Thermal imaging anomaly detection method and system for battery pack of electric vehicle

Abstract

The invention discloses a thermal imaging anomaly detection method and system for a battery pack of an electric vehicle. According to the method, a temperature field image sequence obtained by a thermal imaging sensor is synchronized with working condition parameter time of a battery management system, an expected temperature field is predicted, and a condition residual field is calculated; constructing a battery cell topological graph, taking the residual time sequence statistics as node characteristics, and extracting a battery cell space-time representation vector by combining a time convolution network and a graph attention propagation network through a condition comparison graph encoder which is subjected to working condition conditional topological comparison self-supervision pre-training; and determining candidate abnormal nodes based on Mahalanobis distance novelty scores, performing anti-fact intervention test by using a thermal-electric structure causal directed acyclic graph to eliminate false positive, and outputting a grading early warning result with an interpretable root cause attribution report. According to the method, the detection performance that the area under a detection curve is not less than 0.88 can be realized without labeling an abnormal sample, the full-life periodic performance is maintained to be stable through a self-supervision reference drift feedback mechanism, and the method is suitable for online thermal safety monitoring of battery packs of pure electric and plug-in hybrid electric vehicles.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle safety inspection technology, specifically to a method and system for detecting abnormalities in electric vehicle battery packs using thermal imaging. Background Technology

[0002] Thermal runaway of electric vehicle battery packs is one of the main causes of electric vehicle fires and explosions. Thermal runaway is usually triggered by electrochemical failure mechanisms such as internal short circuit (ISC), lithium plating, and overcharge thermal runaway. There is a thermal precursor stage of tens of minutes to several hours before the irreversible runaway stage. If it can be accurately detected and intervened in the thermal precursor stage (i.e., at least 30 minutes before thermal runaway occurs), catastrophic accidents can be effectively avoided.

[0003] Thermal imaging technology can acquire a two-dimensional continuous temperature field on the surface of a battery pack in a non-contact manner. Compared with the Battery Management System (BMS), which only collects a small number of discrete temperature measurement points, thermal imaging sensors can provide information on the temperature distribution across the entire space, theoretically enabling earlier detection of the onset and spread of thermal anomalies. However, existing thermal imaging-based battery thermal anomaly detection technologies suffer from the following four core technical shortcomings.

[0004] First, operating condition interference leads to a high false alarm rate. This is due to variations in the battery pack's state of charge (SOC), charge / discharge rate (C-rate), and ambient temperature. The normal thermal distribution patterns differ significantly between the current state (C) and the state of healthy operation (SOH), with the maximum normal temperature difference between adjacent modules reaching over 15 degrees Celsius. Existing methods based on fixed temperature thresholds or simple statistical thresholds have a false alarm rate exceeding 30% under fast charging conditions (charge / discharge rate not lower than 2C) and severely insufficient sensitivity under low-rate discharge conditions (charge / discharge rate not higher than 0.2C), leading to missed detection of weak signals in the early stages of thermal runaway.

[0005] Second, abnormal samples are extremely scarce. Dangerous anomalies such as thermal runaway and internal short circuits are extremely rare in actual vehicle operation, resulting in a severe imbalance between normal and abnormal samples (typically exceeding 10,000:1). This leads to a serious class imbalance problem in supervised deep learning detection models when dealing with anomalies. Furthermore, the thermal properties of different electrochemical systems (lithium iron phosphate, ternary lithium, sodium ion, etc.) vary significantly, and models trained on a single chemical system exhibit extremely poor cross-system generalization, making them unsuitable for direct transfer to other vehicle models or platforms.

[0006] Third, single-modal visual inspection lacks reliability. When relying solely on thermal imaging features for detection, inherent noise in the thermal imaging sensor, external radiation hotspots caused by direct sunlight, false heat sources generated by specular reflection from the metal casing of the battery pack, and normal transient local temperature fluctuations during charging and discharging can all lead to false hotspot detections, resulting in a high false alarm rate. Furthermore, the detection conclusions output by existing methods lack physical interpretability, making it impossible to accurately trace the detection results back to a specific faulty battery cell.

[0007] Fourth, performance continuously degrades throughout the entire life cycle. As the battery ages (internal resistance increases by 30% to 80% after a typical 1000 cycles), the normal thermal distribution pattern of the battery pack under the same operating conditions will systematically shift. Existing detection models lack the ability to sense and adapt online to the drift in normal thermal characteristics caused by aging, resulting in a continuous increase in the false alarm rate as the battery service time extends, and even losing practical value after the third year of service.

[0008] To address the aforementioned four types of technical deficiencies, there is currently no complete technical solution that organically couples dynamic benchmark prediction of operating conditions, unsupervised comparative self-supervised representation learning, graph topology prior modeling, and causal inference of thermal-electrical structures. This results in the inability to simultaneously meet application requirements for false alarm rate and missed alarm rate in actual vehicle engineering scenarios. Summary of the Invention

[0009] Technical Objective: To address the problems of existing electric vehicle battery pack thermal imaging anomaly detection methods, such as high false alarm rate under varying operating conditions, inability to effectively train without anomaly labeled samples, lack of physical interpretability of detection conclusions, and inability to track thermal characteristic drift throughout the battery's life cycle, this invention discloses an electric vehicle battery pack thermal imaging anomaly detection method and system.

