A method and system for early fault diagnosis of power batteries of electric vehicles based on a deep residual network

Abstract

This invention discloses a method and system for early fault diagnosis of electric vehicle power batteries based on deep residual networks, belonging to the fields of new energy vehicle technology, battery management systems, and artificial intelligence applications. The method includes the following core steps: First, the on-board battery management system (BMS) collects real-time time-series data of multi-dimensional physical parameters of each individual battery cell within the power battery pack, including but not limited to voltage, current, and temperature; second, by transforming the time-series data into a two-dimensional matrix resembling an image, this invention fully utilizes the powerful spatial feature extraction capabilities of deep convolutional networks. Simultaneously, by leveraging the residual network structure, it effectively overcomes the gradient degradation problem in deep network training, achieving end-to-end automatic learning and high-precision identification of early battery fault features. This significantly improves the predictability and accuracy of fault diagnosis, providing key technical support for ensuring the safe operation of electric vehicles.

Description

Technical Field

[0001] This invention relates to the core technology of electric vehicles, specifically to the state monitoring and safety management technology of power battery systems, and in particular to a method and related system for high-precision diagnosis and prediction of early and potential faults of lithium-ion power batteries in electric vehicles using deep learning theory. Background Technology

[0002] With the global energy structure transformation and increasingly stringent environmental protection regulations, the new energy vehicle industry, represented by battery electric vehicles (BEVs), is experiencing unprecedented rapid development. As the sole energy source for electric vehicles, the performance, lifespan, and safety of the power battery system directly determine the core competitiveness of the entire vehicle and the user's driving experience. However, a power battery pack, composed of hundreds or thousands of individual cells connected in series and parallel, is a complex electrochemical-thermal-mechanical coupled system. Under harsh real-world conditions such as charging and discharging, vibration, and temperature changes, the battery can experience irreversible performance degradation, and may even induce extreme safety accidents such as thermal runaway, combustion, and explosion, posing a serious threat to the life and property safety of passengers.

[0003] Therefore, the core task of a battery management system (BMS) is to be able to monitor the health status of power batteries in real time and issue timely warnings when faults occur in their early stages, i.e., before parameters exceed normal safety thresholds but show an abnormal evolution trend. This is also a current technical challenge and research hotspot in the industry. Currently, operators (including vehicle drivers, maintenance technicians, and back-end monitoring personnel) mainly face the following specific difficulties when dealing with power battery faults:

[0004] Currently, most commercial BMS systems employ alarm strategies based on fixed thresholds. This method is logically simple: an alarm is triggered only when the instantaneous value of a critical parameter (such as cell voltage or temperature) exceeds a pre-set, absolute safety upper or lower limit. Essentially, this is a passive response mechanism, not proactive prediction. However, the evolution of many serious faults (such as internal short circuits) is very subtle in its early stages, with extremely weak changes in external characteristics such as voltage and temperature, far from reaching the alarm threshold. By the time the parameter finally exceeds the threshold, the fault has often progressed to an irreversible stage, leaving drivers and passengers with very little reaction time, and thermal runaway may have already occurred. For example, Chinese patent CN10XXXXXX5A discloses a judgment method based on voltage and temperature thresholds, but it can only judge under obvious operating conditions such as overcharging, over-discharging, and over-temperature. It is powerless against early faults caused by subtle changes in internal electrochemical characteristics, resulting in a large number of missed alarms in practical applications and limited safety assurance capabilities. Summary of the Invention

