Battery fault prediction model training method and device, prediction method and device

Abstract

The present application relates to a battery fault prediction model training method and device, a prediction method and device, the training method includes obtaining sample data, the sample data includes time domain signals of at least two battery variables, the battery variables include current, voltage or temperature; based on the variable attention mechanism, the interaction characteristics between the time domain signals are processed to obtain a first feature vector; based on the time attention mechanism, the timing characteristics between the time domain signals are processed to obtain a second feature vector: according to the first feature vector and the second feature vector, a cross-dimensional fusion feature vector is obtained by fusion; based on the cross-dimensional fusion feature vector and the fault category label annotated by the time domain signal, the confidence network is iteratively trained to obtain the battery fault prediction model. The model obtained by the training method of the present application can improve the problem of low accuracy and precision in battery fault prediction in the prior art.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a battery fault prediction model training method and device, and a prediction method and device. Background Art

[0002] The battery pack is an important energy source for new energy vehicles, and it plays a vital role in new energy vehicles. If the battery fails, it will affect the normal driving of the car and even endanger the lives of the people in the car. Therefore, predicting battery failure in advance and taking corresponding measures to solve it can effectively ensure the safety of the battery and the efficient and reliable operation of the car.

[0003] Traditional fault diagnosis methods for new energy vehicle batteries mainly include charging status detection and temperature monitoring analysis based on physical principles. Since the battery system of new energy vehicles has highly nonlinear and time-varying characteristics, traditional mathematical models are also difficult to effectively analyze the current and temperature signals of the battery system. In recent years, with the development and application of artificial intelligence technology, machine learning algorithms have begun to be used to analyze battery data and establish a battery health index assessment model to achieve fault prediction. However, most existing models are based on single-dimensional data for prediction, and do not fully utilize the dimensional characteristics of the data, resulting in low fault prediction precision and accuracy. Summary of the invention

[0004] Based on this, a battery fault prediction model training method, prediction method and device are provided to improve the problem of low battery fault prediction precision and accuracy in the prior art.

[0005] On the one hand, a battery failure prediction model training method is provided, comprising:

[0006] Acquire sample data, the sample data comprising time domain signals of at least two battery variables, the battery variables comprising current, voltage or temperature, wherein at least part of the sample data is marked with a fault category label;

[0007] Processing the interaction features between the time domain signals based on a variable attention mechanism to obtain a first feature vector;

[0008] Based on the time attention mechanism, the temporal features between the time domain signals are processed to obtain the second feature vector:

[0009] According to the first feature vector and the second feature vector, fusing to obtain a cross-dimensional fused feature vector;

[0010] Based on the cross-dimensional fusion feature vector and the fault category label annotated by the time domain signal, the belief network is iteratively trained to obtain the battery fault prediction model.

[0011] In one embodiment, the fusing to obtain a cross-dimensional fused feature vector includes:

[0012] According to the first feature vector and the second feature vector, high-dimensional features are mined to obtain variable dimension space features and time dimension space features respectively;

[0013] Extracting the interactive information of the variable dimension spatial features and the time dimension spatial features based on the attention mechanism to obtain interactive features; wherein the interactive features include the features of the time dimension spatial features after being processed by the attention mechanism, and the features of the time dimension spatial features after being spliced, guided to the variable dimension and then processed by the attention mechanism;

[0014] According to the interactive features, feature fusion is performed based on a cross-attention mechanism to obtain the cross-dimensional fused feature vector.

[0015] In one embodiment, the obtaining of sample data further includes:

[0016] Performing time domain division according to the time domain signal of the battery variable to obtain time domain signals of each battery variable in different time domains;

[0017] The fusion obtains a cross-dimensional fusion feature vector, including:

[0018] Fusing the first eigenvector and the second eigenvector corresponding to the time domain signals of each time domain to obtain a cross-dimensional fused eigenvector corresponding to each time domain;

[0019] After the fusion obtains the cross-dimensional fusion feature vector, the method further includes:

[0020] The cross-dimensional fusion feature vectors corresponding to each time domain are linearly fused to obtain a cross-domain time-varying fusion feature vector, so as to train the belief network based on the cross-domain time-varying fusion feature vector.

[0021] In one embodiment, the iterative training of the belief network includes:

[0022] Obtain the position matrix of the time domain signal sensor;

[0023] Processing the position matrix based on a multi-head attention mechanism to obtain a global interaction matrix;

[0024] According to the global interaction matrix and the cross-dimensional fusion feature vector, or according to the global interaction matrix and the cross-domain time-varying fusion feature vector, a graph convolution vector is obtained based on a graph convolution algorithm to iteratively train the belief network based on the graph convolution vector.

[0025] In one embodiment, the iterative training of the belief network further includes:

[0026] Extracting time series features from the graph convolution vector based on one-dimensional convolution to obtain a time series feature vector;

[0027] Using the time series feature vector as an input of a restricted Boltzmann machine, and obtaining a prediction loss according to an output of the restricted Boltzmann machine;

[0028] The battery failure prediction model is obtained by performing iterative training based on the prediction loss.

[0029] In one embodiment, the iterative training of the belief network further includes:

[0030] Training the belief network based on semi-supervised cycle loss to obtain the battery failure prediction model;

[0031] The semi-supervised cycle loss includes an output loss and a distance loss. The output loss is determined according to the predicted value and the target value of the sample data with the fault category label, and the distance loss is determined according to the difference between the predicted values ​​of the sample data and the similarity between samples.

[0032] In one embodiment, the training of the belief network based on a semi-supervised cycle loss comprises:

[0033] The belief network is trained according to the semi-supervised cyclic loss function expressed as follows:

[0034]

[0035] Among them, L semicircle is the semi-supervised cycle loss, f(x i ) is the i-th sample data x i The output after the belief network, y is the target value, For sample data x i and x j The regularization term of the distance in the sample space, For sample data x i and x j The regularization term of the distance between the model output value, s ij For sample data x i and x j The similarity weight between is an adjustable parameter, L is the number of sample data with the fault category label, U is the number of sample data without the fault category label, L s is the Laplace matrix of the adjacency matrix composed of similarity weights, rank represents the rank of the matrix, sts ij ≥0 means the constraint is s ij≥0, c represents the number of connected components of the connected graph constructed by the adjacency matrix.

