A lithium battery thermal runaway early warning method based on a space-time graph neural network

Abstract

The application discloses a kind of vehicle lithium battery thermal runaway early warning method based on space-time graph neural network, belong to lithium battery management technical field, this method includes: the operating state data of vehicle lithium battery is collected and preprocessed, and standardization data sample is obtained;Extract the multivariate characteristics including voltage difference and frequency domain characteristics, and filter out the general feature subset by correlation analysis;Correlation-aware space-time graph learning network is constructed, the spatial features extracted from the general feature subset are residually fused with time features, and the space-time coupling features are output, then the space-time coupling features are mapped into thermal runaway risk prediction label;Finally, the prediction label statistical strategy based on sliding window is used for decision fusion, and the final early warning result is output.The application significantly improves the accuracy, robustness and real-time performance of thermal runaway early warning under complex conditions by deeply mining the internal coupling mechanism and early weak abnormal characteristics of the battery pack, and reduces the false alarm rate and the missed detection rate.

Description

Technical Field

[0001] This invention relates to the field of lithium battery management technology, and in particular to a method for early warning of thermal runaway in automotive lithium batteries based on spatiotemporal graph neural networks. Background Technology

[0002] With the widespread adoption of new energy vehicles, lithium-ion batteries, as their core power source, are facing increasingly prominent safety concerns. Under extreme conditions, lithium batteries may experience sporadic thermal runaway accidents without obvious collision characteristics, seriously threatening the lives and property of passengers. Currently, automotive lithium battery thermal runaway warning systems face two major technical bottlenecks: firstly, the onboard embedded computing power is insufficient to support highly complex algorithms; secondly, purely mechanistic models cannot meet the accuracy requirements of thermal runaway warning systems.

[0003] Currently, the main methods for diagnosing thermal runaway in automotive lithium batteries are as follows: (1) Model-based approach: By constructing a mathematical model that can reproduce the complex electrochemical dynamics and physical evolution of the battery with high fidelity, the internal state that cannot be directly observed is transformed into a mathematical expression that can be calculated. However, this approach sacrifices computational efficiency in pursuit of high fidelity, making it difficult to deploy in real time in the battery management system (BMS). Furthermore, the complex side reactions and multi-physics coupled aging phenomena inside the battery make it difficult for the model to achieve adaptive tracking throughout the entire life cycle, and it is very easy to cause false alarms or missed alarms due to model mismatch.

[0004] (2) Signal processing-based method: This method starts directly from the measurable signals of the battery (such as voltage, current and temperature), and realizes the identification of abnormal states by extracting the time domain and frequency domain features of the signals, such as wavelet transform, empirical mode decomposition or entropy analysis. However, this method is highly dependent on high signal-to-noise ratio and artificial feature engineering. The extracted time and frequency features lack internal physical mechanism interpretation. At the same time, feature aliasing is very easy to occur under complex vehicle noise interference, and the empirically driven fixed diagnostic threshold is difficult to adapt to the long-term aging of the battery, resulting in limited anti-interference ability and generalization performance.

[0005] (3) Knowledge-based approach: A logical reasoning system is constructed to diagnose faults by using expert experience, historical fault data or prior knowledge of the battery system. However, the construction process of this approach is highly dependent on the induction of human experience, which has the bottleneck of strong subjectivity and difficulty in knowledge acquisition. It is difficult to accurately describe the complex nonlinear physical evolution path inside the battery with discrete logical rules. Furthermore, the static reasoning mechanism lacks the ability to generalize and self-learn unknown fault modes. Once the actual operating conditions or fault types exceed the scope of the preset expert knowledge base, the diagnostic system is very likely to miss or fail.

[0006] In summary, despite significant progress in existing research, challenges remain in the reliability of automotive lithium-ion battery thermal runaway early warning technology under complex operating conditions, its ability to identify early and weak features, and its practical engineering application. The key challenge that needs to be addressed in this field is to move away from the extreme reliance on high-fidelity physical mechanisms and artificial prior thresholds, and to develop a deep time-series diagnostic framework that can adaptively capture the implicit coupling relationships of long-span multidimensional signals and strongly resist noise under complex operating conditions. Summary of the Invention

[0007] To address the shortcomings of the prior art, this invention provides a method for early warning of thermal runaway in automotive lithium batteries based on spatiotemporal graph neural networks. By deeply exploring the internal coupling mechanism of the battery pack and the hidden abnormal characteristics in the early stage of thermal runaway, it achieves accurate early warning of thermal runaway under complex operating conditions.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for early warning of thermal runaway in automotive lithium batteries based on spatiotemporal graph neural networks includes the following steps: Step 1: Collect the operating status data of the vehicle lithium battery, preprocess the operating status data to obtain standardized data samples; Step 2: Extract multidimensional features from the standardized data samples, including: direct voltage features, first-order voltage difference sequence and complex frequency domain features, and then screen out a subset of general features through correlation analysis; Step 3: Construct a binary classification early warning model based on a spatiotemporal graph neural network, specifically including: Each monitored variable in the general feature subset is defined as a node in the graph. The dynamic correlation between the monitored variables is adaptively learned from the input data through a graph builder to generate a dynamic adjacency matrix. Temporal features on each node are extracted through a gating mechanism and multi-scale convolution. Based on the adjacency matrix, spatial features between nodes are extracted through bidirectional graph convolution. The spatial features and the temporal features are fused with residuals to output spatiotemporal coupling features. Step 4: Output the thermal runaway risk prediction label for each moment based on the spatiotemporal coupling characteristics; Step 5: Post-process the thermal runaway risk prediction labels using a sliding window-based result statistics and confidence confirmation mechanism. When the proportion of thermal runaway risk prediction labels in the sliding window exceeds a preset threshold, a final thermal runaway warning is issued.

