New energy automobile battery fault identification and early warning method based on hybrid deep learning

Abstract

The invention discloses a new energy automobile battery fault identification and early warning method based on hybrid deep learning. The method comprises the following steps: firstly, defining a fault type and a judgment standard, and generating a label of training data; and then constructing a hybrid deep learning model comprising a convolutional layer model and a bidirectional gating cycle unit layer, respectively extracting spatial features of battery data and a long-term dependency relationship in a time sequence, and encoding state information of the battery. Then, the extracted features are input into a fault recognition module for fine adjustment of model parameters, optimization of a cross entropy loss function of a model and minimization of an error between model prediction and an actual fault, so that the calculation time is shortened, and the training efficiency is improved; and finally, based on the trained hybrid deep learning model, detecting the fault probability of the battery data, predicting target feature data in a period of time in the future by using the model, and generating fault early warning information. According to the method, the accuracy of battery fault identification and the reliability of early warning can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology and relates to a new energy vehicle battery fault identification and early warning method, and specifically relates to a new energy vehicle battery fault identification and early warning method based on hybrid deep learning. Background Art

[0002] As the core component of new energy vehicles, the health status of batteries directly affects the operating efficiency and safety of vehicles. At present, the models and methods used for battery fault identification and early warning mainly include fault classification models based on traditional statistical models, fault classification models based on traditional learning methods such as support vector machines, fault time series identification models based on deep learning, and fault feature extraction and analysis models based on signal processing. However, traditional statistical models are not able to identify complex nonlinear fault modes and cannot process large-scale dynamic data; traditional learning methods such as support vector machines are inefficient in high-dimensional feature spaces and rely on artificial feature engineering; time series models based on single deep learning networks such as RNN and LSTM have limited ability to capture long sequence information, gradient disappearance or explosion, and other problems; analysis methods based on signal processing are limited by feature extraction accuracy and cannot meet the needs of high-precision fault identification. In addition, existing fault identification models generally lack versatility, and their scope of application is mostly limited to specific fault modes, making it difficult to extend to fault scenarios of different vehicles or battery types, and lack adaptability to complex environments.

[0003] In summary, the existing battery fault identification and early warning methods not only rely on a large amount of historical fault data, but are also easily affected by noise during model training, resulting in insufficient accuracy of fault identification and reliability of early warning, and lack of versatility, making it difficult to adapt to the needs of complex fault scenarios. In order to improve the accuracy of new energy vehicle battery fault identification and enhance the early warning capability of potential battery faults, a more advanced and adaptable hybrid deep learning model and early warning strategy are urgently needed to cope with the complex characteristics of multiple types of battery faults and provide a more universal and explainable fault identification and early warning method. Summary of the invention

[0004] In view of the shortcomings of the prior art, the present invention proposes a new energy vehicle battery fault identification and early warning method based on hybrid deep learning. A hybrid deep learning model is constructed using convolutional layers (CNN) and bidirectional gated recurrent units (BiGRU) layers to extract the spatial characteristics of battery data and the long-term dependencies in the time series, encode the battery status information, and improve the accuracy of battery fault identification and the reliability of early warning.

[0005] A new energy vehicle battery fault identification and early warning method based on hybrid deep learning includes the following steps:

[0006] Step 1: Preprocessing and labeling of battery fault feature data

[0007] First, obtain the collected characteristic data of the power battery during the operation of new energy vehicles, including charging status, vehicle speed, cumulative mileage, battery voltage, battery current, SOC, insulation resistance, temperature, maximum voltage, minimum voltage, maximum temperature and minimum temperature. Set the valid value range of each characteristic data, and based on the fault identification method and generation mechanism, select the appropriate characteristic data type from the collected characteristic data, perform preprocessing, and then integrate and generate labels for fault identification:

[0008] s1.1. Outlier removal and missing value processing: Clean the acquired data, remove outliers that exceed the normal range of the features, use the mean or interpolation method to fill in missing values ​​to ensure data integrity, and finally delete duplicate sample rows.

