Method and system for evaluating state of health of lithium-ion battery of electric vehicle based on context-aware relationship graph network model

Abstract

The application discloses a kind of based on context-aware relationship graph network model electric vehicle lithium ion battery health state evaluation method and system, it is related to battery health state evaluation field, this method includes: screening according to real-time monitoring data, obtain multiple charging segments;The reference capacity of each charging segment is calculated using the reverse coulomb calculation method and feature vector is extracted;According to feature vector, using the context-aware relationship graph network model trained to predict, obtain battery current maximum available capacity;Context-aware relationship graph network model includes context encoding layer, embedding layer, stacked context-aware relationship graph network layer and readout module;Context encoding layer and embedding layer are connected with stacked context-aware relationship graph network layer;Stacked context-aware relationship graph network layer is also connected with readout module;According to battery current maximum available capacity and the reference capacity of each charging segment, determine battery health state.The application can improve evaluation accuracy.

Description

Technical Field

[0001] This application relates to the field of battery health status assessment, and in particular to a method and system for assessing the health status of lithium-ion batteries for electric vehicles based on a context-aware graph network model. Background Technology

[0002] Lithium-ion batteries are favored by new energy vehicles due to their high energy efficiency, lack of memory effect, long cycle life, and high power density. However, electrochemical side reactions during operation—such as solid electrolyte interfacial film growth, lithium deposition, and cathode degradation—inevitably lead to battery aging. The aging rate is closely related to user behavior, therefore most new energy vehicles are equipped with a Battery Management System (BMS) to limit harmful behaviors such as fast charging and overcharging. Aging increases safety risks, including internal short circuits and thermal runaway. While aging is unavoidable, its process can be managed. Accurately understanding the battery's State of Health (SOH) is crucial. SOH estimation allows the BMS to set appropriate operating limits, thereby extending battery life, improving safety, and enhancing energy efficiency. However, current methods for assessing battery state of health still need improvement in accuracy. Summary of the Invention

[0003] The purpose of this application is to provide a method and system for assessing the health status of lithium-ion batteries in electric vehicles based on a context-aware graph network model, which can improve the accuracy of lithium-ion battery health status assessment.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for assessing the health status of lithium-ion batteries in electric vehicles based on a context-aware graph network model, including: Acquire real-time monitoring data of electric vehicles during the stationary charging phase; Multiple charging segments are obtained by filtering based on the real-time monitoring data; The reference capacity of each charging segment is calculated using the reverse coulomb method; Extract feature vectors from the charging segment; Based on the feature vector, a trained context-aware graph network model is used to predict the current maximum usable capacity of the battery. The context-aware graph network model includes a context encoding layer, an embedding layer, a stacked context-aware graph network layer, and a readout module. The context encoding layer and the embedding layer are both connected to the stacked context-aware graph network layer. The stacked context-aware graph network layer is also connected to the readout module. The battery health status is determined based on the battery's current maximum available capacity and the reference capacity of each charging segment.

[0005] In one embodiment, multiple charging segments are obtained by filtering based on the real-time monitoring data, specifically including: The real-time monitoring data is preprocessed to obtain preprocessed data; the preprocessing includes: data cleaning, data segmentation, and data reassembly. The preprocessed data is sorted and filtered according to time order to obtain multiple charging segments.

[0006] In one embodiment, the feature vector includes: SOC change, average SOC, average total current, mileage, average temperature, segment number, SOC standard deviation, total current standard deviation, temperature standard deviation, and temperature change rate.

[0007] In one embodiment, the stacked context-aware graph network layer includes two context-aware graph network layers; The context-aware graph network layer includes a point-to-point pairing network, an edge network, a message network, a multi-head attention mechanism, a weighted aggregation module, and a gated loop unit. The point-to-point pairing network is connected to the context coding layer, the embedding layer, the edge network, and the message network, respectively; the edge network and the message network are also connected to the multi-head attention mechanism, respectively; the edge network, the message network, and the multi-head attention mechanism are also connected to the weighted aggregation module; the gated loop unit is connected to the point-to-point pairing network, the weighted aggregation module, and the readout module, respectively.

[0008] In one embodiment, the readout module includes an importance score calculation module, a weighted summation module, a layer normalization module, a hidden transformation module, a nonlinear activation module, and a projection module connected in sequence.

[0009] In one embodiment, the context coding layer is a multilayer perceptron.

[0010] In one embodiment, determining the battery health status based on the battery's current maximum available capacity and the reference capacity of each charging segment specifically includes: The battery health status is obtained by dividing the current maximum available capacity of the battery in each charging segment by the reference capacity of each charging segment.

[0011] Secondly, this application provides a health status assessment system for electric vehicle lithium-ion batteries based on a context-aware graph network model, comprising: The acquisition module is used to acquire real-time monitoring data of electric vehicles during the stationary charging phase. The filtering module is used to filter the real-time monitoring data to obtain multiple charging segments; The calculation module uses the reverse coulomb calculation method to calculate the reference capacity of each charging segment; The extraction module is used to extract feature vectors based on the charging segment; The prediction module is used to predict the current maximum usable capacity of the battery based on the feature vector using a trained context-aware graph network model. The context-aware graph network model includes a context encoding layer, an embedding layer, a stacked context-aware graph network layer, and a readout module. The context encoding layer and the embedding layer are both connected to the stacked context-aware graph network layer. The stacked context-aware graph network layer is also connected to the readout module. The battery health status determination module is used to determine the battery health status based on the current maximum available capacity of the battery and the reference capacity of each charging segment.

[0012] In one embodiment, the filtering module specifically includes: The preprocessing submodule is used to preprocess the real-time monitoring data to obtain preprocessed data; the preprocessing includes: data cleaning, data segmentation, and data reassembly. The sorting and filtering module is used to sort and filter the preprocessed data according to time order to obtain multiple charging segments.

[0013] In one embodiment, the feature vector includes: SOC change, average SOC, average total current, mileage, average temperature, segment number, SOC standard deviation, total current standard deviation, temperature standard deviation, and temperature change rate.

