Battery system fault assessment and safety early warning method based on pre-training architecture

Abstract

The invention relates to the technical field of battery system fault assessment, in particular to a pre-training architecture-based battery system fault assessment and safety early warning method, which comprises the following steps of: S1, selecting charging and discharging data fragments from a cloud data platform; s2, deleting abnormal characters and invalid data; s3, segmenting the vehicle operation data to form time blocks; s4, performing dimension expansion on the segmented time blocks to form an input matrix X; s5, inputting the X into an encoder to obtain a potential representation feature Z; s6, the potential representation feature Z is flattened and then input into a prediction head; s7, performing large-scale pre-training on the encoder and the prediction head and fixing weight parameters; s8, designing a battery system fault evaluation model, and training the evaluation model by using the fault information label data; and S9, obtaining a fault classification result by using the battery system fault evaluation model, and performing fault judgment and early warning. According to the invention, the accuracy and efficiency of battery system fault diagnosis are improved, the calculation is light, and the generalization ability is strong.

Description

Technical Field

[0001] The present invention relates to the technical field of battery system fault assessment, and particularly to a battery system fault assessment and safety warning method based on a pre-trained architecture. Background Art

[0002] With the rapid development of electric vehicles and energy storage systems, the safety issues of battery systems have become increasingly prominent. Existing battery fault diagnosis methods mainly rely on electrochemical principles and the structure of battery systems, and conduct fault diagnosis by simulating the electrochemical reactions inside the battery. However, this method has many deficiencies: First, it is difficult to comprehensively simulate all the electrochemical reactions inside the battery, resulting in inaccurate diagnostic results; second, the fault diagnosis process is complex and requires a large amount of professional knowledge and time investment; finally, existing technologies have deficiencies in real-time monitoring and warning, and often cannot identify the abnormal state of the battery system in time, thus unable to effectively prevent the occurrence of battery faults and safety accidents.

[0003] With the continuous development of artificial intelligence technology, more and more intelligent technologies have made great progress in the automotive field. Therefore, developing an intelligent battery system fault warning method that can achieve real-time monitoring and warning and improve the accuracy and efficiency of diagnosis is of great significance for enhancing the safety of battery systems. Summary of the Invention

[0004] The present invention provides a battery system fault assessment and safety warning method based on a pre-trained architecture, aiming to improve the accuracy and efficiency of battery system fault diagnosis, and having advantages such as efficient data utilization, lightweight calculation, strong interpretability, and strong generalization ability, which can provide an effective solution for the battery system fault assessment task.

[0005] The present application provides the following technical solutions:

[0006] A battery system fault assessment and safety warning method based on a pre-trained architecture, comprising the following steps:

[0007] S1. Select data segments from the vehicle operation data in the cloud data platform;

[0008] S2. Delete abnormal characters and invalid data in the data segments;

[0009] S3. Segment the data segments to form several time blocks;

[0010] S4. Expand the dimensions of the segmented time blocks to form an input matrix X;

[0011] S5. Input X into an encoder to obtain a latent representation feature Z, and the encoder is based on the GPT-2 structure;

[0012] S6. Flatten the latent representation feature Z and input it into the prediction head to obtain the predicted time series;

[0013] S7. Conduct large-scale pre-training on the encoder and the prediction head and fix the weight parameters;

[0014] S8. Design a battery system fault assessment model and train the assessment model using a small amount of labeled data containing fault information;

[0015] S9. Use the battery system fault assessment model to obtain the fault classification result and the fault probability value, and perform early warning push according to the judgment rule.

[0016] Technical principle: Based on the vehicle operation data received by the cloud data platform, extract the charge and discharge data segments, obtain the input feature matrix through data preprocessing and dimension expansion, use the encoder to extract the latent representation features of the time series data, obtain the predicted time series through the prediction head, conduct large-scale pre-training on the architecture composed of the encoder and the prediction head and fix the weights, add a trained classifier to form a fault prediction model, and conduct fault early warning according to the preset judgment rule.