[0005] This invention aims to overcome the aforementioned shortcomings of existing technologies and provide a method and system for early fault diagnosis of electric vehicle power batteries based on deep residual networks. This method addresses core technical problems in existing technologies, such as poor predictability of threshold-based methods, strong dependence of model-based methods on model accuracy and operating conditions, weak feature extraction capabilities of traditional machine learning methods, and reliance on manual design. The core objective of this invention is to create an intelligent diagnostic model capable of automatically learning and extracting deep, nonlinear, and spatiotemporally coupled features closely related to early faults from raw multidimensional time-series data collected by the battery management system (BMS) end-to-end. This enables highly sensitive, accurate, and robust early identification and warning of various potential faults in power batteries.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] A method for early fault diagnosis of electric vehicle power batteries based on deep residual networks includes the following steps: Step 1: Data acquisition and preprocessing, acquiring multi-dimensional state parameter time-series data during the operation of the power battery system in real time. The multi-dimensional state parameter time-series data includes the voltage sequence, temperature sequence, and total current sequence of a preset number of battery cells in the battery pack; cleaning and normalizing the acquired multi-dimensional state parameter time-series data, and using a fixed-length sliding time window to truncate the processed data sequence to generate multiple data sample segments; Step 2: Feature matrix reconstruction, for each The data sample fragments are reconstructed from a multi-channel one-dimensional time-series format into a two-dimensional feature matrix according to a preset mapping rule. The mapping rule is as follows: T consecutive sampling times within a time window are used as columns of the matrix, and M different state parameter channels (such as the voltage or temperature of M individual units) are used as rows of the matrix, thereby generating a two-dimensional feature matrix with dimension M×T. Step 3: Construct a deep residual diagnostic network. Design and construct a deep residual network model that takes the two-dimensional feature matrix as input. The network model structure includes, in sequence: an initial convolutional module, at least one residual learning stage, and a final convolutional module. The classification output module; wherein each residual learning stage consists of multiple cascaded residual blocks, the core of each residual block is a nonlinear transformation unit F(x) containing shortcut connections, and its output H(x) satisfies the relation H(x)=F(x)+x, where x is the input of the residual block; Step 4: Offline training of the model, using a historical dataset with accurate fault category labels to perform supervised training of the deep residual diagnosis network; by setting the cross-entropy loss function as the optimization objective, and using the gradient descent optimization algorithm to iteratively update the network weight parameters until the value of the loss function converges to Within a preset tolerance range, a trained fault diagnosis model is obtained; Step 5: Online fault diagnosis and early warning. The two-dimensional feature matrix generated by the vehicle system in real time and processed by Step 1 and Step 2 is input into the fault diagnosis model trained in Step 4 for forward inference calculation; The model outputs a vector, where each element of the vector corresponds to the confidence probability of the battery system being in a normal state or each preset early fault category; The confidence probability is compared with the preset alarm threshold, and when the probability of any fault category exceeds its corresponding threshold, the corresponding early fault diagnosis and early warning signal is output.

[0008] Preferably, the multidimensional state parameter time series data in step one further includes the total voltage, state of charge (SOC) estimation sequence, and insulation resistance value sequence of the battery system; the data cleaning includes interpolation filling for missing values ​​and outlier removal for abnormal values; the numerical normalization process adopts the min-max normalization method to linearly map all parameter values ​​to the interval [0,1], and its calculation formula is: X_norm=(X-X_min) / (X_max-X_min), where X is the original data, and X_min and X_max are the minimum and maximum values ​​of the parameter in the training dataset, respectively.

[0009] Preferably, in the construction of the two-dimensional feature matrix M×T in step two, M is the product of the total number of battery cells N and the number of selected features k for each cell, i.e., M=N×k; when k=2, the M rows are arranged in a predetermined order, for example, the first N rows are the voltage values ​​of N cells and the last N rows are the temperature values ​​of N cells; T is the length of the sliding time window, which ranges from 50 to 200 sampling points.

[0010] Preferably, in step three, the nonlinear transformation unit F(x) within the residual block is composed of two or three stacked convolutional layers, with each convolutional layer followed by a batch normalization layer and a rectified linear unit (ReLU) activation function layer; the shortcut connection is an identity mapping, and when the output dimension of F(x) does not match the dimension of the input x, a 1×1 convolution operation is applied on the shortcut connection path to perform dimension matching and linear projection.

[0011] Preferably, the convolutional layers within the residual block use a 3×3 kernel size and a stride of 1, and employ padding to maintain the spatial dimension of the feature maps before and after convolution. The batch normalization layer is calculated as follows: y=γ*((x-μ_B) / sqrt(σ_B^2+ε))+β, where μ_B and σ_B^2 are the mean and variance of the current batch data, respectively, γ and β are learnable scaling and translation parameters, and ε is a small constant to prevent division by zero.