[0014] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method and system for assessing the health status of lithium-ion batteries in electric vehicles based on a context-aware graph network model. It filters real-time monitoring data from when the electric vehicle is in a stationary charging phase to obtain charging segments, effectively isolating the battery health status analysis from the influence of different driving behaviors, thereby improving the accuracy of the assessment. The system uses a trained context-aware graph network model to predict the battery's current maximum available capacity, and then combines this with the reference capacity of each charging segment to determine the battery health status, further improving the accuracy of the battery health status assessment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a framework diagram for end-to-end capacity estimation based on CARGN. Figure 2 A schematic diagram showing the raw data of a single vehicle throughout its entire lifecycle. Figure 3 This is a schematic diagram of parameter changes within a typical charging segment; Figure 4 Flowchart for data preprocessing and reference capacity calculation; Figure 5 A schematic diagram of the reference capacity degradation trajectory after screening representative vehicles; Figure 6 This is a schematic diagram of the CARGN model architecture; Figure 7 This is a diagram illustrating the process of creating a graph. Figure 8 This is a schematic diagram illustrating the update of the first node under the edge weight matrix of the first head; Figure 9 This is a schematic diagram of the multi-head aggregation process for the first node; Figure 10 This is a schematic diagram of the node update process in a GRU cell; Figure 11 This is a diagram of the overall experimental architecture for capacity estimation. Figure 12 A diagram showing the comparison of the number of parameters for all models; Figure 13 A comparison chart of capacity estimation results under different test sets and algorithms; Figure 14 A comparison chart of capacity point estimation metrics for all models on the full test set; Figure 15 Comparison plot of capacity point estimates for all models on a given vehicle; Figure 16 A schematic diagram illustrating the global feature importance analysis using multiple methods; Figure 17 This is a schematic diagram illustrating the importance of local features for a single sample. Figure 18 The figure shows the robustness test results of the model under Gaussian noise interference; Figure 19 A schematic diagram illustrating the real-time operation and analysis of all models; Figure 20A flowchart illustrating a method for assessing the health status of electric vehicle lithium-ion batteries using a context-aware relational graph network model, provided as an embodiment of this application; Figure 21 A schematic diagram of functional modules of an electric vehicle lithium-ion battery health status assessment system based on a context-aware relational graph network model, provided in an embodiment of this application; Figure 22 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] In one exemplary embodiment, such as Figure 20 As shown, a method for assessing the health status of lithium-ion batteries in electric vehicles based on a context-aware graph network model is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method includes the following steps.

[0020] Step 101: Obtain real-time monitoring data of the electric vehicle during the stationary charging phase.

[0021] Step 102: Filter the data based on the real-time monitoring data to obtain multiple charging segments.

[0022] Step 103: Calculate the reference capacity for each charging segment using the reverse coulomb method.

[0023] Step 104: Extract feature vectors based on the charging segment.

[0024] Step 105: Based on the feature vector, use the trained context-aware graph network model to predict the current maximum usable capacity of the battery; the context-aware graph network model includes a context encoding layer, an embedding layer, a stacked context-aware graph network layer, and a readout module; the context encoding layer and the embedding layer are both connected to the stacked context-aware graph network layer; the stacked context-aware graph network layer is also connected to the readout module.

[0025] Step 106: Determine the battery health status based on the current maximum available capacity of the battery and the reference capacity of each charging segment.

[0026] In an exemplary embodiment, multiple charging segments are obtained by filtering the real-time monitoring data. Specifically, this includes: preprocessing the real-time monitoring data to obtain preprocessed data; the preprocessing includes: data cleaning, data segmentation, and data reassembly; and sorting and filtering the preprocessed data in chronological order to obtain multiple charging segments.

[0027] In an exemplary embodiment, the feature vector includes: SOC change, average SOC, average total current, mileage, average temperature, segment number, SOC standard deviation, total current standard deviation, temperature standard deviation, and temperature change rate.

[0028] In an exemplary embodiment, the stacked context-aware graph network layer includes two context-aware graph network layers; the context-aware graph network layer includes a point-to-point pairing network, an edge network, a message network, a multi-head attention mechanism, a weighted aggregation module, and a gated loop unit; the point-to-point pairing network is connected to the context encoding layer, the embedding layer, the edge network, and the message network, respectively; the edge network and the message network are also connected to the multi-head attention mechanism, respectively; the edge network, the message network, and the multi-head attention mechanism are all connected to the weighted aggregation module; the gated loop unit is connected to the point-to-point pairing network, the weighted aggregation module, and the readout module, respectively.

[0029] In an exemplary embodiment, the readout module includes an importance score calculation module, a weighted summation module, a layer normalization module, a hidden transformation module, a nonlinear activation module, and a projection module connected in sequence.

[0030] In practical applications, the context coding layer is a multilayer perceptron.

[0031] In an exemplary embodiment, the battery health status is determined based on the current maximum available capacity of the battery and the reference capacity of each charging segment, specifically by dividing the current maximum available capacity of the battery in each charging segment by the reference capacity of each charging segment to obtain the battery health status.

[0032] In another exemplary embodiment, a specific process from modeling, training to application of a context-aware graph network model-based method for assessing the health status of lithium-ion batteries in electric vehicles is also provided in practical applications. Figure 1 and Figure 11 As shown, it includes the following steps.

[0033] S1: Acquire vehicle-related data and process it, then divide the dataset.

[0034] S2: Feature Extraction. This application selects ten features. These features, as shown in numerous studies, can indicate battery degradation. Their selection helps the model identify battery aging characteristics, thereby more accurately estimating the remaining maximum capacity.

[0035] S3: Construct the Context-Aware Relational Graph Network (CARGN) model. The model was trained and optimized using the prepared training set. In addition to the proposed CARGN model, six other state-of-the-art models were introduced for comparison. These models underwent the same training settings and optimization procedures to ensure that all models achieved optimal performance at similar levels of complexity. Training hyperparameter settings are shown in Table 5.

[0036] S4: Model evaluation using a test set. Each optimized and tuned model was tested on the same test set, and their performance in the capacity estimation task was compared under completely unseen operating conditions. Additionally, this application conducted noise robustness testing, interpretability analysis, and real-time operational efficiency evaluation on the same test set.

[0037] S5: After obtaining the model verified by the above evaluation, the multi-dimensional feature vector extracted from the battery charging stage is used as input, and the estimated battery capacity value is used as output. The output result, compared with the battery's nominal capacity, represents the current health status of the battery. The estimated battery capacity value is the current maximum usable capacity.

[0038] In the data acquisition and processing section S1, which involves partitioning the dataset, methods for segmenting and preprocessing large datasets are utilized. The main objective is to mitigate the adverse effects of data noise and abnormal fluctuations on the analysis results, and a data segmentation strategy is defined. To improve the accuracy of capacity estimation, this application uses data segments from the vehicle's stationary charging process. This method effectively isolates the analysis from the influence of different driving behaviors, thus providing a more reliable baseline for estimating the maximum remaining capacity. Furthermore, this section details the calculation method for the reference capacity, the data filtering process, and the dataset partitioning strategy.

