Lithium battery fault diagnosis method and device, system, and storage medium

Abstract

The application discloses a lithium battery fault diagnosis method and device, system and storage medium, one-dimensional time sequence monitoring signals are upgraded and coded into two-dimensional state transition feature images through a Markov transition field, convolutional neural networks are used to block embedding and construct space-time feature maps for the feature images, a graph attention network is introduced to dynamically weight and model the correlation between image block nodes, and finally, global pooling and a classifier are used to output fault diagnosis results, so that the extraction capability of early weak fault features is improved, and accurate diagnosis of various fault types of lithium batteries is realized.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery technology, specifically relating to a lithium battery fault diagnosis method, device, system, and storage medium. Background Technology

[0002] Lithium-ion batteries, due to their high energy density and long cycle life, have been widely used in electric vehicles, energy storage power stations, and consumer electronics. However, with the continuous improvement of battery system scale and power density, batteries may experience various faults during operation, such as internal short circuits, external short circuits, overcharging, and over-discharging. If these faults are not detected and diagnosed in a timely manner, they may accumulate to a certain extent and trigger thermal runaway, leading to serious safety accidents such as fires or explosions. Therefore, researching efficient and accurate lithium-ion battery fault diagnosis methods is of great significance for ensuring the safe and reliable operation of battery systems.

[0003] Current fault diagnosis methods for lithium batteries can be mainly divided into three categories: signal processing-based methods, model-based methods, and data-driven methods. Signal processing-based methods extract fault features through mathematical means such as wavelet transform, entropy calculation, and correlation coefficient analysis, but their ability to characterize early and subtle faults is limited. Model-based methods construct models based on the battery's electrical, thermal, and electrochemical mechanisms and achieve fault diagnosis through residuals; however, they rely on precise physical parameter identification and complex modeling, making them difficult to adapt to varying operating conditions. Data-driven methods automatically learn fault modes by collecting massive amounts of historical operating data, making them highly versatile and suitable for fault diagnosis research in big data scenarios.

[0004] In the early, latent stage, internal short-circuit faults in lithium batteries exhibit extremely weak and insidious fault signals. Their representation in conventional monitoring signals is very similar to that of a normal state, making them difficult to identify effectively. Therefore, they are among the most challenging faults to diagnose in individual lithium batteries. Existing deep learning methods mostly focus on directly modeling one-dimensional time-series monitoring signals, failing to fully extract fault information and resulting in the difficulty in effectively identifying these weak early fault signals. Summary of the Invention

[0005] To address the problems existing in the prior art, the present invention provides a lithium battery fault diagnosis method, device, system, and storage medium.

[0006] To achieve the above objectives, the present invention provides the following solution: A lithium battery fault diagnosis method, comprising: Step S1: Collect multi-source one-dimensional time-series monitoring signals of voltage, current, and temperature of the lithium battery under different operating conditions; Step S2: Perform data preprocessing on the acquired multi-source one-dimensional time-series monitoring signals to form a multi-channel Markov transfer field feature image; Step S3: Decompose the multi-channel Markov transfer field feature image into a series of non-overlapping image blocks, use a convolutional neural network to map each image block into a high-dimensional feature vector carrying local state transition information, and superimpose learnable position codes to obtain a node feature sequence. Step S4: Using each image patch as a node in the graph, and stacking the node feature vectors to form a node feature matrix, the temporal samples are transformed into a structured spatiotemporal feature map. Step S5: Input the spatiotemporal feature map into the graph attention network, and obtain the node-level high-order feature representation through the self-attention mechanism and the multi-head attention mechanism; Step S6: Perform global average pooling on the node-level high-order features output by the graph attention network to obtain a graph-level global feature vector. Map the state probability of this global feature vector to a Softmax classifier through a fully connected layer, and take the category with the highest probability as the fault diagnosis result of the lithium battery to be diagnosed.