[0009] Further, the operational status data is preprocessed, including: The operational status data is cleaned to remove outliers; then, a sliding window technique that does not cross charging and discharging segments is used to process the cleaned operational status data to generate standardized data samples.

[0010] Furthermore, the sliding window technique that does not span charging / discharging segments includes: The start of consecutive charge / discharge segments is padded, and the length of the padded segment is the same as the length of the sliding window. Sliding sampling is performed within the labeled independent charge / discharge segments, and windows that span different charge / discharge segments are deleted.

[0011] Furthermore, the correlation analysis specifically includes: The Pearson correlation coefficient is used to quantify the degree of linear correlation, and mutual information is used to capture nonlinear dependencies. A correlation strength matrix between multiple features and voltage signals is constructed, and then feature selection is performed based on the preset correlation threshold and feature redundancy test results.

[0012] Furthermore, a dynamic adjacency matrix is ​​generated using a graph builder, specifically as follows: The binary classification early warning model passes adp Function to generate adjacency matrix The calculation formula is: ; in, gc Represents a graph constructor. , indicating the subgraph number.

[0013] Furthermore, temporal features at each node are extracted through gating mechanisms and multi-scale convolutions, including: Based on the gating mechanism of Gated Recurrent Unit (GRU), node features are split into multiple groups, and convolution operations are performed along the time dimension using convolution kernels of different sizes. Adaptive feature selection is achieved through value gate and update gate. The value gate compresses the feature values ​​to the [-1,1] interval using the hyperbolic tangent activation function to preserve the positive and negative changes of the features. The update gate generates weights in the [0,1] interval using the sigmoid activation function to control the retention ratio of features at each position.

[0014] Furthermore, step 3 also includes: training the binary classification early warning model, specifically, using a class-weighted binary cross-entropy loss function as the optimization objective for model training, the loss function being: ; in, N For the sample size, This represents the true label of the i-th sample. The model is passed through the Sigmoid function The output predicted probability, β These are the weighting coefficients.

[0015] Furthermore, step 5 specifically includes: Given a time sliding window of length W, at each sampling time t, backtrack all predicted labels from the past W time points within the window and calculate the proportion of labels representing fault risk appearing within the window. , ; In the formula, This represents the i-th predicted label, only when... A thermal runaway warning will only be issued at the current time t if the value is ≥0.5.

[0016] This invention also provides a thermal runaway early warning system for automotive lithium batteries based on spatiotemporal graph neural networks. This system is used to implement the aforementioned early warning method, including: The data acquisition and preprocessing module is used to acquire the operating status data of the vehicle lithium battery and preprocess the operating status data. The feature extraction module is used to extract multi-dimensional features from standardized data samples and perform correlation analysis and screening. The model building module is used to build a binary classification early warning model based on a spatiotemporal graph neural network; The online early warning module is used to process the spatiotemporal coupling feature data output by the model in real time and output thermal runaway early warning signals based on the sliding window mechanism.

[0017] Compared with existing technologies, the beneficial technical effects of this invention are as follows: The binary classification early warning model based on spatiotemporal graph neural networks constructed in this invention breaks through the limitation of traditional spatiotemporal graph neural networks that rely solely on fixed graph structures. By introducing a multivariate time series correlation learning module, the model can automatically and explicitly learn and capture the pairwise correlations of dynamic changes among various monitoring variables of the battery system from the data. This invention introduces first-order voltage difference sequences and complex frequency domain features, and constructs a general feature subset through multi-dimensional analysis that integrates Pearson correlation coefficients and mutual information, providing highly representative inputs for capturing early, weak signs.

