Battery fault diagnosis method based on graph structure and LSTM

Abstract

The invention discloses a battery fault diagnosis method based on a graph structure and an LSTM, and belongs to the field of battery fault diagnosis, and the method comprises the steps: converting battery blocks of a battery module and a connection relation thereof into a graph model based on a battery pack space connection information representation mode of the graph structure, so as to effectively capture the space dependence between the battery blocks; meanwhile, by combining the graph neural network and the long-short-term memory network, the spatial-temporal characteristics of the battery state can be accurately extracted, the processing capacity of time sequence data is enhanced, and therefore the accuracy and comprehensiveness of fault diagnosis are improved; besides, aiming at the problem of lack of a confidence evaluation mechanism in the prior art, an accurate confidence evaluation module is designed, and the reliability and transparency of the decision are effectively improved by calculating the confidence score of the diagnosis result.

Description

Technical Field

[0001] The present invention belongs to the field of battery fault diagnosis, and particularly relates to a battery fault diagnosis method based on a graph structure and LSTM. Background Art

[0002] The existing electric vehicle battery fault diagnosis technologies mainly have the following problems: First, the spatial connection information of the battery pack is not fully utilized. Traditional methods mostly rely on single feature data or simple clustering analysis, and do not fully consider the complex spatial topological structure between battery blocks, resulting in the failure to effectively model the spatial dependence relationship. Second, the existing technologies have relatively limited capabilities in spatio-temporal feature extraction. Most methods ignore the interaction of battery data in time and space, and do not comprehensively consider the dynamic changes of battery states over time and the influence between different battery blocks. Finally, the confidence evaluation system for the diagnosis results of electric vehicle battery faults is not perfect. Many methods lack a confidence evaluation mechanism for the diagnosis results and cannot effectively measure the credibility of the diagnosis results, leading to difficulties in making reliable judgments in the face of uncertain or ambiguous fault situations. Summary of the Invention

[0003] To solve the above technical problems, the present invention provides a battery fault diagnosis method based on a graph structure and LSTM, including:

[0004] Acquire the time-series feature information of each battery block in the battery module of the electric vehicle and the state of the battery pack based on the battery management system, and construct a time-series data set based on the time-series feature information of each battery block and the state of the battery pack;

[0005] Construct graph structure data based on the hardware connection relationship of the battery blocks in the battery module;

[0006] Construct a spatio-temporal feature extraction module for battery block information based on the graph structure data and the long short-term memory network model;

[0007] Construct a calculation model based on the spatio-temporal feature extraction module and the multi-head attention mechanism, and train the calculation model through the time-series data set to obtain an electric vehicle battery fault diagnosis model;

[0008] Deploy the electric vehicle battery fault diagnosis model to the electric vehicle BMS platform to diagnose the faults of the battery.

[0009] Preferably, the time-series feature information of each battery block includes battery voltage, current, temperature, state of health of the battery, state of charge of the battery, internal resistance, and charge cycle.

[0010] Preferably, the process of constructing the time-series data set based on the time-series feature information of each battery block and the state of the battery pack includes:

[0011] Pre-define the characteristic information of each battery block in the predefined battery module, collect the characteristic information in each battery block, and obtain the complete input characteristic data of the battery pack;

[0012] Perform battery pack status annotation on the complete input characteristic data of the battery pack, integrate the complete input characteristic data of the battery pack and the annotated battery pack status, and obtain the time series data set for battery fault diagnosis.

[0013] Preferably, the process of constructing the graph structure data based on the hardware connection relationship of the battery blocks in the battery module includes:

[0014] Take each battery block as a node in the graph, and take the connection relationship between the battery blocks as the edge between the nodes to construct a graph data structure including spatial information;

[0015] Take the characteristic information of each battery block at a single moment as the characteristic of the corresponding node, and construct the graph structure data;

[0016] Among them, each set of input characteristic data of the battery pack should be graph data at 12 time points, and the expression is:

[0017] G m =[G1,G2,...,G 12 =Graph(X m );

[0018] Among them, G m represents the graph data corresponding to the input characteristic data embedded based on the graph structure representation method, X m is the input characteristic data, G1 is the graph data at the first time point, G2 is the graph data at the second time point, G 12 is the graph data at the twelfth time point, and Graph(·) is the graph representation and feature embedding operation.

