Fault diagnosis and early warning method for intelligent thermal management system of power battery

Abstract

The invention discloses a fault diagnosis and early warning method for an intelligent thermal management system of a power battery, and relates to the technical field of new energy automobile power battery safety. Comprising the following steps executed in sequence: S1) based on real-time monitoring data of a distributed temperature sensor in a battery module, constructing a dynamic space-time neural field model through an adaptive graph neural network, and generating a thermal field fingerprint representing a thermal propagation topology mode; according to the fault diagnosis and early warning method for the intelligent thermal management system of the power battery, a dynamic space-time neural field model is constructed, a sub-DEG C weak thermal abnormal signal is accurately captured, and a micro short circuit fault is recognized in advance in combination with electric heating delay analysis, so that the problem that a fault precursor signal is submerged by a complex working condition is solved; a double-channel verification mechanism depends on double mutual verification of physical law rigid constraint and intelligent residual feature recognition, the misjudgment risk caused by cooling system fluctuation and sensor noise is eliminated, and the early warning accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of new energy vehicle power battery safety technology, and specifically to a fault diagnosis and early warning method for a power battery intelligent thermal management system. Background Art

[0002] In the new energy vehicle sector, power battery safety is a core concern. Thermal runaway, triggered by hidden defects such as internal micro-shorts, is a major source of destructive risk. In their initial stages, these internal faults often produce only extremely weak and transient temperature anomaly signals, such as recurring transient temperature spikes or subtle temperature distribution pattern shifts in specific cells or localized areas. However, existing fault diagnosis and early warning technologies based on thermal management systems are significantly inadequate in identifying these early, weak precursors to faults. Mainstream methods rely heavily on preset absolute temperature thresholds or temperature rise rate thresholds for judgment. This single threshold mechanism faces significant challenges under complex real-world operating conditions. First, during normal operation of a power battery intelligent thermal management system, the start-stop operation of cooling and heating actuators (such as water pumps, fans, and PTCs) inevitably introduces significant localized rapid temperature fluctuations, whose amplitude and characteristics may be similar to the weak anomaly signals generated in the early stages of a micro-short. Second, inherent measurement noise and environmental interference from temperature sensors further contaminate these already weak fault signals. Furthermore, the Joule heating effect generated by the battery itself during normal high-current charging and discharging can also mask small local abnormal temperature rises. The combined effect of these factors makes it difficult for diagnostic systems based on simple thresholds to effectively distinguish between true internal fault precursors and normal system dynamics or noise interference, resulting in a serious lag in the early warning of progressive thermal faults such as internal micro-shorts. The alarm is often triggered only when the fault develops to a more serious stage, greatly compressing the response time window of the safety system and may even cause false alarms or missed alarms, failing to meet the urgent need for early and accurate warning of battery safety. Therefore, there is an urgent need for an intelligent fault diagnosis and warning method that can effectively improve the ability to early identify and highly accurately warn of weak, specific pattern thermal anomaly signals caused by internal micro-shorts in power batteries. Summary of the Invention

[0003] To achieve the above objectives, the present invention is implemented through the following technical solutions: A fault diagnosis and early warning method for a power battery intelligent thermal management system, comprising the following steps performed in sequence:

[0004] S1) Based on real-time monitoring data from distributed temperature sensors within the battery module, a dynamic spatiotemporal neural field model is constructed through an adaptive graph neural network to generate a thermal field fingerprint that characterizes the topological pattern of heat propagation.

[0005] S2) aligning the thermal field fingerprint with a theoretical Joule heat power distribution calculated based on the real-time current and internal resistance parameters of the battery in time and space, and analyzing the difference in delay characteristics between the actual temperature rise and the theoretical value through cross-correlation delay detection;

[0006] S3) performs dual-channel verification on the abnormal signal identified in step S2):

[0007] Channel 1: Calculate the theoretical solution boundary of the heat diffusion equation in the abnormal area to verify whether the signal meets the material thermodynamic parameter constraints;

[0008] Channel 2: Generate a background noise signal using a generative adversarial network, and use the discriminator to identify the residual characteristic patterns in the real signal;

[0009] S4) Trigger a graded warning response based on the verification results.