[0018] This invention employs a sliding window technique that does not span charging / discharging segments to fill in the beginnings of consecutive segments. The length of the filled segment is the same as the length of the sliding window, ensuring that the window covers all the original data information during sliding. This invention also proposes a sliding window diagnostic technique based on predictive tags, further improving the robustness of the diagnostic system and reducing false alarms caused by transient noise or occasional interference. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the vehicle lithium battery thermal runaway early warning method of the present invention; Figure 2 This is a schematic diagram of the sliding window technology of the present invention that does not span charging and discharging segments; Figure 3 This is a thermogram showing the correlation between battery parameters in this invention. Figure 4 This is a diagram of the core structure of the relevance-aware spatiotemporal graph learning network of the present invention; Figure 5 This is a model architecture diagram of the relevance-aware spatiotemporal graph learning network of the present invention; Figure 6 This is a statistical diagram illustrating the thermal runaway early warning results based on the sliding window of the present invention. Figure 7 The diagram shows the early warning results of normal and faulty vehicles based on sliding window decision-making in this invention. (a) is the early warning output curve of normal vehicle No. 8, (b) is the early warning output curve of normal vehicle No. 29, (c) is the early warning output curve of faulty vehicle No. 54, and (d) is the early warning output curve of faulty vehicle No. 57. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] This invention provides a method for early warning of thermal runaway in automotive lithium batteries based on spatiotemporal graph neural networks, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect the operating status data of the vehicle lithium battery, preprocess the operating status data to obtain standardized data samples; In this embodiment, the sample data used is battery operating status data collected from 60 real vehicles. Each vehicle battery pack integrates 88 individual battery cells. At each preset sampling time point, the data acquisition module synchronously acquires real-time characteristic parameters including: the voltage values ​​of each of the 88 individual battery cells, the total bus current of the battery pack, the state of charge (SOC) of the battery system, and the highest and lowest battery temperatures monitored by a temperature sensor network. The aforementioned multidimensional characteristic data is arranged in a time series and used as input parameters for subsequent prediction models. Secondly, the data cleaning module identifies and removes outliers and abnormal values ​​caused by transmission interference in the original signal using a combination of sliding window statistics and quantile detection. For key signals such as voltage and current, the mean and standard deviation are calculated within the sliding window, and data points exceeding the range of "mean ± 3 times standard deviation (3σ)" are considered transient outliers and smoothed or removed.

[0024] For example, for a sliding window of length L The formula for the mean of a signal sequence within a window is: ; Indicates the first i Signal sample values ​​of a window sequence; The formula for the standard deviation within a window is: ; A data point is considered a transient outlier if it meets the following 3σ condition: ; For signals that change slowly, such as temperature, the box plot principle is used to identify data points that exceed 1.5 times the interquartile range above or below the upper and lower quartiles as outliers and make appropriate corrections. This can eliminate the interference of noise impulses on model training.

[0025] For example, for a temperature sequence within a window of a given dataset Its upper and lower quartiles, Q3 and Q1, are the 75th and 25th percentiles, respectively. Therefore, the interquartile range is: ; If data points A point is considered a statistical outlier if it meets the following conditions: ; Building upon this, to generate standard samples that can be directly input into the model, this embodiment also proposes a sliding window technique that does not span charging and discharging segments. (See attached document) Figure 2The technique involves padding the beginning of consecutive segments with a length equal to the sliding window length, ensuring the window covers all original data during sliding. Then, each window is evaluated; if it spans different charge / discharge segments, the window is deleted. This technique performs sliding sampling within labeled, independent charge / discharge segments, extracting consecutive subsequences as samples through a fixed-length sliding window. This design ensures that each training sample originates from a complete, physically meaningful microscopic process, effectively avoiding erroneous dynamic patterns introduced by windows crossing charge / discharge transition points, significantly improving sample purity and model learning efficiency.

[0026] The data preprocessing described above can transform raw, coarse vehicle data into a series of clean, continuous, clearly labeled, and physically meaningful standardized data samples. This provides a reliable, high-quality data foundation for subsequent model training and validation, and is an indispensable key step in building a high-precision fault diagnosis system.

[0027] Step 2: Extract multidimensional features from the standardized data samples, including: (1) Direct voltage feature extraction.

[0028] Direct voltage characteristics are the core foundational features for power battery state monitoring, encompassing the total voltage of the battery pack, the time series of individual cell voltages, and statistical representations (mean, standard deviation, range, coefficient of variation, etc.) extracted from these two types of signals. These features directly reflect the overall energy state of the battery pack and the consistency level of individual cells, providing a stable and physically meaningful input basis for subsequent diagnostic networks.

[0029] (2) Extraction of temporal and frequency domain features.

[0030] From a physical perspective, the voltage differential sequence directly characterizes the dynamic response capability of a power battery to external load fluctuations and sudden changes in charging current. Under transient conditions such as rapid acceleration, deceleration, and start-stop switching, the battery output current changes drastically, causing significant instantaneous fluctuations in the voltage differential sequence. The amplitude of these fluctuations is closely related to the battery's internal resistance and polarization characteristics. To achieve a quantitative characterization of the instantaneous rate of voltage change, the formula for calculating the first-order voltage differential sequence is: ; in, The first-order voltage difference at time k directly corresponds to the slope of the voltage time series curve at that time. This represents the sampled value of the battery voltage at time k. For the first The battery voltage sample value at time t, where This represents the time interval between two consecutive voltage samples.

[0031] Simultaneously, complex frequency domain feature extraction was performed on the voltage values. The time-domain voltage signal was mapped to the frequency domain using discrete Fourier transform to uncover the hidden periodic fluctuation patterns within the signal. For a length of... N Discrete voltage time series V The discrete Fourier transform is calculated using the following formula: ; ; in, j The imaginary unit, X ( m ) is the first m The complex value in the frequency domain corresponding to each frequency point, with the real part and imaginary part reflecting the in-phase component and quadrature component of that frequency component, respectively; V ( k () represents the time-domain voltage sample value at the k-th sampling time. N This represents the total number of sampling points for the time-domain voltage sequence. The maximum amplitude value in the frequency domain spectrum represents the intensity of the strongest periodic fluctuation in the voltage signal, and its corresponding frequency is the characteristic frequency of the voltage signal.

[0032] (3) Other feature screening and correlation analysis.