[0010] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0011] A method for detecting anomalies in thermal imaging of an electric vehicle battery pack includes the following steps: acquiring a thermal imaging image sequence of the battery pack and a condition parameter vector synchronized with the thermal imaging image sequence in time, wherein the condition parameter vector includes state of charge, charge / discharge rate, ambient temperature, and health status; inputting the condition parameter vector into a pre-trained sparse Gaussian process regression model to obtain the expected temperature field and standard deviation prediction under the current condition; standardizing the difference between the current frame temperature field and the expected temperature field using the standard deviation prediction to obtain the conditional residual field; constructing a cell topology graph with each cell in the battery pack as a node, the heat conduction path between cells as heat conduction edges, and the electrical connection relationship between cells as electrical connection edges; using the conditional residual field statistics sequence of the corresponding cell within the historical time window as the node feature; and then... The topology graph is input into a pre-trained conditional comparison graph encoder. Temporal embeddings of each node are extracted through a temporal convolutional network, and neighborhood information is aggregated on the cell topology graph through a graph attention propagation network, outputting the representation vector of each node. The Mahalanobis distance score of each node's representation vector relative to the normal operation representation distribution is calculated, and nodes with Mahalanobis distance scores exceeding the adaptive detection threshold are identified as candidate abnormal nodes. Structural causal inference is performed on the candidate abnormal nodes. Based on a pre-constructed thermal-electrical causal directed acyclic graph, the do operator is applied to the thermal imaging abnormal feature variables to truncate the directed edges between the operating condition disturbance variables and the thermal imaging abnormal feature variables while keeping the observed values ​​of the operating condition disturbance variables unchanged. It is then determined whether the causal effect of the thermal imaging abnormal feature variables is independent of the operating condition disturbance variables, and a graded early warning result is output.

[0012] Furthermore, the sparse Gaussian process regression model adopts the Nyström sparse approximation method, takes the operating condition parameter vector as input, and outputs the mean prediction and standard deviation prediction of the temperature at each pixel position; the conditional residual field is obtained by subtracting the mean prediction from the temperature field of the current frame and dividing by the standard deviation prediction, and the region in the conditional residual field with an absolute value exceeding 2.0 constitutes the residual saliency region.

[0013] Furthermore, the weight of the thermally conductive edge is the reciprocal of the contact thermal resistance between adjacent cells, and the weight of the electrical connection edge is the electrothermal coupling coefficient between the corresponding cells; the graph attention propagation network sets independent attention calculation heads for the thermally conductive edge and the electrical connection edge respectively, and the outputs of each attention calculation head are spliced ​​and then linearly transformed to obtain the final aggregated features of each node.

[0014] Furthermore, the pre-training of the conditional contrast graph encoder employs a conditional topology contrast strategy. Positive sample pairs are formed by normalizing the conditional residual field node feature sequences of the same cell under different operating conditions using a sparse Gaussian process regression model. Negative sample pairs are formed by combining node feature sequence pairs of different cells at the same time and node frame pairs containing the Mahalanobis distance score exceeding the initial threshold. A topology-aware graph enhancement operation is applied to the cell topology graph. This operation includes randomly masking some heat-conducting edges according to their weight ratio and sampling subgraphs of nodes based on their thermal contribution. The conditional contrast graph encoder is optimized using a conditional contrast loss function to maximize the cosine similarity between node representation vectors of positive sample pairs and minimize the cosine similarity between node representation vectors of negative sample pairs.

[0015] Furthermore, the temporal convolutional network adopts a multi-layer dilated causal convolutional structure, with the dilation factor of each layer increasing in powers of 2, and convolves the conditional residual field statistics sequence within a historical time window of 30 frames in length; the graph attention propagation network adopts the GraphSAGE aggregation framework, and the aggregation function introduces attention coefficients based on the heat-conducting edge weights.

[0016] Furthermore, the normal operation representation distribution is obtained by fitting a multivariate Gaussian distribution to the node representation vectors of confirmed normal frames that have not triggered warnings in the past 90 days through maximum likelihood estimation; the adaptive detection threshold is the 99th percentile of the Mahalanobis distance score of all normal confirmed frame nodes within a 90-day sliding window, and is automatically updated every 24 hours.

[0017] Furthermore, the thermal-electric causal directed acyclic graph is obtained by applying the PC algorithm to perform structural learning on historical operating data and controlled injection fault data; when the absolute value of the difference between the causal effect of the thermal imaging anomaly feature variable after the do operator is applied and the observed effect before the application exceeds the preset judgment threshold, it is determined that there is a real thermal anomaly independent of the operating condition disturbance variable and the warning level is upgraded; otherwise, the warning output is suppressed.