[0039] Data source and description.

[0040] The dataset used in this application comes from the National Big Data Alliance for New Energy Vehicles. It contains real-time monitoring data collected during actual vehicle operation, ensuring its authenticity and reliability. This dataset is very comprehensive, covering a wide range of parameters. In addition to battery-specific data, it includes information such as vehicle operating status (e.g., driving, parking, stationary charging), cumulative mileage, timestamps, and various temperature readings. Furthermore, the maximum recorded mileage and data volume for each vehicle show significant differences in usage patterns. Some vehicles are used less frequently, while others are used frequently, meaning our analysis covers vehicles with different charging frequencies. Figure 2 The raw data for the parameters selected for a single vehicle throughout its operating history is displayed. The data used, along with battery-related information, is shown in Table 1.

[0041] Table 1 Vehicle and Battery Information

[0042] Data preprocessing.

[0043] Due to sensor noise or severe vibrations during vehicle use, raw data may exhibit significant anomalies, such as sudden fluctuations or data corruption. Therefore, preprocessing the raw data is crucial. The initial preprocessing stage mainly includes data cleaning, segmentation, and data reorganization. First, the raw data is sorted chronologically. Then, driving and charging segments are identified based on vehicle status, charging status, and speed. The segmented segments are reorganized and aggregated into their respective files. Although the amount of data varies between different vehicles, resulting in inconsistent data volume after cleaning, only data from vehicles with larger driving distances is ultimately used to ensure the practical significance of the results. Figure 3 This shows the trends in SOC, range, and total current during a typical charging period.

[0044] Capacity calculation and dataset partitioning.

[0045] This application uses the inverse coulomb method to calculate the reference capacity. Appropriate charging segment data is selected, and this method is applied to obtain the fully charged battery capacity. Subsequently, the interquartile range (IQR) method is used to remove outliers generated by abnormal segments to prevent deviations in the capacity degradation trend.

[0046] Reference capacity calculation.

[0047] This application adopts a capacity-based SOH evaluation method because it is directly related to the basic energy storage capacity of the battery. All formulas used for reference capacity calculation and screening are summarized in Table 2. The SOH definition based on capacity is shown in Formula (1).

[0048] Directly using laboratory charge-discharge tests for capacity measurement is not feasible in actual vehicles. Therefore, the ampere-hour integration method is adopted, which calculates the SOC change by integrating the current over time (Equation (2)). This application uses the inverse ampere-hour integration method for capacity estimation, that is, calculating the ratio of the charge change to the SOC change during the charging event, as shown in Equation (3). For discrete data samples, the discrete form of the equation is used, as shown in Equation (4).

[0049] like Figure 4 As shown, to ensure accuracy, the charging segments are carefully screened. Boundary conditions based on the initial SOC and the range of SOC variation are applied, as shown in Equations (5) and (6). To reduce errors caused by fluctuations, the capacity change rate is constrained by Equation (7). After calculating the reference capacity of all charging segments for a vehicle, outliers are removed using the IQR method according to Equation (8). Before performing linear fitting, the capacity data is smoothed using a Savitzky-Golay filter (window length = 5, polynomial order = 3) to reduce noise while preserving the trend. Finally, linear fitting is applied to verify whether the degradation slope conforms to the normal aging pattern, as shown in Equation (9), to determine whether the vehicle data should be retained for further analysis. The reference capacity degradation trajectory of the representative vehicle after screening is shown in Equation (9). Figure 5 As shown.

[0050] Table 2 Reference Capacity Calculation Formula

[0051] in, This represents the battery's current maximum remaining capacity. Nominal capacity The SOC value at time t; Initial SOC for charging; This is the charging current. ΔQ represents the depth of charge; k represents the change in battery charge during charging; and k represents the end time of a single charging segment. Δt represents the charging segment duration; Δt represents the time difference between data points. Let k be the charging current at time k; Let SOC be the value at time k. To estimate the convergence of the results; The acceptable convergence range is set at 5% in this application. It is the first quartile; It is the second quartile; is the third quartile; IQR is the interquartile range; b is the slope of the linear fit between the reference capacity and the charging segment number; x is the charging segment number. Ψ: Reference capacity; n: Total number of charging segments; Ψ: Maximum slope threshold. t0: Initial time. The maximum available capacity of this segment (calculated discretely). Let j be the discrete current value at a certain moment. For the initial time SOC, This represents the change in SOC during the charging segment. For a certain moment, SOC The maximum available capacity of a certain segment. This represents the maximum usable capacity of the previous segment. Combined with the preceding symbol, it indicates that the capacity difference between two adjacent segments should not be too large, conforming to the normal battery variation pattern. This is the average of the charging segment numbers within a battery usage cycle. This represents the average maximum usable capacity of a single charge segment within a battery life cycle. To define a slope limit, the formula calculates a first-order fitted slope for capacity changes with the charging segment sequence. To filter usable data, the overall data trend must conform to the battery aging trend. It is set to -0.005 here.

[0052] Dataset partitioning.

[0053] To verify the effectiveness and generalization ability of the model, a batch-based sampling strategy was used to partition the complete dataset containing 60 vehicles in a 5:1 ratio. The dataset comes from 5 predefined vehicle batches, each with a different number of vehicles. To evaluate the performance of the algorithm across all batches, the test set contains vehicles from each batch, with 2 vehicles randomly selected from each batch for testing. Specific data partitioning is shown in Table 3. The final result is a training set of 50 vehicles and an independent test set of 10 vehicles.

[0054] Table 3 Dataset Partitioning

[0055] S2: Feature Extraction. This application selects ten features. These features, as shown in numerous studies, can indicate battery degradation. Their selection helps the model identify battery aging characteristics, thereby more accurately estimating the remaining maximum capacity.

[0056] Feature engineering and normalization.

[0057] Feature extraction.

[0058] This application selects ten key feature parameters from multiple dimensions to estimate the remaining capacity of lithium-ion batteries. These features include SOC change (ΔSOC), average SOC (SOC̄), average total current (Current̄), mileage, average temperature (Tem̄), segment number, SOC standard deviation (SOC_σ), total current standard deviation (Current_σ), temperature standard deviation (Tem_σ), segment index within a single vehicle, and temperature change rate (r_T). The selection of these features is closely related to the battery aging mechanism: features related to SOC and current are directly associated with active lithium loss, electrode material phase transition stability, and side reaction rates; temperature-related features are key factors affecting electrolyte decomposition, SEI growth, and internal resistance changes, and their fluctuations can lead to mechanical fatigue of electrode materials; operational record features provide cumulative usage load and time context information, helping to capture long-term capacity degradation trends. This comprehensive feature engineering strategy ensures that the model can accurately identify battery aging characteristics, laying a solid foundation for accurately estimating the remaining maximum capacity. Among all features, mileage and segment number can be directly extracted. All the above averages are arithmetic means.