[0019] Preferably, the process of constructing the spatio-temporal feature extraction module of the battery block information based on the graph structure data and the long short-term memory network model includes:

[0020] Construct a spatio-temporal feature extraction module of the battery block information based on the graph neural network and the long short-term memory network model;

[0021] Among them, the graph neural network extracts spatio-temporal features from the graph structure data to obtain global embedding features;

[0022] The long short-term memory network extracts time features from the global embedding features to obtain intermediate features after time feature embedding.

[0023] Preferably, in the graph neural network, the expression for feature extraction by the first-layer graph convolutional layer is:

[0024]

[0025] Among them, W1 represents the learnable weight matrix of the first graph convolutional layer; σ(·) is the activation function, G1 is the graph data, A is the adjacency matrix corresponding to the graph data, is the normalized adjacency matrix; D is the degree matrix, and D is a diagonal matrix, and the values of the elements on its diagonal are the number of adjacent nodes of the corresponding nodes. Emb1 is the global embedding feature obtained through the graph convolutional layer;

[0026] The expression for the intermediate feature after obtaining the time feature embedding is:

[0027] H1 = LSTM(Emb1, Q1);

[0028] Among them, Q1 represents the learnable weight matrix of the first LSTM layer, H1 represents the intermediate feature after time feature embedding, and LSTM(·) represents the LSTM layer processing function.

[0029] Preferably, it further includes that in the graph neural network, the expression for feature extraction by the second graph convolutional layer is:

[0030]

[0031] Among them, W2 represents the learnable weight matrix of the second graph convolutional layer, and Full(·) represents the full connection layer processing function.

[0032] Preferably, the process of training the calculation model through the time series data set to obtain the electric vehicle battery fault diagnosis model includes:

[0033] Using the Adam optimizer to optimize the cross-entropy loss function, and updating the network parameters through backpropagation and iterating multiple times to optimize the calculation model until the cross-entropy loss function tends to be the smallest and the model converges and then outputs to obtain the electric vehicle battery fault diagnosis model.

[0034] Preferably, the process of diagnosing the battery fault based on the electric vehicle battery fault diagnosis model includes:

[0035] Obtaining the global embedding feature containing spatio-temporal information based on the spatio-temporal feature extraction module;

[0036] Capturing the temporal dependence and spatial relationship in the global embedding feature of the spatio-temporal information based on the multi-head attention mechanism to obtain the attention output;

[0037] Performing weighted fusion on the attention output to obtain the final multi-head attention output;

[0038] Calculate the final output of the multi-head attention through a fully connected layer and a residual connection to obtain processed feature data;

[0039] Calculate the processed feature data through a Dropout layer to obtain output features;

[0040] Use the Softmax layer as the output layer to convert the output features into probability values for classification decisions to obtain the final prediction result.

[0041] Preferably, after obtaining the final prediction result, it further includes:

[0042] Construct a confidence evaluation method for the output result of the battery fault diagnosis model, evaluate the final prediction result based on the confidence evaluation method, and obtain the accuracy rate of each diagnosis type;

[0043] Calculate the mean and variance of all prediction probabilities under the fault type based on the accuracy rate of each diagnosis type, and obtain the confidence score and confidence threshold based on the mean and variance;

[0044] If the confidence score is greater than the confidence threshold, output the diagnosis result and the corresponding confidence score. If the confidence score is less than the confidence threshold, perform manual intervention.

[0045] Compared with the prior art, the present invention has the following advantages and technical effects:

[0046] The spatio-temporal feature extraction method combining graph neural network and LSTM can accurately model the spatial and temporal dependencies in the battery system, improving the accuracy of fault diagnosis; especially for the mutual influence between different battery blocks in the battery pack, the present invention can effectively capture it, avoiding the defect of ignoring the spatial structure in the traditional method;

[0047] By modeling the spatial relationship between battery blocks and processing time series information, the present invention can accurately identify the fault state of the battery under different working conditions, making the battery fault diagnosis more reliable;

[0048] The present invention designs a confidence evaluation mechanism, which can evaluate the credibility of the results of battery fault diagnosis; through the confidence scoring of the diagnosis results, users can obtain more accurate and reliable fault judgments, especially when facing fuzzy or highly uncertain fault situations; deploying the fault diagnosis model into the battery management system (BMS) of electric vehicles can achieve real-time diagnosis and early warning of battery faults, timely detect potential faults, reduce safety hazards and ensure the safe operation of electric vehicles. Description of the Drawings

