

[0080] The feature fusion layer 430 concatenates the feature vectors extracted from the three sources: 128 (vehicle) + 128 (user) + 256 (battery timing) = 512-dimensional fused features. This integrates vehicle, user, and battery timing information into a single feature space, enabling joint modeling of multi-source information.

[0081] The shared hidden layer 440 consists of three stacked fully connected layers, with all prediction tasks sharing the weights of this part of the network. The first layer, FC(512→256)+ReLU+Dropout(0.3), takes the 512-dimensional fusion features (vehicle features + user behavior + battery time-series concatenation) output from the feature fusion layer as input. The fully connected FC layer compresses the dimensions and performs preliminary filtering of redundant information; ReLU is used to uncover the complex nonlinear relationships between battery aging, driving habits, and operating conditions; Dropout(0.3) suppresses noise in the training set and improves generalization. The second layer, FC(256→128)+ReLU+Dropout(0.3), takes the 256-dimensional features output from the first layer as input. This layer is a shared backbone, with all four prediction tasks (SOH, decay rate, decay mode, and remaining life) sharing its weights. The expert network learns the underlying battery aging patterns common to all tasks here. ReLU continues to retain nonlinearity; Dropout continuously regularizes to avoid overfitting. The third layer is FC (128→64)+ReLU. The input of this layer is the 128-dimensional feature output of the second layer, and the output is a 64-dimensional shared feature vector. Since it is close to the task branch entry and the feature dimension is low, continuing Dropout will easily lose effective information. Therefore, this layer is only paired with ReLU and no more Dropout is added. The final output 64-dimensional vector will be simultaneously distributed to the four independent task branch layers and is the common feature base of the entire multi-task model.

[0082] The task branch layer has 450 layers, including multiple task branches. These branches share a unified structure: FC(64→32)→FC(32→1), meaning dimensionality reduction is achieved through two fully connected layers. The third branch ends with an igmoid (SO) activation (because its output is binary). Each branch independently learns its own task-specific mapping.

[0083] Output layer 460 outputs the following content: SOH (%): Battery health regression prediction; Degradation rate (% / 10,000 km): Average annual / per 10,000 km battery capacity degradation rate; Degradation mode category (corresponding to 4 types of degradation mode classification: linear degradation, early rapid degradation, late-stage rapid degradation, and oscillating type); Remaining service life RUL (10,000 km): Remaining driving range before the battery reaches the failure threshold.

[0084] In one implementation, a multi-task network can be omitted, using only a single task (battery time-series features) + a lightweight regression model (random forest / XGBoost), retaining unsupervised clustering and transfer learning mechanisms. This approach is suitable for scenarios with limited computing power in vehicles, and the evaluation latency can be reduced to less than 30ms. Alternatively, instead of using deep learning methods, a rule base can be built based on the clustering results (e.g., "Type A → linear regression model, Type B → exponential decay model, Type C → piecewise linear model") to achieve more interpretable evaluations, suitable for application scenarios requiring interpretability (such as insurance claims). Another option is to use a temporal convolutional network (TCN) instead of LSTM to process battery time-series data, which allows for parallel computation, faster training speed, and is suitable for large-scale data scenarios.

[0085] S207. Based on the source domain predicted battery health decay data and source domain battery health decay data corresponding to the source domain training samples, determine the source domain predicted loss corresponding to the source domain training samples.

[0086] For example, taking any expert network as an example, its loss function is L = λ1·L_SOH + λ2·L_rate + λ3·L_jump + λ4·L_RUL, where λ1, λ2, λ3, and λ4 are the optimized weights of the expert network for different task branches, L_SOH is the predicted loss for task branch A, L_rate is the predicted loss for task branch B, L_jump is the predicted loss for task branch C, and L_RUL is the predicted loss for task branch D. For the expert network corresponding to type A (linear decay), λ1 = 0.4, λ2 = 0.2, λ3 = 0.2, and λ4 = 0.2 can be set.

[0087] S208. Update the parameters and clustering feature vectors of the embedding model based on the source domain prediction loss corresponding to multiple source domain training samples to obtain the target model.

[0088] Specifically, the backpropagation algorithm is used to calculate the gradient of the loss with respect to each parameter, and then the parameter θ is updated through an optimizer (such as Adam, SGD, etc.). t+1 =θ t η θ L.

[0089] Where L is the source domain prediction loss, and θ contains all network parameters and clustering feature vectors. η The learning rate is used. After multiple epochs of iterative training, the target model is obtained when the loss converges or a preset stopping condition is met.

[0090] Furthermore, the update mechanism for cluster feature vectors is explained in more detail. When the model identifies the input vehicle operation data, it determines which decay mode it is closer to and assigns higher weights to the corresponding expert network. If there is a large loss value between the model's final prediction and the true label, the reasons can usually be attributed to two aspects: first, the location of the cluster centers is not set reasonably; second, the internal parameters of the model have not been updated sufficiently.

[0091] To address this, one implementation employs a training strategy that dynamically adjusts the update weight: in the early stages of iteration, the focus is on updating model parameters; while in the later stages of iteration, the weight of updating cluster feature vectors is gradually increased. The logic behind this strategy is that when training progresses to the later stages and the prediction results are still inaccurate, it often indicates a deviation in the model's scene recognition of the input data, leading to an unreasonable weight allocation. In this case, increasing the update intensity of the cluster feature vectors can induce an adaptive shift in the location of the cluster centers. This shift guides the input vehicle operation data to be re-identified and matched to a more suitable decay mode, thereby improving the weight allocation of the corresponding expert network and ultimately effectively improving the overall prediction accuracy of the model.