[0059] The method for calculating the change in SOC is as follows: .

[0060] Where ΔSOC is the change in SOC. It is the maximum SOC value of the current segment. It is the minimum SOC value of the current segment.

[0061] The method for calculating the rate of temperature change is as follows: .

[0062] in The rate of temperature change is obtained through linear regression using the least squares method, where i is the index of the temperature data point (1,2,…,m). It is the temperature value of the i-th data point. It is the average of the indices. is the mean temperature value, and m is the number of data points used for fitting.

[0063] Normalization.

[0064] Data normalization eliminates bias caused by differences in feature units and numerical ranges. By normalizing the input features and target capacity values ​​to the [0,1] interval, the contribution of each feature to model training is more balanced. This application uses the Min-Max method to normalize the features and reference capacity separately. The specific normalization process is as follows: The scaler learns only the minimum and maximum values ​​for each feature or target on the training set. Then, the transformation is applied according to the following formula: .

[0065] Where X is the original data, It is the minimum value of this type of data in the training set. It is the maximum value. It is the transformed data, whose value is mapped to the interval [0,1].

[0066] S3 includes S31 and S32.

[0067] S31: The process of building the model and the model content are as follows: The core innovation of CARGN lies in dynamically constructing a graph structure from multi-source heterogeneous features acquired during battery operation and using learnable context vectors to achieve global awareness of the system state. These vectors guide the dynamic evolution of relationships between nodes in the graph and subsequent information aggregation. The model ultimately adaptively generates weights based on real-time battery state, extracts the most critical information from the features, and uses this information to predict battery capacity. The CARGN model is shown in Table 4. Figure 6 As shown. The embedding layer is the input embedding layer.

[0068] Table 4 Overall Architecture of CARGN Model

[0069] S311. Overall network architecture.

[0070] The overall network model adopts a four-layer architecture, such as Figure 6 As shown, its specific structure is shown in Table 4. The input is a multi-dimensional feature vector extracted from the battery charging section, and the output is the estimated battery capacity value. The four basic components of the model are described below.

[0071] The top layer is the context encoding layer. This module acts as a feature compressor, mapping the input high-dimensional feature vector to a low-dimensional, dense context vector. Given the highly nonlinear relationship between the input features and the target capacity, the mapping network is constructed from alternating linear and nonlinear layers. The resulting context vector encapsulates the global operating state of the battery system, providing necessary contextual information for subsequent node feature fusion. Simultaneously, the input embedding layer independently performs high-dimensional mapping on each scalar feature. This process effectively enriches the representation of each feature by projecting it into a higher-dimensional space. Specifically, this layer transforms each feature into a high-dimensional representation, forming a set of initial node feature vectors. This prepares a rich feature representation for subsequent graph structure computation.

[0072] The core computations of the model are performed by stacked CARGN layers. This application sets the number of CARGN layers to two to balance model complexity and representational power. A single layer is insufficient to capture complex feature interactions, while more layers increase the risk of overfitting without providing a significant performance improvement.

[0073] Within these layers, the model performs dual message passing and node state updates. The process begins with dynamic relation construction: initial node feature pairs are combined with context vectors, and the input edge network computes weight matrices representing the relationships between feature pairs, thus forming a dynamic graph. Simultaneously, feature nodes are transformed into messages through linear transformations for subsequent data updates. Subsequently, a multi-head mechanism (four heads in this application) is used, with each head independently processing the messages using computed edge weights, producing vectors. The output vectors of the four heads are combined using attention weights generated based on each head's output vector to obtain final aggregate nodes that integrate the global state and the relationships between features. Finally, the initial node states and aggregate nodes pass through a Gated Recurrent Unit (GRU) to filter information and update the node states. These nodes are then passed to the next CARGN layer to continue the above process for further refinement.

[0074] The final output complex nodes contain comprehensive and refined information after double updates. This information is crucial for accurate capacity prediction, making the CARGN layers key to model accuracy. These layers integrate feature relationships, global context, and initial features. More importantly, the connection strength between nodes is not fixed but dynamically calculated based on context vectors and current node features. This mechanism enables the model to adaptively capture key information under complex and variable operating conditions and external disturbances, achieving dynamic context-aware relationship reasoning, significantly improving prediction accuracy and providing significant robustness.

[0075] In the final stage of the architecture, the updated node features are passed to the readout module. This module first uses an attention mechanism to automatically weigh the importance of different nodes, and then aggregates the information from all nodes to generate a unified graph representation vector. This vector is finally transformed into the final battery capacity prediction through a feedforward neural network.

[0076] S312. Core Components and Principles.

[0077] Multilayer perceptron.

[0078] The process begins with a context encoding mechanism based on a Multi-Layer Perceptron (MLP). This component acts as a feature compressor. An MLP is a basic feedforward neural network consisting of an input layer, at least one hidden layer, and an output layer. The formula for the forward propagation of the l-th layer is as follows.

[0079] .

[0080] .

[0081] in It is the weight matrix of the l-th layer. It is a bias vector. It is the activation value of the previous layer. σ is the activation value of the current layer, and σ is the activation function. This is the result of weighted summation and biasing of all inputs to the current layer.

[0082] The hidden layers contain non-linear activation functions, enabling the model to capture complex non-linear patterns and interactions between input features. The generated context vector effectively encapsulates the global operating state of the battery system. This vector is injected into each subsequent relation layer, dynamically adjusting the calculation of relationships between features based on the overall battery state. This design significantly enhances the model's expressive power and adaptability to different operating scenarios.

[0083] The graph network consists of the first four rows of the CARGN layer.

[0084] In the dynamic relationship graph construction and information aggregation stages, each original feature point is transformed into a high-dimensional node feature vector through the input embedding layer, thereby enriching its representational power and effectively extending a single feature to a multi-dimensional representation. Each feature point is independently mapped to a node. All nodes are input into the edge network to create a graph that records fine-grained interactions between nodes within a head. The specific process is as follows: Figure 7 As shown, the model dynamically calculates the connection strength (edge ​​weight) between any two nodes i and j through an edge network. This network incorporates the features of the source nodes. Target node features The concatenation of the global context vector and the global context vector is used as input, as shown in the following formula.

[0085] .