[0017] Beneficial effects: The generative architecture can be pre-trained by using a small amount of labeled data and a large amount of unlabeled data, so as to achieve efficient data utilization. In the field of battery system fault assessment, the labeled fault data is relatively small, but the unlabeled data can be effectively used for pre-training through contrastive learning, improving the generalization ability and accuracy of the model.

[0018] The input sequence is segmented into sequence segments of a certain length at a certain stride, so that the number of input tokens is greatly reduced, allowing the model to view a longer historical sequence when the computing resources are limited. At the same time, through time block processing, the local features and global trends in the time series data can be understood more clearly, which is very valuable for decision-making support in practical applications.

[0019] The entire architecture fully considers the characteristics of time series such as autocorrelation and periodicity during the training process, enabling the model to have strong generalization ability and be able to handle time series prediction tasks in different scenarios.

[0020] Furthermore, the vehicle operation data includes the single-cell voltage sequence, current, SOC, and temperature.

[0021] Furthermore, in S2, abnormal characters are identified by the Z-score or moving window statistical method, invalid data is identified by checking null values or numerical ranges, and linear interpolation is used for correction after deleting abnormal characters and invalid data.

[0022] Further, the encoder in S5 is based on the GPT-2 architecture and includes an input layer and GPT-2 Transformer blocks. The input layer adds positional encoding to the input matrix X to introduce the order information of the input sequence. The GPT-2 Transformer blocks include a multi-head self-attention layer, a feed-forward sub-layer, and a normalization layer, and the normalization layer is arranged before the multi-head self-attention layer and the feed-forward sub-layer.

[0023] Beneficial effects: The encoder design can effectively capture complex patterns and long-range dependencies in time series data when processing time series data, and ensure the stability and efficiency of the training process through the normalization layer, providing high-quality feature representations for downstream tasks, and greatly improving the overall performance and application scope of the model.

[0024] Further, S6 includes:

[0025] S6-1. Flatten the latent representation feature Z;

[0026] S6-2. Input the flattened latent feature into the prediction head to obtain the predicted time series.

[0027] Further, the prediction head includes a first layer, an intermediate layer, and an output layer. The first layer maps the flattened latent representation feature to a new feature space and uses ReLU as the activation function; the intermediate layer is used to provide stronger feature extraction ability, and the output layer uses the softmax function to generate the prediction result, and the prediction result is the fault probability distribution.

[0028] Beneficial effects: The prediction head can effectively convert the latent representation feature extracted by the encoder into the fault probability distribution, providing clear and easy-to-understand data results.

[0029] Further, S7 extracts a large amount of historical operation data of normal vehicles from the cloud data platform, preprocesses it into time series data according to the steps of S1 to S3, and uses data augmentation technology to increase the diversity of data to form a large-scale data set; trains the encoder and the prediction head on the large-scale data set and fixes the weight parameters.

[0030] Beneficial effects: The prediction architecture trained using the large-scale data set has high accuracy and stability, and can provide reliable fault warnings and diagnosis results; at the same time, this method is not only applicable to the current task requirements, but also has flexibility and scalability. When facing new specific tasks, the fixed encoder and prediction head can be fine-tuned or directly used without retraining the entire model, shortening the development cycle.

[0031] Further, S8 includes:

[0032] S8-1. Design a battery system fault classifier;

[0033] S8-2. Train the classifier using the labeled data containing fault information and fix the weights of the classifier.

[0034] Beneficial effects: This method can not only make full use of the advantages brought by large-scale pre-training, but also achieve efficient fault detection and classification under limited resources. As more labeled data accumulates, the classifier can be retrained and fine-tuned regularly to continuously optimize the model performance.