[0033] Other characteristics include environmental and load variables such as current, state of charge, temperature, and vehicle speed, which provide a complete operating context for voltage behavior. These characteristics are not isolated from the voltage signal, but rather form a strong coupling relationship through the electrochemical characteristics of the power battery and the external working mechanism. Based on this, this embodiment performs a multi-dimensional correlation analysis on these related operating condition characteristics and the derived time-series characteristics, complex frequency domain characteristics, and voltage values ​​calculated above.

[0034] For details, please refer to the appendix. Figure 3 In this embodiment, the Pearson correlation coefficient (PCC) is used to quantify the degree of linear correlation, and mutual information is combined to capture nonlinear dependencies, constructing a correlation strength matrix between multiple features and voltage signals. Subsequently, based on a preset correlation threshold (eliminating weakly correlated features) and feature redundancy test results, a subset of features that are highly correlated with voltage behavior patterns and have complementary information is further selected. This subset is then used as the input features for the subsequent diagnostic network, thereby improving the model's accuracy in predicting thermal runaway.

[0035] For thermal runaway faults, considering the complexity and nonlinearity of their evolution, basic electrical quantities alone are insufficient to capture early, subtle signs. Therefore, while retaining highly correlated SOC, mean temperature, and extreme values, we further introduce the first-order voltage difference, which reflects dynamic response characteristics, and frequency domain features (frequency centroid and root-mean-square frequency), which have been shown to be highly coupled with the long-term voltage state in the aforementioned analysis. This results in a general feature subset for thermal runaway early warning.

[0036] Step 3: Construction of a correlation-aware spatiotemporal graph learning network.

[0037] Reference Appendix Figure 4 A binary classification early warning model based on a spatiotemporal graph neural network is constructed. A temporal feature extraction module (such as a gated recurrent unit or a temporal convolutional network) independently encodes the temporal features of each node to capture its dynamic evolution. Based on this, a spatial feature extraction module (such as a graph convolutional network or a graph attention network) aggregates information from neighboring nodes according to a predefined system topology, effectively perceiving the spatial diffusion and mutual influence of fault features. Simultaneously, the network introduces a multi-scale temporal aggregation mechanism, processing features from different time windows in parallel or hierarchically to collaboratively perceive rapidly changing electrical signal anomalies and slowly changing thermal accumulation trends, thus comprehensively covering the multi-rate characteristics of thermal runaway evolution. This network is constructed as an end-to-end binary classifier, with its final output being 0 representing a normal system and 1 representing a detected risk of thermal runaway.

[0038] For details, please refer to the appendix. Figure 5 Conv2d represents a two-dimensional convolution operation used to extract local temporal features of the input sequence; filter represents a filtering branch used to generate candidate features; gate represents a gate branch used to generate gate weights between 0 and 1, thereby controlling the retention ratio of candidate features; Nx indicates that the structure is repeated N times. The initial node features before graph convolution. The node aggregation features are those propagated through the adjacency matrix; α is the feature fusion weight, used to balance the node's own information with the information propagated from its neighboring nodes, i.e., through... We use weighted fusion to preserve the original features and fuse spatially related features, thereby avoiding the problem of excessive smoothing during graph convolution.

[0039] First, for the power battery fault diagnosis task, this model first performs input dimension reconstruction and receptive field adaptation, reconstructing the input data dimension of the original regression task into the standard input dimension required by the spatio-temporal graph neural network model. To ensure the normal progress of the convolution operation and avoid convolution kernel out-of-bounds, when the length of the input sequence is less than the receptive field of the model, zero-padding is performed on the left side of the input sequence, and the padding length is the difference between the two, so that the input sequence length adapts to the receptive field requirements. The receptive field is calculated based on the concept of dilated convolution, reflecting the range of the time window that the model can capture. The calculation formula of the receptive field is: ; where K is the convolution kernel size, D is the dilation rate growth base, L 1 is the number of convolutional layers in the network. The "1" on the left side of the formula represents the initial receptive field length of a single time step itself before the convolution operation, l represents the convolutional layer index. In this embodiment, L 1 = 3. When T < RF, zero-padding is performed on the left side of the input sequence, and the padding length is RF - T to meet the minimum window requirement. The dilated convolution mechanism realizes the efficient capture of long-term dependencies by adjusting the dilation rate d: in the standard convolution (d = 1), the convolution kernel covers K consecutive time steps each time; when d = 2, the sampling interval of the convolution kernel is 1 time step, realizing the extraction of sparse but key long-range features.

[0040] Secondly, perform graph structure component and node embedding configuration. Under the framework of the graph convolutional network, the model generates the adjacency matrix adp through the function, and its calculation formula is: ; where gc represents the graph constructor, , represents the number of the subgraph, which can be manually adjusted to achieve subgraph modeling without global modeling of all sensor nodes. Combining the characteristics of the fixed geographical location of sensors in the power battery fault diagnosis scenario, the model uses a static embedding method to construct node relationships, which is suitable for scenarios where node attributes are known and relationships are fixed and does not involve dynamic node relationship updates of dynamic embedding. The core configurable parameters of the model include three: the subgraph scale is used to control the influence range of the neighbors of a single node. When the subgraph scale is equal to the total number of nodes, a fully connected graph is constructed, and all nodes affect each other; the node dimension represents the dimension of the node embedding, that is, the length of the feature vector corresponding to each node; the hyperbolic tangent coefficient α is used to adjust the activation function characteristics during the node embedding process. The construction of the adjacency matrix only focuses on the core configurable parameters and function implementation, and does not involve the specific internal matrix operation process, providing a basis for subsequent spatial feature extraction.