[0018] Furthermore, the method also includes: calculating the offset of the mean of the Mahalanobis distance score distribution of all normal confirmation frame nodes within the current 30-day sliding window relative to the historical baseline mean; when the offset exceeds the preset drift detection threshold, triggering the online update of the hyperparameters of the sparse Gaussian process regression model, and recalculating the conditional residual field with the updated sparse Gaussian process regression model, while updating the normal operation characterization distribution and the adaptive detection threshold.

[0019] Furthermore, the graded early warning results include an early warning level identifier and a root cause attribution report. The root cause attribution report is based on the attention weight matrix of each layer of the graph attention propagation network. It traces back the contribution of each node to the Mahalanobis distance score along the direction of heat diffusion propagation, and identifies the cell corresponding to the node with the highest contribution as the root cause cell. The root cause cell number, the battery module number, the historical temperature rise rate curve, and the causal contribution score are output in the root cause attribution report.

[0020] This invention also provides a thermal imaging anomaly detection system for electric vehicle battery packs, comprising: a data acquisition and synchronization module configured to acquire a sequence of thermal imaging images of the battery pack and align the operating condition parameter vector collected by the battery management system with the thermal imaging image sequence using timestamps; an operating condition benchmark prediction module configured to input the operating condition parameter vector into a sparse Gaussian process regression model, output the expected temperature field and standard deviation prediction, and calculate the conditional residual field; a topology graph construction module configured to construct a cell topology graph with each cell as a node and assign node features; a conditional comparison graph encoding module configured to output the representation vector of each node; an anomaly scoring module configured to determine candidate anomaly nodes based on Mahalanobis distance scoring; and a causal inference early warning module configured to output graded early warning results.

[0021] Beneficial Effects: The electric vehicle battery pack thermal imaging anomaly detection method and system provided by this invention have the following beneficial effects:

[0022] 1. By establishing a mapping from multidimensional operating condition parameters to the desired temperature field through a sparse Gaussian process regression model, a conditional residual field with standardized operating conditions is generated. Within the full operating condition range of 5% to 100% state of charge, 0.1C to 3C charge / discharge rate, and -30°C to 55°C ambient temperature, the false alarm rate under normal operating conditions is controlled to below 3%, solving the problem of false alarms caused by changes in operating conditions in existing technologies.

[0023] 2. By adopting a conditional topology comparison self-supervised pre-training strategy, a compact normal representation manifold is learned using unlabeled data from normal operation. Under zero anomaly labeling conditions, the anomaly detection area under the curve (AUROC) is no less than 0.88, which solves the bottleneck problem of scarce anomaly samples.

[0024] 3. The prior physical topology of the battery pack is injected into the detection network through heterogeneous graph modeling, enabling the model to perceive the topological propagation law of heat diffusion. Compared with the pure vision solution, the detection recall rate of early hot spots of diffuse internal short circuits is improved by no less than 30 percentage points.

[0025] 4. By using a structural causal inference engine to perform counterfactual verification on candidate anomalies, the overall false alarm rate is reduced to below 1%, and a root cause attribution report that can be traced back to the specific cell number is generated, meeting the safety regulatory compliance requirements.

[0026] 5. By using the Mahalanobis distance score overall distribution shift monitoring mechanism, the thermal characteristic drift caused by battery aging is detected in real time and online adaptive updates are triggered, so that the detection performance can be maintained stably without manual intervention throughout the entire battery life cycle. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0028] Figure 1 This is a schematic diagram of the overall architecture of the electric vehicle battery pack thermal imaging anomaly detection system according to an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the conditional comparison graph encoder structure and the conditional topology comparison pre-training strategy according to an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of the workflow of the thermo-electric structure causal inference early warning engine according to an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of the self-supervised benchmark drift detection and adaptive update process according to an embodiment of the present invention. Detailed Implementation

[0032] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0033] Reference Figure 1 The electric vehicle battery pack thermal imaging anomaly detection system of the present invention includes: a thermal imaging sensor array, a battery management system, a data acquisition and synchronization module, a working condition benchmark prediction module, a topology graph construction module, a condition comparison graph encoding module, an anomaly scoring module, and a causal inference early warning module.

[0034] A thermal imaging sensor array is deployed inside or outside the battery pack to continuously acquire two-dimensional temperature field images T(x,y,t) of the cell surface at a frame rate of 25 Hz, with a temperature resolution of no less than 0.05℃ and an image resolution of no less than 320×240 pixels. The battery management system acquires operating parameters at a sampling frequency of 1 Hz: State of Charge (SOC) (%), Charge / Discharge Rate (C-rate) (C), and ambient temperature. (Unit: degrees Celsius) and State of Health (SOH) (unit: %). The data acquisition and synchronization module upsamples the operating condition parameter vector output by the battery management system to 25 Hz through linear interpolation, so that each frame of thermal imaging image corresponds to a simultaneous operating condition parameter vector, with a time alignment error of no more than 20 milliseconds.