[0092] Based on the implementation methods of S201-S208 described above, various battery health degradation modes are effectively identified by clustering the source domain vehicle operation data, and a unique clustering feature vector is generated for each mode. On this basis, an embedding model containing a classification network and multiple expert networks is constructed. The clustering feature vectors are used to accurately identify the degradation mode of the input sample, and the weights of each expert network are dynamically determined. By weighted fusion of the sub-prediction results of each expert network, the model can adaptively output high-precision prediction results for different degradation modes. Finally, based on the loss feedback between predicted and real data, the model parameters and clustering feature vectors are updated synchronously, enabling the model to continuously optimize its fitting ability to complex degradation modes, thereby significantly improving the accuracy of battery health degradation prediction and the model's generalization ability.

[0093] Furthermore, the source domain data originates from a specific vehicle model / battery type, while practical applications often require predictions for a different vehicle model or battery type. Due to differences in powertrain parameters (such as battery capacity, thermal management system design, and motor characteristics) and chemical properties of different battery types (such as ternary lithium and lithium iron phosphate), their battery degradation patterns will also differ. Directly applying the source domain model to the target domain may lead to decreased prediction accuracy due to domain shift. Therefore, transfer learning of the target model based on target domain data is necessary.

[0094] In one implementation, the method further includes S209-210 after S208: S209. Obtain the target domain training sample set.

[0095] The target domain refers to a new field that needs to be adapted, and the labeled data is usually relatively limited. In the embodiments of this application, the target domain is the data field corresponding to a different vehicle model and / or battery type than the source domain; the target domain training samples are data samples from the target domain, and each sample contains vehicle operation data of the target domain and its corresponding battery health degradation label.

[0096] Specifically, a certain amount of actual operating data and corresponding State of Health (SOH) calibration data are collected from vehicles of the target model or battery type. Due to the high cost of acquiring data in the target domain, its sample size is typically much smaller than that of the source domain. The data also requires preprocessing operations such as cleaning, alignment, and standardization to ensure its format is consistent with the source domain data.

[0097] S210. Transfer learning is performed on the target model based on the source domain training sample set and the target domain training sample set to obtain the transfer model.

[0098] While the target model performs well in the source domain, its cluster centers, expert network parameters, and classification network parameters are all learned based on the source domain data distribution. When faced with target domain data, the following problems arise: Cluster center shift: The cluster centers in the source domain reflect the operating patterns of vehicle models in the source domain, but the operating characteristics of vehicles in the target domain may not be accurately covered by the existing cluster centers. Expert network mismatch: The expert network in the source domain learns the decay patterns of the source domain, but the decay curve shape in the target domain may be different. Weight allocation bias: Due to the cluster center shift, the classification network's weight allocation for target domain data may be unreasonable. Therefore, transfer learning is needed to adapt the model using target domain data while retaining the general knowledge learned in the source domain, enabling it to accurately predict battery health degradation in the target domain.

[0099] The source domain training sample set is obtained based on a vehicle model and / or battery, while the target domain training sample set is obtained based on a different vehicle model and / or battery than the source domain training sample set. The target domain training sample set includes multiple target domain training samples, which include target domain vehicle operation data and corresponding target domain battery health degradation data.

[0100] Specifically, transfer learning methods that can be adopted include: Option A: Parameter transfer (fine-tuning) uses the pre-trained model parameters in the source domain (existing vehicle models) as initial values, and fine-tunes them using a small amount of data in the target domain (recommended ≥100 complete charge-discharge cycles). The first two shared layers are frozen (to extract general features), and only the last three layers (task-specific layers) are updated.

[0101] Option B: Feature transfer. Use source domain data to train an autoencoder as a feature extractor to map target domain data to a shared feature space. Build an SOH regression model in the shared space without retraining the entire network.

[0102] Option C: Domain adaptation, which uses a domain adversarial neural network to add a domain classifier after the feature extractor, and uses adversarial training to make the feature distributions of the source domain and the target domain consistent, or uses maximum mean difference alignment: minimizing the distance between the feature distributions of the source domain and the target domain.

[0103] Figure 5 This application provides a schematic diagram of a transfer learning process, as illustrated in the embodiments of this application. Figure 5As shown, the target model combines target domain data (target domain training sample set) and source domain data (source domain training sample set) for transfer learning. During the transfer process, three transfer methods can be selected, including parameter fine-tuning (freezing the first few layers and updating the last few layers), feature transfer (autoencoder + regression), and domain adaptation (domain-adversarial neural network (DANN) or maximum mean discrepancy (MMD) alignment). Finally, after completing the transfer learning, the transfer model is obtained.

[0104] Among these methods, parameter fine-tuning can freeze the first two shared layers and fine-tune the last three task layers, performing 10-20 rounds of small iterations. This transfer method is suitable when the number of target domain data samples is constant (sample number ≥ 100). Feature transfer trains a source domain autoencoder and maps the target domain data to a shared feature space to train a lightweight regressor. This transfer method is suitable when the amount of target domain data is small (sample number < 100). Domain adaptation constructs an adversarial neural network and adds a domain classifier for adversarial training to minimize the MMD distance. This transfer method is suitable when the target domain data distribution is large.

