Lithium iron phosphate battery multi-class fault diagnosis method based on deep learning

Abstract

The invention discloses a lithium iron phosphate battery multi-class fault diagnosis method based on deep learning, and belongs to the field of lithium battery fault diagnosis. The sparrow search algorithm optimized by the eagle-Cauchy variant is used for self-adaptive updating of variational mode decomposition parameters, and the lightweight fault diagnosis framework fusing the convolutional neural network and Transform is applied to fault diagnosis, so that local and global features in battery time sequence data can be effectively captured; complex modes and interdependence existing in fault features are skillfully captured, diagnosis and detection of various tiny faults of the lithium battery are facilitated, the recognition capability of the tiny faults is enhanced, fault information can be timely and accurately recognized, and fault misdiagnosis is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery fault diagnosis, and in particular to a multi-type fault diagnosis method for lithium iron phosphate batteries based on deep learning. Background Art

[0002] The research on fault diagnosis of lithium batteries has become an important direction that has attracted much attention. In the past, many research works have focused on the development of lithium battery fault diagnosis technology. Traditional lithium battery fault diagnosis methods are mainly based on physical models and signal processing methods. By building models for lithium battery cells, including electrical models, thermal models and multi-physics field models, the models are combined for fault detection and fault type identification. However, a single model has high computational complexity, relies on accurate physical parameter estimation, is only applicable to a single specific fault diagnosis, and is difficult to adapt to different types of faults.

[0003] In recent years, data-driven lithium battery fault diagnosis methods have received widespread attention. These methods can automatically extract features and perform fault diagnosis by learning from a large amount of lithium battery working data. These methods can effectively classify and predict battery fault states, but require complex feature engineering processing, such as manual selection, extraction, and combination of features, which increases computational costs. In practical applications, the difficulty of lithium battery fault diagnosis lies in finding a feature that can accurately reflect fault information. The ideal effect of the above methods depends on sufficient training data. For faults with less obvious degree and less feature data, misdiagnosis and missed diagnosis may occur, and the effect on improving the accuracy of fault diagnosis is limited. Over time, minor defects and faults in batteries may evolve into major faults, so accurate diagnosis is crucial to the durability and reliability of lithium batteries. Summary of the invention

[0004] The present invention provides a multi-category fault diagnosis method for lithium iron phosphate batteries based on deep learning to solve the problems of low diagnostic accuracy, fault confusion and misdiagnosis in diagnosing and identifying multi-category minor faults.

[0005] The embodiment of the present invention provides a multi-type fault diagnosis method for lithium iron phosphate batteries based on deep learning, comprising the following steps:

[0006] Step 1, obtaining voltage time series data of a lithium battery cell charge and discharge test, and preprocessing the voltage time series data;

[0007] Step 2, input the preprocessed voltage time series data into the improved sparrow search algorithm integrating the egret optimization algorithm and the Cauchy variation to update the variational mode decomposition model, adaptively update the parameters of the variational mode decomposition model, decompose and reconstruct the voltage time series data, and extract the fault feature data;

[0008] Step 3: input the fault feature data into a CNN-Transformer model for training, and use the trained CNN-Transformer model to perform multi-type fault diagnosis on the lithium battery.

[0009] Optionally, in one embodiment of the present invention, step 2 specifically includes:

[0010] Step 2-1, set the decomposition layer number K and penalty factor α range of the initial variational mode decomposition, set the initial parameter sparrow population size and the maximum number of iterations, and use the minimum envelope entropy as the fitness function:

[0011]

[0012] Where ε(i) is the envelope signal of the original signal after Hilbert demodulation; h(i) is the normalized form of ε(i); N is the length of the signal;

[0013] Step 2-2, continuously update the population positions of the discoverer, follower, and sentinel, iteratively optimize the decomposition layer number K and penalty factor α, and calculate the corresponding fitness value;

[0014] The formula for the discoverer to update the position is:

[0015]

[0016] Among them, x n ∈(0,1), ρ is the control parameter;

[0017]

[0018] RB=randn(1,Dim)

[0019] RL=0.5×Levy(Dim)