[0086] in and is the node feature vector, c is the context vector, and || indicates concatenation. Let be the edge weights between the two nodes.

[0087] Crucially, all nodes participate in this process, with each node acting as both a source and a target in an exhaustive pairing and combination process to form a complete set of relationships. Each node is connected to all other nodes (including self-connections), forming a comprehensive graph structure. The inclusion of context vectors is to contextualize relationships in the current global state, enabling the model to adaptively evaluate the influence between nodes under different environmental conditions. For example, in high-rate battery discharge scenarios, context vectors help capture enhanced synergistic changes between temperature and SOC by adjusting edge weights accordingly.

[0088] The calculated edge weights are normalized over the entire graph using the Softmax function to obtain the attention coefficients. , The edge weight is denoted as .

[0089] .

[0090] Multi-head attention mechanism.

[0091] To capture diverse relationship patterns, the model employs a multi-head mechanism. Each head k independently computes a unique set of edge weights. Message aggregation is performed from different perspectives. Subsequently, a head attention mechanism is introduced to autonomously learn the importance weights of each head. The outputs of all k heads are weighted and summed to generate the final aggregated information for each node. This process is controlled by the following formula. Figure 8 This explains how the first node is updated based on the edge weight matrix of the first head. Figure 9 The multi-head aggregation process of the first node is shown.

[0092] .

[0093] in It is the feature vector of the i-th node output by the h-th head. It is the weight of the i-th node in the h-th head. It is the final aggregate vector of the i-th node. H is the total number of heads.

[0094] Gated loop unit.

[0095] After aggregation, the model is integrated using GRU. and To update node states. In CARGN, GRU-based updates offer dynamic modeling compared to simpler methods. By combining gating mechanisms and memory persistence, it captures dependencies in battery capacity data to achieve accurate state transitions, enhancing the learning of complex feature interactions.

[0096] In this application, the use of GRU differs from traditional paradigms. The original nodes from the input embedding layer serve as the previous hidden state, while the input is an aggregated node with global and relational features. Therefore, unlike traditional GRU, the hidden state and input come from the same computational step. The update mechanism is structurally similar: traditional GRU updates the previous hidden state using the current input, while this application uses aggregated nodes to refine the features of the original nodes. GRU gating can selectively preserve historical information and seamlessly integrate new information.

[0097] The basic update equation is shown in the following formula. The information update process in GRU is as follows: Figure 10 As shown, GRU's gating mechanism effectively filters and integrates new and old information.

[0098] .

[0099] Specifically, resetting the gate r is done through The calculation determines how much hidden state information to discard when calculating candidate states. The update gate z uses... Calculate and control the mixing ratio between newly calculated candidate states and hidden states. Candidate hidden states Subsequently according to Derivation. Finally, the GRU unit outputs the updated node features by performing linear interpolation between the hidden state and the candidate state. like definition.

[0100] Where r is the output vector of the reset gate, and σ is the sigmoid activation function. It is the weight matrix of the reset gate input vector. It is the bias term of the reset gate input vector. It is the weight matrix for resetting the hidden state of the door. This is the bias term for resetting the hidden state of the door. z is the output vector for updating the door. It is the weight matrix for updating the gate input vector. It is the bias term for updating the gate input vector. It is the weight matrix for updating the hidden state of the gate. It is the bias term for updating the hidden state of the door. These are candidate hidden states, and tanh is the hyperbolic tangent activation function. It is the weight matrix of the input vector when calculating the candidate hidden state. These are the corresponding bias terms. It is the weight matrix of the hidden state. ⊙ represents the corresponding bias term, and ⊙ indicates element-wise multiplication. It is the final output node that integrates the original node features and aggregated information. This represents new information "absorbed" from the current candidate state. This indicates information "inherited" from a previous state. This is the initial node feature vector.

[0101] Read module.

[0102] In the final stage, the model generates a capacity estimate through a hierarchical readout and prediction module. This module employs an attention mechanism for each updated node feature. Calculate importance score These scores are normalized to weights using the Softmax function. Then, the features of all nodes are weighted and summed to generate a global graph representation vector F. This synthesized graph vector is then passed to the prediction network. The prediction network finally transforms the processed global representation into a final battery capacity prediction through a non-linear mapping. .

[0103] .

[0104] S32: Model optimization and evaluation metrics.

[0105] Table 5 Training Hyperparameter Settings

[0106] S32.1. Hyperparameter adjustment.

[0107] To ensure accurate estimation, this application requires hyperparameter tuning during training. To ensure a fair comparison of different model architectures, all models maintain similar parameter complexity and are not individually tuned. Optimization settings are shown in Table 6.

[0108] The training and validation process for each model is treated as a black-box objective function. It takes the initial learning rate as input and the final MAPE (Matrices Per Scale) of the output model on the training set as the performance metric. Learning rates that are too large or too small can hinder convergence or optimization. The optimal initial learning rate for CARGN was determined to be 0.0006. The model with the lowest training MAPE is selected as the final model. The same search strategy is applied to the comparison models.

[0109] Table 6 Optimization Settings

[0110] S32.2. Evaluation Indicators.

[0111] To comprehensively evaluate prediction accuracy, this application uses the following evaluation metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination R0. 2 Mean Absolute Percentage Error (MAPE). The formulas for calculating these indicators are as follows: .

[0112] .

[0113] .

[0114] .

[0115] Where n is the number of data points. This is the actual value. It is a predicted value. It is the arithmetic mean of the true values.

[0116] S4: Evaluate the models using a test set. Each optimized and tuned model is tested on the same test set, and their performance in the capacity estimation task is compared under completely unseen operating conditions.

[0117] Figure 13 The overall estimation results of the CARGN model and the contrastive model on the test set are presented. A comparative analysis is performed with six well-known regression models in the field, including Convolutional Neural Networks (CNN), LSTM, Gated Recurrent Units (GRU), Graph Attention Networks (GAT), Transformer, and Temporal Convolutional Networks (TCN).

[0118] These six baseline models were chosen because they represent popular deep learning architectures in the battery SOH estimation literature. CNN and TCN are standard methods for spatial-temporal feature extraction, while LSTM and GRU are widely used for sequence data modeling. GAT shares a graph-based foundation with CARGN but employs a different relation learning mechanism. This selection establishes a comprehensive benchmark against established methods, ensuring that the comparison covers both traditional and state-of-the-art methods commonly used in capacity prediction, thus providing a meaningful performance context. All model comparisons are performed with similar numbers of parameters, and a detailed comparison of the number of parameters for all models is as follows: Figure 12 As shown. Among them, Figure 12 In the diagram, A stands for CNN, B for GRU, C for Transformer, D for LSTM, E for GAT, F for TCN, and G for CARGN.

