Lithium battery soh evaluation method and system based on graph structured micro-voltage fragments

Abstract

The application provides a lithium battery SOH evaluation method and system based on a graph structured micro-voltage segment, and belongs to the technical field of lithium battery health state evaluation and artificial intelligence. The method comprises the following steps: preprocessing voltage data of lithium battery cyclic charging and discharging, screening and aligning micro-voltage segments which are highly sensitive to battery degradation; reconstructing the time and space features of the preprocessed battery micro-voltage segment data through a trained deep learning model; abstracting the feature information and its sequence relationship into a graph structure containing topological structure and node features; expanding the node receptive field through a stacked multi-layer graph neural network, and obtaining a graph-level feature vector through pooling; performing nonlinear mapping on the obtained graph-level feature vector, and outputting the SOH evaluation result of the lithium battery. The system comprises a data preprocessing unit, a deep learning evaluation model unit and a result output unit. The application does not require additional feature engineering, only inputs the original charging voltage data, and realizes high-precision evaluation through simple calculation.

Description

Technical Field

[0001] This invention relates to the fields of lithium battery state of health (SOH) assessment and artificial intelligence technology, and in particular to a lithium battery SOH assessment method and system based on graph-structured micro-voltage segments. Background Technology

[0002] Lithium-ion batteries, due to their superior performance, have been widely used in many important engineering fields such as electric vehicles, energy storage systems, and aerospace, and dominate the secondary battery market. However, during long-term operation, the state of health (SOH) of lithium-ion batteries gradually deteriorates with changes in time and operating conditions, specifically manifested in a continuous decrease in maximum discharge capacity and a gradual shortening of remaining lifespan. Typically, the state of health is defined as the ratio of the actual usable capacity of the current cycle to the battery's nominal capacity. Therefore, accurate assessment of SOH is crucial for effectively determining the battery's aging stage, developing precise battery management strategies, and extending the service life of the entire energy storage system.

[0003] Currently, methods for evaluating the state of harm (SOH) of lithium-ion batteries can be mainly divided into two types: model-based methods and data-driven methods. Model-based methods are highly dependent on the reliability of the model and require a large number of experiments. At the same time, the ideal battery model is easily affected by actual operating conditions. Traditional machine learning has limited ability to process time-series data, especially long-term dependencies, which limits the adaptability of the above methods in practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a lithium battery SOH evaluation method and system based on graph-structured micro-voltage segments. Through incremental capacity analysis, highly sensitive micro-voltage segments are adaptively selected and graph data models are constructed to mine the potential information between their graph-structured micro-voltage data, thereby achieving end-to-end SOH evaluation and solving the problems of large RMSE and MAE and excessive input data requirements of existing models.

[0005] To achieve the above objectives, this invention proposes a lithium battery SOH evaluation method based on graph-structured micro-voltage segments, comprising the following steps:

[0006] Step S1: Preprocess the voltage data of the lithium battery cycle charge and discharge, filter out the small voltage segments that are highly sensitive to battery degradation, and align the data to make the data length the same.

[0007] Step S2: Reconstruct the spatiotemporal features of the preprocessed small voltage segments of the battery using a trained deep learning model;

[0008] Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features;

[0009] Step S4: By stacking multiple layers of graph neural networks, the receptive field of nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling.

[0010] Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the SOH evaluation results of the lithium battery.

[0011] Preferably, step S1 includes the following steps:

[0012] Step S11: Analyze the incremental capacity curve of a complete charging cycle of a 100% SOH lithium battery to determine the time point corresponding to the voltage at which the most violent internal reaction occurs. The formula for the incremental capacity curve is:

[0013] ;

[0014] in IC For incremental capacity, Q For battery capacity, V Battery voltage, T It is the sampling time. Q T+1 Sampling time T The battery capacity corresponding to +1 Q T Sampling time T The corresponding battery capacity at that time V T+1 Sampling time T The battery voltage corresponding to +1 V T Sampling time T The corresponding battery voltage at that time;

[0015] Step S12: Using the peak time point as the starting point of the region, adaptively select a voltage data within a micro-window as the highly sensitive region. The calculation formula is as follows:

[0016] ;

[0017] in, V p This is the voltage corresponding to the peak point of the IC. X This refers to the voltage step size;

[0018] Step S13: Perform linear interpolation on the highly sensitive region, resample to generate time-series voltage data of consistent length, and the interpolation function formula is:

[0019] ;

[0020] ;

[0021] ;

[0022] in, F ( T () represents any point in time. T Voltage at point, T t , T t+1 For two adjacent time points in the original data, V t , V t+1 For the corresponding time point in the original data t , t Voltage value at +1 T s This is a time-point sequence generated by resampling. ΔT The sampling time interval, K For index number, m The maximum index of the resampled sequence. T E The sampling termination time, T 0 This is the sampling start time.

[0023] Preferably, step S2 includes the following steps:

[0024] Step S21: Perform a one-dimensional convolution operation on the time series voltage data of the same length to obtain the local spatial dependency;

[0025] Step S22: Establish long-term temporal dependencies through the gating mechanism of the Long Short-Term Memory network, fuse the local spatial features and temporal dependency information obtained by convolution, and reconstruct them into spatiotemporal features.