[0009] s1.2, feature data smoothing: After completing the removal of outliers and filling in missing values, the collected feature data X = [x1, x2, ..., x n ]∈R c×n The Savitzky-Golay filter is used to smooth the data. Generate the i-th feature smoothing data at time t t=1, 2, ..., n, i=1, 2, ..., c, c is the total number of sampled battery feature types, and n is the number of sampling time points, that is, the number of samples.

[0010] s1.3, feature data standardization and normalization: After completing the smoothing process, the smoothed data set Standardization and normalization operations are performed so that all eigenvalues ​​related to new energy vehicle battery failure fall within the same numerical range [0,1], with a mean of 0 and a variance of 1.

[0011]

[0012] In the formula, μ i and σ i are the mean and standard deviation of the i-th feature, X i，min and X i，max is the minimum and maximum value of the i-th feature, represents the data after the i-th feature is standardized, It represents the data set after the normalization of the i-th feature. Standardization and normalization make the values ​​of each feature in the same dimension, which is conducive to improving the convergence speed and stability of the model.

[0013] s1.4, Battery Fault Marking: During its operation cycle, the power battery of new energy vehicles is in a relatively complex and changeable environment, and is easily challenged by various situations. There are many factors that affect the safety of power battery charging, and the resulting new energy vehicle battery failures also present diverse characteristics. After completing data preprocessing, through in-depth analysis of the factors affecting the safety of new energy vehicle charging, the collected feature data is classified to extract the target feature set of the fault j = 1, 2, ..., m, where m represents the number of fault types, and each type of fault is labeled accordingly:

[0014] ① Abnormal self-discharge: The energy of the battery does not enter the external circuit through discharge, but is lost in other ways, causing a single or multiple battery cells to be in an obvious outlier state and pose a safety hazard. The self-discharge rate is calculated by the voltage change of adjacent sampling points:

[0015]

[0016] Where V t and V t-1 They are the voltage values ​​at the current moment and the previous moment respectively. When the self-discharge rate SDR>3%, it is marked as an abnormal state.

[0017] ②Sudden internal short circuit: The phenomenon that the battery suddenly falls into a thermal runaway state due to abnormal conduction of the positive and negative electrodes inside the battery cell, and the specific manifestations include a sharp rise in the temperature of the battery cell or an abnormal increase in the voltage. The sudden internal short circuit anomaly is divided into sudden internal short circuit induced by high temperature and sudden internal short circuit caused by high voltage. When the battery temperature exceeds 38°C, the voltage is greater than 4.1V, or the voltage is less than 2.5V, it is marked as a sudden internal short circuit.

[0018] ③Insulation failure: The insulation resistance of the battery system is lower than the limit due to vehicle wading, battery leakage, insulation layer aging or other reasons. The insulation ratio IRR is an important basis for judging insulation failure:

[0019]

[0020] Where IR is the insulation resistance, U SUM It is expressed as the total voltage. When the battery is in the charging state, IRR ≤ 100Ω / V, or IRR ≤ 500Ω / V in other states, it is marked as insulation failure.

[0021] According to the above-mentioned fault judgment indicators, the fault data is marked to form label data Y = {y t，j |y t，j ∈{0,1}}∈R n×m , where m represents the number of fault types, y t，jIndicates whether the j-th type of fault occurs at time point t, 1 indicates the presence of this type of fault, 0 indicates the absence of this type of fault, j=1, 2, ..., m.

[0022] Step 2: Extraction of key features of battery failure

[0023] After marking the fault type in step 1, we deeply analyze the relevant operating status of the new energy vehicle during operation and extract key features that help accurately diagnose battery faults. In order to quantify the impact of each feature on different fault types, we use the Pearson correlation coefficient to perform correlation analysis, quantify the linear relationship between the feature and the fault type label, and select the key feature variables that are most indicative of fault identification:

[0024]

[0025] In the formula, Represents the feature data vector after smoothing of the i-th feature, y j is the fault label vector of the jth type of fault; and They are and j The mean of The value range is [-1,1], The closer the absolute value is to 1, the more characteristic and faulty j The stronger the correlation between them.