[0049] The accompanying drawings, which form a part of this application, are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation of this application. In the drawings:

[0050] Figure 1 is the overall technical flow chart of the embodiment of the present invention;

[0051] Figure 2 is the schematic diagram of the graph construction method based on hardware electrical connection in the embodiment of the present invention;

[0052] Figure 3 is the schematic diagram of graph representation and feature embedding operation in the embodiment of the present invention;

[0053] Figure 4 is the structural block diagram of the spatio-temporal feature extraction module in the embodiment of the present invention;

[0054] Figure 5 is the structural block diagram of the electric vehicle battery fault diagnosis model in the embodiment of the present invention. Detailed Embodiments

[0055] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the accompanying drawings and combine with the embodiments to detail this application.

[0056] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0057] Embodiment 1

[0058] As Figure 1 shown, this embodiment provides a battery fault diagnosis method based on graph structure and LSTM, including:

[0059] Based on the battery management system, obtain the temporal feature information of each battery block in the battery module of the electric vehicle and the state of the battery pack, and construct a temporal data set based on the temporal feature information of each battery block and the state of the battery pack;

[0060] Construct graph structure data based on the hardware connection relationship of the battery blocks in the battery module;

[0061] Construct a spatio-temporal feature extraction module for battery block information based on the graph structure data and the long short-term memory network model;

[0062] A computational model is constructed based on the spatio-temporal feature extraction module and the multi-head attention mechanism, and the computational model is trained with the time series dataset to obtain an electric vehicle battery fault diagnosis model;

[0063] The electric vehicle battery fault diagnosis model is deployed on the electric vehicle BMS platform to diagnose battery faults.

[0064] Specifically, it includes:

[0065] S1. Construction of the time series dataset for battery fault diagnosis. The time series feature information of each battery block in the electric vehicle battery module (including battery voltage, current, temperature, battery states SOC and SOH, internal resistance, charging cycle) and the state of the battery pack (including normal battery and faults such as overcharge, over-discharge, short circuit, capacity degradation, etc.) are collected together to construct the dataset;

[0066] The goal of this embodiment is to perform real-time diagnosis on the state of the overall battery pack based on the feature information of each battery block in the electric vehicle battery module. To achieve this goal, it is first necessary to collect the time series dataset for training the battery fault diagnosis model. Since the battery management system (BMS) can continuously provide the characteristic measurement data of the electric vehicle battery, this embodiment therefore relies on the BMS to construct the time series dataset. The specific steps include:

[0067] S1-1. Pre-define the feature information of each battery block in the battery module, including battery voltage V, current I, temperature T, battery states SOC and SOH, R, and charging cycle C. That is, for the one-time sampling feature information x of the battery block:

[0068] x = [V, I, T, SOC, SOH, R, C];

[0069] Collect 12 times at 5s time intervals to obtain a set of time series data x of the i-th battery block i as:

[0070]

[0071] where, represents the feature information (including voltage V, current I, temperature T, battery states SOC and SOH, internal resistance R, and charging cycle C) of the i-th battery block at the j-th time point; and the total number of battery blocks in the battery module is N, then i ∈ [1, N], j ∈ [1, 12];

[0072] S1-2. For each battery block, perform the data collection operation in S1-1 to obtain a set of complete input feature data X of the battery pack m :

[0073] X m= [x1, x2,..., x N ;

[0074] where x N represents the timing feature data of the Nth battery block;

[0075] S1-3. Perform battery pack state Y m annotation on the input feature data X, including battery normal and fault categories such as overcharge, over-discharge, short circuit, capacity fade, etc., and combine the battery pack input feature data X m with the battery pack state Y m to jointly form a dataset:

[0076] data m = [X m , Y m ;

[0077] S1-4. Repeat steps S1-1 to S1-3 to obtain a total of M groups of battery fault diagnosis timing datasets D:

[0078] D = [data1, data2,..., data M ;

[0079] Any group of data data m contains the feature information of the battery pack and the corresponding battery state, and m ∈ [1, M]; and divide the training set and the test set. The training set is used as the basis for subsequent model training, while the test set is used for model confidence evaluation and fine-tuning.