[0026] Preferably, step S3 includes the following steps:

[0027] Step S31: Divide the time series voltage data into nodes, and use the reconstructed spatiotemporal features as the feature vectors of the corresponding nodes;

[0028] Step S32: Construct a directed graph with unidirectional edges connected in sparse matrix form and labeled with the SOH value of the corresponding cycle, based on the edge connection method of the time series.

[0029] Step S33: Using the actual SOH value of the corresponding cycle as the graph label, construct a data model that includes topological structure and node features.

[0030] Preferably, step S4 includes the following steps:

[0031] Step S41: By stacking multiple layers of graph neural networks, the receptive field of the nodes is gradually expanded. The network layer operation formula is as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] in, For the first The node feature matrix of the layer, For an adjacency matrix with self-loops, I It is the identity matrix. for The degree matrix, For the first The node feature matrix of the layer, For the first The layer's learnable weight matrix, where σ is the activation function. For nodes and Connection weights between them The elements of the degree matrix, The row index represents the target node. This is a column index, representing the source node;

[0036] Step S42: Aggregate neighborhood node information through graph convolutional neural networks, mine potential dependency information between adjacent nodes, and generate globally aware node embeddings;

[0037] Step S43: Convert node features into graph-level feature vectors through global average pooling.

[0038] Preferably, step S5 includes the following steps:

[0039] Step S51: Gradually extract aging information through multi-level nonlinear transformations and scale the dimensions of the graph-level feature vectors.

[0040] Step S52: Use the fully connected layer to output the SOH value of the tested battery.

[0041] Preferably, the training parameters of the above model are determined by minimizing the mean square error, as shown in the formula:

[0042] ;

[0043] in, β Let argmin be the set of model weight parameters, and argmin(·) be the parameters that minimize the objective function. n The number of training samples.p For sample index, y p For the first p The actual SOH value of the battery in the next cycle. For the first p Battery SOH estimate for the next cycle.

[0044] The present invention also provides a lithium battery SOH evaluation system based on graph-structured micro-voltage segments, comprising: a data preprocessing unit, a deep learning evaluation model unit, and a result output unit;

[0045] The data preprocessing unit is used to preprocess the lithium battery cycle charge and discharge voltage data, filter out highly sensitive small voltage segments, and align the data lengths.

[0046] The deep learning evaluation model unit includes a feature reconstruction module, a graph data model module, and a graph embedding and regression module, which are used to reconstruct the feature information of the preprocessed highly sensitive micro voltage fragment data and construct a graph data model to obtain graph-level feature vectors.

[0047] The result output unit is used to perform nonlinear mapping and output the SOH evaluation results of lithium batteries.

[0048] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0049] Step S1: Preprocess the voltage data of the lithium battery cycle charge and discharge, filter out the small voltage segments that are highly sensitive to battery degradation, and align the data to make the data length the same.

[0050] Step S2: Reconstruct the spatiotemporal features of the preprocessed small voltage segments of the battery using a trained deep learning model;

[0051] Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features;

[0052] Step S4: By stacking multiple layers of graph neural networks, the receptive field of nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling.

[0053] Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the SOH evaluation results of the lithium battery.

[0054] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, performs the following steps:

[0055] Step S1: Preprocess the voltage data of the lithium battery cycle charge and discharge, filter out the small voltage segments that are highly sensitive to battery degradation, and align the data to make the data length the same.

[0056] Step S2: Reconstruct the spatiotemporal features of the preprocessed small voltage segments of the battery using a trained deep learning model;

[0057] Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features;

[0058] Step S4: By stacking multiple layers of graph neural networks, the receptive field of nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling.

[0059] Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the SOH evaluation results of the lithium battery.

[0060] Therefore, this invention proposes a method and system for evaluating the state of harmonics (SOH) of lithium batteries based on graph-structured micro-voltage segments, with the following advantages:

[0061] (1) The present invention can quickly determine the voltage-sensitive region by using the complete charging voltage data of the battery with 100% SOH, adaptively select the voltage data within the small window, and realize the battery SOH evaluation by using only the small voltage segment data, which greatly reduces the amount of data required.

[0062] (2) This invention uses a combination model of one-dimensional convolutional neural network and long short-term memory network to fully extract the local spatial features and long-term temporal dependencies of voltage data, innovatively abstracts the time series data into a graph structure, transforms the sequence relationship into the topological connection between graph nodes, and uses graph neural network to mine the potential correlation information in the data, which significantly improves the feature expression ability.

[0063] (3) This invention does not require manual design of feature engineering. It directly uses the original charging voltage data as input, and forms an end-to-end process through preprocessing, feature extraction, graph structure modeling to SOH output, which simplifies the calculation process and improves the versatility of the method.

[0064] (4) This invention achieves deep fusion of local features and global information through the feature aggregation capability of multi-layer GCN, effectively explores the potential laws of battery aging, and significantly reduces evaluation error. Attached Figure Description

[0065] Figure 1 This is a flowchart of a lithium battery SOH evaluation method based on graph-structured micro-voltage segments;

[0066] Figure 2 This is a schematic diagram illustrating the principle of Long Short-Term Memory (LSTM) networks.

[0067] Figure 3 A block diagram of the internal structure of a computer device;

[0068] Figure 4 The diagram shows the SOH evaluation results for Cell 7 and Cell 8 in the Oxford Battery Degradation Data Set, where: (a) is the SOH evaluation result for Cell 7, and (b) is the SOH evaluation result for Cell 8.