[0012] Preferably, the overall architecture of the deep residual diagnostic network in step three is as follows: the input two-dimensional feature matrix first passes through an initial convolutional module with a kernel size of 7×7 and a stride of 2 for preliminary feature extraction and dimensionality reduction; then it passes through four residual learning stages in sequence, with each stage containing 3, 4, 6, and 3 residual blocks respectively, and the number of feature map channels in each stage increases by a factor of two; finally, a global average pooling layer is used to transform the high-dimensional feature map into a one-dimensional feature vector, which is then connected to a fully connected layer and a Softmax activation function to form the classification output module.

[0013] Preferably, the calculation formula for the cross-entropy loss function L in step four is: L(y, p) = - Σ(i=1 to C) [y_i * log(p_i)], where C is the total number of fault categories, y is the true label vector in one-hot encoding form, p is the predicted probability vector output by the model's Softmax layer, and y_i and p_i are the true value and predicted probability of the i-th category, respectively.

[0014] Preferably, the gradient descent optimization algorithm in step four is an adaptive moment estimation (Adam) optimizer, whose parameter update rule combines first-order moment estimation (momentum) and second-order moment estimation (adaptive learning rate). Its core update formula is: θ_t = θ_(t-1) - η * m_t / (sqrt(v_t) + ε), where θ is the network weight to be updated, η is the learning rate, and m_t and v_t are the bias correction estimates of momentum and velocity terms, respectively.

[0015] Preferably, the preset early fault categories in step five include at least: single cell overvoltage fault, single cell undervoltage fault, single cell overtemperature fault, abnormal temperature rise rate fault, increased battery pack inconsistency fault, and early internal short circuit characteristic fault; each fault category is assigned an independent alarm threshold P_th, which is optimized based on receiver operating characteristic curve (ROC) analysis to maximize the fault detection rate under an acceptable false alarm rate.

[0016] Preferably, the online fault diagnosis and early warning in step five further includes a multi-level alarm mechanism: when the output probability p_i of a certain fault category exceeds its alarm threshold P_th_i for the first time, a level one (indicative) alarm is triggered; if the output probability of the fault category continues to be higher than its threshold in the subsequent K consecutive diagnostic cycles, the alarm level is raised to level two (warning); if the probability value exceeds a higher severity threshold P_severe_i, the highest level (dangerous) alarm is directly triggered, and the vehicle's power limiting or safety protection strategy is linked.

[0017] Beneficial effects: Compared with the prior art, the method provided by the present invention brings one or more of the following significant and substantial beneficial effects:

[0018] In this invention, multidimensional time-series data is innovatively reconstructed into a two-dimensional feature matrix containing spatiotemporal information, and end-to-end feature learning is performed using a deep residual network. This enables the automatic and efficient extraction of deep, nonlinear abstract features related to fault mechanisms from massive, high-dimensional raw BMS data. As a result, the accuracy and sensitivity of early hidden fault identification in batteries can be significantly improved without relying on expert knowledge for tedious manual feature engineering.

[0019] This invention innovatively introduces residual blocks containing shortcut connections as the basic building blocks of the network, effectively solving the problems of gradient vanishing or exploding and network performance degradation that are prone to occur when the number of layers increases in traditional deep neural networks. It enables the successful construction and training of deeper network models to capture more subtle and long-term dynamic dependencies in battery state data, thereby significantly improving the generalization ability of the diagnostic model and the diagnostic robustness of batteries under different operating conditions and with different aging levels.

[0020] In this invention, by deploying a trained lightweight deep learning model on an in-vehicle terminal or cloud platform, efficient online analysis and forward inference of real-time data streams are performed, enabling continuous and forward-looking health monitoring of the entire life cycle of the power battery system. It can issue accurate warnings at the nascent stage when a fault changes from quantitative to qualitative, thereby transforming the traditional "passive response" alarm into "proactive prevention" diagnosis, buying valuable time for drivers and passengers to handle the situation, and greatly improving the overall operational safety of electric vehicles. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an early fault diagnosis method and system for electric vehicle power batteries based on deep residual networks, as proposed in this invention. Detailed Implementation

[0022] To more clearly and completely illustrate the technical solution of the present invention, the present invention will be further described below with reference to the accompanying drawings.

[0023] Example

[0024] like Figure 1 The diagram illustrates a complete implementation process of an early fault diagnosis method for a power battery pack in a pure electric passenger vehicle, as proposed in one embodiment of the present invention. The battery pack consists of 96 ternary lithium-ion battery cells with a rated capacity of 60Ah connected in series.