[0036] On the other hand, a battery failure prediction method is provided, comprising:

[0037] Acquire time domain data of at least two battery variables, wherein the battery variables include current, voltage or temperature;

[0038] Inputting the time domain data into a battery fault prediction model to obtain a fault prediction result;

[0039] The battery failure prediction model is trained according to the battery failure prediction model training method.

[0040] In another aspect, a battery failure prediction model training device is provided, comprising:

[0041] A training data acquisition module, used to acquire sample data, wherein the sample data includes time domain signals of at least two battery variables, wherein the battery variables include at least two of current, voltage or temperature, wherein at least part of the sample data is marked with a fault category label;

[0042] The model training module is used to process the interaction features between the time domain signals based on the variable attention mechanism to obtain a first feature vector; process the timing features between the time domain signals based on the time attention mechanism to obtain a second feature vector: according to the first feature vector and the second feature vector, a cross-dimensional fused feature vector is obtained by fusing; based on the cross-dimensional fused feature vector and the fault category label annotated by the time domain signal, a confidence network is iteratively trained to obtain the battery fault prediction model.

[0043] A battery failure prediction device is also provided, comprising:

[0044] A battery data acquisition module, used to acquire time domain data of at least two battery variables, wherein the battery variables include at least two of current, voltage or temperature;

[0045] A fault prediction module is used to input the time domain data into a battery fault prediction model to obtain a fault prediction result; the battery fault prediction model is trained according to the battery fault prediction model training method.

[0046] The above-mentioned battery fault prediction model training method obtains sample data with multiple variables. On the one hand, it adopts the variable attention mechanism to mine the interactive information between signals, and on the other hand, it adopts the time attention mechanism to mine the timing characteristics of signals between different moments and perform cross-dimensional fusion. It effectively utilizes the time-varying characteristics of sample data, enhances the exchange of information of different dimensions, fuses features of different dimensions, eliminates redundant information between features of different dimensions, and improves the precision and accuracy of model fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a structural block diagram of a trust network in one embodiment;

[0048] Figure 2 A schematic diagram of a flow chart of a battery failure prediction model training method in one embodiment;

[0049] Figure 3 is a structural block diagram of a time-varying fusion encoder in an embodiment;

[0050] Figure 4 It is a flowchart of a cross-dimensional fusion step in one embodiment;

[0051] Figure 5 It is a structural block diagram of a battery fault prediction model training device in one embodiment;

[0052] Figure 6 is an internal structure diagram of a computer device in one embodiment;

[0053] Figure 7 A schematic diagram of a process for predicting battery failure in one embodiment;

[0054] Figure 8 is a structural block diagram of a battery failure prediction device in one embodiment;

[0055] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in another embodiment. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0057] As the core component of electric vehicles, power batteries play a vital role. Due to the use environment and complexity of power batteries, their failure rate is relatively high.

[0058] In the existing battery fault prediction technology, it is still based on data analysis. The historical operation of the car battery is analyzed and the impact of the car battery parameters is judged through historical operation analysis. However, the battery system of new energy vehicles has highly nonlinear and time-varying characteristics. Therefore, traditional mathematical models are actually difficult to mine the characteristic information in the data and effectively analyze the current and temperature signals of the battery system.

[0059] Machine Learning (ML) aims to enable computers to learn and improve performance through data without explicit programming instructions. Machine learning technology allows computers to discover patterns, laws and trends by learning and analyzing large amounts of data, and make predictions and decisions based on these findings.

[0060] Machine learning techniques are usually divided into three main types: supervised learning, unsupervised learning, and reinforcement learning. In the supervised learning training process, the model learns the mapping relationship between input and output by matching the input data with the corresponding labels. Unsupervised learning means that during the training process, the model can only learn based on the statistical characteristics of the input data without the assistance of label information. Reinforcement learning means that the model learns through interaction with the environment, and maximizes the cumulative reward by trying different actions in the process of trial and error.

[0061] Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and other technologies.

[0062] This application is based on a belief network model of cross-domain time-varying coding, hereinafter referred to as a belief network, and provides a battery fault prediction model training method to effectively mine the time-varying and spatial characteristics of battery data, thereby improving the accuracy of predicting new energy vehicle battery failures.

[0063] Belief network Figure 1 As shown in the figure, it includes a time-varying fusion module, a cross-domain fusion spatiotemporal signal topology map and a flexible graph confidence module. Time and variable attention can mine the timing features between different signals and the correlation features between signals. Flexible graph convolution can adaptively capture the interactive features of graph signals and analyze the topological relationship between different nodes. The original signal sample can obtain a vector containing time-varying features after time and variable attention. The fusion features are mined through the time-varying fusion encoder in the time-varying fusion module, and the sample feature fusion is realized in the vector space. The fused feature vector is combined with the sensor position relationship to form a cross-domain fusion spatiotemporal signal topology map, and then the map is imported into the flexible graph confidence module. The flexible graph convolution is used to capture the graph topology features, and the convolutional Boltzmann machine is used to process high-dimensional data. The potential features in the data are learned and the model parameters are adaptively adjusted to improve the performance of the model.

[0064] In one embodiment, if Figure 2 As shown, the battery fault prediction model training method includes at least the following steps:

[0065] Step 100: Acquire sample data, where the sample data includes time domain signals of at least two battery variables.

[0066] In the actual implementation process, sensors are used to collect sample data of new energy vehicle battery packs. The sample data may include multiple variables of the battery, such as current, voltage, and temperature. In addition, the collected current, voltage, and temperature signals may be divided into multiple categories according to different time domains. For example, the signals are divided into three categories of data according to the minute, hour, and day domains, one part of which is labeled and the other part is not labeled. All of this data is imported into the confidence network, and the network parameters are updated through iterative training to achieve accurate prediction of new energy battery failures.