[0080] S2. Design a graph structure representation method. Construct a graph representation through the hardware connection relationship between battery blocks in the battery module, and characterize the spatial characteristics of the battery module through the graph data, so as to obtain specific spatial information for different battery connection methods;

[0081] In this embodiment, based on the hardware connection relationship between battery blocks in the battery module, a graph representation method is constructed to capture the spatial characteristics of the battery module. The battery blocks in the battery module are usually connected in series and parallel. Therefore, during fault diagnosis, this embodiment designs a graph structure representation method for the battery blocks based on this characteristic, so that the battery fault diagnosis model can utilize the spatial relationship between the battery blocks to improve the accuracy of fault diagnosis. The specific steps include:

[0082] S2-1. Construct a graph structure. Since series and parallel connections are usually adopted in the battery module to balance the voltage and capacity requirements of the batteries; the series connection increases the total voltage of the battery pack, while the parallel connection increases the capacity and current-carrying capacity of the battery pack. Therefore, based on this principle, in this embodiment, each battery block in the battery module is regarded as a node (Node) in the graph, and the hardware connection relationships (including series and parallel connections) between the battery blocks are used as the edges (Edges) between the nodes, as Figure 2 shown. In this way, the spatial topological relationship of the battery module is graph-structured to form a graph data structure of spatial information.

[0083] S2-2. Design the node feature representation in the graph. Through the graph structure representation method in S2-1, each battery block can be embedded into the overall graph structure. Therefore, the feature information of each battery block at a single moment is used as the feature of its corresponding node to realize the design of the node feature representation in the graph; in addition, since the time-series data of the battery pack collected in S1 contains the feature information of 12 time points, each group of input feature data X m of the battery pack corresponds to the graph data of 12 time points:

[0084] G m =[G1, G2,..., G 12 =Graph(X m );

[0085] where, G m represents the graph data corresponding to the input feature data X m embedded based on the graph structure representation method, and it includes the graph data of 12 time points in total; Graph(·) is the graph representation and feature embedding operation in this embodiment, and this process is as Figure 3 shown.

[0086] S2-3. Based on the steps of S2-1 and S2-2, all the input feature data collected in S1 are embedded as graph data and respectively form a group of data with the corresponding battery state categories.

[0087] S3. Construct a spatio-temporal feature extraction module for battery block information based on the graph structure data and LSTM. This spatio-temporal feature extraction module can fuse the spatio-temporal information of the battery feature data to provide a more powerful global embedding feature;

[0088] In this embodiment, a spatio-temporal feature extraction module is constructed based on the graph structure data and LSTM (Long Short-Term Memory Network) to extract the spatio-temporal features of the battery block information. The purpose of this module is to fuse the spatio-temporal information of the battery feature data, thereby providing stronger expression ability and generating global embedding features, so that the subsequent model can accurately predict the overall fault state of the battery pack. The specific steps include:

[0089] S3-1 Design the basic structure of the spatio-temporal feature extraction module, and extract the spatio-temporal features of battery blocks by combining graph neural network (GNN) and LSTM. The structure of this module is as Figure 4 shown (the number of graph convolutional layers is consistent with the number of time points of the sampled data. If the number of sampling time points is changed, the structure of the spatio-temporal feature extraction module can also be flexibly adjusted to adapt to different situations).

[0090] Among them, the graph convolutional layer is used to capture the mutual influence relationship between battery blocks in the battery module. The feature information of each battery block is propagated through the graph convolutional layer. In this way, the graph convolution can effectively represent the spatial information of the battery blocks in the battery module; and the feature representation of the battery block at the current moment is incorporated into the graph convolutional layer at subsequent moments through the LSTM layer to ensure that the model pays attention to the influence relationship in the time dimension, and finally obtains the global embedding features at each time point;

[0091] S3-2 Perform spatio-temporal feature extraction on the graph data G m containing 12 time points; First, obtain the adjacency matrix corresponding to the graph data G1 as A, then the global embedding feature Emb1 obtained after passing through the graph convolutional layer is:

[0092]

[0093] Among them, W1 represents the learnable weight matrix of the first graph convolutional layer; σ(·) is the activation function, and the ReLU function is used; G1 is the graph data, A is the adjacency matrix corresponding to the graph data, is the normalized adjacency matrix; D is the degree matrix, and D is a diagonal matrix, and the values of the elements on its diagonal are the number of adjacent nodes of the corresponding nodes; then perform time feature extraction through the LSTM layer:

[0094] H1 = LSTM(Emb1, Q1);

[0095] Among them, Q1 represents the learnable weight matrix of the first LSTM layer, H1 represents the intermediate feature after time feature embedding, and LSTM(·) represents the LSTM layer processing function;

[0096] For the second graph convolutional layer, its input is the graph data G2 at the second time point and the output feature of the first LSTM layer, then this process is:

[0097]

[0098] Among them, W2 represents the learnable weight matrix of the second graph convolutional layer, and Full(·) represents the full connection layer processing function, whose purpose is to improve the time feature expression ability and standardize the data length of H1;

[0099] Therefore, based on this process, all global embedding features that integrate moment information can be obtained, namely: Emb1, Emb2,..., Emb 12 .