[0007] Preferably, in step S2, the collected multi-source one-dimensional time-series monitoring signals are preprocessed, and each time-series signal is encoded into a two-dimensional state transition feature image by Markov transfer field. Then, the multiple two-dimensional feature images are superimposed according to the channel dimension to form a multi-channel Markov transfer field feature image.

[0008] Preferably, in step S4, each image patch is used as a node of the graph, and the node feature vectors are stacked to form a node feature matrix. The Manhattan distance is used to measure the spatial proximity between nodes based on the row and column index of the image patch in the original grid, and the edges between nodes are dynamically constructed through the K-nearest neighbor algorithm, thereby transforming the time series samples into a structured spatiotemporal feature graph.

[0009] Preferably, in step S5, the spatiotemporal feature map is input into a graph attention network, and the normalized attention weights between a node and its neighboring nodes are adaptively calculated through a self-attention mechanism. A multi-head attention mechanism is then used to weighted aggregate the features of neighboring nodes to obtain a node-level high-order feature representation.

[0010] The present invention also provides a lithium battery fault diagnosis device, comprising: The first processing module is used to collect multi-source one-dimensional time-series monitoring signals of voltage, current, and temperature of the lithium battery under different operating conditions; The second processing module is used to preprocess the acquired multi-source one-dimensional time-series monitoring signals to form a multi-channel Markov transfer field feature image. The third processing module is used to decompose the multi-channel Markov transfer field feature image into a series of non-overlapping image blocks, use a convolutional neural network to map each image block into a high-dimensional feature vector carrying local state transition information, and superimpose learnable position codes to obtain a node feature sequence. The fourth processing module is used to transform temporal samples into structured spatiotemporal feature maps by stacking node feature vectors to form a node feature matrix, with each image patch as a node in the graph. The fifth processing module is used to input the spatiotemporal feature map into the graph attention network, and obtain node-level high-order feature representations through self-attention mechanism and multi-head attention mechanism; The sixth processing module is used to perform global average pooling on the node-level high-order features output by the graph attention network to obtain a graph-level global feature vector. This global feature vector is then mapped to the state probability of a Softmax classifier through a fully connected layer, and the class with the highest probability is taken as the fault diagnosis result of the lithium battery to be diagnosed.

[0011] As a preferred embodiment, the second processing module performs data preprocessing on the acquired multi-source one-dimensional time-series monitoring signals, encodes each time-series signal into a two-dimensional state transition feature image through a Markov transfer field, and then superimposes the multiple two-dimensional feature images according to the channel dimension to form a multi-channel Markov transfer field feature image.

[0012] Preferably, the fourth processing module uses each image patch as a node in the graph, stacks node feature vectors to form a node feature matrix, uses Manhattan distance to measure the spatial proximity between nodes based on the row and column indices of the image patch in the original grid, and dynamically constructs edges between nodes through the K-nearest neighbor algorithm, thereby transforming the time series samples into a structured spatiotemporal feature graph.

[0013] Preferably, the fifth processing module inputs the spatiotemporal feature map into the graph attention network, adaptively calculates the normalized attention weights between the node and its neighboring nodes through a self-attention mechanism, and uses a multi-head attention mechanism to weighted aggregate the features of the neighboring nodes to obtain a node-level high-order feature representation.

[0014] The present invention also provides a lithium battery fault diagnosis system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a lithium battery fault diagnosis method when executed by the processor.