[0035] Reference Figure 1 The core component of the benchmark prediction module is a sparse Gaussian process regression model, which uses the Nyström sparse approximation method to reduce the computational complexity from O(n log n). 3 ) decreased to O(nm) 2 (n is the number of training samples, m is the number of induced points, and m is much smaller than n). The model uses a vector of operating condition parameters. The input is the mean and standard deviation of the temperature at each pixel location, and the output is the predicted temperature at each pixel location, where SOC represents the state of charge and SOH represents the healthy state. A composite kernel is used as the kernel function. Hyperparameters are optimized by maximizing the logarithmic marginal likelihood, where It is the radial basis function kernel. Use the Matérn kernel function. Offline training was performed on at least 5000 normal operating condition-heatmap pairs within a wide operating range, including a state of charge of 5% to 100%, a charge / discharge rate of 0.1C to 3C, and an ambient temperature of -30°C to 55°C. 500 induction points were selected, and the online inference time per frame did not exceed 0.5 milliseconds.

[0036] The conditional residual field is calculated using the following formula:

[0037]

[0038] in The conditional residual field represents the normalized temperature deviation of pixel (x, y) at time t, which is dimensionless. Pixel regions with an absolute value exceeding 2.0 constitute the residual saliency region, which is the direct input signal to the downstream topology encoder. t is the discrete-time frame number, with a frame rate of 25 Hz and an interval of 40 milliseconds between adjacent frames. x and y are the pixel abscissa and ordinate of the thermal imaging image, in pixels (px). In this embodiment, the sensor resolution is 320×240, x∈[0,319], y∈[0,239]. The measured temperature value at pixel (x,y) acquired by the thermal imaging sensor array at time t is expressed in degrees Celsius (°C), with a temperature resolution of not less than 0.05°C. This is the mean prediction of the sparse Gaussian process regression model under the condition parameter vector C(t), that is, the normal expected temperature at pixel (x,y) under the current condition, in degrees Celsius (°C). The working condition parameter vector is a four-dimensional real-valued vector. The corresponding values ​​are, in order, state of charge (%), charge / discharge rate (C), ambient temperature (°C), and health status (%). This represents the predicted standard deviation at pixel (x, y) in the output of the sparse Gaussian process regression model, reflecting the uncertainty of normal thermal distribution under the current operating conditions. It is used as the standardized denominator to eliminate differences in operating condition fluctuations. Unit: degrees Celsius (°C).

[0039] The conditional residual field R(x,y,t) eliminates the differences in the background temperature field caused by changes in operating conditions, retaining only the abnormal deviation information relative to the normal expectation of the current operating conditions. During normal operation, the absolute value of the conditional residual field should approximately follow a standard normal distribution (mean 0, standard deviation 1) across the entire map.

[0040] The topology graph construction module maps each cell in the battery pack to a graph node, constructing a directed graph. ,in N is the set of nodes (N is the total number of battery cells). For the heat-conducting edge set, This is the set of electrically connected edges. The weights of the thermally conductive edges are also included. Set as the interfacial thermal resistance between adjacent cells The reciprocal of (unit: W / (m) 2 The weight of the electrical connection edge (K) is obtained through post-assembly thermal testing. Calibrated through electrothermal coupling experiments, the values ​​range from [0,1], reflecting the degree of mutual influence of the thermal states of the two cells along the current path. Each node... eigenvectors From the most recent historical time window The conditional residual field of each frame (1.2 seconds) consists of the mean, peak value and standard deviation of the local region, with a dimension of 90 (3 statistics × 30 frames).

[0041] Reference Figure 2 Positive sample pairs in the conditional comparison graph encoding module are defined as belonging to the same cell node. At two different time points where the Euclidean distance between the operating parameter vectors is greater than 0.5, the conditional residual field node feature sequence pairs are normalized by the sparse Gaussian process regression model; negative sample pairs include feature pairs of different cell nodes at the same time and candidate abnormal frame node feature pairs, with a positive-to-negative sample ratio of 1:7. Topology-aware graph enhancement includes two types of operations: edge masking enhancement (randomly masking 10% to 30% of the heat-conducting edges according to the weight ratio of the heat-conducting edges) and node subgraph sampling (sorting the nodes according to the mean of the conditional residual field and retaining 80% to 95% of the nodes). The conditional contrastive loss function is calculated according to the following formula:

[0042]

[0043] in To conditionally compare the loss values, the average of all nodes within the pre-training batch is used as the optimization objective for backpropagation; minimizing... This allows for compact clustering of normal operating characteristics, while abnormal nodes naturally deviate from the normal manifold. is the cosine similarity function, with a value range of [-1, 1]. The larger the value, the more similar the representations of the two nodes are. This is the representation vector of the query node. . The positive sample representation vector is obtained by encoding the same battery cell under different operating conditions after sparse GP normalization, and is related to... The structure is aligned; the normalization of operating conditions ensures that the characteristics of normal cells under different operating conditions should be similar. The negative sample representation vector contains node features of different cells at the same time (Type 1) and node features of candidate abnormal frames (Type 2); the index k traverses all negative samples in the same mini-batch. The temperature hyperparameter controls the sharpness of the similarity distribution, with a value of 0.07. The smaller the value, the more sensitive the model is to distinguish between positive and negative samples, and the more the gradient is concentrated on difficult negative samples.