[0026] Sort the Pearson correlation coefficients of each feature and the fault label by absolute value, and select the feature variables whose absolute value of the correlation coefficient is higher than the preset threshold as the key feature variables highly correlated with the fault label to form the key feature identification matrix Z, Z = {z ij |z ij ∈{0,1}}∈R c×m , the matrix element z ij Indicates whether feature i is the key feature of fault j. 1 indicates that feature i is the key feature of fault j, and 0 indicates that it is not a key feature.

[0027] Step 3: Battery Fault Type Identification Model

[0028] A hybrid deep learning model combining a convolutional neural network (CNN) layer and a bidirectional gated recurrent unit (BiGRU) layer is constructed. In order to improve the recognition accuracy, the fault recognition tasks of different categories are decomposed into multiple binary classification subtasks. For the subtask Type corresponding to the jth type of fault, j To detect whether a specific type of fault occurs separately, the model input matrix

[0029] The convolutional neural network layer includes a convolutional layer, an activation function and a pooling layer. j Perform feature extraction to capture local patterns and spatial relationships between features; each element in the output matrix C generated by the convolution operation is processed by the ReLU activation function to obtain the matrix C' after nonlinear transformation, which can effectively suppress negative features and make the model focus on positive features, which helps the model learn complex patterns and avoid the problem of gradient disappearance. C' contains multiple channels, each channel represents a different feature map extracted by the convolution kernel, and these channels capture the patterns and features of the input data in the local space; on each channel of the feature map C', use the maximum pooling operation to reduce the dimension of the feature map, extract the maximum value of the channel, and use the maximum value of each channel max(c1), max(c2),..., max(c q ) are combined into a maximum value vector max(c), where q represents the number of channels generated by the convolution operation, that is, the number of convolution kernels.

[0030] The bidirectional gated recurrent unit layer is used to further process the feature data output by the convolutional neural network layer, and uses bidirectional information flow to capture the front-to-back dependencies of sequence data, thereby improving the model's ability to understand and extract time-dependent features. The BiGRU structure processes data in two directions, forward GRU and backward GRU, providing a comprehensive feature representation for each time step, and then performs a forward hidden state analysis on each time step. and the backward hidden state Merge to form the final hidden state representation h t :

[0031]

[0032] The output of BiGRU passes through the fully connected layer, and the hidden states of all time steps captured by BiGRU are summarized to form a high-level feature vector set h = {h1, h2, ..., h n} is mapped to the target fault label vector y j =[y 1，j ,y 2，j , ..., y n，j ], and use the Sigmoid activation function to convert the output into a fault probability value of [0, 1], and then use the fault threshold τ to get the fault label

[0033] y j =W f ·h+b f

[0034]

[0035] Where Wf and b f is the weight matrix and bias of the fully connected layer; define the fault threshold τ, if It is determined that there is a fault j at time point t, otherwise it is determined to be in a normal state.

[0036] Binary cross entropy loss is used as the loss function for model training to handle the fault characteristics of each fault label appearing independently or together:

[0037]

[0038] In the formula, y t，j represents the fault label of the j-th fault at the t-th time point, is the fault label at the tth time point of the jth type of fault identified by the model.

[0039] Step 4: Battery Failure Warning Strategy

[0040] s4.1、Fault warning setting

[0041] Define fault thresholds, use hybrid deep learning models to predict relevant fault parameters, detect whether fault characteristics exceed safety thresholds, and issue early warnings:

[0042] Define a basic threshold e threshold，j , when the prediction value P of the j-th type of fault output by the hybrid deep learning model at time point t t，j Exceeding the basic threshold e threshold，j In order to reduce false alarms and accidental risks, a minimum number of warning points k is set, requiring that the fault warning is officially triggered only when there are at least k data points with fault risks in the continuous time window.

[0043] s4.2. Predict future fault characteristic values

[0044] The goal of the early warning strategy is to predict the temperature conditions of the battery probes in the future in advance so that preventive measures can be taken before the battery temperature becomes abnormal.