[0067] For example, the sampling frequency of the signal is q, which means that the signal is sampled q times in 1s, the current time is t0, and the sampling time step size is T p , which is 30 in this embodiment. The current, voltage and temperature signals are divided into three domains: component sample set, hourly sample set and daily sample set along the time axis, and the domains are respectively represented by χ m , χ h ,x d Indicates that the time lengths for collecting the three types of sample data are respectively represented by T m , T h , T d In this embodiment, T m =60, T h =3600, T d =86400. For example, the mathematical expression of the sample set composed of sample data is as follows:

[0068] Sample set:

[0069]

[0070] in, Represents the data of all signals collected by the sensor at t0, Indicates that the sub-samples are concentrated from t0-T m The first sampling data starts at the moment, N represents the number of signals collected, for example, three sensors collect current signals, three sensors collect voltage signals, and four sensors collect temperature signals, then N is 10, F represents the length of the signal intercepted at this moment, in this embodiment, F is set to 100, indicating that the data collected by the sensors are taken 50 times before and after the sampling point, Represents the field of real numbers.

[0071] Similarly, define the hourly sample set χ h :

[0072]

[0073] in:

[0074]

[0075]

[0076]

[0077] Define the daily sample set χ d :

[0078]

[0079] in:

[0080]

[0081]

[0082]

[0083] Each item in the sample set can be understood as the data of all signals collected by the sensor at a certain moment.

[0084] Step 200, feature mining, includes mining the interactive features between the time domain signals based on a variable attention mechanism to obtain a first feature vector; and mining the timing features between the time domain signals based on a time attention mechanism to obtain a second feature vector.

[0085] The attention mechanism in deep learning allows a neural network or a belief network to focus on relevant parts when processing input data. By introducing the attention mechanism, the neural network or belief network can automatically learn and selectively focus on important information in the input, improving the performance and generalization ability of the model.

[0086] The temporal attention mechanism and the variable attention mechanism can mine the timing features between different signals and the correlation features between signals. The mining process is illustrated using the hourly sample set as an example.

[0087] The process of mining the interaction features between signals by the variable attention mechanism is as follows:

[0088]

[0089]

[0090] in, Represents the data of the hourly sample set input by the r-th layer variable attention mechanism, represents each item in the hourly sample set. The specific number of layers of r should be set in combination with specific data during model training. r-1 Indicates the number of channels of the input data of the rth layer. When r=1, C0=F. T r-1Indicates the length of the time dimension of the rth layer. When r = 1, T0 = T in the hourly sample set h , T0=T in the daily sample set d , T0=T in the sample set m .

[0091] are all weight matrices that can be learned during model training. σ represents the activation function, and S represents the adaptively changeable attention matrix. In S, the element S i,j represents the correlation between signal (i.e., battery variable) i and signal j, S' i,j The Softmax function is used to normalize the weight vectors of each node, so that all output values ​​are between 0 and 1 and their sum is 1, thereby speeding up the model fitting. The first eigenvector obtained by the variable attention mechanism is:

[0092]

[0093] Where S' is S' i,j The vector composed of Sample set The first feature vector after the variable attention mechanism

[0094] In the time dimension, there are also correlations between different moments of the same signal and different moments of different signals. This embodiment uses a time attention mechanism to obtain time series features. The process is as follows:

[0095]

[0096]

[0097] in, It is the weight parameter matrix that the model can learn during the training process. The element E in the time series relationship matrix E i,j Indicates the dependency strength between time i and time j, element E' i,j In order to use the softmax function to regularize E and thus speed up the model fitting, the hourly sample set The second eigenvector obtained by the temporal attention mechanism is:

[0098]

[0099] E' is the element E' i,j Similarly, we can get the daily sample set Sample set The second feature vector obtained by the temporal attention mechanism

[0100] Step 300: Obtain a cross-dimensional fusion feature vector according to the first feature vector and the second feature vector.

[0101] In the actual implementation of this embodiment, the feature vector obtained by the hourly sample set through time attention and variable attention is The time-varying fusion encoder is introduced to fuse the time features and variable features. The structure of the time-varying fusion encoder is as follows: Figure 3 As shown in the figure, it includes BERT (Bidirectional Encoder Representations from Transformer) layer, guided communication module and cross-dimensional fusion module. The BERT layer is used to mine high-dimensional features in the time dimension and variable dimension space. The guided communication module can reduce the heterogeneity of information in different dimensions and enhance the communication of information in different dimensions. The cross-dimensional fusion module can fuse features of different dimensions, eliminate redundant information between features of different dimensions, and improve the accuracy of the model. The feature fusion process is as follows: Figure 4 As shown, the following steps are included:

[0102] Step 301: Obtain variable dimension spatial features and time dimension spatial features based on the BERT model according to the first feature vector and the second feature vector.

[0103] Methods for mining high-dimensional features include graph neural networks, autoencoders, etc. This embodiment uses a BERT model consisting of three BERT layers to mine high-dimensional features, and the inputs of time space and variable space are The process of mining deep feature information in the time dimension space through the BERT model can be expressed as:

[0104]

[0105] The process of mining deep feature information in variable dimension space through the BERT model can be expressed as:

[0106]

[0107] Among them, S pos represents the time dimension feature position embedding matrix, T pos Represents the variable dimension feature position embedding matrix;

[0108] It represents the hidden feature state of the time dimension space output of the l-th BERT layer, that is, the time dimension space feature output by the l-th BERT layer. This is the input of the first layer;

[0109] It represents the hidden feature state of the variable dimension space output of the l-th BERT layer, that is, the variable dimension space feature output by the l-th BERT layer. This is the input of the first layer;

[0110] Since this embodiment uses three BERT layers, l takes values ​​of 1, 2, and 3, and LN represents a regularization operation. is an intermediate value, FFN represents the operation of the forward feedback layer, and MHA represents the multi-head attention mechanism processing.