[0044] Pre-training uses the Adam optimizer with an initial learning rate of 1×10⁻⁶. -3 Cosine annealing decays to 1×10 -5 The batch size is 256, and 200 rounds of pre-training are completed on no less than 50,000 frames of normal thermal imaging data. The momentum encoder momentum coefficient is 0.999.

[0045] The temporal convolutional network employs a four-layer dilated causal convolutional structure with dilation factors of 1, 2, 4, and 8 for each layer. The kernel size is 3, and the activation function is LeaKy ReLU (negative slope coefficient 0.01). The receptive field is 1 + (3-1) × (1 + 2 + 4 + 8) = 29 frames, covering a 30-frame history window (1.2 seconds). A 128-dimensional temporal embedding is extracted from the 90-dimensional node feature sequence. Residual connections are set in each layer to prevent gradient vanishing. The graph attention propagation network uses the GraphSAGE aggregation framework, with heat-conducting edges... and electrical connection edge An independent attention computation head is set up, and node representations are aggregated according to the following formula:

[0046]

[0047] in Let be the hidden representation vector of node i in the l-th layer. When l=0, it is the 128-dimensional temporal embedding output by the temporal convolutional network 12. This is the update representation vector for node i in the (l+1)th layer; the output of the last layer (L=2) is the node representation vector. . Let be the set of neighboring nodes of the heat-conducting edge of node i, defined by the heat conduction path of physical contact between battery cells, and the edge weight is the reciprocal of the contact thermal resistance. Let be the set of neighboring nodes of the electrical connection edge of node i, defined by the series-parallel topology, and the edge weight is the electrothermal coupling coefficient. The thermal conductivity attention coefficient is based on the thermal conductivity edge weight (the inverse of the contact thermal resistance). After a linear transformation of the node-to-feature concatenation, it is normalized by softmax to give the direction of thermal diffusion propagation a higher aggregation weight. The thermal conductivity attention coefficient is based on the thermal conductivity edge weight (the inverse of the contact thermal resistance). After a linear transformation of the node-to-feature concatenation, it is normalized by softmax to give the direction of thermal diffusion propagation a higher aggregation weight. The electrical connection attention coefficient is based on the electrothermal coupling coefficient and is normalized by softmax to capture the cross-module thermal anomaly propagation caused by electrothermal coupling. For vector concatenation operation, the output of the thermally conductive attention head (128-dimensional) and the output of the electrical connection attention head (128-dimensional) are concatenated along the feature dimension to form a 256-dimensional vector. The learnable linear transformation matrix for the l-th layer The 256-dimensional spliced ​​features are projected onto the 128-dimensional output space.

[0048] The graph attention propagation network consists of L=2 layers of GraphSAGE propagation, with a receptive field covering the 2nd order neighboring cells. This allows each node representation to perceive the thermal diffusion dynamics within its 2-hop neighborhood, effectively capturing the cross-cell thermal expansion mode caused by diffuse internal short circuits.

[0049] Reference Figure 1 The anomaly scoring module evaluates each node for anomalies based on Mahalanobis distance. The normal operation representation distribution consists of node representation vectors from confirmed normal frames that have not triggered warnings in the past 90 days. A multivariate Gaussian distribution is fitted using maximum likelihood estimation. Obtain the mean vector Covariance Matrix Apply diagonal loading regularization to Σ Ledoit-Wolf shrinkage estimation is employed to improve the accuracy of high-dimensional estimation. The Mahalanobis distance score for each node is calculated using the following formula:

[0050]

[0051] in The Mahalanobis distance score for node i is a dimensionless real number; when it exceeds the adaptive detection threshold θ, the node is identified as a candidate anomaly node. Let i be the 128-dimensional representation vector of node i output by the conditional comparison graph encoder. It is obtained by joint encoding of a temporal convolutional network and a graph attention propagation network. The mean vector of the distribution representing normal operation. The representation vector of normal frame nodes confirmed over the past 90 days The data is obtained by fitting a multivariate Gaussian distribution using maximum likelihood estimation and is updated every 24 hours. The distribution covariance matrix is ​​used to characterize normal operation. The inverse matrix (precision matrix) is used to measure the correlation of each dimension of the characterization space and to provide anisotropic distance metrics, which are updated every 24 hours. Synchronized updates. The adaptive detection threshold is set to the 99th percentile of the score of the normal confirmed frame node within a 90-day sliding window, and is updated every 24 hours.

[0052] Mahal distance rating The set of nodes that exceed the adaptive detection threshold θ constitutes the candidate abnormal node set A, which enters the causal inference early warning module for counterfactual verification.

[0053] Reference Figure 3 The causal inference and early warning module uses a structural causal model to perform counterfactual verification of candidate anomalies. The construction of the thermal-electrical causal directed acyclic graph (DAG) involves: before system deployment, based on at least 12 months of historical normal operation data and at least 5 types of controlled injection fault experimental data (including at least 4 sets each of internal short-circuit injection, overcharge injection, and external thermal shock), applying the PC algorithm (Peter-Clark constrained algorithm, conditional independence test significance level α=0.05) to learn the directed causal structure between variables. The DAG contains three types of variables: ① thermal imaging anomaly characteristic variables Y (macroscopic distance score of candidate anomaly nodes, hotspot area, thermal diffusion rate); ② BMS electrical parameter variables (voltage deviation of each cell, estimated internal resistance); ③ operating condition disturbance variables Z (charge / discharge rate, ambient temperature, vehicle body vibration acceleration).