[0020] Among them, t is the current iteration number, T is the maximum iteration number, For the i th The new state of the egret in the first stage, X i,j To represent the i th The value of the j-th problem variable of the egret eagle, and is a random candidate solution of the first stage iteration, R1 is an array of dimension 1×Dim randomly generated from the interval [0, 1], Dim is the dimension of the solution space, For the jth th The value of dimension, F i new,P1 is the fit, randn(1,Dim) is an array of dimension 1×Dim randomly generated from a standard normal distribution, xbest is the current optimal value, RB is Brownian motion, x i,j is the current position value of the individual, RL is the weighted Levy flight, X i is the i-th egret variable, F i is the objective function value obtained by the i-th egret, and Levy(Dim) is the Levy flight distribution function, which is calculated by the following formula:

[0021]

[0022] Where s = 0.01, η = 1.5, u and ν are random numbers in the interval [0, 1], and Γ represents the gamma function;

[0023] The formula for the follower to update its position is:

[0024]

[0025] in, is the new position of the i-th follower in the j-th dimension at iteration t+1, X best is the current best position, cauchy(0,1) is the standard Cauchy distribution function;

[0026] The formula for updating the position of the sentinel is:

[0027]

[0028] in, is the new position of the i-th sentinel at the j-th dimension at iteration k+1, is the i-th alerter at the j-th dimension position at iteration k, is the global optimal position under the current iteration number k, β is the step size correction coefficient and obeys the standard normal distribution, f i is the current fitness value of the sparrow, f w and f g are the current overall worst fitness value and the best fitness value, respectively, λ∈[0,1] is the step size control factor, ε=10 -59 is a very small positive number, is the global worst position under the current iteration number k;

[0029] Step 2-3, output the optimal (K, α) and the global optimal fitness function value, use the variational mode decomposition algorithm to decompose and reconstruct the voltage time series data, and extract the fault feature data.

[0030] Optionally, in one embodiment of the present invention, step 3 specifically includes:

[0031] Step 3-1, the CNN network extracts the local features of the feature sequence in the convolution layer. The operation expression of the convolution layer is:

[0032]

[0033] in, For network output, is the output of the jth feature map of the n-1th layer, is the mth convolution kernel, is the bias value, N j is the set of input data, X(·) is the activation function selected independently;

[0034] The selected activation function is the linear rectified unit ReLU function, and the specific formula is:

[0035]

[0036] in, is the activation value of the i-th neuron in the j-th layer, is the linear combination output of the i-th neuron in the j-th layer;

[0037] The pooling function selected by the pooling layer is maximum pooling, and the formula is as follows:

[0038]

[0039] in, is the activation value of the ith neuron in the nth layer, w is the size of the pooling window, j is the index of the output neuron after the pooling operation, is the value of the jth output neuron in the nth layer;

[0040] Step 3-2, the local features output by CNN are transmitted to the Transformer network to obtain the global features of the fault data and increase the mapping of high-dimensional features and lithium-ion battery fault information;

[0041] The Transformer network maps Query (Q), Key (K) and Value (V) through h different linear transformations, concatenates different Attentions and performs another linear transformation. The calculation process of the entire multi-head attention mechanism is:

[0042] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0043] head i =Attention(QWi Q ,KW i K ,VW i V )

[0044]

[0045] Among them, Q, K, V are query matrix, key matrix and value matrix, head i is the output of the i-th attention head, K T is the transpose of the key matrix, the parameter matrix is the parameter matrix for linear transformation, It is a scaling factor used to alleviate the gradient vanishing problem caused by the introduction of softmax;

[0046] The feedforward fully connected layer of Transformer is composed of Fully Connected Layer and ReLU activation function, and the calculation formula is:

[0047] FFN(x)=max(0,xW1+b1)W2+b2

[0048] Among them, W1 is the weight matrix of the first layer of the feedforward network, W2 is the weight matrix of the second layer of the feedforward network, b1 is the bias term of the first layer of the feedforward network, and b2 is the bias term of the second layer of the feedforward network.

[0049] Optionally, in one embodiment of the present invention, in step 1, preprocessing the voltage time series data includes: performing outlier removal and denoising processing on the voltage time series data.