[0111] Among them, the multi-head attention mechanism processes the example data x according to the following process:

[0112]

[0113] Among them, d represents the dimension of hidden embedding, h represents the number of heads of the multi-head attention mechanism, which is set according to the specific data during model training, and W o , is the matrix to be learned in the multi-head attention mechanism, head i Each attention head in the multi-head attention mechanism.

[0114] FFN represents the calculation process of the forward feedback layer, which consists of two layers of linear transformations with Relu as the activation function. The calculation process of FFN for the example data x is as follows:

[0115] FFN(x)=Relu(xW F1 +b F1 )W F2 +b F2

[0116] Among them, W F1 , W F2 is the linear transformation matrix, b F1 、b F2 is the bias matrix, and Relu represents the activation function.

[0117] Step 302: extract the interactive information of the high-dimensional features based on the attention mechanism to obtain interactive features.

[0118] In this embodiment, the time dimension spatial features obtained after mining through the three layers of BERT layer Variable Dimension Space Features The guided communication module is introduced to capture the interactive information between different modal features. The process can be expressed as follows mathematically:

[0119]

[0120] Among them, Attn() represents the operation of the attention mechanism, which is performed according to the existing method. It represents the variable interaction features output by the attention mechanism after the high-dimensional features of the variables are processed. is the operation matrix of the attention mechanism for the high-dimensional features of the variable, determined through model training, The high-dimensional features representing time are output as time interaction features after the attention mechanism. is the operation matrix of the attention mechanism for temporal high-dimensional features, Both represent the concatenation operation of feature vectors, x s express data, x t express The data in .

[0121] When processing high-dimensional features, the high-dimensional feature vectors of the time dimension are guided to the variable dimension space through the splicing operation, reducing the heterogeneity of information in different dimensions and enhancing the communication of information in different dimensions.

[0122] Step 303: Perform feature fusion according to the interactive features to obtain the cross-dimensional fusion feature vector.

[0123] The interactive features obtained by the guided communication module Import the cross-dimensional fusion module to perform feature fusion, and the process can be expressed as:

[0124]

[0125]

[0126] Among them, Cross-attention() represents the processing process of the cross-attention mechanism.

[0127] These are all operation matrices that can be learned during the training process.

[0128] Then put Import the previous feedback layer and Add&Norm (residual connection and layer normalization operation) layer respectively to perform feature fusion and finally obtain the feature vector ζ h The specific process will not be described here.

[0129] After obtaining the fusion features based on the above steps, further model training is performed.

[0130] Step 400: Perform iterative training based on the cross-dimensional fusion feature vector and the fault category label to obtain the battery fault prediction model.

[0131] Generally, the iterative training of a neural network or a belief network includes steps such as data input, calculation of prediction loss, model evaluation and tuning based on prediction loss, etc. In model selection and design, a suitable model architecture is selected according to the nature of the problem and the characteristics of the data, such as a deep neural network, a decision tree, etc. The design of the model includes defining the network structure, selecting the activation function, setting the hyperparameters, etc. This embodiment adopts the following method: Figure 1 The belief network shown.

[0132] In one implementation, an hourly sample set is used as data input to perform feature mining to obtain a cross-dimensional fused feature vector, and a confidence network is further trained based on the cross-dimensional fused feature vector to obtain the required battery fault prediction model.

[0133] In another embodiment, cross-domain fusion is further performed according to the features of domains of different time lengths, and training is performed based on the fused feature vectors. Specifically:

[0134] Based on the same method, the cross-dimensional fusion feature vector ζ of the daily sample set and the sub-sample set after the cross-dimensional fusion module can be obtained d , m .

[0135] Will h , d , m By importing the linear layer, the cross-domain time-varying fusion feature vector ζ can be obtained, and the cross-domain time-varying fusion feature vector ζ is used for subsequent model training. The operation of the linear layer refers to the existing method and will not be repeated here.

[0136] Based on the cross-domain time-varying fusion feature vector ζ, a machine learning algorithm is used for subsequent training to obtain a battery fault prediction model. Compared with the traditional machine learning that requires manual feature extraction, this application divides the original data set into minute, hour, and day domains to generate samples in different domains, which can achieve data enhancement effects. At the same time, the time attention and variable attention are fused, and a time-varying fusion encoder is constructed to mine fusion features, realize the fusion of minute, hour, and day domain sample features, and mine the time-varying and spatial features of the signal.

[0137] In this embodiment, the semi-supervised training method is adopted to reduce the data labeling work, divide the original data set into minute, hour, and day domains, generate samples in different domains, and achieve data enhancement effect, thereby effectively solving the problem of scarce new energy battery failure data set and insufficient sample size. The time-varying characteristics of the signal are mined by combining time attention and attention between variables. A time-varying fusion encoder is constructed to mine the high-dimensional features of the time dimension space and the variable dimension space, reduce the heterogeneity of information of different dimensions, enhance the communication of information of different dimensions, fuse features of different dimensions, eliminate redundant information between features of different dimensions, and improve the accuracy and precision of model prediction of faults.

[0138] In some embodiments, features are continuously captured using a flexible graph convolution algorithm (FGCN).

[0139] FGCN consists of a graph convolution module (GCN) and a multi-head attention module. The multi-head attention module takes into account the different degrees of influence of adjacent nodes and optimizes the graph convolutional neural network. At the same time, it can also adaptively change the value of the adjacency matrix as the model is trained to obtain a global interaction matrix, eliminate inappropriate samples, and greatly improve the efficiency of the model.