[0025] Step 1: Data Acquisition and Preprocessing

[0026] Data Acquisition: The following data is collected in real time via the vehicle's BMS system at a sampling frequency of 1Hz: voltage of 96 battery cells (V1, V2, ..., V96); temperature values ​​from 32 temperature sensors (T1, T2, ..., T32), with each sensor monitoring the temperature of three adjacent cells. For ease of processing, the temperature data is expanded to correspond to the temperature values ​​of the 96 cells using an interpolation algorithm; total battery pack current I (positive for charging, negative for discharging); and battery pack state of charge (SOC).

[0027] In this embodiment, the dataset used for model training and testing includes data accumulated over 5000 hours of operation for the vehicle model under various simulated operating conditions (such as urban congestion, highway cruising, and fast charging) and fault injection experiments (such as simulating local overheating through external heating and simulating early internal short circuits through parallel small resistors). Data cleaning: Outliers and missing values ​​in the data stream are checked. For obvious outliers caused by transient sensor failures (e.g., voltage spikes exceeding the physically possible range), the mean of neighboring points is used for replacement; for a small number of missing values, linear interpolation is used for filling. Data segmentation: The continuous data stream is segmented using a sliding time window of length T = 60 seconds (i.e., 60 sampling points). The time window slides forward by 10 seconds each time to generate data sample fragments with overlapping areas, enhancing data utilization. Data normalization: For each data channel (e.g., V1, V2, ..., I, SOC, etc.), use the maximum and minimum values ​​of that channel calculated in the training dataset to perform min-max normalization on all data (including subsequent online data), scaling the values ​​to the [0,1] interval.

[0028] Step 2: Feature Matrix Reconstruction

[0029] For each preprocessed 60-second data sample segment, a two-dimensional feature matrix is ​​constructed. In this embodiment, the voltage and temperature of each individual unit are selected as core features. Therefore, the total number of feature channels M = 96 (voltage) + 96 (temperature) = 192. The reconstruction rules are as follows: the i-th row of the matrix (1 ≤ i ≤ 96) corresponds to the voltage time series of the i-th individual unit. The j-th row of the matrix (97 ≤ j ≤ 192) corresponds to the temperature time series of the (j-96)-th individual unit. Each column of the matrix corresponds to a sampling time, for a total of 60 columns. Finally, each data sample segment is reconstructed into a 192×60 two-dimensional feature matrix. This structure integrates information from the time dimension and information from different variables (spatial dimension) onto a two-dimensional plane, allowing it to be processed efficiently by convolutional neural networks (CNNs) like images.

[0030] Step 3: Construct a deep residual diagnostic network

[0031] This embodiment employs a deep residual network structure similar to ResNet-34, adapted to the size of the input matrix. The network structure is as follows: Initial convolutional module: a 7×7 convolutional layer with 64 output channels, a stride of (2,2), and padding of (3,3). A batch normalization (BN) layer. A ReLU activation function. A 3×3 max pooling layer with a stride of (2,2). This module is used for preliminary feature extraction of the input matrix and to reduce the size of the feature map. Residual learning stage: The network contains four residual learning stages, each consisting of multiple residual blocks. Stage 1: Contains 3 residual blocks, each with 64 output channels in its convolutional layer. Stage 2: Contains 4 residual blocks, each with 128 output channels in its convolutional layer. The first residual block in this stage has a stride of 2 for downsampling. Stage 3: Contains 6 residual blocks, each with 256 output channels in its convolutional layer. The first residual block in this stage has a stride of 2. Stage 4: Contains 3 residual blocks, each with 512 output channels from its convolutional layer. The first residual block in this stage has a stride of 2. The internal structure of each residual block is: Conv(3x3)->BN->ReLU->Conv(3x3)->BN, then added to the input passed through the shortcut connection, and finally activated by a ReLU. For residual blocks that need downsampling or increased channel count, a 1×1 convolutional layer is added to its shortcut connection path to match the dimension. Classification output module: A global average pooling layer averages the feature map (size 512×6×2) output from the last residual stage in the spatial dimension to obtain a 512-dimensional feature vector. A fully connected layer maps the 512-dimensional vector to the dimension of the number of classes C. A Softmax activation function outputs the predicted probability for each class.