[0054] Counterfactual intervention test: For each candidate node in the candidate anomaly node set A, apply the do operator to truncate all directed edges between the operating condition disturbance variable Z and the thermal imaging anomaly feature variable Y, and re-infer the causal effect on the mutilated graph. The tiered early warning decision is executed according to the following formula:

[0055]

[0056] in The causal intervention effect difference is the absolute value of the difference between the causal effect and the observed effect after intervention; exceeding the judgment threshold. If a real thermal anomaly is detected, an early warning escalation is triggered; otherwise, the early warning output is suppressed. Y is the outcome variable, i.e., the thermal imaging anomaly characteristic variable, which is the explained variable in the thermal-electric causal directed acyclic graph, including the Mahalanobis distance score of candidate anomaly nodes, hotspot area, and thermal diffusion rate. X is the intervention variable, which is the thermal imaging anomaly characteristic observation value corresponding to the candidate anomaly node, and is set to a fixed value x under the action of the do operator. The Pearl do operator is used to truncate all directed edges pointing from the operating condition disturbance variable Z to X in a causal directed acyclic graph, while keeping the actual observed value of Z unchanged, to simulate probabilistic inference under external forced intervention conditions. To intervene in the distribution, the conditional probability of Y is calculated on the multiplied graph, reflecting the pure causal effect of removing thermal imaging anomalies after removing mixed paths of operating conditions. The conditional probability of Y, calculated on the original causal graph based on the observed data, is for the observed distribution. It includes spurious correlation components of the operating condition disturbance variable Z, which are introduced through the confounding path. The decision threshold is initially set to 0.15; mapping: Δ causal ∈[0.15,0.40)→Level-1, [0.40,0.75)→Level-2, ≥0.75→Level-3 warning.

[0057] Root cause cell localization: After the early warning is output, the attention weight paths of the last two layers of the graph attention propagation network, starting from the candidate anomaly nodes, are extracted. Then, the final Mahalanobis distance score of each source node to the candidate anomaly node is calculated layer by layer in reverse along the direction of heat diffusion propagation. The causal contribution is used to identify the cell corresponding to the node with the highest contribution as the root cause cell, and a root cause attribution report is generated, which includes the root cause cell number, the battery module number, the historical temperature rise rate curve, and the causal contribution score.

[0058] Reference Figure 4 The drift detection unit calculates the Mahalanobis distance score of all normally confirmed frame nodes within the current 30-day sliding window every 24 hours. arithmetic mean Compared with the historical benchmark average established during system initialization The offset is calculated using the following formula for comparison:

[0059]

[0060] in The mean offset of the Mahalanobis distance score distribution is a dimensionless real number; when it exceeds the drift detection threshold, the online update unit is triggered to sense the systematic drift of normal thermal characteristics caused by battery aging. The average of the current statistical window – the Mahalanobis distance score of all normally confirmed frame nodes within the sliding window over the past 30 days. The arithmetic mean reflects the normal scoring center under the current battery thermal characteristics. The historical baseline average is the average normal rating established during system initialization (the first 30 days after new vehicle delivery), and is updated after each online update. refresh. The drift detection threshold is set to an initial value of 0.3, which corresponds to a relative deviation of approximately 10% from the average score of normal frames. If this threshold is exceeded, a three-step adaptive update is triggered during the low-power window of static charging, with a single update taking no more than 500ms.

[0061] Online update steps: Step D1, calculate every 24 hours and Compare; Step D2, if (Initial value 0.3) triggers the online update unit; Step D3: Collect normal confirmation frames from the last 30 days (no less than 200 frames), and perform online gradient update on the hyperparameters of the sparse Gaussian process regression model (step size is 0.1 times the initial learning rate, no more than 50 steps); Step D4: Recalculate the conditional residual field with the updated model, extract the new node representation vector, and update μ and Σ with maximum likelihood estimation; Step D5: Recalculate the Mahalanobis distance score distribution of new nodes within the 90-day window, and update the adaptive detection threshold θ; Step D6: ... refresh Then, reset Δμ=0. All steps are completed within the vehicle's stationary charging window (total time not exceeding 500 milliseconds) and do not affect normal driving.

[0062] The thermal imaging sensor array employs an uncooled focal plane infrared detector with specifications of at least 320×240 pixels, a frame rate of 25 Hz, a temperature measurement range of -20°C to 200°C, a temperature resolution of at least 0.05°C, and a spectral response range of 8 to 14 micrometers. For large battery packs exceeding 1 meter in length, four thermal imaging sensors can be deployed for spliced ​​coverage. Onboard computing unit requirements: main processor clock speed at least 1.8GHz (ARM Cortex-A55 or equivalent); memory at least 2GB; optional NPU acceleration. On an onboard computing unit meeting the above requirements, the INT8 quantized conditional comparison map encoding module achieves a single-frame inference latency of no more than 80 milliseconds and a model file size of no more than 5MB. Software implementation: The pre-training framework is based on PyTorch and PyTorch Geometric; onboard inference uses the ONNX Runtime framework (supporting ARM NEON and NPU acceleration); the cloud management service provides a REST API based on Python Flask, communicating with the onboard unit via an encrypted HTTPS channel, with each update communication data volume not exceeding 50MB.

