CN-121973633-A - Thermal imaging anomaly detection method and system for battery pack of electric automobile

CN121973633ACN 121973633 ACN121973633 ACN 121973633ACN-121973633-A

Abstract

The invention discloses a thermal imaging anomaly detection method and a thermal imaging anomaly detection system for an electric automobile battery pack. The method comprises the steps of synchronizing a temperature field image sequence obtained by a thermal imaging sensor with working condition parameters of a battery management system in time, predicting an expected temperature field and calculating a condition residual field, constructing a battery cell topological graph, taking residual time sequence statistics as node characteristics, extracting a battery cell space-time characterization vector by combining a time convolution network and a graph attention propagation network through a condition topology comparison self-supervision pre-training condition comparison graph encoder, determining candidate abnormal nodes based on a Markov distance novelty score, and performing inverse fact intervention test by utilizing a thermo-electric structure causal directed acyclic graph to exclude false positives, and outputting a grading early warning result with an interpretable root cause attribution report. The invention can realize the detection performance of the area under the detection curve not lower than 0.88 without abnormal sample labeling, maintains the stability of the full life cycle performance through a self-supervision reference drift feedback mechanism, and is suitable for online thermal safety monitoring of battery packs of pure electric and plug-in hybrid electric vehicles.

Inventors

SUN RAN
ZHANG CHAO
LI BING
DENG LIN
Wen Jishu

Assignees

六安职业技术学院

Dates

Publication Date: 20260505
Application Date: 20260327

Claims (10)