[0069] Figure 5 The diagram shows the SOH evaluation results of 2C-7 and 2C-8 in the XJTU battery degradation dataset, where: (a) is the SOH evaluation result diagram of 2C-7, and (b) is the SOH evaluation result diagram of 2C-8;

[0070] Figure 6 Segment voltage diagrams for the Oxford battery degradation dataset and the XJTU battery degradation dataset, where: (a) is a segment voltage diagram for the Oxford battery degradation dataset, and (b) is a segment voltage diagram for the XJTU battery degradation dataset;

[0071] Figure 7 A schematic diagram of SOH evaluation for different segments of the Oxford dataset;

[0072] Figure 8 This is a schematic diagram of the SOH evaluation of different segment voltages in the XJTU dataset. Detailed Implementation

[0073] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0074] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0075] Example 1

[0076] like Figure 1 As shown, this invention provides a method for evaluating the state of harm (SOH) of lithium batteries based on graph-structured micro-voltage segments, comprising the following steps:

[0077] Step S1: Preprocess the voltage data from the lithium battery's cyclic charge and discharge to identify small voltage segments highly sensitive to battery degradation, and align the data to ensure they are of equal length. This includes the following steps:

[0078] Step S11: Record the charge and discharge voltage data of the lithium battery in different cycles using a voltage sensor. Select the voltage data of historical charging cycles as input. Perform incremental capacity analysis on the complete charging voltage data of the battery at 100% SOH. Adaptively select a micro-window of voltage data based on the voltage time point corresponding to the peak of the incremental capacity curve, which is a highly sensitive micro-voltage segment. Among them, incremental capacity analysis is performed based on the charging voltage data of the lithium battery in the constant current stage. The calculation formula is as follows:

[0079] ;

[0080] in IC For incremental capacity, Q For battery capacity, V Battery voltage, T It is the sampling time. Q T+1 Sampling time T The battery capacity corresponding to +1 Q T Sampling time T The corresponding battery capacity at that time V T+1 Sampling time T The battery voltage corresponding to +1 V T Sampling time T The corresponding battery voltage at that time;

[0081] Step S12: Using the peak time point as the starting point of the region, adaptively select a voltage data within a micro-window as the highly sensitive region. The calculation formula is as follows:

[0082] ;

[0083] in, V p This is the voltage corresponding to the peak point of the IC. X This refers to the voltage step size;

[0084] Step S13: To reduce noise and obtain the desired characteristics from the IC curve, an incremental capacity calculation method based on linear interpolation resampling is used to perform linear interpolation on the highly sensitive region, resampling to generate time-series voltage data of consistent length. The interpolation function formula is:

[0085] ;

[0086] ;

[0087] ;

[0088] in, F (T () represents any point in time. T Voltage at point, T t , T t+1 For two adjacent time points in the original data, V t , V t+1 For the corresponding time point in the original data t , t Voltage value at +1 T s This is a time-point sequence generated by resampling. ΔT The sampling time interval, K For index number, m The maximum index of the resampled sequence. T E The sampling termination time, T 0 This is the sampling start time;

[0089] Subsequently, Z-score normalization (standard deviation normalization) was used to normalize the highly sensitive small voltage segments that had been resampled by linear interpolation in order to improve the model's convergence speed and generalization performance.

[0090] Step S2: Reconstruct the spatiotemporal feature information of the preprocessed small voltage segments of the battery using a trained deep learning model, including the following steps:

[0091] Step S21: A one-dimensional convolutional neural network (CNN) is used to perform one-dimensional convolution operations on time-series voltage data of uniform length to obtain local spatial dependencies. A one-dimensional CNN typically consists of an input layer, convolutional layers, and activation functions. This network, with its local perception characteristics and weight-sharing mechanism, exhibits significant advantages in processing one-dimensional time-series data such as battery voltage and current. The network input is one-dimensional voltage time-series data, and sliding window convolution operations are performed along the sequence direction using one-dimensional convolution kernels. The specific process is as follows:

[0092] ;

[0093] in, x For one-dimensional time series data, w For convolution kernel weights, For position t Convolution operation, k The kernel size is [size]. x ( t+i ) represents the position of the input sequence. t+i The value at that location, w ( i ) is the location iWeight at each location;

[0094] The convolution operation computes and generates new feature maps bit by bit, and outputs them. Each feature unit corresponds to the feature representation of a local receptive field in the input voltage sequence, as shown in the formula:

[0095] ;

[0096] in, Y The convolutional layer outputs a feature map. For activation function, x For input data, b c For the bias term of the convolutional layer, This is a one-dimensional convolution operation;

[0097] Step S22: Establish long-term temporal dependencies through the gating mechanism of the Long Short-Term Memory network, such as... Figure 2 As shown, Long Short-Term Memory (LSTM) networks mainly consist of forget gates, input gates, and output gates. They are network models specifically designed for processing time-series data. Their gating mechanism enables the network to effectively learn long-term dependencies, overcoming the gradient vanishing and exploding problems in traditional recurrent neural networks. Therefore, they are particularly suitable for capturing cross-cycle aging information during battery cycling. Furthermore, this network is often combined with convolutional neural networks to collaboratively extract local spatial features and global temporal dependencies, achieving spatiotemporal feature extraction of the battery aging process. The process involves fusing the local spatial features and temporal dependency information obtained from convolution to reconstruct spatiotemporal features, including the following steps:

[0098] Step S221: The forget gate determines which information the LSTM discards using the sigmoid function. The specific operation is as follows:

[0099] ;

[0100] in, The output of the forget gate is σ, which is the sigmoid function. For the Gate of Oblivion H t-1 for t The hidden state at time -1 (the previous time step). S t The current network input value. For the offset of the forget gate;

[0101] Step S222: The input gate is jointly operated by the sigmoid and tanh functions. First, the sigmoid function determines which information to update, and then the tanh function is used to obtain the candidate cell states. Through these two steps, new information is selectively recorded into the cell states, as shown in the following formula:

[0102] ;

[0103] ;

[0104] in, i t The output of the input gate, W i For input gate, This is the bias of the input gate. Let tanh represent the candidate cell state, and tanh be the tanh function. W c Candidate cell state, b c Bias for candidate cell states;

[0105] Step S223: Update the cell state. The update is completed by determining the information ratio of the cell state at the previous time step and the current time step through the forget gate and the input gate. The formula is:

[0106] ;

[0107] in, Represents the current state of the cell. Represents the cell state at the previous moment;

[0108] Step S224: The output gate determines which cell states to output. The decision criteria are obtained through a sigmoid layer. These criteria are then multiplied by the vector obtained from the layer to obtain the final output. The formula is as follows:

[0109] ;

[0110] ;

[0111] in, Represents the output of the output gate. This is the weight matrix of the output gate. This represents the offset of the output gate. represent t The output value of the LSTM at time (current time);

[0112] Step S3: Abstract the feature information and its sequence relationships into a graph structure, creating a data model that includes topological structure and node features, including the following steps:

[0113] Step S31: Divide the time series voltage data into... N Each node uses its reconstructed spatiotemporal features as its feature vector, forming a node feature matrix. X , ,in, R For the real number field,N Represents the number of nodes. F Indicates feature dimension;

[0114] Step S32: Construct a graph based on the edge connection method of the time series. Construct a directed graph with sparse matrix representation and corresponding SOH values ​​labeled by unidirectional edges, specifically by establishing a set of edges. E The adjacency matrix of a graph is represented in sparse matrix form. A , , corresponding to a directed graph A ij Indicates the connection node x i arrive x j The edge, A ij =1 indicates a node x i arrive x j There is a border between them. A ij =0 indicates a node x i arrive x j There are no edges between them; besides the adjacency matrix representation, the graph can also be represented by the Laplacian matrix, as shown in the formula:

[0115] ;

[0116] in, L It is a Laplacian matrix. D For degree matrix, A It is an adjacency matrix;

[0117] Step S33: Using the actual SOH value of the corresponding cycle as the graph label, construct a data model that includes topological structure and node features.

[0118] Step S4: By stacking multiple layers of graph neural networks, the receptive field of nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling. This includes the following steps:

[0119] Step S41: Graph convolutional neural networks typically consist of four key components: message passing, feature aggregation, node update, and multi-layer stacking. In this network, each node generates new node features by aggregating the features of its neighbors and fusing them with its own features. By stacking multiple layers, the receptive field of each node gradually expands, enabling it to learn a wider range of graph information. Based on spectral graph theory, graph convolutional neural networks define graph convolution operations in the frequency domain and achieve efficient computation in the spatial domain through first-order Chebyshev polynomial approximation. The core computation process is shown in the following equation:

[0120] ;

[0121] ;

[0122] ;

[0123] in, For the first The node feature matrix of the layer, For an adjacency matrix with self-loops, I It is the identity matrix. for The degree matrix, For the first The node feature matrix of the layer, For the first The layer's learnable weight matrix, where σ is the activation function. For nodes and Connection weights between them The elements of the degree matrix, The row index represents the target node. This is a column index, representing the source node;

[0124] Step S42: Aggregate neighborhood node information through graph convolutional neural networks, mine potential dependency information between adjacent nodes, and generate globally aware node embeddings;

[0125] Step S43: Convert node features into graph-level feature vectors through global average pooling.

[0126] Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the lithium battery SOH evaluation results, including the following steps:

[0127] First, a large amount of source domain data is used to train the deep learning evaluation model, specifically based on ROI micro-voltage segment data obtained by ICA, with the SOH value serving as the training label. The optimal training model parameters are obtained by minimizing the mean square error, as shown in the formula:

[0128] ;

[0129] in, This represents the actual SOH value of the battery in this cycle. The estimated SOH value of the battery for this cycle is obtained. Then, fix the shared parameters. The model is validated using a validation set, and its effectiveness is determined by the root mean square error (RMSE) and mean absolute error (MAE). After obtaining the optimal model parameters, they are fixed as shared parameters and tested on a test set.

[0130] Step S51: Gradually extract aging information through multi-level nonlinear transformations and scale the dimensions of the graph-level feature vectors.

[0131] Step S52: Utilize the nonlinear transformation of the fully connected layer to output the SOH value of the tested battery.

[0132] When building a deep learning evaluation model, the network model can be optimized based on experience and experimental results to determine the final model structure. The learning rate, number of iterations, and appropriate activation functions, loss functions, optimizers, etc. can be set according to actual needs.

