Light power battery health state estimation method based on GAT-BiGRU-Res model

Abstract

The invention discloses a light power battery health state estimation method based on a GAT-BiGRU-Res model, and the method comprises the steps: collecting the historical charging data of a light power battery, and dynamically selecting a 0.1 V optimal voltage window based on an IC peak value to obtain voltage data and capacity data; then, the voltage sequence and the capacity sequence are aligned through an interpolation method, quartering is conducted on the voltage sequence and the capacity sequence to capture local degradation characteristics, and time sequence data are converted into graph structure data through graph structure modeling; and finally, modeling a spatial relationship of non-adjacent local segments by using GAT, modeling time sequence dependence by using BiGRU, optimizing a training process by using Res, and outputting a discharge capacity estimation value by using a full connection layer. The method can effectively solve the problems of difficulty in obtaining complete data of the light power battery, uncertainty of use habits of users and interpretability of feature sequence selection at present, and aims to realize accurate perception of SOH of the light power battery, improve the use safety of the light power battery and reduce safety accidents caused by the light power battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of light power battery safety management, and particularly relates to a light power battery health state estimation method based on a GAT-BiGRU-Res model. BACKGROUND

[0002] Under the background of global climate governance and energy low-carbon transformation, green transformation in the transportation field has become a key measure to achieve the "double carbon" goal. China, as the world's largest new energy vehicle market, has formed an electric vehicle + light traffic tool dual driving electric pattern. Lithium ion batteries, with their high energy density, low self-discharge rate, long cycle life, no memory effect and wide working range, have become the key driving force for the development of the industry. However, the safety problems caused by performance degradation need to be broken through, and the development of high-precision and practical battery state of health (SOH) estimation algorithm is the core problem to ensure the safe operation of vehicles and promote the sustainable development of the industry.

[0003] Researchers at home and abroad have done a lot of research on battery SOH evaluation, which can be summarized as direct measurement method, model-driven method and data-driven method. The direct measurement method determines the SOH of the battery by measuring the capacity or internal resistance, including ampere hour counting method and electrochemical impedance spectroscopy method, which is simple and easy to implement, but needs to be carried out in some special environment, and is difficult to apply online. The model-driven method estimates SOH by describing and characterizing the internal electrochemical mechanism or external behavior characteristics of the battery. It mainly includes electrochemical model and equivalent circuit model. The electrochemical model relies on complex partial differential equations to describe the dynamic behavior of the battery, although it can deeply analyze the internal mechanism of battery performance degradation, but the calculation process is complicated, which limits its feasibility in real-time application. The equivalent circuit model is simple in structure, intuitive and easy to implement, and is favored. This kind of model usually uses filtering technology to track state variables and update model parameters. The model-based method has interpretability and logic, and can estimate the battery capacity from the essential reason of aging. However, this kind of method needs to update parameters regularly and has limited generalization ability, which is difficult to apply to different battery types and working conditions.

[0004] In recent years, with the wide application of artificial intelligence, data-driven methods have been widely used in lithium-ion battery SOH estimation. This method does not require understanding of the complex internal mechanism of the battery, and uses historical data extracted from the charging and discharging process or adopts an end-to-end method to achieve SOH estimation. Generally, traditional machine learning algorithms achieve SOH estimation by constructing network models and using artificial feature extraction and correlation analysis. However, these methods rely on manually extracted health indicators and are not sensitive to data changes, resulting in limited generalization. In contrast, deep learning methods can automatically extract features and perform end-to-end learning, avoiding the complexity of manual feature extraction and significantly improving the accuracy of SOH estimation. Although data-driven methods have shown great potential in battery state estimation, their inherent "black box" characteristics have brought significant limitations. This opacity makes it difficult to extract key features from complex models that lack intuitive physical meaning and sufficient explanatory power, and cannot effectively integrate complex chemical and physical degradation processes into the estimation model. ICA, as an important tool for extracting feature variables, converts the voltage platform of the charging curve into a differential capacity curve with significant peaks, realizing the coupling analysis of battery health status degradation process and internal electrochemical mechanism. In actual scenarios, the charging and discharging process of the battery is often carried out at multiple states of charge, making it difficult to obtain complete charging and discharging curve data. This uncertainty directly affects the reliability of SOH estimation methods based on feature extraction. Incomplete charging and discharging data may lead to inaccurate feature extraction, which in turn affects the performance of the SOH estimation model. Therefore, when developing SOH estimation methods, the complexity and randomness of the actual charging and discharging process and the interpretability of the extracted feature sequence must be considered. In addition, most existing estimation methods only consider the temporal information of the sequence, ignoring its spatial information. SUMMARY