[0100] S4. Based on the spatio-temporal feature extraction module constructed in S3, a complete electric vehicle battery fault diagnosis model is designed by combining the multi-head attention mechanism. This model takes the temporal feature information of each battery block in the battery module as input and predicts the state type of the entire battery pack;

[0101] The battery fault diagnosis model of this embodiment is based on the spatio-temporal feature extraction module in S3 and combines the multi-head attention mechanism (Multi-HeadAttention) to construct a complete electric vehicle battery fault diagnosis model. The core task of this model is to use the temporal feature information of each battery block in the battery module to predict and identify the fault state of the entire battery pack. The overall model structure framework is as Figure 5 shown. Specifically, it includes:

[0102] S4-1. First, this model transforms the temporal input feature data X m into graph data G through the feature embedding method based on graph structure representation designed in S2 m ; then uses the spatio-temporal feature extraction module designed in S3 to obtain the global embedding feature Emb = [Emb1, Emb2,..., Emb 12 ;

[0103] S4-2. Then, the multi-head attention mechanism is used to capture the temporal dependencies and spatial relationships in the spatio-temporal feature data of the battery blocks in parallel, so as to model the global state of the battery pack from multiple perspectives. And through the weighted fusion of the relationships between different time points and battery blocks, the expression ability and accuracy of the battery fault diagnosis model are improved.

[0104] Specifically, it includes:

[0105] Construct multi-head Query, Key, and Value matrices through linear transformation of Emb:

[0106]

[0107] Among them, Q l , K l and V l respectively represent the Query, Key, and Value matrices of the l-th attention head; and represent the weight matrices of the l-th attention head; there are a total of L attention heads, then l ∈ [1, L]; then, the weighted sum of the attention outputs of each attention head:

[0108]

[0109] Among them, d is the normalization factor; finally, calculate the final multi-head attention output MultiHd:

[0110] MultiHd = Concat(Hd1, Hd2,..., Hd L )·W O ;

[0111] Among them, Concat(·) is the channel concatenation operation, and W O is the linear transformation weight matrix;

[0112] S4-3. Extract the complex spatio-temporal relationship between battery blocks through the fully connected layer and residual connection, and reduce the risk of gradient disappearance, so as to improve the accuracy and training efficiency of battery fault diagnosis; that is:

[0113] FM = Concat(Full(MultiHd), MultiHd);

[0114] Among them, Full(·) represents the fully connected layer processing function; FM represents the feature data after being processed by the fully connected layer and residual connection; and the Dropout layer is used to prevent overfitting during model training to obtain the output feature DF:

[0115] DF = Dropout(FM);

[0116] Among them, Dropout(·) represents the Dropout layer function, and the Dropout rate is set to 0.3.

[0117] S4-4. Finally, use the Softmax layer as the output layer, which is responsible for converting the final output feature DF of the model into a probability value that can be used for classification decision-making, that is:

[0118] Y' m = Softmax(FM);

[0119] Among them, Y' m is the final prediction result vector; and the category with the highest corresponding probability is the final result of the model's fault diagnosis.

[0120] S5. Design the loss function and training method of the battery fault diagnosis model, and design a fault diagnosis confidence evaluation method to evaluate the credibility of the diagnosis results; finally, deploy the trained battery fault diagnosis model to the electric vehicle BMS platform to achieve real-time battery fault diagnosis and confidence evaluation.