[0010] Preferably, the step S1) includes:

[0011] During the training phase, transient thermal disturbance data within a preset temperature disturbance range simulating micro-short circuit characteristics is injected into the adaptive graph neural network. The preset temperature disturbance range is set based on the micro-short circuit characteristic simulation requirements and the model training accuracy.

[0012] During online operation, dynamic graph convolution is used to extract the spatial correlation strength matrix and time propagation attenuation features.

[0013] Preferably, in step S2):

[0014] When it is detected that the actual local temperature rise is ahead of the theoretical Joule heat power by milliseconds, it is marked as a suspected micro-short circuit signal.

[0015] Preferably, the dual-channel verification in step S3) includes:

[0016] 3a) In channel 1, the theoretical heat conduction boundary is calculated by partitioning the non-uniform material parameters of each battery region, and only the signal within the theoretical solution boundary is retained;

[0017] 3b) In channel 2, the generator reconstructs a synthetic background signal containing system fluctuations based on normal operating data, and the discriminator identifies the spatial resonant frequency characteristics in the real signal through comparison.

[0018] Preferably, the step 3b) further comprises:

[0019] The residual feature pattern output by the discriminator is matched with the historical micro-short circuit case library for similarity, and the fault confidence score is output.

[0020] Preferably, the hierarchical warning response in step S4) includes:

[0021] Level 1 response: Record abnormal pattern characteristics and increase the priority of similar signal identification;

[0022] Secondary response: Reduce battery output power and initiate enhanced temperature monitoring in abnormal areas;

[0023] Level 3 response: triggers the fuse protection circuit and outputs the fault location coordinates.

[0024] Preferably, the process of constructing the dynamic spatiotemporal neural field model includes:

[0025] Mapping sparse temperature sensor data to virtual three-dimensional heat conduction grid nodes;

[0026] The transfer path weights between graph nodes are dynamically updated according to real-time heat flow changes.

[0027] Preferably, the fault diagnosis and early warning method of the power battery intelligent thermal management system further comprises, after step S3):

[0028] The fault signals verified through dual channels are classified into different types to distinguish between micro short circuit, sensor failure or cooling system failure.

[0029] Preferably, the calculation of the theoretical Joule heat power distribution is based on:

[0030] Dynamic variation model of battery cell internal resistance with temperature and state of charge;

[0031] Real-time charge and discharge current data collected by the current sensor.

[0032] A fault diagnosis device for a power battery intelligent thermal management system, comprising:

[0033] Temperature sensor arrays, current and voltage monitoring modules, and embedded processors are arranged in key areas of the battery module;

[0034] The processor is configured to execute the steps of a fault diagnosis and early warning method for a power battery intelligent thermal management system,

[0035] Wherein, the embedded processor includes:

[0036] a spatiotemporal neural field calculation unit, configured to implement step S1);

[0037] an electrothermal delay analysis unit, configured to implement step S2);

[0038] Dual-channel verification engine, used to implement step S3).

[0039] The present invention provides a fault diagnosis and early warning method for a power battery intelligent thermal management system. It has the following beneficial effects:

[0040] The fault diagnosis and early warning method of the power battery intelligent thermal management system accurately captures sub-℃ weak thermal anomaly signals by constructing a dynamic spatiotemporal neural field model, and realizes early identification of micro-short circuit faults by combining electrothermal delay analysis, solving the problem of fault precursor signals being overwhelmed by complex working conditions; the dual-channel verification mechanism relies on the dual mutual verification of the rigid constraints of physical laws and intelligent residual feature recognition to eliminate the risk of misjudgment caused by cooling system fluctuations and sensor noise, and improve the accuracy of early warning; the coordinated implementation of the hierarchical response strategy and the cluster fault direct fuse mechanism builds a closed-loop safety protection system from precise intervention in the risk initiation stage to rapid isolation of the critical point of thermal runaway, effectively extending the safety margin of the battery system.