[0041] In addition, this embodiment also performs input feature transformation and gated temporal feature extraction. The input features are first transformed in dimension by the start_conv convolutional layer. The initial input [64,1,10,50] is expanded by 32 1×1 convolutional kernels, increasing the input dimension from 1 to the same as the residual channel, resulting in a feature map with dimensions [64,32,10,50]. The purpose is to provide rich hidden layer feature support for subsequent complex spatiotemporal convolutions. Subsequently, multi-scale temporal feature extraction is performed based on the GRU gating mechanism. The 32-dimensional features are split into 4 groups (8 dimensions in each group), and convolution operations are performed using 1×k convolutional kernels with kernel sizes of [2,3,6,7] (sliding only along the time dimension, not the node dimension). The output time steps of each convolutional kernel are 49, 48, 45, and 44, respectively. The dimension is unified by uniformly truncating the last 44 time steps. After concatenation, the 32-dimensional features are restored, and the output dimension is [64,32,10,44]. This process achieves adaptive feature selection through value gates and update gates. The value gate uses a hyperbolic tangent activation function to compress feature values ​​to the [-1, 1] interval to preserve the positive and negative changes in features. The update gate uses a sigmoid activation function to generate weights in the [0, 1] interval, controlling the retention ratio of features at each position. The two gate structures share function parameters but the weights are trained independently, and finally, element-wise multiplication is used to achieve gate fusion. Simultaneously, skip connections are set to solve the gradient vanishing problem. The initial skip connection compresses the input [64, 1, 10, 50] to [64, 64, 10, 1] through a convolutional layer, achieving channel expansion and temporal dimension aggregation while preserving global temporal information. Specifically, it is represented as follows:

[0042]

[0043]

[0044] Where F is the value gate output, representing the candidate feature values ​​after feature filtering, and the tanh activation function preserves the positive and negative direction information of the features. Z represents the initial feature tensor of the input sequence after dimensionality reconstruction and zero-padding, i.e., the input features of layer 0 before entering the temporally gated convolutional module. Z is the update gate output, representing the gate weights that control the feature retention ratio. The σ function generates weights from 0 to 1 to control the feature retention ratio. The output is the time feature, representing the final time dimension feature after fusion through the gating mechanism. This indicates element-wise multiplication. Refers to the one-dimensional convolution operation used to extract value gate features. This refers to the one-dimensional convolution operation used to extract the update gate weights. Simultaneously, the initial skip connection compresses the original input to... To preserve the global time context.

[0045] Finally, spatiotemporal feature fusion and output layer design are performed. After temporal feature extraction, the output layer is designed based on a pre-constructed adjacency matrix. adp Spatial features are extracted through bidirectional graph convolution, and two independent hybrid propagation instances, gconv1 and gconv2, are used to process the forward adjacency matrix respectively. adp and reverse adjacency matrix This achieves the fusion of forward (node-to-neighbor) and reverse (neighbor-to-node) spatial dependencies, and outputs... ; In the formula, This represents the reverse matrix of the adjacency matrix A; In the process of Graph Convolutional Network (GCN), self-loops are first added to the adjacency matrix. Row normalization is performed to avoid loss of node information. The graph convolution depth is set to 3, and the original feature retention coefficient is set to... Feature fusion is achieved through multi-hop propagation, and the formula for calculating the i-th hop feature is: ; in, This is the normalized adjacency matrix with self-links. This represents the feature of the previous hop. After concatenation of multi-hop features, the output is transformed by a multilayer perceptron, and the number of input and output channels remains at 32 to meet the requirements of residual connection.

[0046] Residual connections directly add the graph convolution output to the original temporal features, avoiding the loss of high-frequency details caused by deep convolutions. As the number of network layers increases, the receptive field parameters increase and the convolution kernel width decreases. Shallow layers preserve temporal details, while deeper layers capture abstract trends. Finally, the output of the last layer is compressed to [64,64,10,1] through a skip fusion layer and accumulated and fused with the output of the initial skip connections, ultimately outputting spatiotemporal coupled features.

[0047] Step 4: Output thermal runaway risk prediction labels for each moment based on spatiotemporal coupling features. Specifically, the output layer flattens the fused global spatiotemporal features into a one-dimensional vector, performs nonlinear transformation through a multilayer perceptron, and generates thermal runaway risk prediction labels that match the power battery fault diagnosis task, thus completing the mapping from spatiotemporal features to diagnostic results.

[0048] Step 5: Sliding window diagnostic technique based on predicted labels.