[0119] Subsequently, the model's internal attention mechanism is combined with Sobol sensitivity analysis and SHAP analysis to explain the CARGN model's selection of important features during the estimation process, thus clarifying the principles behind its predictions. Furthermore, real-time running analysis of the model is performed to evaluate its practical performance and potential for deployment in automotive BMS. Finally, Gaussian noise is introduced into the test set to assess the model's robustness.

[0120] Model superiority.

[0121] During model testing, the proposed CARGN demonstrated significant superiority and robustness in the battery capacity estimation task. Test results for seven models are presented, including... Figure 14 (Summarizing the estimation results and overall metrics of all vehicles to provide a comprehensive perspective for model evaluation), and Figure 15(As an example, the specific point estimation results for a particular vehicle are shown.) Among them, Figure 14 (a) in the text represents RMSE. Figure 14 (b) in the text represents MAE. Figure 14 (c) in the equation is 1-R 2 , Figure 14 (d) in the text represents MAPE. Figure 15 The capacity point estimates of each model on a specific vehicle in the test set are compared. Figure 15 (a) shows the estimation results of the CNN model. Figure 15 (b) shows the estimation performance of the GRU model. Figure 15 (c) shows the estimation results of the Transformer model. Figure 15 (d) in the figure shows the estimation performance of the LSTM model. Figure 15 (e) in the figure shows the estimation results of the GAT model. Figure 15 (f) in the figure shows the estimation effect of the TCN model. Figure 15 (g) in the figure shows the estimation results of the CARGN model.

[0122] Performance comparisons show that CARGN achieved the best performance across most of the test vehicles. Specifically, it achieved the lowest MAE and MAPE on 8 out of 10 test vehicles, and the lowest RMSE and 1-R on 7 out of 10 vehicles. 2 It demonstrates excellent individual prediction capabilities. In terms of overall estimation accuracy, CARGN achieves an RMSE of 2.24, a MAE of 1.74, a MAPE of 1.68%, and a 1-R... 2 The accuracy is 0.43, which is better than all other benchmark models. This result validates its higher point estimation accuracy.

[0123] The superior performance of CARGN can be attributed to its unique context-aware graph structure, which explicitly models the interactions and dependencies between battery parameters. This allows for more effective utilization of multi-dimensional time-series features. Therefore, CARGN is not only a high-precision tool for battery SOH estimation, but also provides a novel perspective for the joint modeling of complex system parameters, possessing significant engineering application value and theoretical reference significance.

[0124] Interpretability analysis.

[0125] Model interpretability remains a challenge for data-driven approaches. The "black box" nature of neural networks complicates understanding their decision-making mechanisms. This application explains the superior performance of CARGN by analyzing its attention weights and comparing them with SHAP values ​​and Sobol sensitivity analysis results.

[0126] From a global interpretability perspective, the extracted attention weights and the two analysis methods show significant consistency in terms of the most important features. For example... Figure 16 As shown, the analysis of average attention weights clearly reveals the distribution of model attention across different features, where, Figure 16 In the table, (a) represents the overall average attention weight. Figure 16 (b) in the figure represents the overall Sobol analysis. Figure 16 (c) in the diagram represents the overall SHAP analysis. SHAP analysis provides validation from another perspective, quantifying the impact of each feature on the model output. Furthermore, Sobol sensitivity analysis confirms the importance of key features from the perspective of parameter sensitivity. Figure 16 The meanings of the features are as follows: (SOC standard deviation) (Changes within the SOC segment) (Average current within the segment) (Standard deviation of temperature) (SOC average), mileage (Current labeling difference) (Slope of the first-order temperature fit within the segment), segment index (the segment number within a single vehicle), and (Average temperature).

[0127] In summary, these methods demonstrate that SOC_σ and ΔSOC are the two core features assigned the highest weight by the model. This finding is entirely consistent with expertise in the battery technology field, as the degree of SOC fluctuation is indeed a key physical factor affecting battery health. This consistency indicates that the CARGN model successfully captures the main drivers of battery capacity degradation. At a local level, such as Figure 17 As shown, Figure 17 In the diagram, (a) represents the attention weight for sample 0. Figure 17 (b) shows the SHAP waterfall plot for sample 0, providing a more granular analytical perspective on the attention weight distribution for a single sample. Both show that SOC_σ and ΔSOC remain the most important features for that particular sample, despite their varying degrees of contribution. The SHAP waterfall plot further quantifies the specific impact of each feature, offering valuable insights into the model's decision-making process for individual instances.

[0128] The observed discrepancy in secondary feature importance ranking between internal attention weights and external interpretability methods can be attributed to several architectural and training factors within the CARGN model. First, the dynamic nature of the graph attention mechanism allows feature importance to be adjusted based on input sample features, which inherently differs from static global importance metrics. Second, multi-head attention architectures may capture different feature interactions that are not entirely consistent with the linear importance assessment of SHAP. Furthermore, complex feature transformations through multiple CARGN layers may also obscure the direct correspondence between internal attention distributions and external importance metrics.

[0129] This divergence actually reflects a key advantage of the CARGN model's built-in attention mechanism: it can dynamically filter features and adaptively adjust feature importance weights based on the features of each input sample, rather than relying on a fixed, universal ranking. This dynamic characteristic allows the model to flexibly adapt to capacity estimation needs under different operating conditions, highlighting its advantage in intelligent context-aware decision-making.

[0130] Nevertheless, the inconsistency between internal attention weights and external interpretability rankings points to a key direction for optimization. Future work could introduce interpretability-aware regularization to better align the attention distribution with physical priors while maintaining model flexibility. Architectural modifications, such as adding a consistency loss between attention weights and SHAP-like importance metrics, could further enhance decision coherence. These improvements would enhance the reliability and interpretability of CARGN without sacrificing its adaptability.

[0131] Robustness testing.

[0132] Robustness is crucial for maintaining stable performance in practical applications. To quantitatively assess sensitivity to disturbances, zero-mean additive white Gaussian noise (AWGN) was introduced into the current, temperature, SOC, and odometer signals before the model input, with a maximum noise rate of 5%, which determined the standard deviation. The maximum 5% AWGN exceeds typical sensor noise fluctuations.