[0105] Based on the implementation methods described in S209-S210, by introducing target domain samples for transfer learning of the pre-trained model, the differences in data distribution (domain shift) caused by cross-vehicle or battery types are effectively overcome. Specifically, based on the source domain model, a small amount of target domain data is used for fine-tuning, enabling the model to quickly adapt to the unique operating and decay patterns of new scenarios while retaining general feature recognition. This not only significantly reduces the cost of massive data collection and retraining in new scenarios but also corrects the deviations in clustering and weight allocation, ensuring high-precision prediction and engineering practicality of the transfer model in cross-domain scenarios.

[0106] The learning transfer method described in the above steps is applicable to most scenarios. However, in the scenario of predicting the health of vehicle batteries, this application embodiment further proposes a dedicated transfer mechanism for the physical characteristics of vehicle battery aging.

[0107] In one implementation, the source domain training sample set also includes source domain battery electrochemical features, the target domain training sample set also includes target domain battery electrochemical features, and the source domain battery health degradation data includes source domain battery health degradation curves. The above S210 can be implemented via S2101-S2105: S2101 determines the aging rate of the source domain battery based on the source domain battery health degradation curve.

[0108] The source domain battery health decay curve is a continuous function or discrete sequence of battery health (SOH) changing with time, number of cycles, or mileage in the source domain training samples. It intuitively reflects the complete aging trajectory of the battery from its initial state to its current state. The source domain battery aging rate is a quantitative indicator describing the rate of battery health decay. It is usually expressed as the slope or rate of change of the decay curve within a certain time window, reflecting the speed of battery aging.

[0109] Specifically, the source domain battery health degradation curve is numerically differentiated or subjected to sliding window difference processing to calculate the rate of change of SOH with time (or number of cycles), thus obtaining the aging rate of the source domain battery. For example, the ratio of the change in SOH between adjacent sampling points to the time interval can be calculated using the finite difference method, or the slope of each segment can be extracted as the aging rate after piecewise linear fitting of the degradation curve.

[0110] S2102. Input the target domain vehicle operation data into the target model to obtain the target domain predicted battery health degradation data, and determine the target domain battery aging rate based on the target domain predicted battery health degradation data.

[0111] In transfer learning, the labeled data for the target domain is limited, making it impossible to directly obtain the actual aging rate of the battery in the target domain. Therefore, it is necessary to utilize the target model already trained on the source domain to perform forward inference on the vehicle operation data in the target domain, obtaining the model's prediction of the battery health degradation in the target domain. Subsequently, the battery aging rate in the target domain is extracted from the prediction results and compared with the aging rate in the source domain to evaluate whether the model's prediction in the target domain conforms to physical laws.

[0112] Specifically, vehicle operation data in the target domain is input into the target model one by one or in batches. After the classification network identifies the degradation pattern and determines the weights of the expert networks, each expert network outputs its sub-prediction results, which are then weighted and fused to obtain the target domain's predicted battery health degradation data. Subsequently, the same numerical differentiation or sliding window difference method as S2101 is used to calculate the target domain's battery aging rate from the predicted degradation data.

[0113] S2103. Determine physical loss based on the source domain battery aging rate and the target domain battery aging rate.

[0114] The battery aging process follows fundamental physicochemical laws, meaning that the capacity decay rate is monotonically constrained by factors such as accumulated throughput and temperature stress. This application's embodiments introduce an aging rate consistency regularization term into the loss function of transfer learning, constraining the trend of the aging rate predicted by the transferred model in the target domain to remain consistent with the physical laws of the source domain. Therefore, it is necessary to extract the battery aging rate in the source domain and the battery aging rate in the target domain separately, and perform consistency analysis accordingly.

[0115] Physical loss is used to measure the degree of consistency in monotonicity between the source domain battery health decay curve and the target domain predicted battery health decay curve. When the monotonicity of the two is consistent, the physical loss approaches zero; when a monotonicity violation occurs, the physical loss increases, driving the model parameters to be updated in a direction that conforms to physical laws.

[0116] For example, the physical loss function can be: in, For the aging rate of the source domain battery, For the target domain battery aging rate, To accommodate the relaxation factor, the physical loss can be calculated using this formula.

[0117] S2104. Determine physical weights based on the similarity of electrochemical features between source domain battery electrochemical features and target domain battery electrochemical features.

[0118] Significant differences may exist in the electrochemical characteristics of batteries between different vehicle models or battery types (e.g., different cathode materials, different electrolyte formulations, different cell designs, etc.). When the source and target domain batteries are highly similar in electrochemical characteristics, the aging physics of the source domain has strong reference value for the target domain. In this case, physical losses should be given a larger weight to strengthen physical constraints. Conversely, when the electrochemical differences between the two are large, the physical laws of the source domain may not be fully applicable to the target domain. In this case, the physical weights should be reduced to avoid excessive constraints that could lead to a decline in model performance.

[0119] Electrochemical characteristics refer to the set of parameters representing the electrochemical system of a battery, such as the type of positive and negative electrode materials, specific capacity, internal resistance, coulombic efficiency, and electrochemical impedance spectroscopy characteristics. The physical weights are positively correlated with the similarity of electrochemical characteristics between the source domain and target domain batteries.

[0120] Specifically, the similarity of battery electrochemical characteristics can be calculated in the following way: Similarity of cathode material types (Similarity of materials: 1.0; different types of the same material: 0.6; different types: 0.3); Anode material type similarity (Similarity of materials is 1.0, different types of the same type is 0.6, and different types are 0.3); Electrolyte additive system similarity (Calculated based on the overlap of additive types); Normalized difference between nominal voltage and capacity range of battery cells .