[0049] Reference Appendix Figure 6To further enhance the robustness of the diagnostic system and reduce false alarms caused by transient noise or occasional interference, this embodiment introduces a result statistics and confidence confirmation mechanism based on a sliding window. The core idea of ​​this mechanism is that the instantaneous output (single-point predicted label) of the diagnostic model is susceptible to random factors, while a real, continuous fault state will inevitably exhibit a stable abnormal pattern within a certain time range. Therefore, the system performs post-processing on the original predicted label sequence output by the model. Specifically, a time sliding window of length W is defined, and at each sampling time t, all predicted labels from the past W times within the backtracking window are considered. (in (0 represents normal, 1 represents fault risk) Calculate the proportion of labels representing fault risk appearing within the calculation window. , ; This proportion This reflects the confidence level of the recent model's assessment that the system is in a risky state. Subsequently, a decision threshold is set (in this embodiment, the decision threshold is set to 50%), and only when... A thermal runaway warning will only be issued at the current time t if the ratio is ≥0.5. Otherwise, even if the current instantaneous prediction is "1", it will be regarded as occasional noise and suppressed. This "proportional threshold" strategy is equivalent to performing a time-dimensional low-pass filter on the original binary classification results, ensuring that only those continuous and stable abnormal signals can trigger the final alarm, thereby significantly reducing the false alarm rate.

[0050] In its implementation, this invention employs common deep learning frameworks (such as PyTorch) for model training and evaluation to ensure consistency of experimental conditions and comparability of results. To comprehensively evaluate the model's generalization ability and stability under different data distributions, this invention designs a systematic cross-validation experiment. The specific grouping scheme is shown in Table 1, with a total of 10 independent experimental groups. In each group, the test set consists of data from 5 normal vehicles and 1 vehicle confirmed to have experienced thermal runaway; the training set consists of data from all other vehicles in the dataset except for those in the test set. By rotating the combination of faulty and normal vehicles, the test sets of these 10 experimental groups ultimately cover all available vehicles in the dataset, thus achieving an unbiased and comprehensive evaluation of the model's performance. This design effectively simulates the real-world diagnostic scenario when the model faces unknown vehicles and ensures the statistical significance of the evaluation results.

[0051] Table 1. Configuration of Training and Test Sets for Thermal Runaway Early Warning Model

[0052] Referring to Table 2, the selection of hyperparameters during model training followed the principle of seeking the optimal balance between model capacity, training stability, and generalization performance. The input sequence length was set to 50, a length sufficient to cover the key temporal features of the battery system's operation, ensuring the model captures the complete temporal dependencies of the power battery state, while avoiding redundant noise and increased computational burden due to excessively long sequences. The batch size was set to 64, balancing gradient update stability with hardware memory utilization efficiency, resulting in smoother gradient descent during training and reduced training oscillations. The learning rate was set to 0.001, falling within the commonly used range for deep learning, ensuring effective parameter updates in the early stages of training while preventing convergence oscillations due to an excessively large learning rate and slow convergence due to an excessively small learning rate. The dropout rate was set to 0.3, introducing appropriate regularization constraints to the model, mitigating overfitting risks while preserving effective features, and significantly improving the model's generalization ability. Regarding the core architecture parameters of the model, the node feature dimension is set to 40, providing sufficient and suitable capacity for feature representation of multi-source monitoring data of power batteries. This ensures the richness of feature expression while avoiding computational redundancy caused by excessive dimensionality. The GCN depth is set to 3, ensuring that the model can fully capture the spatial correlation between various monitoring variables of the battery system while controlling the computational complexity of graph convolution and preventing the gradient vanishing problem caused by excessively deep networks. The number of spatiotemporal convolutional layers is set to 3. The stacked layers allow the model to fully explore the spatiotemporal coupling features of power battery data and achieve multi-scale feature extraction, while avoiding a surge in model complexity and increased training difficulty due to too many layers. After multiple rounds of pre-tuning, the above parameter combination has been proven to achieve convergent stability, good accuracy, and reliable generalization in the power battery thermal runaway fault diagnosis task.

[0053] Table 2 Hyperparameter Selection Table

[0054] In this embodiment, to construct a binary classification network for thermal runaway early warning, a class-weighted binary cross-entropy loss function is selected as the optimization objective for model training. This design is based on two core considerations: first, the theoretical fit between the loss function itself and the binary classification task; and second, specific optimization for the extreme class imbalance characteristics of this task. The loss function is as follows: ; in, N For the sample size, Indicates the first i The true label of each sample (0 represents normal, 1 represents risk of thermal runaway). The model is passed through the Sigmoid function Output predicted probabilities. Key parameters. This is a minority class weight coefficient specifically designed for balancing the classes; its value is typically inversely proportional to the proportion of minority class samples. The design of the loss function directly measures the predicted probability distribution. With the binary label of reality The difference between them, and its optimization objective is essentially maximum likelihood estimation, driving the model's output probability to be as close as possible to the true state. Its theoretical properties are highly consistent with binary classification tasks. Regardless of the positive class... or negative class Both the samples and the loss function provide symmetrical and explicit gradient signals. When a severe prediction error occurs, the gradient magnitude is large, prompting the model to correct itself quickly; when the prediction is close to correct, the gradient tends to flatten, which is beneficial for stable convergence. This characteristic is particularly important in this task when facing extreme class imbalance (normal samples far outnumber faulty samples), ensuring that the penalty for mispredictions of minority class samples is sufficiently sensitive. This is further enhanced by introducing weighting coefficients. β This function artificially increases the contribution of rare faulty samples to the total loss. This significantly increases the penalty for misclassifying faulty samples, thus effectively combating model bias caused by data skew—that is, the tendency to predict all samples as the majority of normal samples.