1. The thermal imaging anomaly detection method for the battery pack of the electric automobile is characterized by comprising the following steps of: Acquiring a thermal imaging image sequence of a battery pack and a working condition parameter vector which is synchronous with the thermal imaging image sequence in time, wherein the working condition parameter vector comprises a charge state, a charge-discharge multiplying power, an ambient temperature and a health state; Inputting the working condition parameter vector into a pre-trained sparse Gaussian process regression model to obtain an expected temperature field and standard deviation prediction under the current working condition, and normalizing the current frame temperature field and the expected temperature field of the thermal imaging image sequence by the standard deviation prediction after the difference between the current frame temperature field and the expected temperature field to obtain a condition residual field; constructing a battery cell topological graph by taking each battery cell in a battery pack as a node, taking a heat conduction path between the battery cells as a heat conduction edge and taking an electric connection relation between the battery cells as an electric connection edge, and taking a condition residual field statistic sequence of each node corresponding to the battery cell in a historical time window as a node characteristic; inputting the cell topological graph into a pre-trained condition comparison graph encoder, extracting time sequence embedding of each node through a time convolution network, aggregating neighborhood information on the cell topological graph through a graph attention propagation network, and outputting characterization vectors of each node; Calculating the mahalanobis distance score of the characterization vector of each node relative to the normal operation characterization distribution, and determining the node with the mahalanobis distance score exceeding the self-adaptive detection threshold as the candidate abnormal node; And executing structural causal inference on the candidate abnormal nodes, applying do operators to thermal imaging abnormal characteristic variables corresponding to the candidate abnormal nodes based on a pre-constructed thermal-electric causal directed acyclic graph, cutting off directed edges between the working condition disturbance variables and the thermal imaging abnormal characteristic variables, keeping the observation values of the working condition disturbance variables unchanged, judging whether the causal effect of the thermal imaging abnormal characteristic variables is independent of the working condition disturbance variables, and outputting a grading early warning result.
2. The thermal imaging anomaly detection method for the battery pack of the electric automobile is characterized in that a Nystr m sparse approximation method is adopted by a sparse Gaussian process regression model, a working condition parameter vector is taken as input, mean value prediction and standard deviation prediction of temperatures at each pixel position are taken as output, the condition residual field is obtained by dividing the difference between a current frame temperature field and the mean value prediction by the standard deviation prediction, and a region with an absolute value exceeding 2.0 in the condition residual field forms a residual salient region.
3. The method for detecting the thermal imaging abnormality of the battery pack of the electric automobile according to claim 1 is characterized in that the weight of the heat conducting side is the inverse number of the contact thermal resistance between the corresponding adjacent cells, the weight of the electric connecting side is the electric heating coupling coefficient between the corresponding cells, the drawing force transmission network is characterized in that independent attention calculating heads are respectively arranged for the heat conducting side and the electric connecting side, and the final aggregation characteristics of all nodes are obtained through linear transformation after the outputs of all the attention calculating heads are spliced.
4. The thermal imaging anomaly detection method for the battery pack of the electric automobile is characterized in that a multi-layer expansion causal convolution structure is adopted by the time convolution network, expansion factors of all layers are increased by power of 2, a condition residual field statistic sequence in a historical time window with the length of 30 frames is convolved, time sequence embedding comprising a temperature rise rate characteristic and a thermal diffusion acceleration characteristic is extracted, a GRAPHSAGE aggregation framework is adopted by the drawing force propagation network, attention coefficients based on heat conduction edge weights are introduced into an aggregation function, and neighbor node information in the thermal diffusion direction is enabled to obtain higher weights in aggregation.
5. The electric automobile battery pack thermal imaging anomaly detection method is characterized in that a condition topology contrast strategy is adopted in pre-training of a condition contrast graph encoder, the condition residual field node characteristic sequence pairs of the same battery cell after normalization of sparse Gaussian process regression models at different condition moments form positive sample pairs, node characteristic sequence pairs of different battery cells at the same moment and node frame pairs containing the mahalanobis distance score exceeding an initial threshold value form negative sample pairs, topology perception graph enhancement operation is applied to a battery cell topology graph, the topology perception graph enhancement operation comprises sampling nodes according to heat conducting edges of a heat conducting edge weight proportion random mask part and according to node thermal contribution degrees, and the condition contrast graph encoder is optimized according to a condition contrast loss function, and the condition contrast loss function maximizes cosine similarity between node representation vectors of the positive sample pairs and minimizes cosine similarity between node representation vectors of the negative sample pairs.
6. The thermal imaging anomaly detection method for the battery pack of the electric automobile is characterized in that the normal operation representation distribution is obtained by fitting a multi-element Gaussian distribution through maximum likelihood estimation by node representation vectors of normal frames which are not triggered to be early-warned in the past 90 days, the normal operation representation distribution is represented by a mean value vector and a covariance matrix, the self-adaptive detection threshold is the 99 th percentile of the Markov distance scores of all normal frame nodes in a 90-day sliding window, and the self-adaptive detection threshold is automatically updated in a period of 24 hours.
7. The method for detecting the thermal imaging anomalies of the battery pack of the electric automobile according to claim 1, wherein the thermal-electric causal directed acyclic graph is obtained by performing structural learning on historical operation data and controlled injection type fault data by using a PC algorithm, nodes in the directed acyclic graph comprise thermal imaging anomalies characteristic variables, battery management system multidimensional electrical parameter variables and working condition disturbance variables, and when the absolute value of the difference between the causal effect of the thermal imaging anomalies characteristic variables after application of a do operator and the observation effect before application exceeds a preset judgment threshold, the candidate anomalies nodes are judged to have real thermal anomalies independent of the working condition disturbance variables and upgrade early warning grades, otherwise, the candidate anomalies are judged to be non-anomalies caused by the working condition disturbance and early warning output is restrained.
8. The thermal imaging anomaly detection method for the battery pack of the electric automobile is characterized by further comprising the steps of counting the offset of the mean value of the Markov distance score distribution of all normal confirmation frame nodes in a current 30-day sliding window relative to a historical reference mean value, triggering the online updating of the super parameters of a sparse Gaussian process regression model when the offset exceeds a preset drift detection threshold value, recalculating a conditional residual field by the updated sparse Gaussian process regression model, and updating the normal operation characterization distribution and the self-adaptive detection threshold value by newly acquired normal confirmation frame data.
9. The thermal imaging anomaly detection method for the battery pack of the electric automobile is characterized in that the grading early warning result comprises early warning grade identification and a root attribution report, wherein the root attribution report is generated by reversely tracing the contribution degree of each node to the mahalanobis distance score along the thermal diffusion propagation direction based on an attention weight matrix of each layer of a graph attention propagation network, determining a cell corresponding to the node with the highest contribution degree as the root cell, and outputting the root cell number, the number of a battery module where the cell is located, a historical temperature rise rate curve and a causal contribution degree score in the root attribution report.
10. An electric vehicle battery pack thermal imaging anomaly detection system, which is used for realizing the electric vehicle battery pack thermal imaging anomaly detection method according to any one of claims 1-9, comprising: The data acquisition and synchronization module is configured to acquire a thermal imaging image sequence of the battery pack, time stamp align the working condition parameter vector acquired by the battery management system with the thermal imaging image sequence, and output a thermal imaging frame and working condition parameter vector pair after time alignment, wherein the working condition parameter vector comprises a charge state, a charge-discharge multiplying power, an ambient temperature and a health state; The working condition reference prediction module is configured to input the working condition parameter vector into a sparse Gaussian process regression model, output a desired temperature field and standard deviation prediction under the current working condition, and calculate a condition residual error field after the standard deviation prediction is standardized on the difference between the current frame temperature field and the desired temperature field; The topological graph construction module is configured to construct a topological graph of the battery cells by taking each battery cell in the battery pack as a node and taking a heat conduction path and an electric connection relation as an edge, and assign the condition residual field statistic sequence of each node corresponding to each battery cell in a historical time window as a node characteristic; the condition contrast diagram coding module is configured to extract time sequence embedding of each node through a time convolution network, aggregate neighborhood information of the battery cell topological diagram through a diagram attention propagation network and output characterization vectors of each node; The anomaly scoring module is configured to calculate the mahalanobis distance score of the characterization vector of each node relative to the normal operation characterization distribution, and determine the node exceeding the self-adaptive detection threshold as a candidate anomaly node; And the causal inference early warning module is configured to execute anti-fact intervention detection on the candidate abnormal nodes based on the thermo-electric causal directed acyclic graph and output a hierarchical early warning result comprising an early warning level identifier and a root cause attribution report.