[0121] To achieve effective training of the model, a loss function for battery fault diagnosis is designed in this embodiment; and a confidence evaluation method is designed for the diagnosis results of the battery fault diagnosis model to ensure that users can make more accurate judgments based on the output results of the battery fault diagnosis model. The specific steps are as follows:

[0122] S5-1. Design a loss function. Since battery fault diagnosis is a multi-class classification problem, this embodiment selects the cross-entropy loss function to calculate the difference between the model prediction and the true label, that is:

[0123]

[0124] where Y m and Y' m represent the true class label and the model output result vector respectively; L m represents the calculation result of the loss function for the mth group of data, and cls is the total number of classes;

[0125] S5-2. Use the Adam optimizer to optimize the cross-entropy loss function to minimize the difference between the prediction result and the true label, and update the network parameters through backpropagation. Optimize the performance of the model through multiple iterations until the loss function tends to be the minimum and the model converges, and finally obtain the trained battery fault diagnosis model;

[0126] S5-3. Design a confidence evaluation method for the output results of the battery fault diagnosis model, and use the test set to perform multiple diagnoses to obtain its diagnosis results {(x1, y'1), (x2, y'2),..., (x n , y' n )}, where x i is the input feature information, y' i is the model output result, and the corresponding label is y i , and calculate the accuracy of each diagnosis type:

[0127]

[0128] where Acc c represents the diagnosis accuracy of class c, and the total number of classes is cls, that is: c ∈ [1, cls]; Funz(·) is an indicator function, indicating whether the output is consistent with the actual label;

[0129] After that, calculate the mean μ c and variance of all prediction probabilities under fault type c to reflect its confidence distribution:

[0130]

[0131] Among them, P(y' i =c|x i ) represents the probability that the output result is category c on the premise that the input is x i , and n c is the number of samples of fault type c; then a confidence adjustment strategy is established, and the adjusted confidence score AdCon is obtained:

[0132] AdCon = α·P(y' i =c|x i )+(1-α)·μ c ;

[0133] Among them, α is the adjustment coefficient, and its value is dynamically adjusted according to the output accuracy of different categories. If the accuracy is high, α is increased, and vice versa, it is decreased;

[0134] Finally, for each category, the confidence threshold Th c :

[0135] Th c =μ c +k·σ c ;

[0136] Among them, k is a constant that controls the looseness of the threshold; therefore, for the output result of the battery fault diagnosis model, the confidence score AdCon of this result is calculated to reflect the credibility of this diagnosis result, and it is compared with the threshold Th c of this category. If AdCon is greater than Th c , the diagnosis result and the corresponding confidence score are returned. If AdCon is less than Th c , "unable to judge" is output, and it can be further confirmed through manual intervention.

[0137] S5-4. Finally, deploy the trained battery fault diagnosis model to the electric vehicle BMS platform to achieve real-time battery fault diagnosis and confidence evaluation. The specific steps include:

[0138] (1) Based on the BMS, collect the data information of each battery block in the battery module of the electric vehicle in real time, and use the method of S1 to obtain the time series data Text in ;

[0139] (2) Use the graph structure representation method in S2 to represent the feature data of each battery block through graph structuring, obtain the corresponding graph data G in , and input it into the battery fault diagnosis model, and finally output the diagnosis result;

[0140] (3) According to the confidence evaluation method in S5-3, evaluate the credibility of the battery fault diagnosis result. If the confidence of the diagnosis result is higher than the preset threshold, output the fault type and its confidence score; if it is lower than the threshold, prompt "unable to judge" and allow for manual intervention confirmation to ensure the accuracy and safety of the system.

[0141] (4) Repeat steps (1), (2), and (3) to achieve real-time monitoring and fault prediction of the battery state information, ensuring that the electric vehicle can continuously monitor the battery pack during driving and issue warnings in a timely manner when abnormalities occur, guaranteeing the safety and reliability of the battery pack.

[0142] The above is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A battery fault diagnosis method based on graph structure and LSTM, characterized in that Including: Based on the battery management system, obtain the timing feature information of each battery block in the battery module of the electric vehicle and the state of the battery pack, and construct a timing data set based on the timing feature information of each battery block and the state of the battery pack; Construct graph structure data based on the hardware connection relationship of the battery blocks in the battery module; Construct a spatio-temporal feature extraction module for battery block information based on the graph structure data and the long short-term memory network model; Construct a calculation model based on the spatio-temporal feature extraction module and the multi-head attention mechanism, and train the calculation model through the timing data set to obtain an electric vehicle battery fault diagnosis model; Deploy the electric vehicle battery fault diagnosis model to the electric vehicle BMS platform to diagnose battery faults.

2. The method according to claim 1, wherein The timing feature information of each battery block includes battery voltage, current, temperature, battery health state, battery charge state, internal resistance, and charge cycle.