[0140] In the graph convolution module, the position matrix is ​​obtained according to the position and distance between sensors, and the position matrix is ​​set as the adjacency matrix A of the graph convolution, and then A is used to construct the symmetric normalized Laplacian matrix of the graph

[0141]

[0142] Among them, the matrix Medium Element I N is the identity matrix, For the matrix The elements in , L is the Laplace matrix of the knowledge graph regularization, which is obtained from the adjacency matrix A: D is the degree matrix of the graph formed by the adjacency matrix.

[0143] Symmetric Normalized Laplacian Matrix is a symmetric matrix, so Perform eigendecomposition (spectral decomposition):

[0144]

[0145] in, is a column vector belonging to the node vector space H, that is, In this embodiment, U is the global interaction matrix output by the multi-head attention module, which will be described in detail later. is the eigenvalue matrix of L.

[0146] Let the input feature of the i-th layer graph convolution be X gcn_i , (where X gcn_1 =ζ), the output feature is X gcn_i+1 Where X gcn_i ∈R n×k , n is the number of graph nodes, k is the dimension of the point-wise space vector. At the same time, a filter is designed for the graph structure:

[0147]

[0148] where h g As the filter function, the graph convolution can be expressed as:

[0149] g θ *X i =U0g θ U0 -1 X gcn_i

[0150] Perform a linear transformation on each channel of the node in the graph obtained after convolution. The linear transformation matrix is ​​W i ∈R k×k , get the output data Z i ,Right now:

[0151]

[0152] In order to reduce the amount of calculation, according to Chebyshev inequality, Z is approximately estimated i ,Right now:

[0153]

[0154] In the multi-head attention module, its initial input is the adjacency matrix A of the graph convolution module. The process of multi-head attention adaptively updating the adjacency matrix to obtain the global interaction matrix can be described as follows:

[0155]

[0156] Among them, A represents the adjacency matrix of the cross-domain fusion spatiotemporal signal topology graph, A (t) represents the t-th position matrix learned by the multi-head attention mechanism, and They respectively reflect the linear transformation matrices of the multi-head attention mechanism. Q and K represent the feature vectors of the input data after the query operation and key operation in the multi-head attention mechanism. The query operation and key operation are the conventional operations of the multi-head attention mechanism and will not be repeated here. m is the data dimension of the multi-head attention mechanism.

[0157] One of the inputs of the graph convolution module is the global interaction matrix after multi-head attention optimization, namely A (t) , because the multi-head attention mechanism will generate N M A global interaction matrix, forming N M There are different subgraph spaces, here N M Represents the number of attention heads in the multi-head attention module, so N is needed M Two parallel GCNs mine features from different subgraph spaces.

[0158] After N M After mining the graph topology features with graph convolution modules, N M The parallel pooling layer removes redundant information, compresses features, simplifies network complexity, and reduces computational complexity. The pooling process can be described as follows:

[0159] Z pool =∑W×Z i (t) +bias

[0160] Among them, W is the linear transformation matrix of the pooling process, Z i (t) is the output of the i-th GCN, and bias is the bias of the pooling process. Finally, a linear layer is used to fuse N M The linear transformation process of the features mined by the parallel pooling layer can be described as follows:

[0161] Z comb =W comb Z out +b comb

[0162] Among them, Z comb The output of the linear transformation N M The output of the parallel pooling layer, b comb is the bias vector of the linear transformation.

[0163] At the same time, N M The parallel global interaction matrices generate a new global interaction matrix by averaging. The process can be described as:

[0164]

[0165] The new global interaction matrix is ​​transmitted to the next layer of FGCN, and is mined again by the multi-head attention mechanism in the next layer of FGCN. The number of FGCN layers is set according to the actual data.

[0166] In this embodiment, a convolutional Boltzmann machine is used to process high-dimensional data, learn the potential features in the data, and adaptively adjust the model parameters to improve the performance of the model. The main structure of the convolutional Boltzmann machine includes a one-dimensional convolutional neural network and a restricted Boltzmann machine. The one-dimensional convolution extracts time series features by fusing the information of adjacent nodes. The process is as follows:

[0167] ζ a =ReLU(Φ*(ReLU(g θ *Z comb )))

[0168] Among them, * represents the standard convolution operation, g θ *Z comb Represents the one-dimensional convolution process, Φ represents the parameters of the convolution kernel, and ReLU represents the activation function.

[0169] The restricted Boltzmann machine mainly includes a visualization layer V and a hidden layer H. The restricted Boltzmann machine is a probabilistic model that can capture the probabilistic relationship between signal features and faults. Therefore, in this embodiment, the feature vector ζ a The restricted Boltzmann machine is introduced to further mine features and thus improve the accuracy of prediction.

[0170] The restricted Boltzmann machine is an existing model. In this embodiment, h s (·) represents the operation on the eigenvector, then the eigenvector ζ a , the output of the restricted Boltzmann machine can be expressed as h s (ζ a ), which is also the output of the entire model.

[0171] In some embodiments, due to the complexity of actual production process data, some data are difficult to label, but these data also have a strong correlation with faults. In order to utilize unlabeled data, a semi-supervised method is used to define the objective function, thereby effectively alleviating the problem of unlabeled data and improving the accuracy of model fault diagnosis.

[0172] The sample data set of this application is defined as X = {X L ,X U}, X L Represented as a set of sample data with fault labels, X U It is represented as a set of unlabeled sample data, and the fault category label is recorded as the set C = {1, 2, ..., c}. Then the objective function can be expressed as:

[0173]

[0174] Where f(x i ) represents the sample data x iThe output value after the entire model, y represents the target value, w ij represents the similarity weight between samples, which can be learned dynamically during the model training process. L represents the number of sample data with the fault category label, and U represents the number of sample data without the fault category label. The first term of the objective function represents the objective function of the labeled data, and the second term represents the objective function of the labeled and unlabeled data. It can ensure that the label information between adjacent samples is as similar as possible, so that the feature distance of samples of different fault categories in the vector space is farther, and the feature distance of samples of the same fault category in the vector space is closer.