[0015] The present invention also provides a storage medium storing a computer program, which executes a lithium battery fault diagnosis method when running.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes Markov transition fields to encode one-dimensional time-series monitoring signals into two-dimensional state transition feature images. Compared to the original sequence signal, this adds a representation of the relationships between data values ​​at different times, enabling the depiction of the dynamic process of faults and uncovering and amplifying the derivative feature associations of early weak faults over time. This allows for accurate characterization of the temporal evolution of battery faults from their early stages to their manifest phase. A convolutional neural network is used to embed the Markov transition field feature images into blocks. Leveraging the characteristics of local receptive fields and weight sharing, local texture patterns are accurately captured while preserving the topological relationships of pixel spaces within blocks, effectively avoiding the loss of structural information caused by one-dimensional flattening. The spatiotemporal feature map is dynamically constructed using the K-nearest neighbor algorithm and Manhattan distance, flexibly characterizing long-range dependencies that may exist during state transitions. A graph attention network is introduced to dynamically weight and model the associations between image block nodes, adaptively focusing on the state transition regions most sensitive to faults and suppressing interference from noise blocks. This significantly improves the accuracy and robustness of early weak fault feature extraction, enabling accurate diagnosis of various fault types in lithium batteries. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a lithium battery fault diagnosis method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the construction of spatiotemporal feature maps based on Markov transfer fields and convolutional neural networks as described in this invention. Figure 3 This is a flowchart of the feature extraction and fault diagnosis process based on graph attention network described in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1 like Figure 1As shown, the present invention provides a lithium battery fault diagnosis method, comprising: Step S1: Collect multi-source one-dimensional time-series monitoring signals of voltage, current and temperature of the lithium battery under different operating conditions. Perform online segmentation processing on the collected signals through a sliding time window. Use the multiple data segments obtained in each time window as fault data samples. These data include normal state signals and various fault state signals. Step S2: Perform data preprocessing on the acquired multi-source one-dimensional time-series monitoring signals. Encode each time-series signal into a two-dimensional state transition feature image using Markov transfer field. Then, superimpose the multiple two-dimensional feature images according to the channel dimension to form a multi-channel Markov transfer field feature image. Step S3: Decompose the multi-channel Markov transfer field feature image into a series of non-overlapping image blocks, use a convolutional neural network to map each image block into a high-dimensional feature vector carrying local state transition information, and superimpose learnable position codes to obtain a node feature sequence. Step S4: Using each image patch as a node of the graph, the node feature vectors are stacked to form a node feature matrix. The Manhattan distance is used to measure the spatial proximity between nodes based on the row and column indices of the image patch in the original grid. The edges between nodes are dynamically constructed through the K-nearest neighbor algorithm, thereby transforming the time series samples into a structured spatiotemporal feature graph. Step S5: Input the spatiotemporal feature map into the graph attention network, adaptively calculate the normalized attention weights between the node and its neighboring nodes through the self-attention mechanism, and use the multi-head attention mechanism to perform weighted aggregation of the features of the neighboring nodes to obtain the node-level high-order feature representation. Step S6: Perform global average pooling on the node-level high-order features output by the graph attention network to obtain a graph-level global feature vector. Map the state probability of this global feature vector to a Softmax classifier through a fully connected layer, and take the category with the highest probability as the fault diagnosis result of the lithium battery to be diagnosed.

[0022] In one embodiment of the present invention, in step S1, the lithium battery to be diagnosed is first operated in a normal charging and discharging state, and one-dimensional time-series signals such as voltage, current, and temperature are collected from the battery's monitoring nodes as diagnostic signals. Then, different types of fault conditions are applied to the lithium battery, including but not limited to simulating internal short-circuit faults with different resistance values ​​by connecting a parallel variable resistor, applying an external short circuit, and applying overcharge or over-discharge excitation to simulate corresponding faults, thereby collecting monitoring signals under various fault conditions. Based on the fault characteristics, a sliding time window of predefined size is used to segment and process the collected multi-source signals online, and the multiple data segments acquired in each sliding time window are used as fault data samples.