[0035] Furthermore, the battery system fault classifier includes a conversion layer, a feature extraction module, and a fully connected layer. The conversion layer converts the latent feature Z into a shape suitable for two-dimensional convolution operations. The feature extraction module is used to further extract features. The fully connected layer outputs the prediction result. The feature extraction module is based on the CNN+Resnet structure and includes an initial convolution layer, residual blocks, and a global average pooling layer. The residual blocks adopt a bottleneck structure.

[0036] Beneficial effects: This classifier can achieve high-precision real-time classification of various fault types in the battery system. The bottleneck structure and the global average pooling layer can reduce the number of parameters and improve the utilization rate of computing resources, enabling the model to reduce the computational burden while maintaining high performance.

[0037] Furthermore, the fault classification includes self-discharge anomaly, connection anomaly, capacity anomaly, and sampling anomaly. The judgment rule is that when three consecutive data segments are classified as the same type of fault and the fault probabilities are all greater than 0.95, it is determined that the vehicle has a fault and a warning push is made.

[0038] Beneficial effects: Using consecutive data segments for verification instead of relying on data at a single time point can filter out noise or short-term fluctuations caused by accidental factors to a certain extent, thus more accurately capturing actual faults. At the same time, it can effectively prevent false negatives caused by the uncertainty of a single data point, ensuring that all important faults can be detected in a timely manner, which is of great significance for ensuring the safe and reliable operation of the battery system. Description of the Drawings

[0039] Figure 1 It is a flowchart of a battery system fault assessment and safety warning method based on a pre-trained architecture;

[0040] Figure 2 It is a schematic structural diagram of the pre-trained architecture. Detailed Embodiments

[0041] In the "Technical Specification for Electric Vehicle Remote Service and Management System", it is clearly required to establish a three-level new energy vehicle monitoring platform at the national, enterprise and local levels. Vehicle status data is collected through in-vehicle terminals and uploaded to the cloud monitoring platform. The monitoring platform can provide functions such as real-time monitoring of vehicle information, query of vehicle condition data, data analysis and management, and intelligent warning. It supports the query and analysis of various in-vehicle operation data including battery voltage, temperature, BMS, VCU, MCU, fuel cell, engine, etc., providing a strong guarantee for the safe operation of new energy vehicles.

[0042] The present invention provides a method for battery system fault assessment and safety warning based on a pre-trained architecture. With the help of a cloud data platform, it can achieve intelligent battery system fault diagnosis and warning, which is of great significance for improving the safety of new energy vehicles.

[0043] The following is a further detailed description through specific embodiments:

[0044] Embodiment 1

[0045] A method for battery system fault assessment and safety warning based on a pre-trained architecture, as Figure 1 shown, includes the following steps:

[0046] S1. Select charge and discharge segments from the vehicle operation data on the cloud data platform where the voltage-current correlation is greater than 0.8 and the SOC difference is greater than 30;

[0047] The cloud data platform can receive a variety of vehicle real-time operation data collected by vehicle sensors, covering various state monitoring such as charging, driving, and parking. Battery system-related vehicle operation data can be conveniently extracted from the database of the cloud data platform.

[0048] The state of charge SOC is a key indicator to measure the remaining battery charge. When the SOC difference is greater than 30%, it represents a significant change in the charging or discharging state of the battery; at the same time, the high correlation between voltage and current indicates that in these data segments, the working state of the battery is relatively consistent and stable. In the analysis process of the battery system, the main components in the equivalent circuit are regarded as resistors, and there are no other non-linear components such as capacitors, inductors, diodes, etc. At this time, the correlation between voltage data and current data is usually very obvious. In this case, a correlation greater than 0.8 is a reasonable result. Combining the change of SOC and the correlation analysis of voltage and current can comprehensively evaluate the state of the battery and provide the accuracy and reliability of fault warning.

[0049] Therefore, by selecting segments of the charging and discharging phases from the vehicle operation data and retaining the data segments where the difference in SOC is greater than 30 and the correlation between voltage and current is greater than 0.8 as the original data segments for fault warning, it can effectively help capture the key information that is most likely to indicate battery health problems.