[0045] The target characteristic data Q of the j-th type of fault collected t，j Input the trained hybrid deep learning model to get the first prediction result P t+1，j , P t+1，j Add to Q t，j At the end of t+1，j , and input it into the trained hybrid deep learning model again for the next prediction. Repeat the above steps until the prediction is T steps The predicted value for future time steps T steps =N×60 / tgap , t gap represents the sampling interval, and N represents the prediction duration.

[0046] s 4.3. Detecting fault characteristics in future time periods

[0047] In the predicted future time step feature sequence In the above example, for the predicted value of the j-th type of fault at each prediction time point t, an indicator function δ is defined t，j , record whether it exceeds the basic threshold e threshold :

[0048]

[0049] In the formula, δ i，j =1 indicates that the jth fault at the tth prediction time point exceeds the threshold and needs to be recorded as an abnormal fault state, and 0 indicates a normal state.

[0050] s4.4. Generate fault warning information

[0051] For each prediction time point t, according to the indicator function δ t，j Construct abnormal fault set S t , including all abnormal failures that occurred at the corresponding time point:

[0052] S t ={j|δ t，j =1}

[0053] If there is an abnormal fault set An early warning is issued and corresponding warning information Info is generated, including the fault type, fault occurrence time and fault threshold.

[0054] The present invention has the following gain effects:

[0055] The present invention proposes a new energy vehicle battery fault identification and early warning method based on a hybrid deep learning model composed of a convolutional neural network (CNN) layer and a bidirectional gated recurrent unit (BiGRU) layer, which effectively improves the accuracy, real-time and adaptability of multi-category battery fault identification. First, the battery data that meets the fault classification requirements is preprocessed, and the key indicators are further marked with custom feature coding using the effective value range of key feature variables such as voltage and temperature to achieve effective distinction between normal and abnormal states. Through fault labeling and Pearson correlation coefficient screening, the fault identification input features are optimized to ensure that the model has universality and adaptability in multiple types of faults. The hybrid deep learning model of this method uses the CNN layer to extract the spatial features of the battery data, and captures the long-term dependencies in the time series through the BiGRU layer, which enhances the understanding and recognition capabilities of multi-category battery fault modes; in addition, the CNN-BiGRU recognition model of the present invention decomposes each type of fault detection task into two-class subtasks, and combines the attention mechanism to improve the ability to accurately capture fault information. At the same time, threshold setting and recursive prediction methods based on fault characteristics can achieve early warning and issue alarms for potential abnormal fault conditions before a fault occurs, which helps to improve the safety of new energy vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is the flow chart of fault data feature analysis.

[0057] Figure 2 Schematic diagram of the structure of the hybrid deep learning model (CNN-BiGRU) constructed for the embodiment.

[0058] Figure 3 This is a flowchart for selecting model hyperparameters in an embodiment.

[0059] Figure 4 The figure is a flow chart of the battery failure warning method based on the hybrid deep learning model. DETAILED DESCRIPTION

[0060] The present invention will be further explained below with reference to the accompanying drawings;

[0061] A new energy vehicle battery fault identification and early warning method based on hybrid deep learning includes the following steps:

[0062] Step 1: Preprocessing and labeling of battery fault feature data

[0063] like Figure 1As shown in the figure, the battery monitoring data is read from the acquired new energy vehicle driving monitoring data file, including the battery charging status, vehicle speed, cumulative mileage, battery voltage, battery current, SOC, insulation resistance, temperature, maximum voltage value, minimum voltage value, maximum temperature value, and minimum temperature value at different time points; it is converted into a matrix form X∈R that can be used for analysis c ×n , c represents the fault feature dimension, n is the number of sampling time points, that is, the number of samples, and data preprocessing is performed on the data X. The threshold conditions corresponding to the three fault types of self-discharge abnormality, sudden internal short circuit abnormality and insulation failure are defined. The preprocessed sample data is traversed, and the fault label Y={y t，j |y t，j ∈{0,1}}∈R n×m , m represents the number of fault types, and the element y in the matrix t，j Describes the label value of the j-th fault at the t-th time point. 1 indicates that this type of fault exists, and 0 indicates that this type of fault does not exist.