3. The method according to claim 1, wherein The process of constructing a timing data set based on the timing feature information of each battery block and the state of the battery pack includes: Pre-define the feature information of each battery block in the battery module, collect the feature information in each battery block, and obtain the complete input feature data of the battery pack; Perform battery pack state annotation on the complete input feature data of the battery pack, and integrate the complete input feature data of the battery pack and the annotated battery pack state to obtain a timing data set for battery fault diagnosis.

4. The method according to claim 1, wherein The process of constructing graph structure data based on the hardware connection relationship of the battery blocks in the battery module includes: Take each battery block as a node in the graph, and take the connection relationship between the battery blocks as the edge between the nodes to construct a graph data structure including spatial information; Take the feature information of each battery block at a single moment as the feature of the corresponding node to construct graph structure data; Among them, each group of battery pack input feature data should be graph data of 12 time points, and the expression is: G m = [G1, G2,..., G 12 = Graph(X m ); Among them, G m represents the graph data corresponding to the input feature data obtained by embedding based on the graph structure representation method, X m is the input feature data, G1 is the graph data at the first time point, G2 is the graph data at the second time point, and G 12 is the graph data at the twelfth time point, and Graph(·) is the graph representation and feature embedding operation.

5. The method according to claim 1, wherein The process of constructing a spatio-temporal feature extraction module for battery block information based on the graph structure data and the long short-term memory network model includes: Construct a spatio-temporal feature extraction module for battery block information based on the graph neural network and the long short-term memory network model; Among them, the graph neural network performs spatio-temporal feature extraction on the graph structure data to obtain global embedding features; The long short-term memory network performs time feature extraction on the global embedding features to obtain intermediate features after time feature embedding.

6. The method according to claim 5, wherein In the graph neural network, the expression for feature extraction by the first graph convolutional layer is: Among them, W1 represents the learnable weight matrix of the first graph convolutional layer; σ(·) is the activation function, G1 is the graph data, A is the adjacency matrix corresponding to the graph data, is the standardized adjacency matrix; D is the degree matrix, and D is a diagonal matrix, the values of the elements on its diagonal are the number of adjacent nodes of the corresponding nodes, and Emb1 is the global embedding feature obtained through the graph convolutional layer; The expression for obtaining the intermediate features after time feature embedding is: H1 = LSTM(Emb1, Q1); Among them, Q1 represents the learnable weight matrix of the first LSTM layer, H1 represents the intermediate features after time feature embedding, and LSTM(·) represents the LSTM layer processing function.

7. The method according to claim 6, wherein It also includes that in the graph neural network, the expression for feature extraction by the second graph convolutional layer is: Among them, W2 represents the learnable weight matrix of the second graph convolutional layer, and Full(·) represents the full connection layer processing function.

8. The method according to claim 1, wherein The process of training the calculation model through the timing data set to obtain an electric vehicle battery fault diagnosis model includes: The cross - entropy loss function is optimized using the Adam optimizer, and the network parameters are updated through backpropagation for multiple iterations to optimize the calculation model until the cross - entropy loss function tends to be minimized and the model converges and then outputs, obtaining the electric vehicle battery fault diagnosis model.

9. The method according to claim 1, characterized in that The process of diagnosing battery faults based on the electric vehicle battery fault diagnosis model includes: Obtaining global embedding features containing spatio - temporal information based on the spatio - temporal feature extraction module; Capturing the temporal dependencies and spatial relationships in the global embedding features of the spatio - temporal information based on the multi - head attention mechanism to obtain an attention output; Performing weighted fusion on the attention output to obtain the final multi - head attention output; Calculating the processed feature data through a fully - connected layer and a residual connection on the final multi - head attention output; Calculating the output features through a Dropout layer on the processed feature data; Using the Softmax layer as the output layer to convert the output features into probability values of classification decisions to obtain the final prediction result.

10. The method according to claim 9, wherein After obtaining the final prediction result, it further includes: Constructing a confidence evaluation method for the output result of the battery fault diagnosis model, evaluating the final prediction result based on the confidence evaluation method to obtain the accuracy rate of each diagnosis type; Calculating the mean and variance of all prediction probabilities under the fault type based on the accuracy rate of each diagnosis type, and obtaining the confidence score and confidence threshold based on the mean and variance; If the confidence score is greater than the confidence threshold, output the diagnosis result and the corresponding confidence score. If the confidence score is less than the confidence threshold, perform manual intervention.

Images