[0121] A battery electrochemical feature similarity score is constructed by combining the above indicators. , .

[0122] in, , For the weights of each indicator to satisfy .

[0123] Furthermore, the magnitude of model parameter updates directly impacts the transfer learning process. When the electrochemical systems of the source and target domains are highly similar, the knowledge learned in the source domain is highly applicable to the target domain. In this case, a larger range of model parameter adjustments can be allowed to fully absorb the transferred knowledge from the source domain. Conversely, when the electrochemical systems of the two domains differ significantly, the reference value of the source domain knowledge for the target domain is limited. Allowing large parameter adjustments may cause the model to forget the effective knowledge learned in the source domain. In this case, parameter adjustments should be limited to a smaller range, requiring only fine-tuning. Therefore, it is possible to... The target model parameter update range is adjusted. Table 2 shows a migration strategy provided by an embodiment of this application, which is based on different... Determine different migration strategies.

[0124] Table 2

[0125] It is evident that higher similarity allows for a wider range of parameters to be updated. Based on this, in step S2105 below, the parameters of the target model can be selectively updated based on the parameter adjustment range determined by the migration strategy.

[0126] Based on the migration strategy determination method provided in the above embodiments, the parameter adjustment range during transfer learning is dynamically determined by electrochemical feature similarity, achieving fine-grained control over the magnitude of model parameter updates. When the electrochemical systems of the source and target domain batteries are highly similar, the parameter adjustment range is correspondingly expanded, allowing the model to fully absorb the aging knowledge learned in the source domain, accelerating the adaptation process of the target domain, and improving migration efficiency. When the electrochemical differences between the two are large, the parameter adjustment range automatically shrinks, constraining the model parameters to be fine-tuned near the source domain parameter values, effectively preventing catastrophic forgetting caused by over-adjustment and protecting the effective feature representation capabilities accumulated in the source domain.

[0127] S2105. Based on task loss, physical loss and physical weight, determine the transfer loss, and update the parameters and clustering feature vectors of the target model based on multiple transfer losses to obtain the transfer model.

[0128] The core objective of transfer learning is to adapt the model to the data distribution of the target domain while maintaining the predictive ability of the source domain. Task loss measures the difference in prediction accuracy between the source and target domains, ensuring the stability of the basic predictive ability during the transfer process. The introduction of physical loss and physical weights adds guidance from physical mechanisms to the task-driven approach, enabling the model to adaptively inherit physically consistent aging knowledge from the source domain while learning the data distribution of the target domain. These three factors combine to form a comprehensive transfer loss function that simultaneously considers prediction accuracy, physical consistency, and cross-domain transfer efficiency. Finally, backpropagation is used to jointly optimize the model parameters and clustering feature vectors, resulting in a transfer model adapted to the target domain.

[0129] Task loss is a loss function that measures the difference in prediction accuracy between the target model and the source and target domains. Specifically, task loss characterizes the difference between the source domain prediction accuracy of the target model when predicting battery health degradation data based on source domain vehicle operation data, and the target domain prediction accuracy when predicting battery health degradation data based on target domain vehicle operation data. The smaller this difference, the more balanced the model's prediction ability is in both domains, and the better the transfer effect.

[0130] Specifically, the migration loss function is: ,in For mission losses, The physical weights are used. Based on the transfer loss corresponding to training samples from multiple target domains, the classification network parameters, expert network parameters, and clustering feature vectors of the target model are jointly updated using the backpropagation algorithm. After multiple rounds of iterative training, the model converges to obtain the final transfer model.

[0131] Based on the implementation methods of S2101-S2105 described above, by extracting the battery aging rates of the source and target domains and constructing a physical loss, the monotonicity of the predicted degradation curve is effectively constrained, ensuring that the model output conforms to the irreversible physical law of battery aging and avoiding non-physical prediction anomalies. Simultaneously, by dynamically determining physical weights based on electrochemical feature similarity, adaptive adjustment of the physical constraint strength is achieved. This strengthens knowledge transfer when electrochemical systems are similar and reduces constraint interference when differences are significant, improving the specificity and robustness of the transfer. Finally, a transfer loss is constructed by combining the task loss and weighted physical loss, and the model parameters and clustering feature vectors are jointly optimized. While ensuring the predictive capability of the source domain, this significantly improves the model's generalization performance and prediction reliability under limited labeled data in the target domain, achieving efficient and physically interpretable transfer learning across vehicle models and battery platforms.

[0132] The above-mentioned S2101-S2105 incorporates the physical characteristics of battery degradation during the transfer process. Based on this, the embodiments of this application further introduce the transfer learning process by incorporating the battery temperature distribution characteristics during vehicle operation.

[0133] Specifically, the source domain training sample set also includes source domain battery temperature distribution features, and the target domain training sample set also includes target domain battery temperature distribution features; when the source domain training sample set and the target domain training sample set are obtained based on the same battery but different vehicle models, the transfer loss can be obtained through S301-S302: S301. Determine the temperature loss based on the temperature distribution characteristics of the source domain battery and the target domain battery.