[0005] In view of the deficiencies in the prior art, in order to further improve the abnormal early warning capability of the electric bicycle in use, effectively solve the problems of difficulty in obtaining complete data of the light power battery, uncertainty of user usage habits and interpretability of feature sequence selection, and realize accurate perception of the SOH of the light power battery, the present application proposes a light power battery health state estimation method based on a GAT-BiGRU-Res model. First, the historical charging data of the light power battery is collected, and the voltage data and capacity data are obtained based on the dynamic selection of the 0.1V optimal voltage window according to the IC peak value. Then, the interpolation method is used to align the voltage and capacity sequences, and four equal divisions are performed respectively to capture the local degradation characteristics. The time series data is converted into graph structure data through graph structure modeling. Finally, the spatial relationship of non-adjacent local segments is modeled by GAT, the time series dependence is modeled by BiGRU, and the training process is optimized by Res. The discharge capacity estimation value is output by the full connection layer to realize accurate estimation of SOH. Thus, the reliability and safety of the electric bicycle in use are improved, and the occurrence of safety accidents such as thermal runaway is reduced.

[0006] In order to solve the above technical problems, the present application proposes a light power battery health state estimation method based on a GAT-BiGRU-Res model, which comprises the following steps:

[0007] Step 1: Incremental capacity IC calculation is performed according to the collected historical charging data of the light power battery, wherein the historical charging data includes the voltage, current and time of the light power battery in the constant current stage; and the segment data is dynamically selected according to the IC curve peak value, wherein the segment data includes voltage data and capacity data obtained by the 0.1V optimal voltage window;

[0008] Step 2: According to the voltage data and capacity data of the segment data in step 1, the voltage sequence and the capacity sequence are aligned by using the interpolation method, and four equal divisions are performed respectively. The voltage-capacity feature vector of each sub-segment after four equal divisions is taken as a graph node, and the time series data is converted into graph structure data;

[0009] Step 3: According to the graph structure data in step 2, the spatial relationship of non-adjacent local segments is modeled by using the graph attention network (GAT) to capture the local spatial characteristics of the aging process of the light power battery, the time series analysis is performed by using the bidirectional gated recurrent unit (BiGRU) to capture the time sequence characteristics of the aging process of the light power battery, and the residual connection (Res) is used to optimize the training process. Finally, the discharge capacity estimation value is output by the full connection layer to realize accurate estimation of the health state (SOH).

[0010] Further, the light power battery health state estimation method according to the present application comprises the following steps:

[0011] The specific content of step 1 comprises:

[0012] The formula for incremental capacity IC calculation according to the collected historical charging data of light power batteries is as follows:

[0013]

[0014]

[0015] wherein, is the charging capacity, the unit is Ah; is the charging current, the unit is A; is the charging voltage, the unit is V, is the charging time, the unit is s;

[0016] The IC curve is smoothed using Gaussian filtering, and the calculation formula is as follows:

[0017]

[0018] wherein, represents the Gaussian filtering kernel function, is the mean of the Gaussian kernel, which determines the center position of the filtering kernel, is set to 0; is the input variable, which represents the position within the filtering window; is the standard deviation of the Gaussian kernel, which controls the width of the filtering kernel and the strength of the smoothing;