[0055] Experimental results and comparative analysis.

[0056] To further verify the advantages of the spatiotemporal graph neural network model constructed in this invention, an ablation comparison experiment was designed in this embodiment. Only the feature extraction module of the model was replaced, while the rest of the training strategy remained consistent with the experimental settings, and key performance indicators in the binary classification task were statistically analyzed. The diagnostic results of the spatiotemporal graph neural network proposed in this paper are shown in Table 3. For comparison, the Transformer encoder, which has strong temporal feature extraction capabilities, was selected as the baseline network, and its diagnostic results are shown in Table 4. Both tables show the accuracy, false positive rate, and false negative rate for two normal vehicles (vehicles 8 and 29) and two faulty vehicles (vehicles 54 and 57).

[0057] Table 3. Thermal runaway diagnosis performance of spatiotemporal neural networks for different vehicles.

[0058] Table 4 Comparison of thermal runaway diagnosis performance of networks under different vehicles

[0059] Comparative experimental results show that both the spatiotemporal graph neural network constructed in this invention and the contrast network can achieve thermal runaway fault diagnosis to a certain extent, but their overall performance differs significantly. On normal vehicles No. 8 and No. 29, the accuracy of the proposed model reached 0.9883 and 0.9841, respectively, with false alarm rates of only 0.0117 and 0.0159. In contrast, the accuracy of the contrast network was 0.9206 and 0.9162, respectively, with false alarm rates rising to 0.0794 and 0.0838. This indicates that the proposed model is more stable in identifying normal operating conditions and can effectively suppress false alarms. On faulty vehicles No. 54 and No. 57, the false alarm rates of the proposed model were only 0.0051 and 0.0090, far lower than the 0.0619 and 0.0665 of the contrast network, demonstrating a stronger ability to capture early, subtle fault features and significantly reducing the risk of missed thermal runaway detection. Overall, thanks to the advantages of spatiotemporal joint modeling and multi-scale feature fusion, the model provided by this invention outperforms the comparison network in all three key indicators: accuracy, false alarm rate, and false negative rate. It not only ensures diagnostic stability under normal conditions but also improves the identification sensitivity under fault conditions. It effectively overcomes the model bias problem caused by data class imbalance and can better meet the engineering early warning requirements of advanced, reliable, low false alarm, and low false negative rate for thermal runaway of power batteries.

[0060] After completing model inference and obtaining time-series predicted labels, it is necessary to effectively distinguish between faulty and normal vehicles based on the prediction results. This invention adopts the proposed sliding window-based fault discrimination strategy, performing post-processing and decision fusion on the predicted label sequences to improve the stability and anti-interference ability of the diagnostic results. The experiment still selects two normal vehicles (vehicles 8 and 29) and two faulty vehicles (vehicles 54 and 57) for visualization, as shown below. Figure 7 As shown, the curves illustrating the percentage change of the predicted label as 1 for each of the four vehicles within the sliding window are presented. (a) is the warning output curve for vehicle number 8 (normal), (b) is the warning output curve for vehicle number 29 (normal), (c) is the warning output curve for vehicle number 54 (faulty), and (d) is the warning output curve for vehicle number 57 (faulty). Based on the statistical characteristics of the dataset and the engineering warning requirements, the fault determination threshold for the percentage of abnormal labels within the window is set to 50%.

[0061] This shows that the proportion of abnormal vehicles among normal vehicles is consistently significantly lower than the threshold, indicating that the model has an extremely low risk of false alarms under real driving conditions. Faulty vehicles can be effectively identified in the early stages of warning and maintain a high proportion of abnormal vehicles throughout subsequent operation. Although there are slight fluctuations in the proportion of abnormal vehicles within some windows, mainly due to sensor noise, sudden changes in operating conditions, and instantaneous discrimination bias of the model against weak fault features, the sliding window mechanism achieves information smoothing and statistical filtering in the temporal dimension, effectively suppressing disturbances caused by single-point misjudgments, and therefore will not interfere with the final fault determination result.

[0062] To further verify the generality of the strategy, this embodiment also conducted systematic verification on all vehicles in the dataset. Ultimately, none of the normal vehicles triggered the threshold alarm, and all faulty vehicles could be accurately identified. This fully demonstrates that the sliding window-based thermal runaway fault diagnosis scheme proposed in this invention has good robustness, stability, and engineering practicality, and can meet the early, reliable, and low false alarm warning requirements of power batteries.

[0063] In summary, the thermal runaway early warning method for automotive lithium batteries of the present invention has the following advantages: This invention innovatively employs a class-weighted binary cross-entropy loss function during the model training phase. By introducing a minority class weight coefficient, it significantly increases the model's penalty for misclassified fault samples. This mechanism completely overcomes the model prediction bias caused by extreme class imbalance in the power battery dataset, ensuring that the early warning system maintains extremely high diagnostic accuracy and robustness against interference even when facing complex and variable vehicle operating conditions and strong noise interference.

[0064] In the spatial dimension, this invention employs a multi-hop feature propagation graph convolutional network based on a gated mechanism to effectively perceive the potential diffusion path of weak anomalies between adjacent nodes; in the temporal dimension, it utilizes a multi-scale temporal feature extraction module to collaboratively perceive rapidly changing transient electrical signal anomalies and slowly changing long-term thermal accumulation trends.