[0175] In order to effectively utilize unlabeled samples, the present invention uses the similarity weight w between samples. ij Construct an adjacency matrix S s , where s ij =w ij , S s Each item ij Represents the similarity weight between two samples.

[0176] Calculate S s The Laplacian matrix of :

[0177] L s =D s -S s

[0178] Among them, D s Represents the matrix S s The dual matrix of the diagonal matrix formed by the sum of each row of s The number of eigenvalues ​​0 is equal to the adjacency matrix S s The number of connected components of the connected graph is:

[0179] rank(L s )=L+Uc

[0180] Among them, rank represents the rank of the matrix, c represents the adjacency matrix S s The number of connected components of the connected graph. Therefore, the second term of the objective function can be expressed as:

[0181]

[0182] Among them, s ij Initialized randomly at the beginning of model training and adaptively changed as the model is trained, c can be represented by the matrix S s Calculated, st represents the constraint condition, considering that the second term of the original formula is the sample data x i and x j The model output value f(x i)-f(x j ), and here ||x i -x j || 2 Only represents sample x i and x j The distance in the original sample space, so we introduce Make it more similar to the second term of the original objective function, Indicates adjustable parameters.

[0183] Define f i and f j Represents sample x i and x j After the entire model, the mapping vector in the high-dimensional space of the restricted Boltzmann machine can be expressed by the eigenvector ζ a After the restricted Boltzmann machine, it is obtained and d ij Represents sample x i and x j The distance in the original sample space, p ij represents f i and f j The distance in high-dimensional space is:

[0184] d ij =||x i -x j || 2 ;p ij =||p i -p j || 2

[0185] In order to make the model converge faster, the distance d ij and p ij Regularization gives:

[0186]

[0187] Among them, d i =[d i1 ,d i2 ,...,d in ] represents sample x i and the set of distances between other samples in the original sample space, p i =[p i1 ,p i2 ,...,p in ] represents sample x i The set of distances between the output values ​​of other samples after passing through the model, d ij and p ijCombined, this can better reflect the process of model mining features and is more conducive to model convergence, so the second term of the objective function can be further transformed into:

[0188]

[0189] The objective function uses labeled samples and unlabeled samples. It is a semi-supervised objective function. The similarity weight s between samples ij The model is dynamically cyclically changed according to model training, so the final loss function of the model is defined as a semi-supervised cyclic loss function, which can be expressed as:

[0190]

[0191] In the actual training process, the number of samples in each batch is set to 512, the number of iterations is set to 500, the learning rate is set to 0.01, and the loss function is defined as the semi-supervised cyclic loss function L semicircle , the Adam algorithm is used for training. The parameter adjustments of each attention mechanism, graph convolution model, and restricted Boltzmann machine in the actual training process will not be repeated here.

[0192] The training method provided in this application divides the original data set into minutes, hours, and days, generates samples in different domains, and achieves data enhancement effects, thereby effectively solving the problem of scarce new energy battery failure data sets and insufficient sample size. At the same time, the time attention and variable attention are integrated, and a time-varying fusion encoder is constructed to mine fusion features, realize the fusion of minute, hour, and day domain sample features, and mine the time-varying features and spatial features of the signal, thereby improving the accuracy and precision of the model prediction fault. In addition, the flexible graph convolution in the model can use the topological relationship of the graph signal to adaptively mine its spatiotemporal features, remove inappropriate samples, and greatly improve the efficiency of the model. The convolutional Boltzmann machine is used to process high-dimensional data, learn the potential features in the data, and adaptively adjust the model parameters. Finally, considering that some data are not labeled due to conditions, the present invention uses a semi-supervised method to define the objective function, thereby effectively alleviating the problem of unlabeled data and improving the accuracy of model fault diagnosis. By defining a cyclic loss function, the feature distance of samples of different fault categories in the vector space is farther, and the feature distance of samples of the same fault category in the vector space is closer, thereby improving the performance of the model.

[0193] It should be understood that although Figure 2 , Figure 4 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 , Figure 4 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0194] In one embodiment, Figure 5 As shown, a battery fault prediction model training device 1000 is provided, comprising: a training data acquisition module 1001 and a model training module 1002, wherein:

[0195] The training data acquisition module 1001 is used to acquire sample data, wherein the sample data includes time domain signals of at least two battery variables, wherein the battery variables include at least two of current, voltage or temperature, wherein at least part of the sample data is marked with a fault category label;

[0196] The model training module 1002 is used to process the interaction features between the time domain signals based on the variable attention mechanism to obtain a first feature vector; process the timing features between the time domain signals based on the time attention mechanism to obtain a second feature vector: according to the first feature vector and the second feature vector, a cross-dimensional fused feature vector is obtained by fusing; based on the cross-dimensional fused feature vector and the fault category label annotated by the time domain signal, the confidence network is iteratively trained to obtain the battery fault prediction model.

[0197] The training data acquisition module 1001 further performs time domain division according to the time domain signals of the variables to obtain the time domain signals of each variable at different time domains.

[0198] When fusing to obtain a cross-dimensional fused feature vector, the model training module 1002 fuses the first feature vector and the second feature vector corresponding to the time domain signal of each time domain to obtain a cross-dimensional fused feature vector corresponding to each time domain, and linearly fuses the cross-dimensional fused feature vector corresponding to each time domain to obtain a cross-domain time-varying fusion feature vector, so as to train the belief network based on the cross-domain time-varying fusion feature vector.

[0199] Specifically, in this embodiment, the confidence network includes a time-varying fusion module, a cross-domain fusion spatiotemporal signal topology map and a flexible graph confidence module. The time-varying fusion module can mine the timing features between different signals and the correlation features between signals based on time and variable attention, and perform feature fusion. The fused feature vector is combined with the sensor position relationship to form a cross-domain fusion spatiotemporal signal topology map, and then the map is imported into the flexible graph confidence module. The flexible graph convolution is used to capture the graph topology features, and the convolutional Boltzmann machine is used to process high-dimensional data. The potential features in the data are learned, and the model parameters are adaptively adjusted to improve the performance of the model.