[0019] For the filtered IC curve, the voltage segment based on the peak value of the IC curve is dynamically selected within the voltage range of 0.1V, and the capacity data of the segment is defined as:

[0020]

[0021] (5)

[0022]

[0023] wherein, is the voltage segment data selected based on the peak value of the IC curve at the n th cycle, , , , is the voltage at the peak position of the IC curve at the n th cycle; is the capacity within the n th voltage window at the n th cycle, is the capacity when the voltage is , is the capacity when the voltage is , is the capacity when the voltage is , is the capacity when the voltage is is the number of feature points in the segment data, is the capacity data in the corresponding voltage segment;

[0024] In order to eliminate the influence of different data units, the original data of the selected voltage segment and capacity segment are respectively standardized, and the calculation formula is as follows:

[0025]

[0026] Among them, is the value of the original data after standardization, is the mean of the original data, is the standard deviation of the original data.

[0027] The specific content of step 2 includes:

[0028] The voltage data and capacity data of the segment data of step 1 are aligned using the cubic spline interpolation method, and the length of the feature sequence in each cycle is set to 80; In order to capture the local spatial features and time sequence features of the light power battery aging process, the voltage sequence and the capacity sequence are divided into four equal sub-segments, and each sub-segment represents a node in the figure, then the node feature is represented as follows:

[0029]

[0030] Among them, represents the node feature vector, represents vector splicing, represents the node;

[0031] The node feature matrix is represented as follows:

[0032]

[0033] Among them, represents the number of nodes, represents the feature dimension;

[0034] According to the joint similarity of the battery and the capacity, the edge is dynamically constructed, and the nonlinear dependence relationship between the sub-segments is captured, and the process is:

[0035] First, the cosine similarity is used to calculate the similarity matrix of the voltage and the similarity matrix of the capacity :

[0036]

[0037]

[0038] Among them, and denote the voltage matrix and the capacity matrix, respectively, denotes the 2-norm;

[0039] Then, the joint similarity matrix is generated by normalization and weighted fusion :

[0040]

[0041] wherein, is a weight parameter;

[0042] Finally, the edges are extracted and the self-loops are excluded to generate the sparse adjacency matrix :

[0043]

[0044] The graph structure data comprises a node feature matrix and a sparse adjacency matrix .

[0045] The specific content of step 3 comprises:

[0046] 3-1) According to the graph structure data of step 2, the spatial relationship of non-adjacent local fragments is modeled using a graph attention network (GAT), comprising:

[0047] According to the graph structure data of step 2, the node composed of the voltage sequence and the capacity sequence is linearly transformed using a learnable weight matrix in the graph attention layer of the graph attention network (GAT),

[0048]

[0049] wherein, and are the voltage-capacity feature vectors before and after node transformation, respectively;

[0050] The attention coefficient of the target node and its neighbor nodes is calculated through the attention mechanism :

[0051]

[0052] wherein, denotes a learnable parameter vector, denotes the transpose;

[0053] The attention coefficient is normalized using function to obtain the final attention weight :

[0054]

[0055] wherein, denotes the neighbor set of a node ; denotes the attention coefficient of a target node and its neighbor nodes ;

[0056] Then, the multi-head attention mechanism is adopted to learn the node relationship in parallel, and each attention head independently calculates different inter-node weights to generate differentiated node representations from multiple perspectives:

[0057]

[0058] wherein, denotes the output feature representation of a node under the th attention head;

[0059] The node information in the entire graph structure is summarized by averaging the outputs of multiple attention heads; meanwhile, the residual connection (Res) is adopted to relieve the gradient vanishing phenomenon, and the final spatial feature vector is obtained;

[0060] 3-2) Time series analysis is performed by modeling the temporal dependency through a bidirectional gated recurrent unit (BiGRU), and the residual connection (Res) is adopted to optimize the training process. The discharge capacity estimation value is output by the fully connected layer, and the accurate estimation of the state of health (SOH) is realized, including:

[0061] A sequence learning model is constructed by a bidirectional gated recurrent unit (BiGRU), which dynamically regulates the interaction of historical and future information by coupling forward and backward gated recurrent units (GRUs). The forward process hidden layer calculation formula is:

[0062]

[0063] The backward process hidden layer calculation formula is:

[0064]

[0065] The output calculation formula of the bidirectional gated recurrent unit is:

[0066]

[0067] wherein, is the input at time , is the weight of the forward process hidden layer, is the weight of the backward process hidden layer, Bias for output;

[0068] The spatial feature vector processed by the bidirectional gate recurrent unit (BiGRU) is input into a full connection layer to convert the hidden state into a discharge capacity estimation value:

[0069]

[0070] wherein, and are the weight and bias of the full connection layer, respectively;

[0071] The discharge capacity estimation value is obtained, and the accurate estimation of the state of health of the light power battery is realized.

[0072] Compared with the prior art, the beneficial effects of the present application are:

[0073] In order to further improve the abnormal early warning capability in the use process of the electric bicycle, a light power battery state of health estimation method based on GAT-BiGRU-Res model is proposed, the present application obtains voltage data and capacity data based on dynamic selection of 0.1V optimal voltage window from IC peak, explains the selected segment data from the aging mechanism level that contains key aging information, which is helpful for detailed analysis of the battery degradation process; and through GAT dynamic construction of graph structure data, the spatial relationship of non-adjacent local segments is modeled, and BiGRU is used to model the time sequence dependence, and Res is used to optimize the training process, which significantly improves the accuracy of SOH estimation. Thus, the problems of difficulty in obtaining complete data of light power battery, uncertainty of user usage habits and interpretability of feature sequence selection are effectively solved, the reliability and safety of the electric bicycle are improved, and the occurrence of safety accidents such as thermal runaway is reduced. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 is the flow chart of the light power battery state of health estimation method based on GAT-BiGRU-Res model of the present application;

[0075] Figure 2 is the GAT-BiGRU-Res neural network model structure proposed by the present application;

[0076] Figures 3 to 5 are the experimental results of estimating the SOH of the light power battery based on the GAT-BiGRU-Res model of the present application, wherein, Figure 3 is the XJTU-Batch1 experimental result, Figure 4 is the XJTU-Batch2 experimental result, Figure 5 is the CALCE-CX2 experimental result. DETAILED DESCRIPTION

[0077] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention.

[0078] This invention implements a method for estimating the state of health of lightweight power batteries based on the GAT-BiGRU-Res model, such as... Figure 1 As shown, the main steps are as follows:

[0079] Step 1: Collect historical charging data of the lightweight power battery, including current, voltage and time. Calculate the incremental capacity based on the collected historical charging data of the lightweight power battery. Select the voltage and capacity data by dynamically selecting the optimal voltage window of 0.1V based on the peak value of the IC curve.

[0080] The formula for calculating the incremental capacity IC based on the collected historical charging data of lightweight power batteries is as follows:

[0081]

[0082]

[0083] in, It refers to the charging capacity, measured in Ah. This is the charging current, measured in amperes (A). It is the charging voltage, and the unit is V. It is the charging time, measured in seconds (s).

[0084] The IC curve is smoothed using a Gaussian filter, and the calculation formula is as follows:

[0085]

[0086] in, This represents the Gaussian filter kernel function. The mean of the Gaussian kernel determines the center position of the filter kernel. Set to 0; It is an input variable, representing the position within the filter window; It is the standard deviation of the Gaussian kernel, which controls the width of the filter kernel and the strength of the smoothing.