[0063] The effectiveness of this invention is verified using a pure electric passenger vehicle equipped with a 96-series 2-parallel (192 individual cells) lithium iron phosphate battery module. The thermal imaging sensor array consists of four 25 Hz uncooled focal plane infrared sensors (resolution 320×240 pixels), covering all 192 individual cells. The sparse Gaussian process regression model was trained offline on 24 months of driving data (approximately 7500 operating condition-thermal image pairs); the conditional comparison graph encoding module was pre-trained for 200 rounds on 80,000 frames of unlabeled normal operation thermal imaging data; the thermal-electrical causal directed acyclic graph was learned based on 18 months of operating data and 12 sets of controlled fault injection experimental data (4 sets of internal short circuits, 4 sets of overcharging, and 4 sets of external thermal shock).

[0064] On a test set including full-condition normal operation test (2000 frames), fast charging high-temperature test (500 frames), controlled internal short-circuit injection fault test (300 frames, starting 30 minutes before the fault), and external thermal shock interference test (200 frames), the measured key performance indicators were as follows: false alarm rate of fast charging condition was 2.1% (compared to 38.5% by the fixed threshold method, a decrease of 36.4 percentage points); recall rate of controlled internal short-circuit fault detection 30 minutes before the fault was 91.3% (compared to 64.8% by the traditional BMS voltage differential method, an improvement of 26.5 percentage points); false alarm rate of external thermal shock was 0.8% (compared to 22.6% by the single-modal vision method, a decrease of 21.8 percentage points); and root cause cell positioning accuracy was 88.7%. After 12 consecutive months of aging tests (approximately 400 cycles, with battery SOH decreasing from 100% to 89%), the AUROC performance degradation was found to be no more than 2.1 percentage points without manual recalibration, verifying the effectiveness of the self-supervised benchmark drift adaptive update mechanism.

[0065] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting anomalies in electric vehicle battery packs using thermal imaging, characterized in that, Specifically, the following steps are included: Acquire a thermal imaging image sequence of the battery pack and a working condition parameter vector synchronized with the thermal imaging image sequence in time. The working condition parameter vector includes state of charge, charge / discharge rate, ambient temperature, and health status. Input the working condition parameter vector into the pre-trained sparse Gaussian process regression model to obtain the expected temperature field and standard deviation prediction under the current working condition. Then, calculate the difference between the current frame temperature field and the expected temperature field of the thermal imaging image sequence and standardize it using the standard deviation prediction to obtain the conditional residual field. A cell topology graph is constructed with each cell in the battery pack as a node, the heat conduction path between cells as heat conduction edges, and the electrical connection relationship between cells as electrical connection edges. The conditional residual field statistics sequence of the corresponding cell in the historical time window is used as the node feature. The cell topology graph is input into a pre-trained conditional contrast graph encoder. The temporal embeddings of each node are extracted through a temporal convolutional network. The neighborhood information is aggregated on the cell topology graph through a graph attention propagation network, and the representation vector of each node is output. Calculate the Mahalanobis distance score of each node's representation vector relative to the normal operation representation distribution, and identify nodes whose Mahalanobis distance score exceeds the adaptive detection threshold as candidate abnormal nodes. Structural causal inference is performed on candidate anomaly nodes. Based on the pre-constructed thermal-electric causal directed acyclic graph, the do operator is applied to the thermal imaging anomaly feature variables corresponding to the candidate anomaly nodes. The directed edges between the operating condition disturbance variables and the thermal imaging anomaly feature variables are truncated while keeping the observed values ​​of the operating condition disturbance variables unchanged. It is then determined whether the causal effect of the thermal imaging anomaly feature variables is independent of the operating condition disturbance variables, and the graded early warning results are output.

2. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, The sparse Gaussian process regression model uses the Nyström sparse approximation method, takes the operating condition parameter vector as input, and outputs the mean prediction and standard deviation prediction of the temperature at each pixel position; the conditional residual field is obtained by subtracting the mean prediction from the temperature field of the current frame and dividing by the standard deviation prediction, and the region in the conditional residual field with an absolute value exceeding 2.0 constitutes the residual saliency region.

3. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, The weight of the thermally conductive edge is the reciprocal of the contact thermal resistance between adjacent cells, and the weight of the electrical connection edge is the electrothermal coupling coefficient between the corresponding cells. The graph attention propagation network sets independent attention calculation heads for the thermally conductive edge and the electrical connection edge respectively. The outputs of each attention calculation head are concatenated and then linearly transformed to obtain the final aggregated features of each node.

4. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 3, characterized in that, The temporal convolutional network employs a multi-layer dilated causal convolutional structure, with the dilation factor of each layer increasing in powers of 2. It convolves the conditional residual field statistics sequence within a historical time window of 30 frames in length to extract the temporal embedding, which includes features of temperature rise rate and thermal diffusion acceleration. The graph attention propagation network adopts the GraphSAGE aggregation framework, and the aggregation function introduces attention coefficients based on the weight of heat conduction edges, so that the information of neighboring nodes in the direction of thermal diffusion receives higher weight in the aggregation.

5. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, The pre-training of the conditional comparison graph encoder adopts a conditional topology comparison strategy: positive sample pairs are formed by normalizing the conditional residual field node feature sequences of the same cell under different operating conditions using a sparse Gaussian process regression model; negative sample pairs are formed by forming node feature sequence pairs of different cells at the same time and node frame pairs containing the Mahalanobis distance score exceeding the initial threshold; topology-aware graph enhancement operations are applied to the cell topology graph, including randomly masking part of the heat-conducting edge according to the weight ratio of the heat-conducting edge and sampling the subgraph of the node according to the node's thermal contribution. The conditional contrastive loss function is used to optimize the conditional contrastive map encoder. The conditional contrastive loss function maximizes the cosine similarity between the node representation vectors of positive sample pairs and minimizes the cosine similarity between the node representation vectors of negative sample pairs.

6. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, The normal operation representation distribution is obtained by fitting a multivariate Gaussian distribution to the node representation vectors of confirmed normal frames that have not triggered warnings in the past 90 days through maximum likelihood estimation, and is represented by the mean vector and covariance matrix; the adaptive detection threshold is the 99th percentile of the Mahalanobis distance score of all normal confirmed frame nodes within a 90-day sliding window, and is automatically updated every 24 hours.

7. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, The thermal-electric causal directed acyclic graph (DAG) is obtained by applying a PC algorithm to perform structural learning on historical operating data and controlled injection fault data. The nodes in the DAG include thermal imaging anomaly feature variables, battery management system multidimensional electrical parameter variables, and operating condition disturbance variables. When the absolute value of the difference between the causal effect of the thermal imaging anomaly feature variables after the do operator is applied and the observed effect before the application exceeds a preset judgment threshold, it is determined that the candidate anomaly node has a real thermal anomaly independent of the operating condition disturbance variables and the warning level is upgraded. Otherwise, it is determined to be a non-abnormal event caused by the operating condition disturbance and the warning output is suppressed.

8. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, It also includes the following steps: The mean of the Mahalanobis distance score distribution of all normal confirmation frame nodes within the current 30-day sliding window is calculated relative to the historical baseline mean. When the offset exceeds the preset drift detection threshold, the hyperparameters of the sparse Gaussian process regression model are updated online. The conditional residual field is recalculated using the updated sparse Gaussian process regression model, and the normal operation characterization distribution and the adaptive detection threshold are updated using newly collected normal confirmation frame data.

9. The method for detecting anomalies in an electric vehicle battery pack using thermal imaging according to claim 1, characterized in that, The graded early warning results include an early warning level identifier and a root cause attribution report. The root cause attribution report is generated in the following way: based on the attention weight matrix of each layer of the graph attention propagation network, the contribution of each node to the Mahalanobis distance score is traced back along the direction of heat diffusion propagation. The cell corresponding to the node with the highest contribution is identified as the root cause cell. The root cause cell number, the battery module number, the historical temperature rise rate curve, and the causal contribution score are output in the root cause attribution report.

10. A thermal imaging anomaly detection system for electric vehicle battery packs, characterized in that, A method for implementing an anomaly detection method for thermal imaging of an electric vehicle battery pack as described in any one of claims 1-9 includes: The data acquisition and synchronization module is configured to acquire a thermal imaging image sequence of the battery pack, and to timestamp the operating condition parameter vector collected by the battery management system with the thermal imaging image sequence, and output the time-aligned thermal imaging frame and operating condition parameter vector pair, wherein the operating condition parameter vector includes state of charge, charge / discharge rate, ambient temperature and health status. The operating condition baseline prediction module is configured to input the operating condition parameter vector into a sparse Gaussian process regression model, output the expected temperature field and standard deviation prediction under the current operating condition, and calculate the conditional residual field after standardizing the difference between the current frame temperature field and the expected temperature field by the standard deviation prediction. The topology graph construction module is configured to construct a cell topology graph with each cell in the battery pack as a node and the heat conduction path and electrical connection relationship as an edge, and to assign the conditional residual field statistical sequence of the corresponding cell in the historical time window as a node feature. The conditional comparison graph encoding module is configured to extract the temporal embedding of each node through a temporal convolutional network, aggregate the neighborhood information of the cell topology graph through a graph attention propagation network, and output the representation vector of each node. The anomaly scoring module is configured to calculate the Mahalanobis distance score of the representation vector of each node relative to the normal operation representation distribution, and to identify nodes that exceed the adaptive detection threshold as candidate anomaly nodes. The causal inference early warning module is configured to perform counterfactual intervention tests on the candidate abnormal nodes based on the thermal-electric causal directed acyclic graph, and output a graded early warning result including an early warning level identifier and a root cause attribution report.

Images