[0041] The fault diagnosis and early warning method of this power battery intelligent thermal management system is based on precise thermal modeling technology with three-dimensional coupling of temperature, state of charge, and aging, ensuring diagnostic reliability under all life cycle conditions and reducing maintenance costs. The collaborative design of heterogeneous hardware architecture and anti-interference sensor network breaks through the real-time bottleneck, enabling complex analysis processes to meet the stringent latency requirements of the vehicle system. The intelligent closed loop of fault location and maintenance decision-making not only improves the safety level, but also provides high-value data support for battery health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a system module interaction diagram of a fault diagnosis and early warning method for a power battery intelligent thermal management system according to the present invention;

[0043] Figure 2 The figure is a flow chart of a fault diagnosis and early warning method for a power battery intelligent thermal management system according to the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] See also Figure 1 and Figure 2 The present invention provides a technical solution: a fault diagnosis and early warning method for a power battery intelligent thermal management system, comprising the following steps performed in sequence:

[0046] S1) Based on real-time monitoring data from distributed temperature sensors within the battery module, a dynamic spatiotemporal neural field model is constructed through an adaptive graph neural network to generate a thermal field fingerprint that characterizes the topological pattern of heat propagation.

[0047] S2) Temporally and spatially aligning the thermal field fingerprint with the theoretical Joule thermal power distribution calculated based on the battery's real-time current and internal resistance parameters, and analyzing the difference in delay characteristics between the actual temperature rise and the theoretical value through cross-correlation delay detection;

[0048] S3) performs dual-channel verification on the abnormal signal identified in step S2):

[0049] Channel 1: Calculate the theoretical solution boundary of the heat diffusion equation in the abnormal area to verify whether the signal meets the material thermodynamic parameter constraints;

[0050] Channel 2: Generate a background noise signal using a generative adversarial network, and use the discriminator to identify the residual characteristic patterns in the real signal;

[0051] S4) Trigger a graded warning response based on the verification results.

[0052] It should be further explained that, in the specific implementation process, a dynamic spatiotemporal neural field model is constructed through an adaptive graph neural network based on the multi-point temperature data streams collected in real time by distributed temperature sensors in the battery module; during the model training phase, 0.1-0.5℃ transient thermal disturbance data simulating the characteristics of battery micro-short circuit faults are injected to enable the network to autonomously learn the essential differences in topological structures between normal thermal fluctuations and abnormal heat propagation; during online operation, sparse sensor data are mapped to virtual three-dimensional heat conduction grid nodes, and the dynamic graph convolution layer is used to extract the spatial correlation intensity matrix and time propagation attenuation characteristics in real time to generate a thermal field fingerprint with fault sensitivity.

[0053] Then, an electrothermal delay coupling analysis is performed: the spatial distribution of the theoretical Joule thermal power is calculated based on the real-time charge and discharge current of the battery and the internal resistance parameters that change with temperature and state of charge; the thermal field fingerprint and the theoretical thermal power are aligned at the millisecond level, and the delay characteristics of the actual temperature rise and the theoretical value are analyzed through the cross-correlation delay detection algorithm. That is, if the actual temperature rise in a specific local area is detected to be milliseconds ahead of the theoretical Joule thermal power, it is marked as a suspected micro-short circuit signal; the normal operation of the thermal management system or the fluctuations caused by sensor noise are in line with the theoretical thermal inertia delay law.

[0054] Dual-channel verification is performed on the marked abnormal signals: Channel 1 uses non-uniform material parameter partitioning to calculate the theoretical solution boundary of the heat diffusion equation in the abnormal area. If the signal intensity exceeds the physical limit range determined by the specific heat capacity and thermal conductivity of the battery material, it is determined to be a hardware failure and eliminated; Channel 2 uses a generative adversarial network. The generator reconstructs a high-simulation background signal containing system fluctuations and noise based on historical normal operating data. The real signal and the synthetic signal are compared through the discriminator. If spatial resonant frequency characteristics or specific topological mode residues that do not exist in the background signal are identified, it is confirmed as a micro-short circuit fingerprint.

[0055] Finally, a graded response is triggered based on the verification results: for low-confidence anomalies, only characteristic patterns are recorded and monitoring sensitivity is improved; for medium-risk signals, battery output power is limited and local enhanced monitoring is initiated; for high-confidence faults, fuse protection is immediately triggered and three-dimensional positioning coordinates are output.