[0133] Example 2

[0134] This invention also provides a lithium battery SOH evaluation system based on graph-structured micro-voltage segments. Through a three-level architecture of data preprocessing, feature extraction and graph modeling, and result output, it achieves high-precision end-to-end SOH evaluation. Each unit and its subordinate modules have clear division of labor and are progressively advanced, which not only ensures the validity of the data and the in-depth mining of features, but also ensures the accuracy of the evaluation results. The system includes: a data preprocessing unit, a deep learning evaluation model unit, and a result output unit.

[0135] The data preprocessing unit is used to preprocess the lithium battery cycle charge and discharge voltage data, filter out highly sensitive small voltage segments, and align the data lengths.

[0136] The deep learning evaluation model unit includes a feature reconstruction module, a graph data model module, and a graph embedding and regression module, which are used to reconstruct the feature information of the preprocessed highly sensitive micro voltage fragment data and construct a graph data model to obtain graph-level feature vectors.

[0137] The result output unit is used to perform nonlinear mapping and output the SOH evaluation results of lithium batteries.

[0138] The data preprocessing unit filters out the most sensitive micro-voltage segments to battery degradation from the raw lithium battery cyclic charge-discharge voltage data. Simultaneously, it eliminates data discrepancies through standardization, providing high-quality, consistent input data for subsequent feature extraction and model training. This addresses issues of raw data redundancy, insufficient sensitivity, and inconsistent formats. Based on the complete charge-cycle incremental capacity curve (ICA) of a 100% SOH lithium battery, it locates the IC peak point where the internal electrochemical reaction is most intense. Using this as an anchor point, it adaptively selects voltage data within a micro-window, accurately identifying highly sensitive regions directly related to battery aging and eliminating meaningless redundant data to reduce subsequent computation. Through linear interpolation resampling technology, it unifies highly sensitive micro-voltage segments of different cycles and lengths into a fixed-length time series, avoiding feature extraction bias caused by inconsistent data lengths. Z-score normalization and other methods eliminate the influence of dimensionality and numerical fluctuations in voltage data, improving the convergence speed and generalization ability of subsequent deep learning models and ensuring the model's adaptability under different operating conditions and battery types.

[0139] The deep learning evaluation model unit, as the core processing unit of the system, deeply mines the battery aging characteristics contained in highly sensitive micro-voltage segments through multi-module collaboration, transforms time-series data into a structured graph model, and realizes progressive extraction from local features to time-series features and then to global features, providing graph-level feature vectors with strong representation capabilities for the final SOH evaluation. It includes a feature reconstruction module, a graph data model module, and a graph embedding and regression module. The feature reconstruction module extracts local spatial features and long-term temporal features from highly sensitive micro-voltage segments and fuses them into spatiotemporal feature information. This addresses the problem that single feature extraction methods cannot simultaneously consider local correlations and long-term dependencies. Local spatial features are extracted through sliding window convolution operations of a one-dimensional convolutional neural network (1D-CNN), such as capturing local dependencies between adjacent data points in the voltage sequence (e.g., voltage abrupt changes, slope variations, etc.). These local features are early signals of battery aging. The module also effectively memorizes the long-term trend of voltage changes during battery cycling through the gating mechanism (forget gate, input gate, output gate) of a long short-term memory network (LSTM), overcoming the gradient vanishing problem of traditional recurrent neural networks (RNNs), capturing the cumulative aging patterns across cycles, and establishing long-term temporal dependencies. Finally, the local spatial features extracted by 1D-CNN are fused with the long-term temporal features captured by LSTM to form a spatiotemporal feature vector that combines details and trends, comprehensively depicting the dynamic process of battery aging and achieving spatiotemporal feature fusion. The graph data model module abstracts the spatiotemporal feature vectors of one-dimensional time series into a graph structure data model. It transforms the feature attributes and sequence relationships of the data into the node features and topological connections of the graph, mining the hidden structured correlation information in the time series data and improving the expressive power of the features. The preprocessed fixed-length high-sensitivity voltage time series data is evenly divided into N nodes, each node corresponding to a spatiotemporal feature vector of a local voltage data segment. According to the time series order of the voltage data, a directed graph with unidirectional edge connections is constructed, that is, the nodes of the previous time step are connected to the nodes of the next time step. The adjacency relationship is represented by a sparse matrix, which intuitively reflects the temporal logic of the voltage data. The actual SOH value of the battery corresponding to each cycle is used as the label of the graph to establish a mapping relationship between the graph structure data and SOH, providing supervision signals for subsequent graph neural network training.The graph embedding and regression module performs deep feature aggregation on graph-structured data to generate globally perceptive graph-level feature vectors. This addresses the issue that single-node features cannot reflect the overall aging state. By stacking multiple layers of graph convolutional neural networks (GCNs), the receptive field (i.e., the neighborhood range that a node can perceive) of each node is gradually expanded. This ensures that each node not only contains its own features but also aggregates feature information from neighboring nodes, uncovering potential dependencies between nodes (such as the correlation of aging features at different time periods). Through iterative aggregation of multiple GCNs, node embedding vectors with a global perspective are generated. Each node vector contains aging information of the entire voltage sequence, achieving feature enhancement from local to global. Through operations such as global average pooling, the embedding vectors of all nodes are aggregated into a fixed-dimensional graph-level feature vector. This vector integrates the structured aging information of the entire highly sensitive voltage segment, providing the core input for the final SOH assessment.