[0050] The multi-class fault diagnosis method for lithium iron phosphate batteries based on deep learning in the embodiment of the present invention has the following beneficial effects:

[0051] 1. The improved sparrow search algorithm (SCSSA) realizes the adaptive parameter selection of variational mode decomposition (VMD), realizes the adaptive optimization of parameters, and improves the generalization performance of the model.

[0052] 2. Use the improved sparrow search algorithm (SCSSA) in the model to update the VMD part to avoid the algorithm falling into the local optimum, accurately capture the characteristics of different faults, and improve the extraction accuracy of lithium battery fault characteristics.

[0053] 3. The lightweight fault diagnosis framework that combines the advantages of convolutional neural networks (CNN) and Transformer can effectively capture the local and global features in battery time series data, skillfully capture the complex patterns and interdependencies in fault characteristics, and help diagnose and detect various minor faults in lithium batteries. It enhances the ability to identify minor faults, so as to timely and accurately identify fault information and reduce misdiagnosis of faults.

[0054] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0056] Figure 1 A flowchart of a multi-type fault diagnosis method for lithium iron phosphate batteries based on deep learning according to an embodiment of the present invention;

[0057] Figure 2 is a model structure diagram of an embodiment of the present invention;

[0058] Figure 3 This is a model training curve diagram of an embodiment of the present invention;

[0059] Figure 4 A model test curve diagram of an embodiment of the present invention;

[0060] Figure 5 The figure is a confusion matrix diagram of the diagnosis result of the model according to the embodiment of the present invention. DETAILED DESCRIPTION

[0061] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0062] Figure 1 The present invention provides a flowchart of a multi-type fault diagnosis method for lithium iron phosphate batteries based on deep learning according to an embodiment of the present invention.

[0063] like Figure 1 As shown, the multi-class fault diagnosis method of lithium iron phosphate battery based on deep learning includes the following steps:

[0064] Step 1: Obtain voltage time series data of a lithium battery cell charge and discharge test, and preprocess the voltage time series data.

[0065] Specifically, voltage time series data is obtained through a lithium battery cell charge and discharge test, and the data is numbered to remove abnormal values ​​and noise in the voltage time series data.

[0066] Step 2: input the preprocessed voltage time series data into the improved sparrow search algorithm (SCSSA) integrating the egret optimization algorithm and the Cauchy variation to update the variational mode decomposition model, adaptively update the parameters of the variational mode decomposition model, decompose and reconstruct the voltage time series data, and extract the fault feature data.

[0067] In an embodiment of the present invention, step 2 specifically includes:

[0068] Step 2-1, set the decomposition layer number K and penalty factor α range of the initial variational mode decomposition, set the initial parameter sparrow population size and the maximum number of iterations, and use the minimum envelope entropy as the fitness function:

[0069]

[0070] Where ε(i) is the envelope signal of the original signal after Hilbert demodulation; h(i) is the normalized form of ε(i); N is the length of the signal;

[0071] Step 2-2, continuously update the population positions of the discoverer, follower, and sentinel, iteratively optimize the decomposition layer number K and penalty factor α, and calculate the corresponding fitness value;

[0072] The formula for the discoverer to update the position is:

[0073]

[0074] Among them, x n ∈(0,1), ρ is the control parameter;

[0075]

[0076] RB=randn(1,Dim)

[0077] RL=0.5×Levy(Dim)

[0078] Among them, t is the current iteration number, T is the maximum iteration number, For the i th The new state of the egret in the first stage, X i,j To represent the i th The value of the j-th problem variable of the egret eagle, and is a random candidate solution of the first stage iteration, R1 is an array of dimension 1×Dim randomly generated from the interval [0, 1], Dim is the dimension of the solution space, For the jth th The value of dimension, F i new,P1 is the fit, randn(1,Dim) is an array of dimension 1×Dim randomly generated from a standard normal distribution, x best is the current optimal value, RB is Brownian motion, x i,j is the current position value of the individual, RL is the weighted Levy flight, X i is the i-th egret variable, F i is the objective function value obtained by the i-th egret, and Levy(Dim) is the Levy flight distribution function, which is calculated by the following formula:

[0079]

[0080] Where s = 0.01, η = 1.5, u and ν are random numbers in the interval [0, 1], and Γ represents the gamma function;

[0081] The formula for the follower to update its position is:

[0082]

[0083] in, is the new position of the i-th follower in the j-th dimension at iteration t+1, X best is the current best position, cauchy(0,1) is the standard Cauchy distribution function;

[0084] The formula for updating the position of the sentinel is:

[0085]

[0086] in, is the new position of the i-th sentinel at the j-th dimension at iteration k+1, is the i-th alerter at the j-th dimension position at iteration k, is the global optimal position under the current iteration number k, β is the step size correction coefficient and obeys the standard normal distribution, f i is the current fitness value of the sparrow, f w and f g are the current overall worst fitness value and the best fitness value, respectively, λ∈[0,1] is the step size control factor, ε=10 -59 is a very small positive number, is the global worst position under the current iteration number k;

[0087] Step 2-3, output the optimal (K, α) and the global optimal fitness function value, use the variational mode decomposition algorithm to decompose and reconstruct the voltage time series data, and extract the fault feature data.

[0088] Step 3: Input the fault feature data into the CNN-Transformer model for training, and use the trained CNN-Transformer model to perform multi-type fault diagnosis on lithium batteries. Input the test set into the trained model, and output the test results of CNN-Transformer to perform multi-type fault diagnosis on lithium batteries.

[0089] The fault feature data extracted in step 2 is randomly divided into a training set and a test set, and the training set is input into the CNN-Transformer model for learning to obtain a trained CNN-Transformer model, such as Figure 2 shown.

[0090] In an embodiment of the present invention, step 3 specifically includes:

[0091] In step 3-1, the CNN network extracts local features of the feature sequence in the convolution layer, gradually reducing the size and quantity of feature information, achieving accurate capture of local spatial information and significant enhancement of details. The operation expression of the convolution layer of the CNN network is:

[0092]

[0093] in, For network output, is the output of the jth feature map of the n-1th layer, is the mth convolution kernel, is the bias value, N j is the set of input data, X(·) is the activation function selected independently;

[0094] The activation function selected is the linear correction unit ReLU function, which increases the nonlinearity of the network and improves the network's expression ability. The specific formula is:

[0095]

[0096] in, is the activation value of the i-th neuron in the j-th layer, is the linear combination output of the i-th neuron in the j-th layer;

[0097] The pooling layer is used to reduce the spatial size of the feature map, reduce the number of parameters, and prevent overfitting. The pooling function selected in this paper is Max Pooling. The specific formula is as follows:

[0098]

[0099] in, is the activation value of the ith neuron in the nth layer,w is the size of the pooling window, j is the index of the output neuron after the pooling operation, is the value of the jth output neuron in the nth layer;

[0100] Step 3-2, the local features output by CNN are transmitted to the Transformer network to obtain the global features of the fault data and increase the mapping of high-dimensional features and lithium-ion battery fault information;

[0101] The Transformer network maps Query (Q), Key (K) and Value (V) through h different linear transformations, concatenates different Attentions and performs another linear transformation. The calculation process of the entire multi-head attention mechanism is:

[0102] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0103] head i =Attention(QW i Q ,KW i K ,VW i V )

[0104]

[0105] Among them, Q, K, V are the query matrix (Q), key matrix (K) and value matrix (V), head i is the output of the i-th attention head, K T is the transpose of the key matrix, the parameter matrix is the parameter matrix for linear transformation, It is a scaling factor used to alleviate the gradient vanishing problem caused by the introduction of softmax;

[0106] The Transformer's feedforward fully connected layer (FFN) consists of a Fully Connected Layer and a ReLU activation function, and the calculation formula is:

[0107] FFN(x)=max(0,xW1+b1)W2+b2

[0108] Among them, W1 is the weight matrix of the first layer of the feedforward network, W2 is the weight matrix of the second layer of the feedforward network, b1 is the bias term of the first layer of the feedforward network, and b2 is the bias term of the second layer of the feedforward network.

[0109] The multi-type fault diagnosis method of lithium iron phosphate batteries based on deep learning of the present invention is described in detail below through specific embodiments.