[0056] Step S1) includes: injecting transient thermal disturbance data within a preset temperature disturbance range of 0.1-0.5°C to simulate micro-short circuit characteristics into the adaptive graph neural network during the training phase, wherein the preset temperature disturbance range is set based on the micro-short circuit characteristic simulation requirements and the model training accuracy; and extracting the spatial correlation intensity matrix and time propagation attenuation characteristics through dynamic graph convolution during online operation.

[0057] 2 It should be further explained that in the specific implementation process, during the construction of the dynamic spatiotemporal neural field model, a specifically designed simulated fault data set is first injected into the adaptive graph neural network in the offline training stage: this data set is generated by superimposing a transient local temperature rise spike signal of the order of 0.1-0.5°C on the normal battery thermal management operating condition data, where the spike duration simulates the initial characteristics of a real micro-short circuit and is randomly distributed in the three-dimensional spatial coordinates of the battery; the network autonomously explores the essential differences between normal thermal fluctuations including the start-stop disturbance of the cooling system and the injected fault signal through a comparative learning mechanism, focusing on capturing two key features of abnormal heat propagation in the spatial topological structure, namely: the abnormal enhancement of the correlation strength between specific nodes and the sudden change of the heat flow time decay coefficient.

[0058] During the online implementation phase, real-time data streams from sparsely distributed physical temperature sensors are mapped to corresponding nodes in a virtual three-dimensional heat conduction grid. When a battery charge / discharge state switch or cooling system activation is detected, the weights of the transfer paths between graph nodes are dynamically and adaptively updated. The reliability of the heat flow direction is calculated based on the real-time rate of change of the temperature gradients at adjacent nodes. If a path experiences a heat flow reversal for three consecutive sampling periods, its weight coefficient is reduced. Spatial correlation features are aggregated layer by layer through multi-layer graph convolution operations, ultimately outputting a thermal field fingerprint consisting of a spatial correlation strength matrix and a temporal propagation attenuation coefficient. If a local region in this fingerprint contains a combination of spatial correlation strength values ​​exceeding two standard deviations of the historical mean and a temporal attenuation coefficient below the lower limit of the normal range, the subsequent electrothermal delay analysis process is triggered. For areas with insufficient temperature sensor coverage, a forward inference algorithm based on the heat conduction equation is used to supplement the virtual node data to ensure the spatial continuity of the thermal field fingerprint.

[0059] In step S2):

[0060] When it is detected that the actual local temperature rise is ahead of the theoretical Joule heat power by milliseconds, it is marked as a suspected micro-short circuit signal.

[0061] 3 It should be further explained that in the specific implementation process, in the electrothermal delay coupling analysis stage, the spatial distribution model of the theoretical Joule thermal power is first calculated based on the real-time charge and discharge current waveforms collected by the high-precision current sensor, combined with the internal resistance parameters of the battery cell that dynamically changes with temperature and state of charge; at the same time, the generated thermal field fingerprint data stream is received, and the two are synchronized to millisecond-level accuracy through the timestamp alignment module.

[0062] When a step change in current is detected, the delay characteristic analysis is started: at the starting moment of the rising edge of the current pulse, the change trajectories of the theoretical Joule thermal power curve and the actual temperature rise curve are recorded respectively, and the phase offset of the two curves is calculated by the cross-correlation delay detection algorithm. If the actual temperature rise curve of a specific battery area has a systematic lead of 0.5 milliseconds to 5 milliseconds relative to the theoretical curve, and this lead is repeated in three consecutive current pulse cycles, it is marked as a suspected micro-short circuit signal; conversely, if the actual temperature rise lags behind the theoretical value by more than 1 millisecond or shows irregular random offsets, it is judged as normal thermal management system operation or sensor noise interference.

[0063] For the special operating conditions of high-rate battery charge and discharge, a current change rate compensation factor is introduced. When the current change rate exceeds a threshold, the delayed detection window is automatically shortened to one-third of its original length to prevent the thermal accumulation effect under high-current conditions from masking weak lead characteristics. During constant-current operation, long-term correlation analysis is enabled to accumulate statistics on the frequency and spatial concentration of lead events over a ten-minute period. All marked suspected signals are attached with a timestamp, spatial coordinates, and lead feature vector and transmitted to the dual-channel verification module for further judgment.