[0140] The output unit transforms the graph-level feature vectors output by the graph embedding and regression modules into specific lithium battery SOH (State of Health) evaluation values ​​through nonlinear mapping. This achieves the final transformation from feature vectors to evaluation results, ensuring the accuracy and consistency of the output results with physical meaning. Multiple fully connected layers scale the dimensions of the graph-level feature vectors and perform nonlinear mapping to gradually extract aging information strongly correlated with SOH, overcoming the problem that linear models cannot fit the complex feature-SOH mapping relationship. The final fully connected layer outputs a single numerical SOH evaluation value (range 0~1 or 0%~100%), directly reflecting the ratio of the battery's current health state to its nominal state. During the training phase, this unit calculates the mean squared error (MSE) between the evaluation value and the true SOH value, providing a loss signal for parameter updates in the entire deep learning model, driving continuous model optimization and improving evaluation accuracy.

[0141] Example 3

[0142] like Figure 3 As shown, the present invention provides a computer device storing a computer program. When the computer program is executed by a processor, the processor performs steps S1 to S5 of a lithium battery SOH evaluation method based on graph-structured micro-voltage segments:

[0143] Step S1: Preprocess the acquired lithium battery cycle charge and discharge voltage data to identify small voltage segments that are highly sensitive to battery degradation and align the data to have the same length.

[0144] Step S2: Reconstruct the spatiotemporal feature information of the preprocessed battery micro-voltage segments using the trained deep learning evaluation model;

[0145] Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features;

[0146] Step S4: By stacking multiple layers of graph neural networks, the receptive field of the nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling.

[0147] Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the SOH evaluation results of the lithium battery.

[0148] The internal structure of a computer device mainly includes a processor, memory, network interface, input devices, and a display screen connected via a system bus. The memory is divided into non-volatile storage media and internal memory. Non-volatile storage media stores the operating system and computer programs. When these programs are executed by the processor, a lithium battery SOH evaluation method based on graph-structured micro-voltage segments can be implemented. Internal memory can also store corresponding programs for the processor to perform the same tasks. The display screen can be an LCD screen or an e-ink screen, and input devices include integrated touch layers, physical buttons, trackballs, touchpads, or external input devices.

[0149] It should be noted that the execution of steps is not strictly limited by order. The order of some steps can be adjusted, and some steps may contain multiple sub-steps or stages that can be executed alternately. In addition, the above method can be implemented by a computer program, and the relevant program is stored in a non-volatile computer-readable storage medium.

[0150] Example 4

[0151] A computer-readable storage medium storing a computer program, the storage medium including non-volatile memory (such as ROM, PROM, EPROM, EEPROM, flash memory, etc.) and volatile memory (such as various types of RAM), storing steps S1 to S5 of a lithium battery SOH evaluation method based on graph-structured micro-voltage segments:

[0152] Step S1: Preprocess the acquired lithium battery cycle charge and discharge voltage data to identify small voltage segments that are highly sensitive to battery degradation and align the data to have the same length.

[0153] Step S2: Reconstruct the spatiotemporal feature information of the preprocessed battery micro-voltage segments using the trained deep learning evaluation model;

[0154] Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features;

[0155] Step S4: By stacking multiple layers of graph neural networks, the receptive field of the nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling.

[0156] Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the SOH evaluation results of the lithium battery.

[0157] The above-mentioned technical features can be combined in any way. For the sake of brevity, not all possible combinations of the above-mentioned technical features 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.

[0158] The invention will be further illustrated below through specific implementation examples.

[0159] Specific Implementation Case 1

[0160] Commercial lithium-ion battery data was used to evaluate the graph-structured micro-voltage segment-based lithium-ion battery SOH assessment method in this example. This dataset includes aging data from eight Kokam pouch cells. All battery experiments were conducted in a 40°C constant temperature chamber, with the cells cycle-charged to 4.2V at a constant 1C current, followed by dynamic discharge under ARTEMIS urban driving conditions. The specific steps of this example are as follows:

[0161] Step S1: Preprocess the voltage data of the lithium battery cycle charge and discharge, filter out the small voltage segments that are highly sensitive to battery degradation, and align the data to make the data length the same.

[0162] Step S11: Perform incremental capacity analysis on the complete charging voltage data of Cell 1 at 100% SOH, select the time point corresponding to the peak voltage of the incremental capacity curve as the starting point, and adaptively select the voltage data within a micro-window of duration T, where T is 180s;

[0163] Step S12: Using the peak time point as the starting point of the region, adaptively select a voltage data within a micro-window as the highly sensitive region. The calculation formula is as follows:

[0164] ;

[0165] in, V p This is the voltage corresponding to the peak point of the IC. X This refers to the voltage step size;

[0166] Step S13: Use Z-score normalization to resample the linearly interpolated sample with a step size of [missing value]. N The highly sensitive small voltage segments are normalized to improve the model's convergence speed and generalization performance. The calculation formula is as follows:

[0167] ;

[0168] in, μ and Let these represent the mean and standard deviation of the charging voltage data in the training set, respectively. x i Represents the original data. x This is the data after normalization.

[0169] Step S2: Reconstruct the spatiotemporal features of the preprocessed small voltage segments of the battery using a trained deep learning model;

[0170] Step S21: Perform a one-dimensional convolution operation on the time series voltage data of the same length to obtain the local spatial dependency;

[0171] Step S22: Establish long-term temporal dependencies through the gating mechanism of the Long Short-Term Memory network, fuse the local spatial features and temporal dependency information obtained by convolution, and reconstruct them into spatiotemporal features;

[0172] In this step, during training, mean squared error (MSE) is used as the loss function to measure the training error. Then, the ADMA optimizer is used to iteratively update the network parameters to minimize the mean squared error loss function, and the learning rate is set to 0.001. Finally, the training data is input into the deep learning evaluation model for training until the training period ends or it is verified that the loss has not improved within the set period.