[0200] In one embodiment, the first feature vector and the second feature vector are introduced into a time-varying fusion encoder to fuse time features and variable features. The time-varying fusion encoder includes a BERT layer, a guided communication module, and a cross-dimensional fusion module.

[0201] Exemplarily, multi-layer BERT layers are used to mine high-dimensional features to obtain variable dimension spatial features and time dimension spatial features respectively.

[0202] Extracting the interactive information of the variable dimension spatial feature and the time dimension spatial feature based on the attention mechanism to obtain the interactive feature, wherein the interactive feature includes the feature of the time dimension spatial feature after being processed by the attention mechanism, and the feature of the time dimension spatial feature after being guided to the variable dimension and processed by the attention mechanism;

[0203] According to the interactive features, feature fusion is performed based on a cross-attention mechanism to obtain the cross-dimensional fused feature vector.

[0204] When performing iterative training, the model training module 1002 also obtains the position matrix of the time domain signal sensor; uses the multi-head attention mechanism to process the position matrix to obtain a global interaction matrix; the multi-head attention mechanism will generate multiple global interaction matrices to form multiple different sub-graph spaces, and the global interaction matrix is ​​combined with the cross-domain time-varying fusion feature vector output by the time-varying fusion encoder, and is processed using multiple parallel graph convolution modules to obtain a graph convolution vector.

[0205] The graph convolution vector integrates the location information of the sensor, which helps to reduce the error in the prediction results caused by the different locations of the sensors.

[0206] Since there is a certain probabilistic relationship between signal features and faults in essence, in order to better capture this relationship, a convolutional Boltzmann machine module is designed to process high-dimensional data, learn the potential features in the data, and adaptively adjust the model parameters to improve the performance of the model.

[0207] Exemplarily, the model training module 1002 extracts timing features from the graph convolution vector based on one-dimensional convolution to obtain a timing feature vector; uses the timing feature vector as the input of a restricted Boltzmann machine, and obtains a predicted loss based on the output of the restricted Boltzmann machine; and performs iterative training based on the predicted loss to obtain the battery fault prediction model.

[0208] In one embodiment, the loss of the entire model adopts a semi-supervised loop loss, which includes an output loss and a distance loss. The output loss is determined according to the predicted value and the target value of the sample with the fault category label, and the distance loss is determined according to the difference between the predicted values ​​of the samples and the similarity between the samples.

[0209] Specifically, the mathematical expression of the semi-supervised cycle loss function is as follows:

[0210]

[0211] Among them, L semicircle is the semi-supervised cycle loss, f(x i ) is the i-th sample data x i The output after the belief network, y is the target value, For sample data x i and x j The regularization term of the distance in the sample space, For sample data x i and x j The regularization term of the distance between the model output value, s ij is the similarity weight between sample data, is an adjustable parameter, L is the number of sample data with the fault category label, U is the number of sample data without the fault category label, L s is the Laplacian matrix of the adjacency matrix composed of similarity weights, rank represents the rank of the matrix, and c represents the number of connected components of the connected graph composed of the adjacency matrix.

[0212] For the specific limitations of the battery fault prediction model training device 1000, please refer to the limitations of the battery fault prediction model training method above, which will not be repeated here. The various modules in the above-mentioned battery fault prediction model training device 1000 can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0213] In one embodiment, the battery fault prediction model training device 1000 may be a computer device, which may be a server, and its internal structure diagram may be as shown in FIG. Figure 6 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store a trust network. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a battery fault prediction model training method is implemented.

[0214] In one embodiment, a battery failure prediction method is provided. Figure 7 As shown, the following steps are included:

[0215] Step 500, acquiring battery data, wherein the battery data includes time domain data of at least two battery variables, wherein the battery variables include current, voltage or temperature;

[0216] Step 600: input the acquired battery data into a battery fault prediction model to obtain a fault prediction result, wherein the battery fault prediction model is obtained based on the training method provided above.

[0217] The battery fault prediction model adopted by the prediction method includes a time-varying fusion module that has been iteratively trained. In the time-varying fusion module, the battery data mines the interactive features and time series features between time domain signals based on the variable attention mechanism and the time attention mechanism, and fuses the mined features. In particular, the battery data can be divided into minutes, hours, and days to obtain time domain signals of different time domains. In the time-varying fusion module, the data of each time domain is mined and fused across dimensions through the variable attention mechanism and the time attention mechanism. Finally, the feature vectors of multiple time domains are linearly fused to obtain the cross-domain time-varying fusion feature vector. This method effectively solves the problem of scarce new energy battery failure data sets and insufficient sample size.

[0218] In the battery fault prediction model, the location information of the sensor is also combined to perform graph convolution, and the topological relationship of the graph signal is used to adaptively mine its spatiotemporal characteristics.

[0219] The feature vector after graph convolution is input into the convolutional Boltzmann machine, which is used to process high-dimensional data to obtain the final prediction result.

[0220] In one embodiment, Figure 8 As shown, a battery fault prediction device 2000 is provided, comprising a battery data acquisition module 2001 and a fault prediction module 2002;

[0221] A battery data acquisition module 2001 is used to acquire time domain data of at least two battery variables, where the battery variables include at least two of current, voltage or temperature;

[0222] The fault prediction module 2002 is used to input the time domain data into a battery fault prediction model to obtain a fault prediction result; the battery fault prediction model is trained according to the aforementioned battery fault prediction model training method.