[0064] The dual-channel verification in step S3) includes:

[0065] 3a) In channel 1, the theoretical heat conduction boundary is calculated by partitioning the non-uniform material parameters of each battery region, and only the signal within the theoretical solution boundary is retained;

[0066] 3b) In channel 2, the generator reconstructs a synthetic background signal containing system fluctuations based on normal operating data, and the discriminator identifies the spatial resonant frequency characteristics in the real signal through comparison.

[0067] 4. It is important to further clarify that, during the dual-channel verification phase, physical constraint screening and adversarial generation verification are performed collaboratively on abnormal signals flagged by electrothermal delay analysis. Physical constraint screening first locates the spatial coordinates of the abnormal signal and retrieves the measured specific heat capacity and thermal conductivity parameters of the battery material in that area. If material batch differences are detected within the module, a non-uniform partitioning model is used to divide the computational domain. The theoretical temperature rise boundary curve is calculated based on the heat diffusion equation. If the signal intensity peak exceeds 1.5 times the upper limit of the theoretical solution or the duration is shorter than the lower limit of the material's thermal relaxation time, the sensor is deemed to have failed and is removed. Only signals within the theoretical boundary are retained for the next step.

[0068] Adversarial generation and verification are initiated simultaneously: the generator draws on a database of historical normal operating conditions to reconstruct a synthetic background signal consisting of cooling pump vibration interference, current sampling noise, and ambient temperature fluctuations. The discriminator receives a time-frequency mixture of the real system signal and the synthetic signal and identifies abnormal patterns through a three-stage residual analysis. The first stage detects abnormal energy accumulation in the 3-5 Hz characteristic frequency band in the frequency domain. The second stage spatially analyzes whether the abnormal signal forms a concentric temperature gradient distribution centered on the fault point. The third stage compares the topological similarity of the abnormal pattern with historical micro-short circuit cases. If all three stages detect residual features not present in the synthetic background, a spatial resonance frequency fingerprint and a fault confidence score are output. In the event of conflicting conclusions between the physical screening and the adversarial verification, cross-channel arbitration is initiated: if the physical screening passes but the adversarial verification confidence falls below a threshold, the warning is downgraded to a Level 1 alert. If the physical screening fails but the adversarial verification reveals strong spatial resonance features, a sensor calibration procedure is triggered.

[0069] Step 3b) also includes:

[0070] The residual characteristic pattern output by the discriminator is matched against a historical micro-short case library for similarity, and a fault confidence score is output. It should be further explained that during the case matching phase of adversarial generation verification, the residual characteristic pattern output by the discriminator is first decomposed into a three-dimensional topological feature matrix: the first dimension extracts the energy distribution envelope of the 3-5Hz characteristic frequency band in the frequency domain, the second dimension quantifies the attenuation slope of the spatial concentric circle temperature gradient, and the third dimension records the connectivity of abnormal nodes in the heat flow propagation path.

[0071] The historical micro-short circuit case library adopts a dynamic hierarchical storage architecture: the first-level case library stores the full parameter characteristics of standard micro-short circuit events under laboratory conditions; the second-level case library includes medium- and high-risk cases confirmed by dual-channel verification during actual vehicle operation, and divides the storage area according to the degree of battery aging.

[0072] Similarity matching performs triple dynamic weighted fusion: first, the correlation coefficient of the frequency domain envelope is compared, then the spatial attenuation slope deviation is analyzed, and finally the node connectivity difference is verified; when the total weighted score of the three items exceeds 0.85, it is determined to be a strongly similar case and a high confidence score is directly output.

[0073] For medium-matching cases with scores of 0.6-0.85, enhanced verification of the time dimension is initiated: it is retrieved whether similar feature patterns appear at the spatial coordinate point more than three times in the past 24 hours. If so, the confidence increase mechanism is triggered.

[0074] In special cases, when a frequency domain signature match is detected but the spatial distribution is abnormal, an interference source elimination process is activated. The current operating condition is compared with the background spectrum signatures of high-frequency vibration or electromagnetic interference periods of the cooling system in the case library. If there are more than five overlapping interference peaks, the confidence level is lowered. The final output module converts the matching results into a continuous fault confidence score from 0 to 1.0. Signals exceeding 0.9 are marked in the red channel and directly passed to the third level response, avoiding delays in the tiered decision-making.