[0064] Step 2: Extraction of key features of battery failure

[0065] The Pearson correlation coefficient is used for correlation analysis. According to the results of the correlation analysis, the absolute value of the correlation with the fault label is selected. p，q |More than r threshold The key feature set F is obtained key :

[0066] F key ={f1, f2, ..., f m}

[0067]

[0068] F key ={f1, f2, ..., f m} is the key feature sampling data set, where m represents the number of fault types, is the sampled data after preprocessing and smoothing of the i-th type of feature data, r i，j is the correlation coefficient between the i-th feature and the j-th fault. Finally, the key feature set of different fault type labels is shown in Table 1:

[0069] Table 1 Key feature sets of different fault type labels

[0070]

[0071]

[0072] Step 3: Battery Fault Type Identification Model

[0073] like Figure 2 As shown in the figure, a hybrid deep learning model combining a convolutional neural network (CNN) layer and a bidirectional gated recurrent unit (BiGRU) layer is constructed to decompose different types of fault identification tasks into multiple binary classification subtasks; each subtask Type j Detect whether the jth type of fault occurs separately and set F key f j Input parameter trainingType j The model of the subtask. Figure 3 As shown, in order to ensure the best performance of the model, a grid search method is used to select the hyperparameters of the model, including the number of filters in the convolutional layer and the number of neural units in the bidirectional gated recurrent unit layer.

[0074] The number of filters in the convolutional layer reflects the richness of the feature patterns that the model can learn during the convolution process. A smaller number of filters may not be enough to fully capture the complex features of the data, while too many filters may introduce unnecessary computational burden and increase the risk of overfitting. This embodiment selects three representative values ​​of 32, 64, and 128 to find the best balance by comparing their performance in different situations.

[0075] The number of neural units in the bidirectional gated recurrent unit layer is a key parameter that determines the model's ability to process time series data. A sufficient number of units ensures that the model has enough memory capacity to capture long-term dependencies in the sequence, but too many units will also increase the complexity of the model, which may cause the training process to become unstable and prone to overfitting. Set three different unit number options of 50, 100, and 150, and compare the model performance under these options to determine the bidirectional gated recurrent unit layer configuration that best suits the current task.

[0076] The grid search method is used to systematically test the performance of different hyperparameter combinations, and the accuracy is used as the criterion for evaluating the quality of the parameters. By traversing all possible combinations and evaluating the accuracy of each combination, we can fully understand the impact of each hyperparameter on the model performance and determine the optimal hyperparameter combination. The specific steps are as follows:

[0077] s3.1. Construct a hyperparameter combination space, including the number of filters Filter size and the number of GRU units GRUunits:

[0078] Filter size∈{32,64,128},GRU units∈{50,100,150}

[0079] s3.2. Perform model training for each hyperparameter combination. In order to prevent the model from overfitting, a regularization term is introduced during the training process:

[0080]

[0081] Among them, λ is the regularization strength, w g is the model weight under the g-th hyperparameter combination, and S represents the total number of all trainable weight parameters in the model. Therefore, the size of S is directly related to the structure and complexity of the model. By adjusting the value of λ, L reg It is applied to the original loss function to balance the model's fitting ability and generalization ability. Then the accuracy of different combinations is evaluated, and the hyperparameter combination with the highest accuracy is selected as the final configuration. By comparison, it is found that when the number of filters is set to 64 and the number of GRU units is set to 100, the model's accuracy finally reaches 0.9948, so this combination is used as the final configuration.

[0082] In addition, the batch size will also affect the efficiency of training and the convergence speed of the model. A smaller batch size can improve the generalization ability of training, but it will increase the calculation time. In this embodiment, the best batch size is selected from 16, 32 and 64 as the final setting.

[0083] Step 4: Battery Failure Warning Strategy

[0084] like Figure 4 As shown, firstly, the basic threshold is set for each type of fault, as shown in Table 2:

[0085] Table 2 Basic threshold settings for various types of faults

[0086]

[0087] Where SDR represents the self-discharge rate of the battery, T u represents the temperature of the u-th temperature probe, U o It represents the voltage of the oth cell in the battery, IRR represents the insulation ratio of the battery, k1 is the total number of probes for measuring the temperature of the battery pack, and k2 is the number of battery cells.