[0032] Step 4: Offline Model Training

[0033] Dataset and Labels: The dataset collected and processed in Step 1 was used. Data labels were added by battery experts based on experimental records and data analysis. This embodiment defines 5 categories (C=5): Category 0: Normal; Category 1: Cell overheating (the temperature of a cell is consistently slightly higher than the average ambient temperature by 3-5°C); Category 2: Increased battery pack inconsistency (the voltage difference within the battery pack continues to widen under specific operating conditions); Category 3: Early internal short circuit (the open-circuit voltage of a cell decreases at a small but stable rate); Category 4: Loose connection (the voltage of a cell exhibits abnormal high-frequency fluctuations during charging and discharging); The ratio of training set, validation set, and test set is 7:1.5:1.5. Training Process: Loss Function: Cross-entropy loss function was used. Optimizer: Adam optimizer was used, with an initial learning rate set to 0.001, momentum parameters β1=0.9, β2=0.999. Training Platform: A server with an NVIDIA Tesla V100 GPU was used, and the model was trained based on the PyTorch deep learning framework. Training parameters: Batch size set to 128, training for a total of 100 epochs. During training, monitor the loss and accuracy on the validation set, and save the weights of the model with the best performance.

[0034] Step 5: Online Fault Diagnosis and Early Warning

[0035] Model Deployment: The trained model weight files are optimized and quantized, converted to the lightweight ONNX format, and deployed in the vehicle domain controller or an in-vehicle information system (T-BOX) with edge computing capabilities. Online Diagnosis: The in-vehicle system executes steps one and two in real time, generating a 192×60 feature matrix, which is then input into the deployed diagnostic model every 10 seconds. The forward inference time of the model is optimized to less than 50 milliseconds on the in-vehicle chip, meeting real-time requirements. Warning Mechanism: Alarm threshold settings: P_th_1 (overtemperature) = 0.85; P_th_2 (inconsistency) = 0.90; P_th_3 (internal short circuit) = 0.80; P_th_4 (loose connection) = 0.95; these thresholds are obtained through ROC curve analysis of the test set, aiming to achieve a low false alarm rate with a high detection rate. Multi-level Alarm: Taking early internal short circuit as an example. Level 1 Alarm: When the model outputs a Class 3 probability p_3 exceeding 0.80 for the first time, the BMS backend system records a "suspected early internal short circuit" alert log and focuses attention on the vehicle. Level 2 Alarm: If p_3 remains greater than 0.80 for the next 5 diagnostic cycles (50 seconds), the system will issue a "Power Battery System Needs Inspection" warning to the driver via the vehicle's infotainment screen and send a Level 2 fault code to the 4S store or monitoring center. Level 3 Alarm: If p_3 exceeds the severe threshold P_severe_3=0.98 at any time, the system immediately triggers the highest level alarm, informing the driver to "stop safely and move away from the vehicle immediately" through sound and prominent visual prompts, while automatically limiting the vehicle's output power and executing the BMS's safety protection logic. To verify the effectiveness of the method of this invention, it was compared with three typical methods mentioned in the background technology. The test dataset was an independent dataset that had never been used for training or validation. Comparison Method: Method 1 (Threshold Method): Based on voltage and temperature safety thresholds set by national standards. Method 2 (ECM + Residual Method): Establish a second-order RC equivalent circuit model, use an extended Kalman filter (EKF) to estimate the terminal voltage, and use the residual between the actual measured voltage and the estimated voltage as a fault indicator. Method 3 (SVM Method): Manually extract 15 statistical features (such as voltage mean, variance, kurtosis, temperature range, etc.) from a 60-second data window, and use a support vector machine for classification.

[0036]

[0037] Results analysis:

[0038] The threshold method is almost entirely ineffective for the "early" faults defined in this embodiment, because at these stages, the absolute values ​​of parameters such as voltage and temperature do not exceed safe ranges. Its relatively high "overall accuracy" is due to the battery being in a normal state most of the time, allowing it to correctly identify "normal," but its fault detection capability (recall rate) is extremely poor, thus losing its significance as a warning system.