[0087] For the filtered IC curve, a voltage segment within a 0.1V voltage range and its capacity data are dynamically selected based on the IC curve peak value, and are defined as follows:

[0088]

[0089]

[0090]

[0091] wherein, is the voltage segment data selected based on the IC peak value at the nth cycle, , , is the voltage at the peak position of the IC curve at the nth cycle; is the capacity within the first voltage window at the nth cycle, is the capacity within the second voltage window at the nth cycle, is the capacity within the third voltage window at the nth cycle, is the capacity within the fourth voltage window at the nth cycle, is the capacity when the voltage is V1, is the capacity when the voltage is V2, is the capacity when the voltage is V3, is the capacity when the voltage is V4, is the number of feature points in the segment data, is the capacity data in the corresponding voltage segment;

[0092] To eliminate the influence of different data units, the original data of the selected voltage segment and capacity segment are respectively standardized, and the calculation formula is as follows:

[0093]

[0094] wherein, is the value of the original data after standardization, is the mean of the original data, is the standard deviation of the original data.

[0095] Step 2, according to the voltage data and capacity data of the segment data in step 1, the voltage sequence and the capacity sequence are aligned by using the interpolation method, and are respectively divided into four equal parts, and the voltage-capacity feature vector of each sub-segment after four equal parts is taken as a graph node, and the time sequence data is converted into graph structure data.

[0096] To ensure the consistency of different cycle characteristic sequences, the cubic spline interpolation method is used to keep the data aligned, and the voltage sequence and the capacity sequence are aligned by using the cubic spline interpolation method for the voltage data and the capacity data of the segment data in step 1. In the present application, the length of the characteristic sequence in each cycle is set to 80. In order to capture the local spatial features and time sequence features of the light power battery aging process, it is further subdivided in order to analyze the battery aging more finely. The voltage sequence and the capacity sequence are both divided into four equal sub-segments, which enhances the representation ability of local changes and helps to understand the battery aging process in depth.

[0097] Each sub-segment represents a node in the graph, and the node features are represented as follows:

[0098]

[0099] ​wherein, denotes a node feature vector, denotes vector concatenation, denotes a node;

[0100] a node feature matrix is represented as follows:

[0101]

[0102] wherein, denotes the number of nodes, denotes the feature dimension;

[0103] According to the joint similarity of the battery and the capacity, edges are dynamically constructed, and the nonlinear dependence relationship between subsegments is captured.

[0104] First, the cosine similarity is used to calculate the similarity matrix of the voltage and the similarity matrix of the capacity :

[0105]

[0106]

[0107] wherein, and denote the voltage matrix and the capacity matrix respectively, denotes the 2-norm;

[0108] Then, the joint similarity matrix is generated by normalization and weighted fusion:

[0109]

[0110] wherein, is a weight parameter;

[0111] Finally, the edges are extracted and the self-loop is excluded, and the sparse adjacency matrix is generated:

[0112]

[0113] In the present application, the graph structure data includes the node feature matrix and the sparse adjacency matrix shown in formula (9) and formula (13).

[0114] Step 3: Based on the graph structure data described in Step 2, use Graph Attention Network (GAT) to model the spatial relationships of non-adjacent local segments to capture the local spatial features of the aging process of the lightweight power battery. Use Bidirectional Gated Recurrent Unit (BiGRU) to perform time series analysis to capture the temporal features of the aging process of the lightweight power battery. Use Residual Connection (Res) to optimize the training process. Finally, output the discharge capacity estimate through the fully connected layer to achieve accurate estimation of the State of Health (SOH).

[0115] Graph Attention Array (GAT) is an advanced graph neural network architecture derived from graph convolutional networks. It dynamically calculates the weights (attention coefficients) between a target node and its neighbors through learning, rather than pre-setting connection weights. This process allows GAT to adaptively capture the correlation and importance between nodes and their neighbors, thus more accurately aggregating information from neighboring nodes and generating richer node representations. During battery aging, the degradation rates of different sub-segments exhibit differences and correlations. Therefore, using GAT to extract deep relationships between voltage and capacity sequences from graph-structured data enables accurate SOH estimation.

[0116] 3-1) Based on the graph structure data described in step 2, model the spatial relationships of non-adjacent local segments using a graph attention network (GAT), including:

[0117] Based on the graph structure data described in step 2, a linear transformation is performed on the nodes composed of the voltage sequence and capacity sequence using the learnable weight matrix in the graph attention layer of the graph attention network (GAT).