[0023] In one embodiment of the present invention, in step S2, the core idea of ​​the Markov transfer field is to discretize the one-dimensional time-series monitoring signal into a finite number of states, and construct a transfer probability matrix by analyzing the transfer relationships between these states. This is applicable to non-stationary time-series data such as voltage, current, and temperature. First, the original signal is quantized and its state is divided, and then defined. The state space is constructed using bins, dividing all observations in sequence X into bins according to their value ranges. In each interval, each observation value They are assigned to the corresponding bins based on their numerical value. After determining the state space, the transition probability matrix is ​​calculated by statistically analyzing the transitions of the sequence between states. Elements in the matrix Define the data point at the current time as belonging to the bin. The previous moment belonged to the sorting of containers. The conditional probability is shown in equation (1): (1) In actual processing of lithium battery time series data, this probability value is estimated by maximum likelihood estimation through the transition frequency between states in the statistical sequence. The calculation formula is shown in equation (2): (2) in Represents the state in a time series. Following state Frequency of occurrence. To compensate for the transition matrix. Ignoring the shortcomings of the original signal's temporal distribution, the Markov transfer field remaps the probability values ​​in the transfer matrix W back to the original timestamps, constructing a... Time correlation matrix ,matrix elements in Represents a time step The state transition at point to time step The probability of being in state is shown in equation (3): (3) In practice, voltage, current and temperature are each used to generate a Markov transfer field image independently. The three images are then superimposed along the channel dimension to form a multi-channel Markov transfer field feature image, thereby simultaneously encoding the state transition information of multi-source monitoring signals.

[0024] In one embodiment of the present invention, in step S3, after obtaining the Markov transition field feature image, it is decomposed into a series of non-overlapping image blocks, drawing on the idea of ​​image block segmentation. A convolutional neural network is then used to map each image block into a high-dimensional feature vector with state transition information. First, a 3×3 standard convolutional layer is used to perform preliminary feature extraction on the input multi-channel feature image, capturing basic texture and cross-variable coupling information. Then, the feature map, after activation and normalization processing, is input into the convolutional kernel, with the kernel size and stride both set to a predefined image block size. Two-dimensional convolutional layer The feature map is then subjected to non-overlapping sliding embedding, as shown in equation (4): (4) in and These are the learnable weights and biases of the convolutional layer, respectively. Since the stride of the convolutional kernel is equal to its size, this operation performs a non-overlapping slide on the feature map, with each step adding a... The local region is mapped to a depth of The feature vectors are used to recover the relative positional relationships of each image patch in the original state space. A learnable positional encoding matrix is ​​introduced. , with image patch embedding sequence Adding each element together yields the final node feature sequence. As shown in equation (5): (5) In one embodiment of the present invention, in step S4, after converting the time-series signal into an image block embedding sequence carrying state transition spatial information, graph theory is introduced to formalize the state representation of the lithium battery into a structured feature graph. ,in For inclusion The set of vertices of a node, i.e. Image blocks, This is the set of edges describing the relationships between nodes. The feature vectors of all nodes are stacked to form a node feature matrix. And use an adjacency matrix Encode topological information between nodes. To more flexibly construct topological relationships and capture long-range dependencies that may exist during state transitions, use the row and column indices of image patches in the original grid. As coordinates, for any two nodes and The Manhattan distance is used to measure the spatial proximity between nodes, as shown in Equation (6): (6) Based on this distance metric, for each node in the graph Search for the nearest node in the remaining set of nodes. Each node is designated as its neighbor node, and connections are established between them, as shown in equation (7): (7) After the above calculations, each lithium battery time-series sample was successfully converted into a structured spatiotemporal feature map. This includes both dynamic evolution features along the timeline and the feature similarity space structure extracted by a convolutional neural network, thus effectively capturing weak fault features in one-dimensional time series. For example... Figure 2 The diagram shows the process of constructing the spatiotemporal feature map of a Markov transition field and a convolutional neural network.