[0050] S2. Delete the abnormal characters and invalid data in the vehicle operation data.

[0051] Select the required type of vehicle operation data from the original data segments, including the single-cell voltage sequence, current, SOC, temperature, etc.

[0052] Due to vehicle sensor failures or communication problems, the recorded data may contain abnormal characters, invalid data, and missing values, and these outliers need to be processed. Identify abnormal characters through methods such as Z-score and moving window statistics, delete the abnormal characters, find and delete invalid data by checking for null values and numerical ranges, and use linear interpolation for correction.

[0053] S3. Segment the vehicle operation data to form several time blocks.

[0054] Since the vehicle operation data is presented as time series data, the original time series can be divided into multiple subsequences by defining the length and step size of the time block (patch), which helps better adapt to the analysis model.

[0055] During the process of segmenting the time blocks, it can be selected whether the time blocks overlap, that is, whether there is a certain common part between adjacent time blocks. Overlapping time blocks can significantly increase the number of training sample pairs, improve the generalization ability of the model, and better capture the local patterns and features in the time series, but it will increase the demand for computing resources. Non-overlapping time blocks avoid the redundant calculations brought by the overlapping parts, reduce the computational amount, and are suitable for tasks that do not require high precision, but may result in insufficient comprehensive feature capture.

[0056] In this embodiment, the overlapping time block method is adopted. If the length of the time block is P and the step size is S, then the number of time blocks N can be expressed as:

[0057]

[0058] where L is the length of the original data segment.

[0059] In this embodiment, the length of the time block P is set to 100 and the step size S is set to 50.

[0060] S4. Expand the dimension of the segmented time blocks to form the input matrix X.

[0061] The original one-dimensional time series data is converted into a multi-dimensional feature matrix through dimensionality expansion operations, enabling the model to capture more complex patterns and features and improve prediction accuracy. At the same time, the dimensionality expansion operation can ensure that each time block has the same dimension, meeting the input requirements of the model; in addition, the dimensionality expansion operation can generate more training samples, increase the diversity of the data, thereby enhancing the generalization ability of the model and reducing the risk of overfitting.

[0062] The segmented time blocks are passed through a one-dimensional convolution Conv1D to expand the original time series dimension to any dimension, forming an input matrix X, specifically including:

[0063] First, select suitable convolution parameters, including the kernel size, the number of kernels, the stride, and the boundary handling method. To capture local features, a smaller kernel such as 3 or 5 is usually selected. The number of kernels is selected according to the designed output dimension. The sliding stride is usually set to 1 to ensure capturing as many local features as possible. For boundary handling, it can be chosen whether to perform padding to adjust the size of the output feature map. In this embodiment, the selected kernel is 3, the number of kernels is 4, the sliding stride is 1, and the boundary handling selects the padding method.

[0064] According to the convolution parameters, apply the one-dimensional convolution operation to each time block to generate new feature vectors.

[0065] Arrange the dimension-expanded features in order to obtain an input matrix X.

[0066] S5. Input X into the encoder to obtain the latent representation feature Z, and the encoder is based on the GPT-2 structure.

[0067] The encoder based on GPT-2 has strong feature extraction capabilities. It captures the dependencies of time series data through the self-attention mechanism, thereby improving prediction accuracy. The encoder structure includes an input layer and GPT-2 Transformer blocks. Among them,

[0068] The input layer adds positional encoding to the input matrix to introduce the order information of the input sequence, ensuring that the model can understand the time order of the input data;

[0069] The GPT-2 Transformer blocks focus on different parts of the input sequence based on the multi-head self-attention mechanism. The main components include:

[0070] The multi-head self-attention mechanism, which contains multiple hidden layers, allows the model to focus on different parts of the input sequence and capture long-range dependencies;

[0071] Each attention layer is followed by a forward feedback sub-layer for further processing and transforming features;

[0072] The encoder uses the Pre-Norm structure, that is, the normalization layer is set before the multi-head self-attention and feed-forward sub-layers. This structure enables the model to perform normalization before each step of calculation, making the training process more stable, ensuring the stable propagation of gradients, and accelerating the training process.