[0039] Comparative model method: The ECM + residual method has a certain diagnostic capability for faults with obvious external characteristics (such as over-temperature and inconsistency), but for early internal short circuits with extremely weak characteristics, the signal-to-noise ratio of the residual signal is very low due to model error and measurement noise interference, resulting in a significant performance degradation.

[0040] Compared to shallow machine learning methods, the SVM method shows improved performance compared to model-based methods, demonstrating the potential of data-driven approaches. However, its performance is still significantly lower than the method of this invention, especially in identifying complex faults such as early internal short circuits. This fully demonstrates that manually designed statistical features cannot fully capture the deep dynamic patterns of faults, while the deep residual network of this invention can automatically learn more discriminative feature representations.

[0041] Advantages of the method of this invention: As can be seen from the table data, the method of this invention significantly outperforms all comparative methods in Precision, Recall, and F1-Score across all four early fault categories, achieving a comprehensive diagnostic accuracy of 98.5%, demonstrating superior diagnostic performance. Particularly noteworthy is its F1 score of 0.92 in the most difficult-to-diagnose "early internal short circuit" fault, far exceeding other methods, highlighting its significant advantage in identifying concealed and nascent faults.

[0042] Finally, it should be noted that the basic concepts have been described above. Obviously, for those skilled in the art, the detailed disclosure above is merely illustrative and does not constitute a limitation of this specification. Although not explicitly stated here, those skilled in the art may make various modifications, improvements, and corrections to this specification. Such modifications, improvements, and corrections are suggested in this specification, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this specification. Furthermore, this specification uses specific terms to describe embodiments of this specification. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a feature, structure, or characteristic associated with at least one embodiment of this specification. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different locations in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this specification can be appropriately combined. Moreover, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this specification are not intended to limit the order of the processes and methods of this specification.

[0043] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early fault diagnosis of electric vehicle power batteries based on deep residual networks, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing. Real-time acquisition of multi-dimensional state parameter time-series data during the operation of the power battery system. This multi-dimensional state parameter time-series data includes the voltage sequence, temperature sequence, and total current sequence of a predetermined number of battery cells within the battery pack. The acquired multi-dimensional state parameter time-series data is then cleaned and normalized. A fixed-length sliding time window is used to truncate the processed data sequence, generating multiple data sample segments. Step 2: Feature matrix reconstruction. For each data sample segment, reconstruct it from a multi-channel one-dimensional time series form into a two-dimensional feature matrix according to a preset mapping rule. The mapping rule is as follows: take T consecutive sampling times within the time window as the columns of the matrix, and take M different state parameter channels (such as the voltage or temperature of M individual units) as the rows of the matrix, thereby generating a two-dimensional feature matrix with dimension M×T. Step 3: Construct a deep residual diagnostic network. Design and construct a deep residual network model that takes the two-dimensional feature matrix as input. The network model structure sequentially includes: an initial convolutional module, at least one residual learning stage, and a final classification output module; wherein each residual learning stage consists of multiple cascaded residual blocks, and the core of each residual block is a nonlinear transformation unit F(x) containing shortcut connections, whose output H(x) satisfies the relation H(x) = F(x) + x; Where x is the input of the residual block; Step 4: Offline training of the model. The deep residual diagnostic network is trained in a supervised manner using a historical dataset with accurate fault category labels. The cross-entropy loss function is set as the optimization objective, and the gradient descent optimization algorithm is used to iteratively update the network weight parameters until the value of the loss function converges to the preset tolerance range, thereby obtaining a trained fault diagnosis model. Step 5: Online fault diagnosis and early warning. The two-dimensional feature matrix generated by the vehicle system in real time and processed by Step 1 and Step 2 is input into the fault diagnosis model trained in Step 4 for forward inference calculation. The model outputs a vector, where each element of the vector corresponds to the confidence probability of the battery system being in a normal state or to each preset early fault category. The confidence probability is compared with a preset alarm threshold. When the probability of any fault category exceeds its corresponding threshold, the corresponding early fault diagnosis and early warning signal is output.