[0025] As one embodiment of the present invention, in step S5, since the contribution of state transition information carried by different nodes to fault judgment varies significantly, this step introduces a graph attention network. Using the constructed spatiotemporal feature map as input, it adaptively assigns differentiated weights to the contributions of nodes under different fault modes. For example... Figure 3 The diagram illustrates the feature extraction and fault diagnosis process based on graph attention networks. The graph attention network first utilizes a shared linear transformation matrix... The input node features are mapped to a high-dimensional space, and then an attention mechanism is introduced to compute the nodes. with neighboring nodes The original correlation coefficient between them is shown in equation (8): (8) in is a learnable weight vector, ‖ denotes vector concatenation, and LeakyReLU is a non-linear activation function. To ensure comparability of coefficients between different nodes, the Softmax function is used to normalize the attention coefficients of all neighbors of the target node, resulting in normalized attention weights. As shown in equation (9): (9) Based on this weight, the node The output features are obtained by weighted summation of the transformed features of all nodes in the neighborhood followed by nonlinear activation. To further enhance the stability of the training process and capture feature correlations in different subspaces, a multi-head attention mechanism is introduced for parallel operation. The attention calculation process is performed independently for each group, and the results are concatenated. The final node output features are shown in Equation (10): (10) in Indicates the first The normalization coefficients calculated by each attention head This is the corresponding linear transformation matrix. This mechanism allows the model to focus on specific state transition regions in the spatiotemporal feature map that are most sensitive to faults, based on the dynamic changes in the input data, and to suppress the interference of noise blocks, thereby significantly improving the accuracy and robustness of feature extraction.

[0026] In one embodiment of the present invention, in step S6, to achieve a complete mapping from feature representation to fault category, after feature extraction by the graph attention network, global average pooling is used to spatially aggregate the output features of all nodes. Let the final node feature matrix output by the graph attention network be... ,in For the number of nodes, To output the feature dimension, the graph-level global feature vector is... The average value of all node features is obtained as shown in equation (11): (11) This operation aggregates fault information from each local state transition region into a global fault representation for the entire time window, eliminating the impact of the number of nodes on the input dimension of the subsequent classification network. Subsequently, the global feature vector undergoes a non-linear transformation through a fully connected layer, compressing and mapping the high-dimensional graph representation to a value corresponding to the number of fault categories. The matching low-dimensional space is used, and the probability distribution of each category is output through the Softmax function, as shown in Equation (12): (12) in and These are the learnable weight matrix and bias vector of the fully connected layer, respectively. The category with the highest probability is ultimately selected as the fault diagnosis output for the lithium battery. During the model training phase, the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label, as shown in equation (13): (13) in The number of training samples. Number of fault categories For real labels, To predict probabilities, the model parameters are optimized by minimizing this loss function, ultimately resulting in a spatiotemporal graph network model capable of accurately diagnosing various fault types in lithium batteries.

[0027] This invention implements a lithium battery fault diagnosis method driven by a spatiotemporal graph network based on Markov transfer fields, convolutional neural networks, and graph attention networks. Specifically, it includes the following steps: (1) Collecting multi-source one-dimensional time-series monitoring signals of voltage, current, and temperature of the lithium battery under different operating states, and performing segmented online processing of the signals through a sliding time window to obtain fault data samples containing normal states and multiple fault states; (2) Upgrading each time-series signal into a two-dimensional state transition feature image through Markov transfer fields, and superimposing multiple images according to channel dimensions to form a multi-channel Markov transfer field feature image; (3) Decomposing the multi-channel feature image into a series of non-overlapping image blocks, and using a convolutional neural network to map each image block into a high-resolution image carrying local state transition information. (3) The feature vector is superimposed with a learnable positional code to obtain the node feature sequence; (4) Each image block is taken as a node, and the Manhattan distance is used to measure the spatial proximity of the nodes according to their row and column indices in the original grid. The edges between nodes are dynamically constructed through the K-nearest neighbor algorithm, and the time series samples are transformed into a structured spatiotemporal feature map; (5) The spatiotemporal feature map is input into the graph attention network, and the normalized attention weight between the node and its neighboring nodes is adaptively calculated through the self-attention mechanism. The neighborhood features are weighted and aggregated through the multi-head attention mechanism to obtain the node-level high-order feature representation; (6) The node-level high-order features are globally averaged to obtain the graph-level global feature vector. The state probability is mapped through the fully connected layer and the Softmax classifier. The category with the highest probability is taken as the fault diagnosis result. This invention can mine and amplify the derivative feature association of early weak faults over time, adaptively focus on the state transition region most sensitive to faults, and quickly and accurately realize the detection and diagnosis of various fault types of lithium batteries.