[0223] In one embodiment, the battery failure prediction device 2000 may be a computer device, which may be a terminal, and its internal structure diagram may be as shown in FIG. Fig. 9 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a battery fault prediction method is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0224] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0225] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0226] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0227] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A battery failure prediction model training method, characterized in that: include: Acquire sample data, the sample data comprising time domain signals of at least two battery variables, the battery variables comprising current, voltage or temperature, wherein at least part of the sample data is marked with a fault category label; Processing the interaction features between the time domain signals based on a variable attention mechanism to obtain a first feature vector; Based on the time attention mechanism, the temporal features between the time domain signals are processed to obtain the second feature vector: According to the first feature vector and the second feature vector, fusing to obtain a cross-dimensional fused feature vector; Iteratively training the belief network based on the cross-dimensional fusion feature vector and the fault category label to obtain the battery fault prediction model; The fusion obtains a cross-dimensional fusion feature vector, including: According to the first feature vector and the second feature vector, high-dimensional features are mined to obtain variable dimension space features and time dimension space features respectively; Extracting the interactive information of the variable dimension spatial features and the time dimension spatial features based on the attention mechanism to obtain interactive features; wherein the interactive features include the features of the time dimension spatial features after being processed by the attention mechanism, and the features of the time dimension spatial features after being spliced, guided to the variable dimension and then processed by the attention mechanism; According to the interactive features, feature fusion is performed based on a cross-attention mechanism to obtain the cross-dimensional fused feature vector.

2. The battery failure prediction model training method according to claim 1, characterized in that: The obtaining of sample data further includes: Performing time domain division according to the time domain signal of the battery variable to obtain time domain signals of each battery variable in different time domains; The fusion obtains a cross-dimensional fusion feature vector, including: Fusing the first eigenvector and the second eigenvector corresponding to the time domain signals of each time domain to obtain a cross-dimensional fused eigenvector corresponding to each time domain; After the fusion obtains the cross-dimensional fusion feature vector, the method further includes: The cross-dimensional fusion feature vectors corresponding to each time domain are linearly fused to obtain a cross-domain time-varying fusion feature vector, so as to train the belief network based on the cross-domain time-varying fusion feature vector.

3. The battery failure prediction model training method according to any one of claims 1-2, characterized in that: The iterative training of the belief network comprises: Obtain the position matrix of the time domain signal sensor; Processing the position matrix based on a multi-head attention mechanism to obtain a global interaction matrix; According to the global interaction matrix and the cross-dimensional fusion feature vector, or according to the global interaction matrix and the cross-domain time-varying fusion feature vector, a graph convolution vector is obtained based on a graph convolution algorithm, so as to iteratively train the belief network based on the graph convolution vector.

4. The battery failure prediction model training method according to claim 3, characterized in that: The iterative training of the belief network further includes: Extracting time series features from the graph convolution vector based on one-dimensional convolution to obtain a time series feature vector; Using the time series feature vector as an input of a restricted Boltzmann machine, and obtaining a prediction loss according to an output of the restricted Boltzmann machine; The battery failure prediction model is obtained by performing iterative training based on the prediction loss.

5. The battery failure prediction model training method according to claim 1, characterized in that: The iterative training of the belief network further includes: Training the belief network based on semi-supervised cycle loss to obtain the battery failure prediction model; The semi-supervised cycle loss includes an output loss and a distance loss. The output loss is determined according to the predicted value and the target value of the sample data with the fault category label, and the distance loss is determined according to the difference between the predicted values ​​of the sample data and the similarity between samples.

6. The battery failure prediction model training method according to claim 5, characterized in that: The belief network is trained based on a semi-supervised cycle loss, comprising: The belief network is trained according to the semi-supervised cyclic loss function expressed as follows: Among them, L semicircle is the semi-supervised cycle loss, f(x i ) is the i-th sample data x i The output after the belief network, y is the target value, For sample data x i and x j The regularization term of the distance in the sample space, For sample data x i and x j The regularization term of the distance between the output values ​​of ij For sample data x i and x j The similarity weight between is an adjustable parameter, L is the number of sample data with the fault category label, U is the number of sample data without the fault category label, L s is the Laplace matrix of the adjacency matrix composed of similarity weights, rank represents the rank of the matrix, sts ij ≥0 means the constraint is s ij ≥0, c represents the number of connected components of the connected graph constructed by the adjacency matrix.

7. A battery failure prediction method, characterized in that: include: Acquire time domain data of at least two battery variables, wherein the battery variables include current, voltage or temperature; Inputting the time domain data into a battery fault prediction model to obtain a fault prediction result; The battery fault prediction model is trained according to the battery fault prediction model training method according to any one of claims 1 to 6.

8. A battery failure prediction model training device, characterized in that: include: A training data acquisition module, used to acquire sample data, wherein the sample data includes time domain signals of at least two battery variables, wherein the battery variables include at least two of current, voltage or temperature, wherein at least part of the sample data is marked with a fault category label; The model training module is used to process the interaction features between the time domain signals based on the variable attention mechanism to obtain a first feature vector; process the time series features between the time domain signals based on the time attention mechanism to obtain a second feature vector: according to the first feature vector and the second feature vector, fuse to obtain a cross-dimensional fused feature vector; based on the cross-dimensional fused feature vector and the fault category label, iteratively train the confidence network to obtain the battery fault prediction model, wherein the fusion to obtain the cross-dimensional fused feature vector includes: According to the first feature vector and the second feature vector, high-dimensional features are mined to obtain variable dimension space features and time dimension space features respectively; Extracting the interactive information of the variable dimension spatial features and the time dimension spatial features based on the attention mechanism to obtain interactive features; wherein the interactive features include the features of the time dimension spatial features after being processed by the attention mechanism, and the features of the time dimension spatial features after being spliced, guided to the variable dimension and then processed by the attention mechanism; According to the interactive features, feature fusion is performed based on a cross-attention mechanism to obtain the cross-dimensional fused feature vector.

9. A battery failure prediction device, characterized in that: include: A battery data acquisition module, used to acquire time domain data of at least two battery variables, wherein the battery variables include at least two of current, voltage or temperature; A fault prediction module is used to input the time domain data into a battery fault prediction model to obtain a fault prediction result; the battery fault prediction model is trained according to the battery fault prediction model training method according to any one of claims 1 to 6.

Images