Description

Thermal imaging anomaly detection method and system for battery pack of electric automobile Technical Field The invention relates to the technical field of electric automobile safety detection, in particular to a thermal imaging anomaly detection method and a thermal imaging anomaly detection system for an electric automobile battery pack. Background Thermal runaway of an electric car battery pack is one of the main causes of fire and explosion accidents of the electric car. Thermal runaway is typically initiated by electrochemical failure mechanisms such as Internal Short Circuit (ISC), lithium evolution, overcharge thermal runaway, and the like, with thermal precursor phases of tens of minutes to hours preceding the irreversible runaway phase. If the thermal precursor can be accurately detected and intervened in the thermal precursor stage (at least 30 minutes before thermal runaway occurs), the catastrophic accident can be effectively avoided. Thermal imaging techniques are capable of acquiring a two-dimensional continuous temperature field of the surface of the battery pack in a non-contact manner. Compared with a mode that a Battery management system (Battery MANAGEMENT SYSTEM, BMS) only collects a small number of discrete temperature measurement points, the thermal imaging sensor can provide full-field space temperature distribution information, and can theoretically sense the initiation and the expansion of thermal anomalies earlier. However, the existing battery thermal anomaly detection technology based on thermal imaging has the following four core technical defects. First, the condition interference leads to a high false alarm rate. The battery pack is charged in different States (SOC), charge-discharge multiplying power (C-rate) and ambient temperature [ ], and the battery pack is connected with the battery pack) And under the state of health (SOH), the difference of the normal thermal distribution forms is obvious, and the normal maximum temperature difference between adjacent modules can reach more than 15 ℃. The prior method based on a fixed temperature threshold or a simple statistical threshold has the false alarm rate of more than 30% under a fast charge condition (the charge-discharge multiplying power is not lower than 2C), and has serious insufficient sensitivity under a low-multiplying power discharge condition (the charge-discharge multiplying power is not higher than 0.2C), so that weak signals in the early stage of thermal runaway are missed to be detected. Second, abnormal samples are extremely rare. Dangerous and abnormal events such as thermal runaway, internal short circuit and the like are very rare in actual vehicle operation, and the proportion of normal samples to abnormal samples is seriously unbalanced (generally exceeds 10000:1), so that a deep learning detection model trained in a supervision mode has a serious class imbalance problem for abnormal classes. Meanwhile, the thermal characteristics of different electrochemical systems (lithium iron phosphate, ternary lithium, sodium ions and the like) are remarkably different, the model trained on a single chemical system is extremely poor in generalization performance across systems, and cannot be directly transferred to other vehicle types or platforms. Third, the single-mode visual inspection is not reliable enough. When the detection is carried out only by relying on thermal imaging image features, inherent noise of a thermal imaging sensor, exogenous radiation hot spots caused by direct sunlight, false heat sources generated by specular reflection of a metal shell of a battery pack and normal transient local temperature fluctuation in the charging and discharging process can cause false hot spot misjudgment, so that the false alarm rate is high. In addition, the detection conclusion output by the existing method lacks physical interpretability, and the detection result cannot be accurately traced to a specific fault cell. Fourth, full life cycle performance continues to degrade. As the battery cycles ages (30% to 80% increase in internal resistance after 1000 typical life cycles), the normal thermal profile of the battery pack will systematically shift under the same conditions. The existing detection model lacks the sensing and online self-adaption capability of normal thermal characteristic drift caused by aging, so that the false alarm rate is continuously increased along with the service time of the battery, and the practical value is lost even after the 3 rd year of service. Aiming at the four technical defects, a complete technical scheme for organically coupling working condition dynamic benchmark prediction, unsupervised comparison self-supervision characterization learning, graph topological structure priori modeling and thermo-electric structure causal inference does not exist in the prior art, so that the false alarm rate and the false alarm rate are difficult to meet application requirements in an actual veh