[0134] Because different vehicle models differ significantly in their thermal management system design, airflow layout, cooling strategies, and overall vehicle energy management logic, even with the same battery, the temperature distribution can vary considerably across different models. Temperature is a key external factor affecting battery aging rate, and differences in battery temperature distribution across different models directly lead to variations in battery health degradation. If this factor is not considered, the transfer model will incorrectly attribute the degradation bias caused by differences in vehicle thermal management to insufficient model predictive ability, thus affecting transfer accuracy. By constructing a temperature loss mechanism, the differences in battery temperature distribution between different vehicle models can be explicitly quantified, providing a basis for subsequent degradation compensation.

[0135] Battery temperature distribution characteristics are a set of parameters that describe the spatial distribution characteristics of the internal or surface temperature of the battery in the source domain vehicle during operation, such as the average temperature, temperature gradient, maximum temperature, minimum temperature, and temperature variance of each region of the battery. They reflect the influence characteristics of the thermal management system of the source domain vehicle on the thermal behavior of the battery. Temperature loss characterizes the temperature distribution difference between the source domain battery temperature distribution characteristics and the target domain battery temperature distribution characteristics.

[0136] Specifically, the temperature loss function is: in, The temperature compensation coefficient for the i-th sample is calculated from the relative entropy between the statistical characteristics of the temperature distribution corresponding to the sample and the reference temperature distribution of the source domain. For the feature mapping function in the reproducing kernel Hilbert space; For regenerating nuclear space.

[0137] S302. Determine the migration loss based on task loss, physical loss, physical weight, and temperature loss.

[0138] Temperature loss is used to instruct the transfer model to compensate for battery health degradation caused by differences in temperature distribution. Specifically, it plays a role in decoupling temperature factors from aging factors in the transfer loss function. Under the same battery but different vehicle models, the aging mechanism of the battery is essentially the same, but the thermal management strategies of different models lead to different battery operating temperatures, resulting in differences in degradation. By explicitly quantifying the differences in temperature distribution, the model can separate the degradation deviation caused by differences in vehicle model thermal management from the total prediction error and compensate for it accordingly. This allows the model to more accurately learn temperature-independent features that truly reflect the battery's intrinsic aging mechanism.

[0139] By adding a temperature loss function to the migration loss function constructed in S2105 above, a migration loss that incorporates temperature loss can be obtained.

[0140] Based on the implementation methods of S301-S302 described above, a temperature loss is constructed by introducing battery temperature distribution characteristics from the source and target domains, effectively quantifying the impact of differences in thermal management systems of different vehicle models on battery temperature distribution. By incorporating the temperature loss into the transfer loss function, the model can explicitly perceive and decouple the degradation deviation caused by differences in vehicle thermal management during the optimization process, providing targeted compensation for the battery health degradation caused by temperature distribution differences, and avoiding incorrectly attributing prediction errors caused by temperature factors to insufficient model capability.

[0141] The aforementioned steps have completed the training of the model. On this basis, before entering the actual reasoning application stage of the model, it is necessary to restrict the output of the model to ensure that the output of the model is logical.

[0142] Specifically, in one implementation, the battery health degradation data includes battery aging rate, remaining lifespan, drop point location, and drop probability. The drop point represents the turning point where battery health changes abruptly. The battery aging rate and remaining lifespan output by the target model satisfy the condition that the difference between the theoretical remaining lifespan determined based on the battery aging rate and the remaining lifespan is less than a preset difference.

[0143] Specifically, there is an integral relationship between the predicted battery aging rate and the predicted remaining life. This application's embodiments introduce a consistency constraint loss to constrain both: .

[0144] in, This is the predicted remaining lifespan. This is a predicted value for the battery aging rate. This is the current SOH estimate. This constraint loss forces the outputs of the two tasks to meet physical consistency, improving the reliability of the joint prediction.

[0145] After obtaining the target model, the method also includes S401-S403: S401, Obtain current vehicle operation data.

[0146] The vehicle's battery operation data can be collected in real time through the vehicle battery management system or vehicle communication terminal. After data preprocessing (such as noise reduction, normalization, and feature extraction), the current vehicle operation data is generated in a format consistent with the model input.

[0147] S402. Input the vehicle operation data into the target model to obtain the current battery health degradation data.

[0148] The preprocessed current vehicle operation data is input into the target model. After the classification network identifies the current battery degradation mode and determines the weight of each expert network, each expert network outputs its sub-prediction results and they are weighted and fused together. Finally, the current battery health degradation data is output, including battery aging rate, remaining life, drop point location and drop probability.

[0149] S403. When the current probability of diving is greater than the preset probability threshold, the decay process corresponding to the remaining lifespan is divided into a steady phase before diving and an acceleration phase after diving, with the current diving point as the dividing point, and the remaining lifespan is predicted separately.

[0150] Battery aging is not always a uniform or smooth decline. Before the critical point of failure, battery health typically exhibits a relatively stable and slow decline (the plateau phase); however, after the critical point, due to internal faults or abrupt changes in the aging mechanism, battery health declines sharply (the accelerated phase). The aging mechanisms and degradation rates of these two phases are fundamentally different. Using uniform model parameters or expert network weights for prediction makes it difficult to simultaneously ensure accurate predictions for both phases. When the probability of failure exceeds a preset threshold, it indicates that the model has a high confidence level that the battery is about to fail. In this case, segmenting the remaining lifespan degradation process with the critical point as the boundary and using different expert network weights in different segments allows the model to refine its modeling of the aging characteristics of each stage, significantly improving the accuracy of remaining lifespan prediction in critical scenarios.

[0151] In this case, the weights of the expert network in the target model are different when predicting the remaining lifetime for different segments.