[0073] Select an appropriate GPT-2 model size according to the complexity of the task and the available resources. Common options include GPT-2 Small (124M parameters), Medium (355M parameters), Large (774M parameters), and XL (1.5B parameters). At the same time, the number of layers and the size of the hidden layers also affect the demand for computing resources. In this embodiment, the GPT-2 Small model and the default 12 layers are selected, and the size of the hidden layer in each layer is 768.

[0074] S6. Flatten the latent representation feature Z and input it into the prediction head to obtain the predicted time series, including the following steps:

[0075] S6-1. Flatten the latent representation feature Z;

[0076] Since the output of GPT-2 is a three-dimensional tensor (batch size, sequence length, hidden state dimension), and subsequent processing such as fully connected layers usually requires two-dimensional input (batch size, number of features), the high-dimensional feature Z generated by the encoder is flattened to adapt to the input format of the subsequent processing layer.

[0077] S6-2. Input the flattened latent feature into the prediction head to obtain the predicted time series.

[0078] The prediction head includes one or more fully connected layers, whose main function is to adjust the output dimension to match the task requirements. Among them, the first layer of the fully connected layer maps the flattened latent representation feature to a new feature space and uses ReLU as the activation function. After the first layer, several intermediate layers are included to provide stronger feature extraction capabilities and use ReLU or other non-linear activation functions. The last layer is used to generate the prediction result. For the time series prediction task, the output result corresponds to the predicted value at a certain or certain future time points. Since this application faces a classification task, the Softmax function is used in the last layer to obtain the probability distribution of different types of faults. In this embodiment, the intermediate layer is set to 1 layer and the ReLU activation function is used.

[0079] Select the number of neurons in each layer according to the complexity of the task, generally set to a value between the number of original input features and the final output dimension. In this embodiment, the original input features include the monomer voltage sequence, current, SOC, and temperature, and the output fault categories include self-discharge anomaly, connection anomaly, capacity anomaly, and sampling anomaly. Therefore, the number of neurons in each layer is set to 4.

[0080] S7. Conduct large-scale pre-training on the encoder and prediction head and fix the weight parameters.

[0081] The entire pre-training architecture is as Figure 2 shown. After slicing and segmenting the original time series, it is input into the encoder module to obtain the latent representation feature Z, which is then flattened and input into the prediction head to obtain the predicted time series. The entire architecture is constrained by the mean squared error loss, and the loss function is shown as follows:

[0082]

[0083] where, y i is the true value, is the predicted value.

[0084] Extract a large amount of historical operation data of normal vehicles from the cloud data platform, preprocess it according to the steps of S1 to S3 to obtain time series data, and use data augmentation techniques to increase the diversity of the data to form a large-scale data set.

[0085] Configure the hyperparameters according to the selected pre-training architecture, define the objective function of self-supervised learning, train the selected model architecture on the large-scale data set, regularly check the change of the loss function and the performance of the model on the validation set, prevent overfitting and adjust hyperparameters such as the learning rate.

[0086] When the exit condition is met, end the pre-training step and fix the weight parameters in the pre-training architecture.

[0087] S8. Design a battery system fault assessment model and train the assessment model using a small amount of labeled data containing fault information. Specifically, it includes:

[0088] S8-1. Design a battery system fault classifier.

[0089] The structure of this classifier includes a conversion layer, a feature extraction module based on CNN+Resnet, and a fully connected layer for classification. The input latent feature Z is obtained using a fixed-weight encoder. The conversion layer first converts the latent feature Z into a shape suitable for two-dimensional convolution operations, further extracts features through the feature extraction module, and finally outputs the prediction result through the fully connected layer.