[0133] All models exhibited only a slight performance degradation under AWGN, confirming their noise resilience. CARGN demonstrated superior performance at any noise level. In this setting, CARGN effectively suppressed noise propagation, and as... Figure 18 As shown, it maintained the highest accuracy among all models, highlighting its potential for practical application. Among them, Figure 18 (a) represents a 1% noise level. Figure 18 (b) represents a 2% noise level. Figure 18 (c) represents a 3% noise level. Figure 18 (d) represents a 4% noise level. Figure 18 (e) represents a 5% noise level.

[0134] Real-time operational analysis of CARGN.

[0135] Since this application aims to develop an algorithm that can be deployed in a real vehicle BMS, evaluating the algorithm's memory usage and estimation time is crucial. Experiments were conducted on a platform equipped with an Intel(R) Core(TM) i9-14900HX processor (2.20 GHz) and an RTX 4060 GPU. The algorithm's memory usage and average inference time were measured, and the results are as follows: Figure 19 As shown.

[0136] Figure 19 The comparison shows model memory usage and average inference time. It can be seen that CARGN performs well in both aspects, at 55.74 KB and 1.46 ms respectively. Compared to traditional models, CARGN requires less memory and its average inference time is typically less than 100 ms per estimate, meeting the requirements of BMS applications. Figure 19 In the figure, (a) represents the average reasoning time. Figure 19 In the diagram, (b) represents the model's memory usage. The letters represent: A (CNN); B (GRU); C (Transformer); D (LSTM); E (GAT); F (TCN); G (CARGN).

[0137] S5: After obtaining the model validated by the above evaluation, firstly, real-time monitoring data of the vehicle during the stationary charging phase is obtained from the electric vehicle's battery management system, including parameters such as voltage, current, temperature, cumulative mileage, and state of charge. This raw data is preprocessed: sorted by time, complete charging segments are selected, and the reference capacity corresponding to each charging segment is calculated using the inverse coulomb counting method as the baseline true value of the state of health (SOH). Simultaneously, the data needs to be cleaned to remove abnormal fluctuations and outliers, and methods such as sliding window filtering are used to smooth out capacity degradation trends.

[0138] Next, ten key features are extracted from each valid charging segment as input to the model. These features include: SOC change (the difference between the maximum and minimum SOC within the segment), average SOC, average charging current, cumulative mileage, average temperature, charging segment number, SOC standard deviation, current standard deviation, temperature standard deviation, and temperature change rate (obtained through linear regression fitting). All features and the target capacity value are normalized using a min-max normalization method to map the values ​​to the [0,1] interval to eliminate the influence of dimensions.

[0139] Then, the ten normalized feature vectors are input into the trained CARGN model. Internally, the model first compresses the input features into a low-dimensional, dense context vector through a context encoding layer, representing the battery's current overall operating state. Simultaneously, each original feature point is independently mapped to a high-dimensional node feature. In the stacked two-layer graph network, the model dynamically constructs a relational graph: concatenating the features of any two nodes with the context vector, calculating the connection strength (i.e., edge weights) between them through an edge network, forming an asymmetric adjacency matrix. A multi-head attention mechanism is used to aggregate information between nodes from multiple perspectives, and the state of each node is updated through a gated recurrent unit (GRU). After two layers of graph inference, the readout module uses an attention mechanism to weighted summation of the importance of each node, obtaining a global graph representation vector, which is finally output as a scalar value via a feedforward neural network.

[0140] This output value is the model's estimated maximum usable capacity of the battery (in ampere-hours). Dividing the estimated maximum usable capacity by the battery's nominal capacity or the baseline true value, and then multiplying by 100%, yields the battery's State of Health (SOH) as a percentage. For example, if the nominal capacity is 120 Ah and the model output is 96 Ah, then the SOH is 80%. This SOH value reflects the current degree of battery degradation and is a crucial basis for the battery management system to set operating strategies, ensure vehicle safety, and extend battery life.

[0141] This application offers a potential solution to key challenges in SOH estimation of lithium-ion batteries for electric vehicles, including the disconnect between laboratory data and real-world operating conditions, insufficient modeling of dynamic interactions between multiple physical features, and poor interpretability of data-driven models. A standardized processing workflow is established based on large-scale real-world vehicle operating data. The model employs a graph structure to dynamically capture relationships between battery features and is enhanced through a context-aware mechanism to better characterize complex aging patterns.

[0142] Experimental results show that CARGN achieves an RMSE of 2.24, a MAE of 1.74, a MAPE of 1.68%, and a 1-R score on the independent test set. 2 The accuracy was 0.43, outperforming the baseline model and demonstrating best-in-class or near-best-in-class estimation performance on most test vehicles. Interpretability analysis combining attention mechanisms, SHAP values, and Sobol sensitivity analysis confirmed that CARGN effectively identifies key features consistent with battery aging mechanisms, thus partially mitigating the "black box" problem of data-driven models. Under robustness testing with 5% AWGN, the model maintained the highest estimation accuracy. To evaluate the operational efficiency of a real-world BMS deployment, CARGN showed a memory footprint of 55.74 KB and an average inference time of 1.46 ms, meeting BMS requirements.

[0143] This application proposes a Context-Aware Graph Network (CARGN) model for SOH estimation and makes specific contributions to dataset construction, model design, runtime efficiency, and interpretability analysis. The proposed model demonstrates superior performance, robustness, and practical deployability on a large real-world electric vehicle dataset. An overall overview of this application is shown in the figure below. Figure 1 As shown, an end-to-end battery capacity estimation framework based on CARGN is illustrated, and the tests to be performed to validate the model's performance are outlined. The main contributions of this application are as follows: (1) This application is the first to apply a graph network that dynamically and explicitly models the interactions between physical features to SOH estimation. The context-aware mechanism adaptively adjusts the feature relationships based on the context vectors derived from all inputs, thereby enabling a deeper understanding of the battery aging mechanism.

[0144] (2) The model was validated on a large-scale dataset of real vehicles, covering a variety of operating conditions. Its superiority in accuracy and robustness was demonstrated through extensive comparisons with six state-of-the-art models.

[0145] (3) It provides a comprehensive interpretability analysis that combines internal attention mechanisms with external technologies (including Sobol sensitivity analysis and SHAP), which alleviates the "black box" problem and improves the confidence in SOH estimation.

[0146] (4) The model demonstrates excellent operating efficiency and minimal resource requirements, which promotes its deployability in actual vehicle BMS.

[0147] Based on the same inventive concept, this application also provides an electric vehicle lithium-ion battery health status assessment device based on a context-aware graph network model for implementing the above-described method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the electric vehicle lithium-ion battery health status assessment device based on a context-aware graph network model provided below can be found in the limitations of the electric vehicle lithium-ion battery health status assessment method based on a context-aware graph network model described above, and will not be repeated here.