[0065] This invention introduces a sliding time window mechanism in the temporal feature extraction stage and employs a temporal smoothing strategy at the prediction output. This multi-verification mechanism effectively filters out instantaneous anomaly judgments caused by sensor glitches or occasional large current surges, greatly reducing the false alarm rate in applications.

[0066] This invention innovatively introduces a dilated convolution mechanism and a GRU gating structure into the multi-scale temporal feature extraction module. By adjusting the dilation rate of the dilated convolution, the network can effectively expand its receptive field without increasing the number of parameters, thus achieving accurate capture of long-range temporal features that are sparse but critical with large sampling intervals.

[0067] In another embodiment of the present invention, a thermal runaway warning system for automotive lithium batteries based on a spatiotemporal graph neural network is also provided. This warning system is used to implement the aforementioned warning method, including: The data acquisition and preprocessing module is used to acquire the operating status data of the vehicle lithium battery and preprocess the operating status data. The feature extraction module is used to extract multi-dimensional features from standardized data samples and perform correlation analysis and screening. The model building module is used to build a binary classification early warning model based on a spatiotemporal graph neural network; The online early warning module is used to process the spatiotemporal coupling feature data output by the model in real time and output thermal runaway early warning signals based on the sliding window mechanism.

[0068] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for early warning of thermal runaway in automotive lithium batteries based on spatiotemporal graph neural networks, characterized in that, Includes the following steps: Step 1: Collect the operating status data of the vehicle lithium battery, preprocess the operating status data to obtain standardized data samples; Step 2: Extract multidimensional features from the standardized data samples, including: direct voltage features, first-order voltage difference sequence and complex frequency domain features, and then screen out a subset of general features through correlation analysis; Step 3: Construct a binary classification early warning model based on a spatiotemporal graph neural network, specifically including: Each monitored variable in the general feature subset is defined as a node in the graph. The dynamic correlation between the monitored variables is adaptively learned from the input data by the graph builder to generate a dynamic adjacency matrix. Temporal features at each node are extracted using gating mechanisms and multi-scale convolutions. Based on the adjacency matrix, spatial features between nodes are extracted through bidirectional graph convolution; The spatial features and the temporal features are residually fused to output spatiotemporal coupled features. Step 4: Output the thermal runaway risk prediction label for each moment based on the spatiotemporal coupling characteristics; Step 5: Post-process the thermal runaway risk prediction labels using a sliding window-based result statistics and confidence confirmation mechanism. When the proportion of thermal runaway risk prediction labels in the sliding window exceeds a preset threshold, a final thermal runaway warning is issued.

2. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 1, characterized in that, Step 1 involves preprocessing the operating status data, including: The operational status data is cleaned to remove outliers; then, a sliding window technique that does not cross charging and discharging segments is used to process the cleaned operational status data to generate standardized data samples.

3. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 2, characterized in that, The sliding window technique that does not span charging / discharging segments includes: The start of consecutive charge / discharge segments is padded, and the length of the padded segment is the same as the length of the sliding window. Sliding sampling is performed within the labeled independent charge / discharge segments, and windows that span different charge / discharge segments are deleted.

4. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 1, characterized in that, In step 2, the correlation analysis specifically includes: The Pearson correlation coefficient is used to quantify the degree of linear correlation, and mutual information is used to capture nonlinear dependencies. A correlation strength matrix between multiple features and voltage signals is constructed, and then feature selection is performed based on the preset correlation threshold and feature redundancy test results.

5. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 1, characterized in that, In step 3, a dynamic adjacency matrix is ​​generated using a graph builder, specifically as follows: The binary classification early warning model passes adp Function to generate adjacency matrix The calculation formula is: ； in, gc Represents a graph constructor. , indicating the subgraph number.

6. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 5, characterized in that, In step 3, temporal features at each node are extracted using a gating mechanism and multi-scale convolution, including: Based on the gating mechanism of Gated Recurrent Unit (GRU), node features are split into multiple groups, and convolution operations are performed along the time dimension using convolution kernels of different sizes. Adaptive feature selection is achieved through value gate and update gate. The value gate compresses the feature values ​​to the [-1,1] interval using the hyperbolic tangent activation function to preserve the positive and negative changes of the features. The update gate generates weights in the [0,1] interval using the sigmoid activation function to control the retention ratio of features at each position.

7. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 6, characterized in that, Step 3 further includes: training the binary classification early warning model. Specifically, a class-weighted binary cross-entropy loss function is used as the optimization objective for model training. The loss function is: ； in, N For the sample size, This represents the true label of the i-th sample. The model is passed through the Sigmoid function The output predicted probability, β These are the weighting coefficients.

8. The method for early warning of thermal runaway of automotive lithium batteries based on spatiotemporal graph neural networks according to claim 1, characterized in that, Step 5 specifically includes: Given a time sliding window of length W, at each sampling time t, backtrack all predicted labels from the past W time points within the window and calculate the proportion of labels representing fault risk appearing within the window. , ； In the formula, This represents the i-th predicted label, only when... A thermal runaway warning will only be issued at the current time t if the value is ≥0.5.

Images