[0090] Among them, the feature extraction module includes an initial convolutional layer, residual blocks, and a global average pooling layer. The initial convolutional layer uses one or more initial convolutional layers to initially extract features. Behind each convolutional layer, there is a Batch Normalization layer and a ReLU activation function. The convolutional kernel size is set to 3x3, the stride is set to 1, and the padding is selected as SAME to keep the output size consistent. The feature extraction module contains multiple residual blocks. Each residual block includes 2 to 3 convolutional layers and corresponding Batch Normalization and ReLU activation functions. A bottleneck structure is adopted within each residual block to reduce the computational amount, that is, first use a 1x1 convolution to reduce the dimension, then use a 3x3 convolution for processing, and finally use a 1x1 convolution to restore the dimension. And the output of the last convolutional layer will be added to the original input fed into this block to help solve the problem of gradient disappearance. After the last layer of convolution, a global average pooling layer is applied to compress the spatial dimension of the feature map to 1x1 while retaining the channel dimension, which helps reduce the risk of overfitting. In this embodiment, the number of residual blocks is set to 4, and each residual block contains 2 convolutional layers inside.

[0091] Apply a fully connected layer to flatten the output after global average pooling and perform final classification. The fully connected layer uses the softmax function to output the probability values of various types of faults.

[0092] This design makes full use of the advantages of CNN and ResNet, can not only effectively extract complex patterns but also avoid the problem of gradient disappearance in deep networks, and is applicable to multi-class classification tasks such as battery system fault assessment. By reasonably setting hyperparameters, the model performance can be optimized according to specific application scenarios.

[0093] S8-2: Use a small amount of labeled data containing fault information to train the classifier and fix the weights of the classifier.

[0094] Select the operation data of faulty vehicles from the cloud database. These operation data contain various types of fault information, such as abnormal self-discharge, abnormal connection, abnormal capacity, and abnormal sampling. After preprocessing these data containing fault information, a fault data set is formed.

[0095] Use the fault data set to train the classifier, output the fault classification output result and the probability value of each class. After evaluating the performance of the classifier, fix the weights of the classifier to complete the construction of the entire battery system fault assessment model.

[0096] S9: Use the battery system fault assessment model to obtain the fault classification result and the fault probability value, and perform early warning push according to the judgment rule.

[0097] Use the battery system fault assessment model, take the preprocessed vehicle operation data segment as the input data, and output the fault classification result and the classification probability value of each type of fault.

[0098] The fault classification includes self-discharge anomaly, connection anomaly, capacity anomaly, and sampling anomaly, among which:

[0099] Self-discharge anomaly means that the power loss rate of the battery during static storage significantly accelerates and exceeds the normal range;

[0100] Connection anomaly refers to problems such as open circuit, short circuit, or poor contact in the electrical connection between individual cells in the battery pack, which affects the overall performance and safety of the battery pack;

[0101] Capacity anomaly means that the actual available capacity of the battery deviates from its nominal capacity, usually manifested as too fast capacity attenuation or obvious capacity inconsistency;

[0102] Sampling anomaly refers to the phenomenon of inaccurate or missing data when the battery management system (BMS) collects battery parameters (such as voltage, current, temperature, etc.), which may lead to incorrect system decisions.

[0103] Set judgment rules according to the fault classification results of data segments to give early warnings for vehicle faults. In this embodiment, when three consecutive segments are classified as the same type of fault and the fault probabilities are all greater than 0.95, it is determined that the vehicle has a fault and an early warning push is made.

[0104] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this embodiment case. Common knowledge such as specific structures and characteristics known in the solutions is not described in detail here. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be subject to the content of its claims, and the specific implementation manners described in the specification can be used to explain the content of the claims.