[0075] The graded warning response in step S4) includes:

[0076] Level 1 response: Record abnormal pattern characteristics and increase the priority of similar signal identification;

[0077] Secondary response: Reduce battery output power and initiate enhanced temperature monitoring in abnormal areas;

[0078] Level 3 response: triggers the fuse protection circuit and outputs the fault location coordinates.

[0079] 6 It should be further explained that, in the specific implementation process, after obtaining the fault confidence score of the dual-channel verification output, the system triggers differentiated responses according to the score threshold range: a first-level response is performed for low-confidence signals with a score of 0.3 to 0.6, extracting the frequency domain envelope features and spatial topological fingerprints of the abnormal signals, encrypting and storing them in the historical feature library and generating dynamic learning labels, so that the subsequent recognition weight of similar signals is increased to 3 times the basic value; a second-level response is initiated for medium-risk signals with a score of 0.6 to 0.9: the output current is limited to less than 50% of the rated value through the battery management system, and at the same time A redundant temperature sensor array is activated within a 5-centimeter radius around the abnormal coordinate point, and the sampling frequency is increased from the conventional 1Hz to 20Hz for enhanced monitoring. If the temperature rise rate in the area exceeds the safety threshold or the confidence level rises to above 0.9 within 10 minutes, the system automatically upgrades to a Level 3 response. For high-risk signals exceeding 0.9, a Level 3 response is immediately implemented: a fuse instruction code and positioning information containing the module number, stacking position, and three-dimensional coordinates are sent to the main control unit. While disconnecting the faulty battery cell circuit, the positioning data is compressed and encrypted via the on-board communication module and uploaded to the cloud diagnostic platform. Under special operating conditions, if multiple adjacent cells are detected to have signals of 0.7 or above at the same time, the Level 2 response is skipped and the regional fuse protection is directly triggered, completing fault isolation within 0.2 seconds.

[0080] The construction process of the dynamic spatiotemporal neural field model includes:

[0081] Mapping sparse temperature sensor data to virtual three-dimensional heat conduction grid nodes;

[0082] The transfer path weights between graph nodes are dynamically updated according to real-time heat flow changes.

[0083] 7 It needs to be further explained that, in the specific implementation process, during the online construction of the dynamic spatiotemporal neural field model, when the real-time data of sparsely distributed physical temperature sensors are mapped to the virtual three-dimensional heat conduction grid nodes, the heat conduction forward inference algorithm is used for areas with insufficient sensor coverage: with three or more adjacent sensor nodes as boundary conditions, the theoretical temperature change trajectory of the virtual node is calculated based on the non-steady-state heat conduction equation. When the fitting error between the inferred value and the measured boundary data is lower than the allowable threshold for five consecutive cycles, the virtual node is activated to participate in the graph convolution operation.

[0084] The mechanism for dynamically updating the transfer path weights between graph nodes includes dual triggering conditions: when the battery switches between charge and discharge states, the heat flow direction sensitivity coefficient is adjusted according to the current change rate; when the cooling system actuator action is detected, a compensation factor of the actuator's characteristic spectrum is automatically injected.

[0085] The weight update logic follows the irreversible principle of thermodynamics: if a path detects that heat flow flows from a low-temperature node to a high-temperature node and the temperature gradient exceeds the critical value of the material, the weight of the path is immediately reset to zero and marked as an abnormal path; for conventional paths, the weight value is dynamically adjusted according to the correlation coefficient of the temperature change rate of adjacent nodes, and the weight of the path with a correlation coefficient lower than 0.5 is attenuated to 30% of the baseline value.

[0086] The graph network that completes the weight update operates through three layers of heterogeneous convolution: the first layer extracts the spatial correlation strength matrix between nodes, the second layer analyzes the time decay characteristics of heat flow propagation, and the third layer integrates spatiotemporal features to output a thermal field fingerprint. If a region in the output fingerprint simultaneously meets the conditions where the spatial correlation strength value is greater than two standard deviations of the historical mean and the time decay coefficient remains consistently below the normal range, a high-priority electrothermal delay analysis is automatically triggered.