[0152] Specifically, first, it is determined whether the current probability of a battery drain is greater than a preset probability threshold. If it is below the threshold, the model's predicted remaining lifespan is directly output as the final result. If it exceeds the threshold, the decay process corresponding to the remaining lifespan is divided into a stable segment before the drain and an accelerated segment after the drain, using the drain point as the boundary. For the stable segment before the drain, a set of expert network weights (such as weights determined by the classification network based on the characteristics of the stable segment) are used to weight and fuse the sub-prediction results of each expert network, outputting the remaining lifespan prediction value for the stable segment. For the accelerated segment after the drain, another set of expert network weights (such as weights determined by the classification network based on the characteristics of the accelerated segment) are used to weight and fuse, outputting the remaining lifespan prediction value for the accelerated segment. Finally, the prediction results of the two segments are combined to obtain a complete segmented remaining lifespan prediction result, providing users with more refined battery life assessment information.

[0153] For example, when the probability of diving When the risk threshold is reached, the following linkage mechanism will be triggered: (a) The remaining lifetime prediction branch automatically switches to the plunge mode and adopts a segmented prediction strategy: the remaining lifetime is divided into two segments, the steady segment before the plunge and the acceleration segment after the plunge, and predicted separately, and the comprehensive remaining lifetime estimate and its confidence interval are output. (b) Electronic devices automatically generate high-risk warning push notifications, including the expected time window and confidence level of the plunge, and send battery maintenance reminders through the user terminal; (c) The model automatically increases the subsequent data collection frequency of the vehicle (from the default 10-second sampling period to 1-second sampling period) to capture the fine changes before and after the dive, which will be used for subsequent model iteration and optimization.

[0154] In addition, different maintenance suggestions can be provided to users based on different output results of the model. Table 3 shows the correspondence between model output and maintenance decisions.

[0155] Table 3

[0156] Based on the implementation methods of S401-S403 described above, the identification and early warning of sudden changes in battery health risk are achieved by introducing the location and probability of the water drop point. The aging rate and remaining life output by the model must meet the constraint that the difference between the theoretical remaining life and the predicted value is less than a preset difference, ensuring the physical rationality of the prediction results. During the inference phase, after real-time vehicle operation data is input into the target model, when the probability of water drop exceeds a preset threshold, a segmented prediction mechanism is automatically triggered. The remaining life decay process is divided into a stable segment before water drop and an accelerated segment after water drop, with the water drop point as the boundary. Different expert network weights are used for prediction in different segments. Since the aging mechanism and decay rate of the stable segment and the accelerated segment are fundamentally different, the setting of differentiated weights allows the model to perform refined modeling of the aging characteristics of each stage, avoiding the underfitting problem of the rapid decay stage after water drop under uniform weights.

[0157] This application embodiment can divide the following device into functional modules based on the above method example. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. Optionally, the module division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0158] In some embodiments, this application also provides a model training apparatus. This apparatus may include one or more functional modules for implementing the methods described in the above embodiments. For example, Figure 6 This is a schematic diagram of the structure of a model training device 600 provided in an embodiment of this application. Figure 6 As shown, the model training device 600 includes: The acquisition module 610 is used to acquire the source domain training sample set; the source domain training sample set includes multiple source domain training samples, and the source domain training samples include source domain vehicle operation data and corresponding source domain battery health degradation data. Processing module 620 is used to perform clustering processing on source domain vehicle operation data in multiple source domain training samples to obtain multiple clusters; The processing module 620 is also used to determine the battery health degradation mode of each cluster based on the source domain battery health degradation data corresponding to each cluster. The processing module 620 is also used to generate cluster feature vectors corresponding to each cluster based on the cluster center corresponding to each cluster. The processing module 620 is also used to embed the clustering feature vector corresponding to each cluster into the classification network of the model to be trained to obtain the embedding model. The embedding model includes a classification network and multiple expert networks. The classification network is used to identify clusters of the input vehicle operation data and determine the weights of each expert network. Each expert network corresponds to a battery health degradation mode and is used to process the input vehicle operation data and output the corresponding battery health degradation prediction results. The training module 630 is used to input the source domain vehicle operation data in the source domain training samples into the embedding model, and to perform weighted fusion of the sub-prediction results output by each expert network based on the weights of each expert network output by the classification network, so as to obtain the source domain predicted battery health degradation data corresponding to the source domain training samples. The training module 630 is also used to determine the source domain prediction loss corresponding to the source domain training sample based on the source domain predicted battery health decay data and the source domain battery health decay data corresponding to the source domain training sample. The training module 630 is also used to update the parameters and clustering feature vectors of the embedded model based on the source domain prediction loss corresponding to multiple source domain training samples, so as to obtain the target model.

[0159] In one implementation, the weights of each expert network are determined as follows: for any source domain training sample, the feature distance between the source domain vehicle operation data and each cluster feature vector in the source domain training sample is determined, and the feature distance is normalized to obtain the weights of each expert network corresponding to the source domain training sample.

[0160] In one implementation, the training module 630 is further used to obtain a target domain training sample set; and to perform transfer learning on the target model based on the source domain training sample set and the target domain training sample set to obtain a transfer model; wherein the source domain training sample set is obtained based on a vehicle model and / or battery, and the target domain training sample set is obtained based on a different vehicle model and / or battery than the source domain training sample set, and the target domain training sample set includes multiple target domain training samples, which include target domain vehicle operation data and corresponding target domain battery health degradation data.