[0088] Using the hybrid deep learning model trained in step 3, recursively predict the fault data in the next N minutes and generate the future T steps The predicted sequence of the jth type of fault at time step.

[0089] In the predicted future fault sequence, construct the indicator function δ t，j To mark whether abnormal data occurs, and according to δ t，j Construct abnormal fault set S t , when there is S t When it is not an empty set, a warning message is generated.

Claims

1. A new energy vehicle battery fault identification and early warning method based on hybrid deep learning collects the C-type feature data of the power battery during the operation of the new energy vehicle and performs preprocessing, which is characterized by: The following steps are involved: Step 1: Battery Fault Feature Marking Define the fault type of new energy vehicle battery, set the valid value range and fault threshold of each type of characteristic data, and generate the fault label data Y={y t，j |y t，j ∈{0,1}}∈R n×m , where m represents the number of fault types, n is the number of sampling time points, i.e. the number of samples; y t，j Indicates whether the jth type of fault occurs at time point t, 1 indicates the presence of such fault, 0 indicates the absence of such fault, t = 1, 2, ..., n, j = 1, 2, ..., m; Step 2: Extraction of key features of battery failure Calculate the feature data vector of the i-th category feature after preprocessing The fault label vector y of the j-th fault j Pearson correlation coefficient between Select the absolute value of the correlation coefficient Feature variables above the preset threshold As the key feature variable highly correlated with the j-th fault label, the key feature identification matrix Z = {z ij |z ij ∈{0,1}}∈R c×m , z ij =1 means that the i-th feature is the key feature of the j-th fault, and 0 means it is not a key feature; Step 3: Battery Fault Type Identification Model Construct a hybrid deep learning model that combines a convolutional neural network layer with a bidirectional gated recurrent unit layer, and input a matrix consisting of the key characteristic variables of the jth type of fault into the hybrid deep learning model. Predict target fault label vector y j =[y 1，j ,y 2，j , ..., y n，j ], and use the Sigmoid activation function to convert the output into a fault probability value of [0,1], and then use the fault threshold τ to get the fault label Binary cross entropy loss is used as the loss function for model training to handle the fault characteristics of each fault label appearing independently or together; Step 4: Battery Failure Warning Strategy s4.1、Fault warning setting Define a basic threshold e threshold,j , when the prediction value P of the j-th type of fault output by the hybrid deep learning model at time point t t，j Exceeding the basic threshold e threshold，j When , it is considered to have a failure risk; then a minimum warning point number k is set, requiring that when there is a failure risk at least k data points in the continuous time window, the failure warning is triggered; s4.

2. Predict future fault characteristic values The target characteristic data Q of the j-th type of fault collected t，j Input the trained hybrid deep learning model to get the first prediction result P t+1，j , P t+1，j Add to Q t，j At the end of t+1，j , input the trained hybrid deep learning model again for the next prediction; repeat the above steps until the prediction is T steps The predicted value for future time steps T steps =N×60 / t gap , t gap represents the sampling interval, and N represents the prediction duration; s4.

3. Detecting fault characteristics in future time periods In the predicted future time step feature sequence In the above example, for the predicted value of the j-th type of fault at each prediction time point t, an indicator function δ is defined t，j , record whether it exceeds the basic threshold e threshold : In the formula, δ t，j =1 means that the jth fault at the tth prediction time point exceeds the threshold and needs to be recorded as an abnormal fault state, and 0 means a normal state; s4.

4. Generate fault warning information For each prediction time point t, according to the indicator function δ t，j Construct abnormal fault set S t , including all abnormal failures that occurred at the corresponding time point: S t ={j|δ t，j =1} If there is an abnormal fault set An early warning is issued and corresponding warning information Info is generated, including the fault type, fault occurrence time and fault threshold.

2. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 1, characterized in that: Obtain the charging status, vehicle speed, accumulated mileage, battery voltage, battery current, SOC, insulation resistance, temperature, maximum voltage, minimum voltage, maximum temperature and minimum temperature of the power battery during the operation of new energy vehicles.

3. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 1, characterized in that: The specific steps of the pretreatment are: s1.

1. Clean the acquired data, remove outliers that exceed the normal range of the features, use the mean or interpolation method to fill in missing values ​​to ensure data integrity, and finally delete duplicate sample rows; s1.2, after the outliers are removed and the missing values ​​are processed, the feature data X = [x1, x2, ..., x n ]∈R c×n Use Savitzky-Golay filter to smooth the data; generate the smooth data of the i-th feature at time t i = 1, 2, ..., c, where c is the total number of sampled battery feature types; s1.3, smooth the data set Standardization and normalization operations are performed so that all eigenvalues ​​related to new energy vehicle battery failure fall within the same numerical range [0,1], with a mean of 0 and a variance of 1.

4. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 1, characterized in that: The failure types of new energy vehicle batteries are defined as including abnormal self-discharge, sudden internal short circuit caused by high voltage, sudden internal short circuit caused by high temperature and insulation failure.

5. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 4, characterized in that: When the self-discharge rate SDR>3%, it is marked as abnormal self-discharge state; When the battery temperature exceeds 38°C or the voltage is less than 2.5V, it is marked as a sudden internal short circuit abnormal state caused by high temperature; When the voltage is greater than 4.1V, it is marked as a sudden internal short circuit abnormal state caused by high voltage; When the insulation ratio IRR ≤ ≤ 100Ω / V when the battery is in a charging state, or when the insulation ratio IRR ≤ ≤ 500Ω / V in other states, it is marked as an insulation failure state.

6. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 1, characterized in that: The convolutional neural network layer includes a convolutional layer, an activation function and a pooling layer. j Perform feature extraction to capture local patterns and spatial relationships between features; each element in the output matrix C generated by the convolution operation is processed by the ReLU activation function to obtain the matrix C' after nonlinear transformation, which can effectively suppress negative features and make the model focus on positive features, which helps the model learn complex patterns and avoid the problem of gradient disappearance; C' contains multiple channels, each channel represents a different feature map extracted by the convolution kernel, and these channels capture the patterns and features of the input data in the local space; on each channel of the feature map C', use the maximum pooling operation to reduce the dimension of the feature map, extract the maximum value of the channel, and use the maximum value of each channel max(c1), max(c2),..., max(c q ) are combined into a maximum value vector max(c), where q represents the number of channels generated by the convolution operation, that is, the number of convolution kernels; The bidirectional gated recurrent unit layer is used to further process the feature data output by the convolutional neural network layer, and uses the bidirectional information flow to capture the front-to-backward dependency of the sequence data, thereby improving the model's ability to understand and extract time-dependent features; the BiGRU structure processes the data in two directions, forward GRU and backward GRU, providing a comprehensive feature representation for each time step, and then performs a forward hidden state analysis on each time step. and the backward hidden state Merge to form the final hidden state representation h t : The output of BiGRU passes through the fully connected layer, and the hidden states of all time steps captured by BiGRU are summarized to form a high-level feature vector set h = {h1, h2, ..., h n } is mapped to the target fault label vector y j =[y 1，j ,y 2，j , ..., y n，j ], use the Sigmoid activation function to convert the output into a fault probability value of [0, 1], and then use the fault threshold τ to get the fault label 7. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 1, characterized in that: Set the number of filters in the convolutional layer to 32, 64, or 128, the number of neural units in the bidirectional gated recurrent unit layer to 50, 100, or 150, and the batch size to 16, 32, or 64. Use the grid search method to test the accuracy of the hybrid model under different hyperparameter combinations, and select the hyperparameter combination with the highest hybrid model accuracy as the final model hyperparameter.

8. The new energy vehicle battery fault identification and early warning method based on hybrid deep learning as claimed in claim 7, characterized in that: Set the number of filters to 64 and the number of GRU units to 100.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 8.

Images