[0110] Taking the Ningde New Energy GC-WDF-C5D2E7-010L / H iron phosphate battery as an example, the voltage data of the standard charge and discharge cycle and different degrees of overcharge cycle, the voltage data of different degrees of over-discharge cycle and different degrees of internal short circuit (ISC) fault cycle data of a single lithium battery are obtained. The data of each fault type consists of 2000 samples, each sample contains 1024 data points, and a matrix of 18000×1024 is obtained. In the experimental device, 1400 samples are allocated for each fault for model training and 600 samples for model testing. The data parameter description is shown in Table 2 below.

[0111] Table 1

[0112]

[0113] The VMD decomposition layer K is set to the range of [1,20], the penalty factor α is set to the range of [1,10000], the initial parameter sparrow population size P = 20, and the maximum number of iterations = 30. After the SCSSA-VMD feature extraction process, the 9 states are represented by 2000 samples respectively, and a 18000×9 data matrix is ​​obtained. The matrix is ​​labeled with numbers 1 to 9 to represent different fault types. The hyperparameters of the model training are shown in Table 2.

[0114] Table 2

[0115]

[0116] Figure 3 The accuracy curve of the fault diagnosis model is given, showing the excellent diagnostic performance of the model in the training and testing stages. After about 40 iterations, the model performance stabilized and achieved a test accuracy of up to 99.43%. Figure 4 The loss function curve of the fault diagnosis model is given. The model shows excellent performance, and both the training loss and the test loss drop rapidly and stabilize at a low value after about 20 iterations. The diagnostic method proposed in the present invention can clearly identify faults of various degrees and categories, and the diagnostic accuracy of all fault types is close to 100%, and the overall accuracy of the model is 99.43%.

[0117] Figure 5The figure is a confusion matrix diagram of the diagnosis results of the fault diagnosis model. The diagonal elements in the confusion matrix are all dark blue. Each category of fault is clearly diagnosed, and the recall rate is maintained above 95%. The proposed method is far superior to other multi-fault diagnosis methods. It is worth noting that in the early stages, the voltage feature difference between minor faults and normal data is small, and it is difficult to distinguish different minor faults by whether the voltage threshold is exceeded. Therefore, the SCSSA-VMD feature processing module proposed in the present invention can help the CNN-Transformer module avoid feature confusion and reduce the probability of sample misjudgment and omission.

[0118] According to the multi-class fault diagnosis method for lithium iron phosphate batteries based on deep learning proposed in an embodiment of the present invention, the adaptive parameter selection of VMD is realized through the improved sparrow search algorithm, the adaptive optimization of parameters is realized, and the generalization performance of the model is improved. The improved sparrow search algorithm (SCSSA) in the model is used to update the VMD part to avoid the algorithm from falling into the local optimum, and accurately capture the characteristics of different faults, thereby improving the extraction accuracy of lithium battery fault characteristics. The lightweight fault diagnosis framework that combines the advantages of CNN and Transformer can effectively capture the local and global features in battery time series data, and skillfully capture the complex patterns and interdependencies in fault characteristics, which is helpful for the diagnosis and detection of various minor faults of lithium batteries, and enhances the recognition ability of minor faults, so as to timely and accurately identify fault information and reduce misdiagnosis of faults.

[0119] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0120] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0121] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or N executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

Claims

1. A multi-class fault diagnosis method for lithium iron phosphate batteries based on deep learning, characterized in that: The following steps are involved: Step 1, obtaining voltage time series data of a lithium battery cell charge and discharge test, and preprocessing the voltage time series data; Step 2, input the preprocessed voltage time series data into the improved sparrow search algorithm integrating the egret optimization algorithm and the Cauchy variation to update the variational mode decomposition model, adaptively update the parameters of the variational mode decomposition model, decompose and reconstruct the voltage time series data, and extract the fault feature data; Step 3: input the fault feature data into a CNN-Transformer model for training, and use the trained CNN-Transformer model to perform multi-type fault diagnosis on the lithium battery.