[0161] In one implementation, the source domain training sample set further includes source domain battery electrochemical features, the target domain training sample set further includes target domain battery electrochemical features, and the source domain battery health degradation data includes source domain battery health degradation curves. The training module 630 is also used to determine the source domain battery aging rate based on the source domain battery health degradation curves; input target domain vehicle operation data into the target model to obtain target domain predicted battery health degradation data, and determine the target domain battery aging rate based on the target domain predicted battery health degradation data; determine physical loss based on the source domain battery aging rate and the target domain battery aging rate; the physical loss characterizes the relationship between the source domain battery health degradation curve and the target domain predicted battery health degradation curve. Whether the monotonicity of the reduction curve is consistent; determining the physical weights based on the similarity of electrochemical features between the source domain battery electrochemical features and the target domain battery electrochemical features; the physical weights are positively correlated with the similarity of electrochemical features between the source domain battery electrochemical features and the target domain battery electrochemical features; determining the transfer loss based on the task loss, physical loss, and physical weights, and updating the parameters and clustering feature vectors of the target model based on multiple transfer losses to obtain the transfer model; the task loss characterizes the difference between the source domain prediction accuracy when the target model predicts battery health degradation data based on source domain vehicle operation data and the target domain prediction accuracy when predicting battery health degradation data based on target domain vehicle operation data.

[0162] In one implementation, the training module 630 is further used to determine the parameter adjustment range when performing transfer learning on the target model based on electrochemical feature similarity; the parameter adjustment range is positively correlated with electrochemical feature similarity; based on multiple transfer losses, the parameters and clustering feature vectors of the target model within the parameter adjustment range are updated to obtain the transfer model.

[0163] In one implementation, the training module 630 is further configured to determine a temperature loss based on the source domain battery temperature distribution characteristics and the target domain battery temperature distribution characteristics; the temperature loss characterizes the temperature distribution difference between the source domain battery temperature distribution characteristics and the target domain battery temperature distribution characteristics; the transfer loss is determined based on the task loss, physical loss, physical weights and temperature loss; the temperature loss is used to instruct the transfer model to compensate for the amount of battery health degradation caused by the temperature distribution difference.

[0164] In one implementation, the battery health degradation data includes battery aging rate, remaining lifespan, drop point location, and drop probability. The drop point represents the turning point where battery health undergoes a sudden change. The battery aging rate and remaining lifespan output by the target model satisfy the condition that the difference between the theoretical remaining lifespan determined based on the battery aging rate and the actual remaining lifespan is less than a preset difference. The training module 630 is also used to acquire current vehicle operation data. The vehicle operation data is input into the target model to obtain the current battery health degradation data. When the current drop probability is greater than a preset probability threshold, the degradation process corresponding to the remaining lifespan is divided into a stable segment before the drop and an acceleration segment after the drop, with the current drop point location as the dividing point, and the remaining lifespan is predicted separately. Among these, the weights of the expert network in the target model are different when predicting the remaining lifespan for different segments.

[0165] When the methods of the above embodiments are implemented in hardware, this embodiment of the invention provides a possible structural schematic diagram of the electronic device involved in the above embodiments. For example... Figure 7 As shown, the electronic device 700 includes: a processor 702, a communication interface 703, and a bus 704. Optionally, the electronic device 700 may also include a memory 701.

[0166] Processor 702 may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 702 may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 702 may also be a combination that implements computing functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0167] The communication interface 703 is used to connect to other devices via a communication network. This communication network can be Ethernet, wireless access network, wireless local area network (WLAN), etc.

[0168] The memory 701 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0169] In one possible implementation, the memory 701 can exist independently of the processor 702. The memory 701 can be connected to the processor 702 via a bus 704 and is used to store instructions or program code. When the processor 702 calls and executes the instructions or program code stored in the memory 701, it can implement the method provided in the embodiments of the present invention.

[0170] In another possible implementation, the memory 701 can also be integrated with the processor 702.

[0171] The 704 bus can be an Extended Industry Standard Architecture (EISA) bus, etc. The 704 bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The symbol is represented by only one line, but this does not mean that there is only one bus or one type of bus.

[0172] Through the above description of the embodiments, those skilled in the art can clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the above device is divided into different functional modules to complete all or part of the functions described above.

[0173] This application also provides a computer-readable storage medium. All or part of the processes in the above method embodiments can be executed by computer instructions instructing related hardware. The program can be stored in the computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. The computer-readable storage medium can be any of the foregoing embodiments or memory. The computer-readable storage medium can also be an external storage device of the above-described device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the above-described device. Further, the computer-readable storage medium can include both internal storage units and external storage devices of the above-described device. The computer-readable storage medium is used to store the above-described computer program and other programs and data required by the above-described device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0174] This application also provides a computer program product comprising a computer program that, when run on a computer, causes the computer to perform the methods provided in the above embodiments.