[0118]

[0119] in, and These are the voltage-capacity feature vectors before and after the node transformation;

[0120] To capture the dynamic relationships between nodes, an attention mechanism is used to compute the target node. Its neighboring nodes Attention coefficient :

[0121]

[0122] in, Represents a learnable parameter vector. Indicates transpose;

[0123] To eliminate the influence of node degree differences, this invention employs... The function normalizes the attention coefficients to obtain the final attention weights. :

[0124]

[0125] where, denotes the neighbor set of node ; denotes the attention coefficient of target node and its neighbor node ;

[0126] Then, in order to enhance the model's representation ability, a multi-head attention mechanism is adopted to learn node relationships in parallel. Each attention head independently calculates different inter-node weights, generating differentiated node representations from multiple perspectives:

[0127]

[0128] where, denotes the output feature representation of node under the th attention head;

[0129] Finally, by averaging the outputs of multiple attention heads, the node information in the entire graph structure is summarized. To prevent information loss caused by the GAT structure, Res is used to alleviate the gradient vanishing phenomenon, enhancing the model's ability to depict the nonlinear degradation trajectory of the battery, and obtaining the final spatial feature vector.

[0130] 3-2) Time series analysis is performed by modeling the temporal dependency through a bidirectional gated recurrent unit (BiGRU), and residual connection (Res) is used to optimize the training process. The full connection layer outputs the discharge capacity estimation value, achieving accurate estimation of the state of health (SOH).

[0131] In order to analyze the spatial feature vectors extracted by the GAT layer and deeply mine the time series trends of voltage and capacity sequences, a sequence learning model is constructed using a bidirectional gated recurrent unit (BiGRU). This model dynamically controls the interaction of historical and future information by coupling forward and backward gated recurrent (GRU) units. The forward process hidden layer calculation formula is:

[0132]

[0133] The backward process hidden layer calculation formula is:

[0134]

[0135] The output calculation formula of the bidirectional gated recurrent unit is:

[0136]

[0137] where, For a moment Input, The weights of the hidden layers in the forward pass. The weights of the hidden layer in the backward process. For the output bias;

[0138] The spatial feature vector processed by the Bidirectional Gated Recurrent Unit (BiGRU) is input into the fully connected layer to convert the hidden state into a discharge capacity estimate:

[0139]

[0140] in, and These are the weights and biases of the fully connected layer, respectively.

[0141] This yields an estimated discharge capacity, enabling an accurate estimation of the health status of lightweight power batteries.

[0142] The GAT-BiGRU-Res neural network model structure proposed in this invention is as follows: Figure 2 As shown, the model includes two GAT layers, one BiGRU layer, one residual module, one pooling layer, and three fully connected layers. The model input is graph-structured data constructed based on voltage and capacity sequences, and the output is an estimated discharge capacity. By executing the steps described above in this invention, accurate estimation of the state of discharge (SOH) of a lightweight power battery can be achieved.

[0143] Research Materials

[0144] This invention selected the XJTU and CALCE datasets for validation analysis. The XJTU dataset includes six sets of battery data, which underwent aging experiments under different charge-discharge strategies. The batteries were 18650 nickel-cobalt-manganese lithium batteries with a nominal capacity of 2 Ah, a nominal voltage of 3.6 V, and charging and discharging cutoff voltages of 4.2 V and 2.5 V, respectively. The entire experiment was conducted at room temperature. This invention selected the first two sets of batteries for analysis, namely Batch 1 and Batch 2, containing 8 and 15 batteries, respectively. The CALCE dataset was released by the Advanced Life Cycle Engineering Center at the University of Maryland. This invention selected four prism batteries from the CX2 series for analysis. The material was LiCoO2, and the batteries were charged at 0.5 C under constant current and constant voltage and discharged at 1 C under constant current at room temperature. The sampling frequency was 1 / 30 Hz.