[0148] In one exemplary embodiment, such as Figure 21 As shown, a health status assessment system for electric vehicle lithium-ion batteries based on a context-aware graph network model is provided, including: The acquisition module is used to acquire real-time monitoring data of electric vehicles during the stationary charging phase.

[0149] The filtering module is used to filter the real-time monitoring data to obtain multiple charging segments.

[0150] The calculation module uses the reverse coulomb calculation method to calculate the reference capacity of each charging segment.

[0151] An extraction module is used to extract feature vectors based on the charging segment.

[0152] The prediction module is used to predict the current maximum usable capacity of the battery based on the feature vector using a trained context-aware graph network model. The context-aware graph network model includes a context encoding layer, an embedding layer, a stacked context-aware graph network layer, and a readout module. The context encoding layer and the embedding layer are both connected to the stacked context-aware graph network layer. The stacked context-aware graph network layer is also connected to the readout module.

[0153] The battery health status determination module is used to determine the battery health status based on the current maximum available capacity of the battery and the reference capacity of each charging segment.

[0154] In one exemplary embodiment, the filtering module specifically includes: The preprocessing submodule is used to preprocess the real-time monitoring data to obtain preprocessed data; the preprocessing includes: data cleaning, data segmentation and data reassembly.

[0155] The sorting and filtering module is used to sort and filter the preprocessed data according to time order to obtain multiple charging segments.

[0156] In an exemplary embodiment, the feature vector includes: SOC change, average SOC, average total current, mileage, average temperature, segment number, SOC standard deviation, total current standard deviation, temperature standard deviation, and temperature change rate.

[0157] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 22As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores health status assessment data for electric vehicle lithium-ion batteries based on a context-aware graph network model. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model.

[0158] Those skilled in the art will understand that Figure 22 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment to which the present application is applied. Specific computer equipment may include, for example, [the following is a list of possible additional structures]. Figure 22 The embodiments show more or fewer components, combinations of certain components, or different component arrangements. In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the above-described method embodiments.

[0159] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the above-described method embodiments.

[0160] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method embodiments.

[0161] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0162] In this application, all actions to acquire signals, information, or data are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with the authorization granted by the owner of the relevant device.

[0163] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0164] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0165] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0166] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for assessing the health status of lithium-ion batteries in electric vehicles based on a context-aware graph network model, characterized in that, include: Acquire real-time monitoring data of electric vehicles during the stationary charging phase; Multiple charging segments are obtained by filtering based on the real-time monitoring data; The reference capacity of each charging segment is calculated using the reverse coulomb method; Extract feature vectors from the charging segment; Based on the feature vector, a trained context-aware graph network model is used to predict the current maximum usable capacity of the battery. The context-aware graph network model includes a context encoding layer, an embedding layer, a stacked context-aware graph network layer, and a readout module; the context encoding layer and the embedding layer are both connected to the stacked context-aware graph network layer; the stacked context-aware graph network layer is also connected to the readout module; The battery health status is determined based on the battery's current maximum available capacity and the reference capacity of each charging segment.

2. The method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model according to claim 1, characterized in that, Based on the real-time monitoring data, multiple charging segments are obtained through filtering, specifically including: The real-time monitoring data is preprocessed to obtain preprocessed data; the preprocessing includes: data cleaning, data segmentation, and data reassembly. The preprocessed data is sorted and filtered according to time order to obtain multiple charging segments.

3. The method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model according to claim 1, characterized in that, The feature vector includes: SOC change, average SOC, average total current, mileage, average temperature, segment number, SOC standard deviation, total current standard deviation, temperature standard deviation, and temperature change rate.

4. The method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model according to claim 1, characterized in that, The stacked context-aware graph network layer includes two context-aware graph network layers; The context-aware graph network layer includes a point-to-point pairing network, an edge network, a message network, a multi-head attention mechanism, a weighted aggregation module, and a gated loop unit. The point-to-point pairing network is connected to the context coding layer, the embedding layer, the edge network, and the message network, respectively; the edge network and the message network are also connected to the multi-head attention mechanism, respectively; the edge network, the message network, and the multi-head attention mechanism are also connected to the weighted aggregation module; the gated loop unit is connected to the point-to-point pairing network, the weighted aggregation module, and the readout module, respectively.

5. The method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model according to claim 1, characterized in that, The readout module includes an importance score calculation module, a weighted summation module, a layer normalization module, a hidden transformation module, a nonlinear activation module, and a projection module, which are connected in sequence.

6. The method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model according to claim 1, characterized in that, The context coding layer is a multilayer perceptron.

7. The method for assessing the health status of electric vehicle lithium-ion batteries based on a context-aware graph network model according to claim 1, characterized in that, The battery health status is determined based on the battery's current maximum available capacity and the reference capacity of each charging segment, specifically including: The battery health status is obtained by dividing the current maximum available capacity of the battery in each charging segment by the reference capacity of each charging segment.

8. A health status assessment system for electric vehicle lithium-ion batteries based on a context-aware graph network model, characterized in that, include: The acquisition module is used to acquire real-time monitoring data of electric vehicles during the stationary charging phase. The filtering module is used to filter the real-time monitoring data to obtain multiple charging segments; The calculation module uses the reverse coulomb calculation method to calculate the reference capacity of each charging segment; The extraction module is used to extract feature vectors based on the charging segment; The prediction module is used to predict the current maximum usable capacity of the battery based on the feature vector using a trained context-aware graph network model. The context-aware graph network model includes a context encoding layer, an embedding layer, a stacked context-aware graph network layer, and a readout module; the context encoding layer and the embedding layer are both connected to the stacked context-aware graph network layer; the stacked context-aware graph network layer is also connected to the readout module; The battery health status determination module is used to determine the battery health status based on the current maximum available capacity of the battery and the reference capacity of each charging segment.

9. The electric vehicle lithium-ion battery health status assessment system based on a context-aware graph network model according to claim 8, characterized in that, The filtering module specifically includes: The preprocessing submodule is used to preprocess the real-time monitoring data to obtain preprocessed data; the preprocessing includes: data cleaning, data segmentation, and data reassembly. The sorting and filtering module is used to sort and filter the preprocessed data according to time order to obtain multiple charging segments.

10. The electric vehicle lithium-ion battery health status assessment system based on a context-aware graph network model according to claim 8, characterized in that, The feature vector includes: SOC change, average SOC, average total current, mileage, average temperature, segment number, SOC standard deviation, total current standard deviation, temperature standard deviation, and temperature change rate.

Images