[0028] Example 2 The present invention also provides a lithium battery fault diagnosis device, comprising: The first processing module is used to collect multi-source one-dimensional time-series monitoring signals of voltage, current, and temperature of the lithium battery under different operating conditions; The second processing module is used to preprocess the acquired multi-source one-dimensional time-series monitoring signals to form a multi-channel Markov transfer field feature image. The third processing module is used to decompose the multi-channel Markov transfer field feature image into a series of non-overlapping image blocks, use a convolutional neural network to map each image block into a high-dimensional feature vector carrying local state transition information, and superimpose learnable position codes to obtain a node feature sequence. The fourth processing module is used to transform temporal samples into structured spatiotemporal feature maps by stacking node feature vectors to form a node feature matrix, with each image patch as a node in the graph. The fifth processing module is used to input the spatiotemporal feature map into the graph attention network, and obtain node-level high-order feature representations through self-attention mechanism and multi-head attention mechanism; The sixth processing module is used to perform global average pooling on the node-level high-order features output by the graph attention network to obtain a graph-level global feature vector. This global feature vector is then mapped to the state probability of a Softmax classifier through a fully connected layer, and the class with the highest probability is taken as the fault diagnosis result of the lithium battery to be diagnosed.

[0029] As a preferred embodiment, the second processing module performs data preprocessing on the acquired multi-source one-dimensional time-series monitoring signals, encodes each time-series signal into a two-dimensional state transition feature image through a Markov transfer field, and then superimposes the multiple two-dimensional feature images according to the channel dimension to form a multi-channel Markov transfer field feature image.

[0030] As one embodiment of the present invention, the fourth processing module uses each image block as a node of the graph, stacks node feature vectors to form a node feature matrix, uses Manhattan distance to measure the spatial proximity between nodes based on the row and column index of the image block in the original grid, and dynamically constructs the edges between nodes through the K-nearest neighbor algorithm, thereby transforming the time series samples into a structured spatiotemporal feature graph.

[0031] As one embodiment of the present invention, the fifth processing module inputs the spatiotemporal feature map into the graph attention network, adaptively calculates the normalized attention weights between the node and its neighboring nodes through a self-attention mechanism, and uses a multi-head attention mechanism to weighted aggregate the features of the neighboring nodes to obtain a node-level high-order feature representation.

[0032] Example 3 The present invention also provides a lithium battery fault diagnosis system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a lithium battery fault diagnosis method when executed by the processor.

[0033] Example 4 The present invention also provides a storage medium storing a computer program, which executes a lithium battery fault diagnosis method when running.

[0034] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for diagnosing lithium battery faults, characterized in that, include: Step S1: Collect multi-source one-dimensional time-series monitoring signals of voltage, current, and temperature of the lithium battery under different operating conditions; Step S2: Perform data preprocessing on the acquired multi-source one-dimensional time-series monitoring signals to form a multi-channel Markov transfer field feature image; Step S3: Decompose the multi-channel Markov transfer field feature image into a series of non-overlapping image blocks, use a convolutional neural network to map each image block into a high-dimensional feature vector carrying local state transition information, and superimpose learnable position codes to obtain a node feature sequence. Step S4: Using each image patch as a node of the graph, the node feature vectors are stacked to form a node feature matrix, and the time series samples are transformed into a structured spatiotemporal feature map. Step S5: Input the spatiotemporal feature map into the graph attention network, and obtain the node-level high-order feature representation through the self-attention mechanism and the multi-head attention mechanism; Step S6: Perform global average pooling on the node-level high-order features output by the graph attention network to obtain a graph-level global feature vector. Map the state probability of this global feature vector to a Softmax classifier through a fully connected layer, and take the category with the highest probability as the fault diagnosis result of the lithium battery to be diagnosed.