2. The method according to claim 1, characterized in that, The multidimensional state parameter time series data in step one further includes the battery system's total voltage, estimated state of charge (SOC) sequence, and insulation resistance value sequence; the data cleaning includes interpolation filling for missing values ​​and outlier removal for abnormal values; the numerical normalization process uses the min-max normalization method to linearly map all parameter values ​​to the [0,1] interval, and its calculation formula is: X_norm=(X-X_min) / (X_max-X_min); Where X represents the original data, and X_min and X_max represent the minimum and maximum values ​​of the parameter in the training dataset, respectively.

3. The method according to claim 1 or 2, characterized in that, In the construction of the two-dimensional feature matrix M×T in step two, M is the product of the total number of battery cells N and the number of selected features k for each cell, i.e., M=N×k; when k=2, the M rows are arranged in a predetermined order, for example, the first N rows are the voltage values ​​of N cells and the last N rows are the temperature values ​​of N cells. T is the length of the sliding time window, which ranges from 50 to 200 sampling points.

4. The method according to claim 1, characterized in that, In step three, the nonlinear transformation unit F(x) within the residual block is composed of two or three stacked convolutional layers, with each convolutional layer followed by a batch normalization layer and a rectified linear unit (ReLU) activation function layer; the shortcut connection is an identity mapping, and when the output dimension of F(x) does not match the dimension of the input x, a 1×1 convolution operation is applied on the shortcut connection path to perform dimension matching and linear projection.

5. The method according to claim 4, characterized in that, The convolutional layers within the residual block use a 3×3 kernel size and a stride of 1, and employ padding to maintain the spatial dimension of the feature maps before and after convolution; the batch normalization layer is calculated using the following formula: y=γ*((x-μ_B) / sqrt(σ_B^2+ε))+β; Where μ_B and σ_B^2 are the mean and variance of the current batch of data, respectively, γ and β are learnable scaling and translation parameters, and ε is a small constant to prevent division by zero.

6. The method according to claim 1, characterized in that, The overall architecture of the deep residual diagnostic network in step three is as follows: the input two-dimensional feature matrix first passes through an initial convolutional module with a kernel size of 7×7 and a stride of 2 for preliminary feature extraction and dimensionality reduction; then it passes through four residual learning stages in sequence, with each stage containing 3, 4, 6, and 3 residual blocks respectively, and the number of feature map channels in each stage increases by a factor of two; finally, a global average pooling layer is used to transform the high-dimensional feature map into a one-dimensional feature vector, which is then connected to a fully connected layer and a Softmax activation function to form the classification output module.

7. The method according to claim 1, characterized in that, The formula for calculating the cross-entropy loss function L in step four is as follows: L(y,p)=-Σ(i=1toC)[y_i*log(p_i)]; Where C is the total number of fault categories, y is the true label vector in one-hot encoding form, p is the predicted probability vector output by the model's Softmax layer, and y_i and p_i are the true value and predicted probability of the i-th category, respectively.

8. The method according to claim 1 or 7, characterized in that, The gradient descent optimization algorithm in step four is the Adam adaptive moment estimator, whose parameter update rule combines first-order moment estimation (momentum) and second-order moment estimation (adaptive learning rate). Its core update formula is: θ_t=θ_(t-1)-η*m_t / (sqrt(v_t)+ε); Where θ is the network weight to be updated, η is the learning rate, and m_t and v_t are the bias correction estimates for momentum and velocity terms, respectively.

9. The method according to claim 1, characterized in that, The early fault categories preset in step five include at least: single cell overvoltage fault, single cell undervoltage fault, single cell overtemperature fault, abnormal temperature rise rate fault, increased battery pack inconsistency fault, and early internal short circuit characteristic fault; each fault category is assigned an independent alarm threshold P_th, which is optimized based on receiver operating characteristic curve (ROC) analysis to maximize the fault detection rate under an acceptable false alarm rate.

10. The method according to claim 1, characterized in that, The online fault diagnosis and early warning in step five further includes a multi-level alarm mechanism: when the output probability p_i of a certain fault category exceeds its alarm threshold P_th_i for the first time, a first-level (indicative) alarm is triggered; If the output probability of this fault category continues to be higher than its threshold in the subsequent K consecutive diagnostic cycles, the alarm level will be raised to level two (warning); if the probability value exceeds a higher severity threshold P_severe_i, the highest level (dangerous) alarm will be triggered directly, and the vehicle's power limiting or safety protection strategy will be linked.

Images