2. The method according to claim 1, characterized in that: Step 2 specifically includes: Step 2-1, set the decomposition layer number K and penalty factor α range of the initial variational mode decomposition, set the initial parameter sparrow population size and the maximum number of iterations, and use the minimum envelope entropy as the fitness function: Where ε(i) is the envelope signal of the original signal after Hilbert demodulation; h(i) is the normalized form of ε(i); N is the length of the signal; Step 2-2, continuously update the population positions of the discoverer, follower, and sentinel, iteratively optimize the decomposition layer number K and penalty factor α, and calculate the corresponding fitness value; The formula for the discoverer to update the position is: Among them, x n ∈(0,1), ρ is the control parameter; RB=randn(1,Dim) RL=0.5×Levy(Dim) Among them, t is the current iteration number, T is the maximum iteration number, For the i th The new state of the egret in the first stage, X i,j To represent the i th The value of the j-th problem variable of the egret eagle, and is a random candidate solution of the first stage iteration, R1 is an array of dimension 1×Dim randomly generated from the interval [0, 1], Dim is the dimension of the solution space, For the jth th The value of dimension, F i new,P1 is the fit, randn(1,Dim) is an array of dimension 1×Dim randomly generated from a standard normal distribution, x best is the current optimal value, RB is Brownian motion, x i,j is the current position value of the individual, RL is the weighted Levy flight, X i is the i-th egret variable, F i is the objective function value obtained by the i-th egret, and Levy(Dim) is the Levy flight distribution function, which is calculated by the following formula: Where s = 0.01, η = 1.5, u and ν are random numbers in the interval [0, 1], and Γ represents the gamma function; The formula for the follower to update its position is: in, is the new position of the i-th follower in the j-th dimension at iteration t+1, X best is the current best position, cauchy(0,1) is the standard Cauchy distribution function; The formula for updating the position of the sentinel is: in, is the new position of the i-th sentinel at the j-th dimension at iteration k+1, is the i-th alerter at the j-th dimension position at iteration k, is the global optimal position under the current iteration number k, β is the step size correction coefficient and obeys the standard normal distribution, f i is the current fitness value of the sparrow, f w and f g are the current overall worst fitness value and the best fitness value, respectively, λ∈[0,1] is the step size control factor, ε=10 -59 is a very small positive number, is the global worst position under the current iteration number k; Step 2-3, output the optimal (K, α) and the global optimal fitness function value, use the variational mode decomposition algorithm to decompose and reconstruct the voltage time series data, and extract the fault feature data.

3. The method according to claim 1, characterized in that Step 3 specifically includes: Step 3-1, the CNN network extracts the local features of the feature sequence in the convolution layer. The operation expression of the convolution layer is: in, For network output, is the output of the jth feature map of the n-1th layer, is the mth convolution kernel, is the bias value, N j is the set of input data, X(·) is the activation function selected independently; The selected activation function is the linear rectified unit ReLU function, and the specific formula is: in, is the activation value of the i-th neuron in the j-th layer, is the linear combination output of the ith neuron in the jth layer. The pooling function selected by the pooling layer is maximum pooling, and the formula is as follows: in, is the activation value of the ith neuron in the nth layer, w is the size of the pooling window, j is the index of the output neuron after the pooling operation, is the value of the jth output neuron in the nth layer; Step 3-2, the local features output by CNN are transmitted to the Transformer network to obtain the global features of the fault data and increase the mapping of high-dimensional features and lithium-ion battery fault information; The Transformer network maps Query (Q), Key (K) and Value (V) through h different linear transformations, concatenates different Attentions and performs another linear transformation. The calculation process of the entire multi-head attention mechanism is: MultiHead(Q,K,V)=Concat(head1,...,head h )W O head i =Attention(QW i Q ,KW i K ,VW i V ) Among them, Q, K, V are query matrix, key matrix and value matrix, head i is the output of the i-th attention head, K T is the transpose of the key matrix, the parameter matrix is the parameter matrix for linear transformation, It is a scaling factor used to alleviate the gradient vanishing problem caused by the introduction of softmax; The feedforward fully connected layer of Transformer is composed of Fully Connected Layer and ReLU activation function, and the calculation formula is: FFN(x)=max(0,xW1+b1)W2+b2 Among them, W1 is the weight matrix of the first layer of the feedforward network, W2 is the weight matrix of the second layer of the feedforward network, b1 is the bias term of the first layer of the feedforward network, and b2 is the bias term of the second layer of the feedforward network.

4. The method according to claim 1, characterized in that In step 1, preprocessing the voltage time series data includes: performing outlier removal and denoising processing on the voltage time series data.

Images