[0145] The parameters mentioned in the materials of this study are shown in Table 1.

[0146] Table 1. Parameter settings for the GAT-BiGRU-Res model

[0147] Structure name Parameter setting GAT1 in_channels = 40, out_channels = 160, heads = 4 GAT2 in_channels = 640, out_channels = 320, heads = 1 Res in_features = 40, out_features = 320 BiGRU input_size = 320, hidden_size = 80 Dropout dropout = 0.2 Pool hidden_size = 160 FC1 in_features = 160, out_features = 64 FC2 in_features = 64, out_features = 32 FC3 in_features = 32, out_features = 1

[0148] Figure 3 are the experimental results of XJTU-Batch1 taking the method of the present application, Figure 4 are the experimental results of XJTU-Batch2 taking the method of the present application, Figure 5 are the experimental results of CALCE-CX2 taking the method of the present application. According to Figures 3 to 5 It can be concluded that the method proposed in the present application has good estimation effect on SOH in different working conditions and different types of batteries, the scatter points are densely distributed near the diagonal line, the estimated value is highly consistent with the actual value, the error is small and basically symmetrically distributed, indicating that the proposed method has high accuracy and reliability. The error probability density diagram further shows that the error distribution presents a unimodal shape, the peak value is significantly concentrated near zero error, indicating that the estimation result of the proposed method for SOH has the statistical characteristics of low deviation and high concentration.

[0149] Although the present application has been described above with reference to the accompanying drawings, the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative rather than limiting, and many modifications can be made by those of ordinary skill in the art without departing from the spirit of the present application, and these all belong to the protection of the present application.

Claims

1. A light power battery state of health estimation method based on a GAT-BiGRU-Res model, characterized in that, The method comprises the following steps: Step 1: Incremental capacity IC calculation is performed according to the collected historical charging data of the light power battery, wherein the historical charging data comprises the voltage, current and time of the light power battery in the constant current stage; the peak value of the IC curve is dynamically selected to obtain the voltage data and capacity data in a 0.1V optimal voltage window; Step 2: According to the voltage data and capacity data of the segment data in step 1, the voltage sequence and the capacity sequence are aligned by using an interpolation method, and are divided into four equal parts respectively, and the voltage-capacity feature vector of each sub-segment after the four divisions is taken as a graph node, and the time sequence data is converted into graph structure data; Step 3: According to the graph structure data in step 2, the spatial relationship of non-adjacent local segments is modeled by using a graph attention network (GAT) to capture the local spatial features of the aging process of the light power battery, the time sequence analysis is performed by using a bidirectional gated recurrent unit (BiGRU) to capture the time sequence features of the aging process of the light power battery, and a residual connection (Res) is adopted to optimize the training process, and finally the discharge capacity estimation value is output by a full connection layer to realize accurate estimation of the state of health (SOH).

2. The method of claim 1, wherein, The specific content of step 1 comprises: The formula for performing incremental capacity IC calculation according to the collected historical charging data of the light power battery is as follows: （1） ； （2） ； wherein, is the charge capacity in Ah; is the charge current in A; is the charge voltage in V, is the charge time in s; The IC curve is smoothed by using a Gaussian filter, and the calculation formula is as follows: （3） ； wherein, denotes a Gaussian filter kernel function, is the mean of the Gaussian kernel, which determines the center position of the filter kernel, is set to 0; is the input variable, which represents the position within the filter window; is the standard deviation of the Gaussian kernel, which controls the width of the filter kernel and the strength of the smoothing; For the filtered IC curve, the voltage segment in the 0.1V voltage range and the capacity data of the segment are dynamically selected based on the IC curve peak value, and are respectively defined as: （4） ； （5） ； （6） ； wherein, is the voltage segment data selected based on the IC peak value at the n-th cycle, , , is the voltage at the position of the IC curve peak value at the n-th cycle; is the capacity within the n-th voltage window at the n-th cycle, is the capacity at the voltage of is the capacity at the voltage of is the number of feature points in the segment data, is the capacity data in the corresponding voltage segment;​​​​​​ In order to eliminate the influence of different data units, the original data of the selected voltage segment and the capacity segment are standardized, and the calculation formula is as follows: （7） ； wherein, is the value of the normalized raw data, is the mean of the raw data, is the standard deviation of the raw data.