[0173] Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features;

[0174] Step S31: Divide the time series voltage data into nodes, and use the reconstructed spatiotemporal features as the feature vectors of the corresponding nodes;

[0175] Step S32: Construct a directed graph with unidirectional edges connected in sparse matrix form and labeled with the SOH value of the corresponding cycle, based on the edge connection method of the time series.

[0176] Step S33: Using the actual SOH values ​​of the corresponding cycles as graph labels, construct a data model that includes topological structure and node features;

[0177] In this step, the root mean square error (RMSE) and the mean absolute error (MAE) are used to measure the model's performance. The formula is as follows:

[0178] ;

[0179] ;

[0180] in, n The number of loops. y i This represents the actual SOH value of the battery in this cycle. This is the estimated SOH value for the battery in this cycle.

[0181] The metrics for the two test sets mentioned above are shown in Table 1:

[0182] Table 1. Results of this method on the test set.

[0183]

[0184] Furthermore, to visually represent the error between the evaluated value and the true value, the SOH evaluation results of the Oxford dataset are as follows: Figure 4 As shown in the figure. (a) is the SOH evaluation result graph of Cell 7, and (b) is the SOH evaluation result graph of Cell 8. The horizontal axis of the evaluation graph represents the number of cycles divided by 100 during the test, and the vertical axis represents the SOH value under the corresponding cycle. The dark line represents the actual SOH value, and the light line represents the SOH evaluation value of the evaluation model. Specific Implementation Example 2

[0186] The XJTU batch 1 battery dataset, containing 8 batteries numbered 2C-1 to 2C-8 with a rated capacity of 2Ah, was used. Charge and discharge tests were conducted at room temperature using a CTS-5V10AGGS-D, charging to 4.2V at a constant current of 2.0C and then discharging to 2.5V at 1C.

[0187] The processing method for this dataset is the same as in Specific Implementation 1. The material comparison of the two battery datasets is shown in Table 2:

[0188] Table 2 Comparison of the two lithium battery datasets

[0189]

[0190] The process for processing this dataset is as follows: Figure 1 As shown, the operation is exactly the same as in Specific Example 1, and the model parameter settings are shown in Table 3:

[0191] Table 3 Model Parameter Settings

[0192]

[0193] The dataset is divided as shown in Table 4:

[0194] Table 4 Dataset Partitioning

[0195]

[0196] The above-described lithium battery SOH evaluation method based on graph-structured micro-voltage segments has the following performance indicators on two test sets:

[0197] Table 5 Results of this method on the test set

[0198]

[0199] Meanwhile, to visually represent the error between the evaluated value and the true value, the SOH evaluation results of the XJTU dataset are as follows: Figure 5 As shown in the figure, (a) is the SOH evaluation result graph for 2C-7, and (b) is the SOH evaluation result graph for 2C-8. The horizontal axis of the evaluation graph represents the number of cycles during the test, and the vertical axis represents the SOH value under the corresponding cycle. The dark line represents the actual SOH value, and the light line represents the SOH evaluation value of the evaluation model.

[0200] From Table 1, Table 5 and Figure 4 , Figure 5 As can be seen, the method in this embodiment can effectively evaluate the SOH value of lithium batteries, thereby arranging appropriate maintenance strategies and operating times based on the evaluated SOH value. This provides users with a suitable decision-making basis for taking preventive measures before lithium battery failure occurs, thus avoiding catastrophic consequences from failure.

[0201] In addition, to highlight the superiority of the proposed model, different voltage ranges A1 to A6 were selected for the two datasets, as shown in Table 6. A comparison of data lengths is also shown in the figure. Figure 6 As shown.

[0202] Table 6 Value range for different voltage intervals

[0203]

[0204] The evaluation was conducted under the same experimental conditions, and the evaluation results are illustrated in the diagram below. Figure 7 and Figure 8 As shown, detailed results and comparisons are shown in Table 7:

[0205] Table 7. Evaluation results of SOH in different voltage ranges

[0206]

[0207] pass Figure 6 As can be clearly seen from Table 7, the method proposed in this invention significantly reduces the amount of data required, while the evaluation results are greatly improved.

[0208] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0209] Therefore, the lithium battery SOH assessment method and system based on graph-structured micro-voltage segments proposed in this invention can achieve end-to-end prediction without prior knowledge. It can directly utilize pre-processed raw charging voltage data without additional feature engineering. Thus, lithium battery SOH assessment can be completed without extensive knowledge of lithium battery aging mechanisms. This end-to-end assessment can directly select highly sensitive micro-voltage window data through incremental capacity analysis, retaining aging-sensitive features with as little voltage data as possible while suppressing insensitive features. Subsequently, the assessment model extracts the spatiotemporal features of aging-sensitive features, deeply exploring their potential degradation information, and achieving accurate SOH assessment for lithium batteries with different compositions.