[0087] The fault diagnosis and early warning method of the power battery intelligent thermal management system further includes, after step S3):

[0088] The fault signals verified through dual channels are classified into different types to distinguish between micro short circuit, sensor failure or cooling system failure.

[0089] 8 It should be further explained that in the specific implementation process, after the dual-channel verification is completed, multi-source feature fusion classification is performed on the confirmed fault signal. First, the material boundary compliance score of the physical constraint screening link and the spatial resonance frequency energy ratio of the adversarial generation verification output are extracted, and the topological similarity of the historical case library matching is combined to form a three-dimensional classification feature vector.

[0090] The classification decision tree adopts a two-level architecture, including the following:

[0091] The first level distinguishes between hardware failure and battery failure. If the material boundary conformity is lower than 0.3 and the spatial resonance energy is weak, it is judged as sensor failure or signal transmission interference, triggering the self-calibration program and freezing the node data for subsequent analysis; if the material boundary conformity is higher than 0.7 and there are significant spatial resonance characteristics, the second level is entered to subdivide the battery failure type.

[0092] The second level implements precise classification of battery faults: When the energy in the 3-5Hz frequency band exceeds 50% of the total spectrum energy and the temperature gradient is distributed in concentric circles, it is confirmed as a micro-short circuit fault. If the spatial resonance feature is weak but broadband vibration energy is detected that is synchronized with the start and stop of the coolant pump, it is classified as a cooling system flow anomaly. Under special operating conditions, when high-rate charging occurs, critical fluctuations in material boundary conformity and no typical resonance features are present, it is marked as poor interface contact caused by battery aging. After classification is completed, a diagnostic report is automatically generated, in which the micro-short circuit fault location accuracy reaches the level of a single battery cell within the module, cooling system faults are traced to specific actuator units, and all non-battery faults trigger system maintenance alarms rather than safety fuses.

[0093] The calculation of the theoretical Joule heating power distribution is based on:

[0094] Dynamic variation model of battery cell internal resistance with temperature and state of charge;

[0095] Real-time charge and discharge current data collected by the current sensor.

[0096] 9 It should be further explained that, in the specific implementation process, in the theoretical Joule heat power distribution calculation, the temperature-state of charge coupled internal resistance dynamic modeling is adopted: a three-dimensional parameter mapping table is established, the first dimension is the battery temperature range, -20℃ to 60℃ is divided into 10 gradients, the second dimension is the state of charge range, 20% to 100% is divided into 8 gradients, and the third dimension is the number of aging cycles, 0-2000 times is divided into 5 stages; the benchmark internal resistance value of each grid node is calibrated through pulse discharge experiments, and during online operation, the instantaneous value of the internal resistance under the current working conditions is dynamically output through the trilinear interpolation algorithm based on the state of charge data provided by the real-time temperature sensor and the battery management system.

[0097] Current data processing uses a bidirectional verification mechanism: the charging and discharging current waveforms collected by the main current sensor are aligned with the equivalent current inferred from the battery cell voltage change rate at the millisecond level. When the deviation between the two exceeds 15% of the rated value, redundant sensor arbitration is initiated. For high-rate pulse currents, phase compensation factors are injected on the rising and falling edges to eliminate measurement delays caused by wire inductance.

[0098] The calculation of thermal power spatial distribution incorporates cell topology corrections: Based on the series-parallel connection relationship of each battery cell within the module and the heat dissipation imbalance coefficient calibrated by an infrared thermal imager, a position weighting factor is added to the theoretical Joule heating formula. The weighting of corner regions is increased to 1.3 times that of the center region. The resulting thermal power distribution map is strictly aligned in time and space with the thermal field fingerprint output by the spatiotemporal neural field model, and the spatial resolution matches the smallest unit of the virtual grid.

[0099] A fault diagnosis device for a power battery intelligent thermal management system, comprising:

[0100] Temperature sensor arrays, current and voltage monitoring modules, and embedded processors are arranged in key areas of the battery module;

[0101] The processor is configured to execute the steps of a fault diagnosis and early warning method for a power battery intelligent thermal management system,

[0102] Among them, embedded processors include:

[0103] a spatiotemporal neural field calculation unit, configured to implement step S1);

[0104] an electrothermal delay analysis unit, configured to implement step S2);

[0105] Dual-channel verification engine, used to implement step S3).