2. The lithium battery fault diagnosis method as described in claim 1, characterized in that, In step S2, the acquired multi-source one-dimensional time-series monitoring signals are preprocessed. Each time-series signal is encoded into a two-dimensional state transition feature image by using a Markov transfer field. The multiple two-dimensional feature images are then superimposed according to the channel dimension to form a multi-channel Markov transfer field feature image.

3. The lithium battery fault diagnosis method as described in claim 2, characterized in that, In step S4, each image patch is used as a node of the graph, and the node feature vectors are stacked to form a node feature matrix. The Manhattan distance is used to measure the spatial proximity between nodes based on the row and column indices of the image patch in the original grid, and the edges between nodes are dynamically constructed through the K-nearest neighbor algorithm, thereby transforming the time series samples into a structured spatiotemporal feature graph.

4. The lithium battery fault diagnosis method as described in claim 3, characterized in that, In step S5, the spatiotemporal feature map is input into the graph attention network. The normalized attention weights between the node and its neighboring nodes are adaptively calculated through the self-attention mechanism. The features of the neighboring nodes are weighted and aggregated using a multi-head attention mechanism to obtain the node-level high-order feature representation.

5. A lithium battery fault diagnosis device, characterized in that, include: The first processing module is used to collect multi-source one-dimensional time-series monitoring signals of voltage, current, and temperature of the lithium battery under different operating conditions; The second processing module is used to preprocess the acquired multi-source one-dimensional time-series monitoring signals to form a multi-channel Markov transfer field feature image. The third processing module is used to decompose the multi-channel Markov transfer field feature image into a series of non-overlapping image blocks, use a convolutional neural network to map each image block into a high-dimensional feature vector carrying local state transition information, and superimpose learnable position codes to obtain a node feature sequence. The fourth processing module is used to transform temporal samples into structured spatiotemporal feature maps by stacking node feature vectors to form a node feature matrix, with each image patch as a node in the graph. The fifth processing module is used to input the spatiotemporal feature map into the graph attention network, and obtain node-level high-order feature representations through self-attention mechanism and multi-head attention mechanism; The sixth processing module is used to perform global average pooling on the node-level high-order features output by the graph attention network to obtain a graph-level global feature vector. This global feature vector is then mapped to the state probability of a Softmax classifier through a fully connected layer, and the class with the highest probability is taken as the fault diagnosis result of the lithium battery to be diagnosed.

6. The lithium battery fault diagnosis device as described in claim 5, characterized in that, The second processing module preprocesses the acquired multi-source one-dimensional time-series monitoring signals, encodes each time-series signal into a two-dimensional state transition feature image through a Markov transfer field, and then superimposes the multiple two-dimensional feature images according to the channel dimension to form a multi-channel Markov transfer field feature image.

7. The lithium battery fault diagnosis device as described in claim 6, characterized in that, The fourth processing module uses each image patch as a node in the graph, stacks node feature vectors to form a node feature matrix, uses Manhattan distance to measure the spatial proximity between nodes based on the row and column indices of the image patch in the original grid, and dynamically constructs edges between nodes through the K-nearest neighbor algorithm, thereby transforming time-series samples into a structured spatiotemporal feature graph.

8. The lithium battery fault diagnosis device as described in claim 7, characterized in that, The fifth processing module inputs the spatiotemporal feature map into the graph attention network, adaptively calculates the normalized attention weights between a node and its neighboring nodes through a self-attention mechanism, and uses a multi-head attention mechanism to weighted aggregate the features of neighboring nodes to obtain a node-level high-order feature representation.

9. A lithium battery fault diagnosis system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the lithium battery fault diagnosis method as described in any one of claims 1-4 when executed by the processor.

10. A storage medium, characterized in that, The storage medium stores a computer program, which executes the lithium battery fault diagnosis method as described in any one of claims 1-4 when it is run.

Images