[0175] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A model training method, characterized in that, The method includes: Obtain a source domain training sample set; the source domain training sample set includes multiple source domain training samples, and the source domain training samples include source domain vehicle operation data and corresponding source domain battery health degradation data; Clustering is performed on the source domain vehicle operation data in the multiple source domain training samples to obtain multiple clusters; Based on the source domain battery health degradation data corresponding to each cluster, the battery health degradation mode of each cluster is determined. Generate a cluster feature vector corresponding to each cluster based on the cluster center of each cluster; The clustering feature vector corresponding to each cluster is embedded into the classification network of the model to be trained to obtain the embedding model. The embedding model includes a classification network and multiple expert networks. The classification network is used to identify clusters of the input vehicle operation data and determine the weights of each expert network. Each expert network corresponds to a battery health degradation mode and is used to process the input vehicle operation data and output the corresponding battery health degradation prediction results. The source domain vehicle operation data in the source domain training samples is input into the embedding model. Based on the weights of each expert network output by the classification network, the sub-prediction results output by each expert network are weighted and fused to obtain the source domain predicted battery health degradation data corresponding to the source domain training samples. Based on the source domain predicted battery health decay data and source domain battery health decay data corresponding to the source domain training samples, the source domain prediction loss corresponding to the source domain training samples is determined. The parameters of the embedding model and the clustering feature vector are updated based on the source domain prediction loss corresponding to multiple source domain training samples to obtain the target model.

2. The method according to claim 1, characterized in that, The weights of each expert network are determined in the following way: For any source domain training sample, determine the feature distance between the source domain vehicle operation data and each clustering feature vector in the source domain training sample, and normalize the feature distance to obtain the weights of each expert network corresponding to the source domain training sample.

3. The method according to claim 1 or 2, characterized in that, After obtaining the target model, the method further includes: Obtain the training sample set for the target domain; The target model is transferred to the source domain training sample set and the target domain training sample set to obtain a transfer model; wherein the source domain training sample set is obtained based on a vehicle model and / or battery, and the target domain training sample set is obtained based on a different vehicle model and / or battery than the source domain training sample set. The target domain training sample set includes multiple target domain training samples, and the target domain training samples include target domain vehicle operation data and corresponding target domain battery health degradation data.

4. The method according to claim 3, characterized in that, The source domain training sample set further includes source domain battery electrochemical features, the target domain training sample set further includes target domain battery electrochemical features, the source domain battery health degradation data includes source domain battery health degradation curves, and the transfer learning of the target model based on the source domain training sample set and the target domain training sample set to obtain a transfer model includes: Based on the source domain battery health degradation curve, the aging rate of the source domain battery is determined. The target domain vehicle operation data is input into the target model to obtain the target domain predicted battery health degradation data, and the target domain battery aging rate is determined based on the target domain predicted battery health degradation data. Physical loss is determined based on the aging rate of the source domain battery and the aging rate of the target domain battery; the physical loss characterizes whether the monotonicity of the health decay curve of the source domain battery is consistent with the predicted health decay curve of the target domain battery. Physical weights are determined based on the electrochemical feature similarity between the source domain battery electrochemical features and the target domain battery electrochemical features; the physical weights are positively correlated with the electrochemical feature similarity between the source domain battery electrochemical features and the target domain battery electrochemical features. The transfer loss is determined based on the task loss, the physical loss, and the physical weights, and the parameters of the target model and the clustering feature vector are updated based on multiple transfer losses to obtain the transfer model; the task loss represents the difference between the source domain prediction accuracy of the target model when predicting the battery health degradation data based on the source domain vehicle operation data and the target domain prediction accuracy when predicting the battery health degradation data based on the target domain vehicle operation data.

5. The method according to claim 4, characterized in that, Before updating the parameters of the target model and the clustering feature vector based on multiple transfer losses to obtain the transfer model, the method further includes: The parameter adjustment range for transfer learning of the target model is determined based on the electrochemical feature similarity; the parameter adjustment range is positively correlated with the electrochemical feature similarity. The process of updating the parameters of the target model and the clustering feature vector based on multiple transfer losses to obtain the transfer model includes: Based on multiple transfer losses, the parameters of the target model and the clustering feature vector within the parameter adjustment range are updated to obtain the transfer model.

6. The method according to claim 4, characterized in that, The source domain training sample set also includes source domain battery temperature distribution features, and the target domain training sample set also includes target domain battery temperature distribution features; when the source domain training sample set and the target domain training sample set are obtained based on different vehicle models with the same vehicle battery, determining the migration loss based on the task loss, physical loss, and physical weights includes: Based on the temperature distribution characteristics of the source domain battery and the temperature distribution characteristics of the target domain battery, the temperature loss is determined; the temperature loss characterizes the temperature distribution difference between the temperature distribution characteristics of the source domain battery and the temperature distribution characteristics of the target domain battery. The migration loss is determined based on the task loss, physical loss, physical weights, and temperature loss; the temperature loss is used to instruct the migration model to compensate for the battery health degradation caused by the temperature distribution differences.

7. The method according to claim 1, characterized in that, The battery health degradation data includes battery aging rate, remaining lifespan, drop point location, and drop probability. The drop point represents the turning point where the battery health changes abruptly. The battery aging rate and remaining lifespan output by the target model satisfy the condition that the difference between the theoretical remaining lifespan determined based on the battery aging rate and the remaining lifespan is less than a preset difference. After obtaining the target model, the method further includes: Obtain current vehicle operation data; The vehicle operation data is input into the target model to obtain the current battery health degradation data; When the current probability of diving is greater than a preset probability threshold, the decay process corresponding to the remaining lifetime is divided into a steady segment before diving and an acceleration segment after diving, with the current diving point as the dividing point, and the remaining lifetime is predicted for each segment. The weights of the expert network in the target model are different when predicting the remaining lifetime for different segments.

8. An electronic device, characterized in that, It includes a processor and a memory, the processor being coupled to the memory; the memory is used to store computer instructions, which are loaded and executed by the processor to enable the computer device to perform the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes computer-executable instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when run on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 7.

Images