3. The method of claim 1, wherein, The specific content of step 2 comprises: The voltage sequence and the capacity sequence are aligned by using a cubic spline interpolation method for the voltage data and the capacity data of the segment data in step 1, and the length of the feature sequence in each cycle is set to 80; in order to capture the local spatial features and the time sequence features of the aging process of the light power battery, the voltage sequence and the capacity sequence are divided into four equal sub-segments, each sub-segment represents a node in the graph, and the node feature is represented as follows: （8） ； wherein, represents a node feature vector, represents vector concatenation, represents a node; node feature matrix is represented as follows: （9） ； wherein, denotes the number of nodes, denotes the feature dimension; According to the joint similarity of the battery and the capacity, the edges are dynamically constructed to capture the nonlinear dependence relationship between the sub-segments, and the process is as follows: First, the cosine similarity is used to calculate the similarity matrix of the voltage and the similarity matrix of the capacity : （10） ； （11） ； wherein, and V and C represent the voltage matrix and the capacity matrix, respectively, || · || represents the 2-norm; The normalized and weighted fusion then generates a joint similarity matrix : （12） ； wherein is a weight parameter; Finally, the extraction edges and exclude self-loops, generating a sparse adjacency matrix : （13） ； The graph structure data includes a node feature matrix and a sparse adjacency matrix .

4. The method of claim 1, wherein, The specific content of step 3 comprises: 3-1) According to the graph structure data in step 2, the spatial relationship of non-adjacent local segments is modeled by using a graph attention network (GAT), which comprises: According to the graph structure data in step 2, the nodes composed of the voltage sequence and the capacity sequence are linearly transformed by using a learnable weight matrix in the graph attention layer of the graph attention network (GAT), （14） ； wherein, and are the voltage-capacity eigenvectors before and after node transformation, respectively; computing attention coefficients for target nodes with their neighbor nodes :​ （15） ； wherein, denotes a learnable parameter vector, denotes a transpose; use The function normalizes the attention coefficients to obtain the final attention weights. : 16） ； wherein, represents a set of neighbors of a node ; represents an attention coefficient of a target node to its neighbor nodes ; Then, a multi-head attention mechanism is adopted to learn the node relationship in parallel, each attention head independently calculates different inter-node weights, and generates differentiated node representations from multiple perspectives: （17） ； in, Represents a node In the Output feature representation under each attention head; The node information in the entire graph structure is summarized by averaging the outputs of multiple attention heads; meanwhile, a residual connection (Res) is adopted to relieve the gradient vanishing phenomenon, and the final spatial feature vector is obtained; 3-2) Time series analysis is performed by modeling time dependence through a bidirectional gated recurrent unit (BiGRU), and a residual connection (Res) is used to optimize the training process. The full connection layer outputs the capacity estimation value, and the state of health (SOH) is accurately estimated, including: A sequence learning model is constructed by a bidirectional gated recurrent unit (BiGRU), which dynamically controls the interaction of historical and future information by coupling forward and backward gated recurrent units (GRU). The forward process hidden layer calculation formula is: （18） ； The backward process hidden layer calculation formula is: （19） ； The output calculation formula of the bidirectional gated recurrent unit is: （20） ； wherein, is the input at time , is the weight of the forward process hidden layer, is the weight of the backward process hidden layer, is the bias of the output; The spatial feature vector processed by the bidirectional gated recurrent unit (BiGRU) is input into the full connection layer to convert the hidden state into the capacity estimation value: （21） ； wherein, and are the weights and biases of the fully connected layer, respectively. Thus, the capacity estimation value is obtained, and the accurate estimation of the state of health of the light power battery is realized.

Images