[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for evaluating the state of harm (SOH) of lithium batteries based on graph-structured micro-voltage segments, characterized in that, Includes the following steps: Step S1: Preprocess the voltage data from the lithium battery's cyclic charge and discharge to identify small voltage segments highly sensitive to battery degradation, and align the data to ensure they are of equal length. This includes the following steps: Step S11: Analyze the incremental capacity curve of a 100% SOH lithium battery during a complete charging cycle to determine the time point corresponding to the IC peak voltage of the most violent reaction inside the battery. The formula for the incremental capacity curve is: ； in IC For incremental capacity, Q For battery capacity, V Battery voltage, T It is the sampling time. Q T+1 Sampling time T The battery capacity corresponding to +1 Q T Sampling time T The corresponding battery capacity at that time V T+1 Sampling time T The battery voltage corresponding to +1 V T Sampling time T The corresponding battery voltage at that time; Step S12: Using the peak time point as the starting point of the region, adaptively select a voltage data within a micro-window as the highly sensitive region. The calculation formula is as follows: ； in, V p This is the voltage corresponding to the peak point of the IC. X This refers to the voltage step size; Step S13: Perform linear interpolation on the highly sensitive region, resample to generate time-series voltage data of consistent length, and the interpolation function formula is: ； ； ； in, F ( T () represents any point in time. T Voltage at point, T t , T t+1 For two adjacent time points in the original data, V t , V t+1 For the corresponding time point in the original data t , t Voltage value at +1 T s This is a time-point sequence generated by resampling. ΔT The sampling time interval, K For index number, m The maximum index of the resampled sequence. T E The sampling termination time, T 0 This is the sampling start time; Step S2: Reconstruct the spatiotemporal features of the preprocessed small voltage segments of the battery using a trained deep learning model; Step S3: Abstract the feature information and its sequence relationship into a graph structure to create a data model that includes topological structure and node features; Step S4: By stacking multiple layers of graph neural networks, the receptive field of nodes is gradually expanded to generate globally perceptive node embeddings, and graph-level feature vectors are obtained through pooling. Step S5: Perform nonlinear mapping on the obtained graph-level feature vectors to output the SOH evaluation results of the lithium battery.

2. The lithium battery SOH evaluation method based on graph-structured micro-voltage segments according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Perform a one-dimensional convolution operation on the time series voltage data of the same length to obtain the local spatial dependency; Step S22: Establish long-term temporal dependencies through the gating mechanism of the Long Short-Term Memory network, fuse the local spatial features and temporal dependency information obtained by convolution, and reconstruct them into spatiotemporal features.

3. The lithium battery SOH evaluation method based on graph-structured micro-voltage segments according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: Divide the time series voltage data into nodes, and use the reconstructed spatiotemporal features as the feature vectors of the corresponding nodes; Step S32: Construct a directed graph with unidirectional edges connected in sparse matrix form and labeled with the SOH value of the corresponding cycle, based on the edge connection method of the time series. Step S33: Using the actual SOH value of the corresponding cycle as the graph label, construct a data model that includes topological structure and node features.

4. The lithium battery SOH evaluation method based on graph-structured micro-voltage segments according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: By stacking multiple layers of graph neural networks, the receptive field of the nodes is gradually expanded. The network layer operation formula is as follows: ； ； ； in, For the first The node feature matrix of the layer, For an adjacency matrix with self-loops, I It is the identity matrix. for The degree matrix, For the first The node feature matrix of the layer, For the first The layer's learnable weight matrix, where σ is the activation function. For nodes and Connection weights between them The elements of the degree matrix, The row index represents the target node. For column indexes, the source node is represented; Step S42: Aggregate neighborhood node information through graph convolutional neural networks, mine potential dependency information between adjacent nodes, and generate globally aware node embeddings; Step S43: Convert node features into graph-level feature vectors through global average pooling.

5. The lithium battery SOH evaluation method based on graph-structured micro-voltage segments according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: Gradually extract aging information through multi-level nonlinear transformations and scale the dimensions of the graph-level feature vectors. Step S52: Use the fully connected layer to output the SOH value of the tested battery.

6. The lithium battery SOH evaluation method based on graph-structured micro-voltage segments according to claim 1, characterized in that: The training parameters of the above model are determined by minimizing the mean square error, using the following formula: ； in, β Let argmin be the set of model weight parameters, and argmin(·) be the parameters that minimize the objective function. n The number of training samples. p For sample index, y p For the first p The actual SOH value of the battery in the next cycle. For the first p Battery SOH estimate for the next cycle.

7. A lithium battery SOH evaluation system based on graph-structured micro-voltage segments, characterized in that, The method for evaluating the state of harmonics (SOH) of a lithium battery based on graph-structured micro-voltage segments as described in claim 1 is applied, comprising: a data preprocessing unit, a deep learning evaluation model unit, and a result output unit; The data preprocessing unit is used to preprocess the lithium battery cycle charge and discharge voltage data, filter out highly sensitive small voltage segments, and align the data lengths. The deep learning evaluation model unit includes a feature reconstruction module, a graph data model module, and a graph embedding and regression module, which are used to reconstruct the feature information of the preprocessed highly sensitive micro voltage fragment data and construct a graph data model to obtain graph-level feature vectors. The result output unit is used to perform nonlinear mapping and output the SOH evaluation results of the lithium battery.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the lithium battery SOH evaluation method based on graph-structured micro-voltage segments as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the lithium battery SOH evaluation method based on graph-structured micro-voltage segments as described in any one of claims 1-6.

Images