Claims

1. A battery system fault assessment and safety warning method based on a pre-training architecture, characterized in that: The following steps are involved: S1. Selecting charging and discharging segments from vehicle operation data from the cloud data platform; S2, deleting abnormal characters and invalid data in the charge and discharge segments; S3, dividing the charging and discharging segments into several time blocks; S4, expanding the dimension of the divided time blocks to form an input matrix X; S5. Input X into the encoder to obtain the potential representation feature Z, wherein the encoder is based on the GPT-2 structure; S6, flatten the potential representation feature Z and input it into the prediction head to obtain the predicted time series; S7, perform large-scale pre-training on the encoder and prediction head and fix the weight parameters; S8. Design a battery system fault assessment model and train the assessment model using label data containing fault information; S9. Use the battery system fault assessment model to obtain the fault classification results and fault probability values, and push warnings based on the judgment rules.

2. The method for battery system fault assessment and safety early warning based on a pre-training architecture according to claim 1, characterized in that: The vehicle operation data includes cell voltage sequence, current, SOC and temperature.

3. The battery system fault assessment and safety warning method based on a pre-training architecture according to claim 1 is characterized in that: The S2 identifies abnormal characters by Z-score or moving window statistics, identifies invalid data by checking null values ​​or numerical ranges, and corrects using linear interpolation after deleting abnormal characters and invalid data.

4. The method for battery system fault assessment and safety early warning based on a pre-training architecture according to claim 1, characterized in that: The encoder in S5 is based on the GPT-2 architecture, including an input layer and a GPT-2Transformer block, wherein the input layer adds position encoding to the input matrix X to introduce sequential information of the input sequence, and the GPT-2Transformer block includes a multi-head self-attention layer, a forward feedback sublayer and a normalization layer, wherein the normalization layer is arranged before the multi-head self-attention layer and the forward feedback sublayer.

5. The method for battery system fault assessment and safety early warning based on a pre-training architecture according to claim 1, characterized in that: The S6 includes: S6-1, flattening the potential representation feature Z; S6-2. Input the flattened latent features into the prediction head to obtain the predicted time series.

6. The method for battery system fault assessment and safety early warning based on a pre-training architecture according to claim 5 is characterized in that: The prediction head includes a first layer, an intermediate layer and an output layer. The first layer maps the flattened potential representation features to a new feature space and uses ReLU as an activation function. The intermediate layer is used to provide stronger feature extraction capabilities. The output layer uses a softmax function to generate a prediction result, which is a fault probability distribution.

7. The method for battery system fault assessment and safety early warning based on a pre-training architecture according to claim 1, characterized in that: The S7 extracts a large amount of historical operation data of normal vehicles from the cloud data platform, preprocesses them according to steps S1 to S3 to obtain time series data, and uses data enhancement technology to increase the diversity of the data to form a large-scale data set; the encoder and prediction head are trained on the large-scale data set and the weight parameters are fixed.

8. The battery system fault assessment and safety warning method based on a pre-training architecture according to claim 1 is characterized in that: The S8 includes: S8-1. Design a battery system fault classifier; S8-2. Use the label data containing fault information to train the classifier and fix the weight of the classifier.

9. A battery system fault assessment and safety warning method based on a pre-training architecture according to claim 8, characterized in that: The battery system fault classifier includes a conversion layer, a feature extraction module and a fully connected layer. The conversion layer converts the potential feature Z into a shape suitable for a two-dimensional convolution operation. The feature extraction module is used to further extract features. The fully connected layer outputs a prediction result. The feature extraction module is based on a CNN+Resnet structure and includes an initial convolution layer, a residual block and a global average pooling layer. The residual block adopts a bottleneck structure.

10. The method for battery system fault assessment and safety early warning based on a pre-training architecture according to claim 1, characterized in that: The fault classification includes self-discharge abnormality, connection abnormality, capacity abnormality, and sampling abnormality; the judgment rule is that when three consecutive data fragments are classified as the same type of fault and the fault probability is greater than 0.95, the battery is judged to be faulty and an early warning is pushed.

Images