[0106] It should be further explained that, in the specific implementation process, the fault diagnosis device is deployed on the surface of the battery module through a distributed temperature sensor array in a honeycomb topology. The layout density in the key area is increased to three times that of the conventional area. Each sensor node integrates a three-wire anti-interference circuit to eliminate electromagnetic crosstalk. The current and voltage monitoring module adopts a dual-mode architecture of shunt and Hall sensor. The main circuit is equipped with a milliohm-level precision shunt to collect charging and discharging currents. The single cell voltage is obtained synchronously through an isolated Hall sensor. The two types of signals are input into the embedded processor after hardware filtering.

[0107] The processor's heterogeneous computing architecture includes three dedicated coprocessors: the spatiotemporal neural field computing unit has a built-in hardware-accelerated graph convolution engine that dynamically loads virtual grid node data and performs real-time weight updates; the electrothermal delay analysis unit is equipped with a high-precision clock synchronization module to achieve microsecond-level alignment of current waveform and temperature data; and the dual-channel verification engine integrates an FPGA hard core for physical constraint screening and a neural network accelerator for adversarial generation verification.

[0108] After the temperature and current data are synchronously collected, the first stage is to generate a thermal field fingerprint in the neural field calculation unit; the second stage is to perform millisecond-level advance detection in the delay analysis unit; the verification stage physically screens the thermodynamic boundaries of the hard-core parallel computing materials, while the neural network accelerator generates background noise and extracts residual features; finally, the dual-channel results of the main control core fusion trigger the hierarchical response instructions.

[0109] The fuse protection circuit adopts an optically isolated drive design. After receiving the third-level response command, it cuts off the target battery cell circuit within 0.5 milliseconds. At the same time, the fault location coordinates are encapsulated as encrypted data frames and uploaded to the cloud diagnosis platform through the CAN-FD bus. The platform automatically links the historical case library to generate a maintenance guidance plan.

[0110] It should be further explained that, in the specific implementation process, during the fault diagnosis and early warning process of the power battery intelligent thermal management system, the distributed temperature sensor devices on the surface of the battery module are first used to obtain multi-point temperature monitoring data streams in real time, and a dynamic spatiotemporal neural field model is constructed through an adaptive graph neural network. During the model training phase, transient thermal disturbance data simulating the characteristics of micro-short-circuit faults inside the battery are injected, and the disturbance amplitude is controlled at an extremely small level, allowing the network to autonomously learn the essential differences in the spatial topological structure between normal thermal fluctuations and abnormal heat propagation. During the online operation phase, the sparse sensor data is mapped to the nodes of a virtual three-dimensional heat conduction grid, and a dynamic graph convolution layer is used to extract the spatial correlation strength characteristics and the time decay characteristics of heat propagation in real time to generate a thermal field feature fingerprint with fault sensitivity.

[0111] Subsequently, an electrothermal delay coupling analysis is performed. Based on the real-time charge and discharge current waveforms collected by the high-precision current sensor device, combined with the internal resistance characteristic parameters of the battery cell that change with temperature and state of charge, the theoretical Joule thermal power spatial distribution model is calculated. The generated thermal field feature fingerprint is aligned with the theoretical thermal power distribution with high precision, and the phase offset characteristics of the actual temperature rise curve and the theoretical curve are analyzed using a cross-correlation delay detection algorithm. When the actual temperature rise of a specific battery area is detected to be systematically ahead of the theoretical calculated value and this phenomenon has repeatable stability, it is marked as a suspected micro-short circuit signal; if a lag or random fluctuation in the temperature rise is detected, it is determined to be normal operation of the thermal management system or environmental interference.

[0112] A dual-channel verification is performed on the marked abnormal signals. The first channel performs physical constraint screening: after locating the spatial coordinates of the signal, the measured thermodynamic parameters of the battery material in the area are called to solve the boundary range of the theoretical solution of the heat diffusion equation. If the signal intensity characteristics exceed the thermal conductivity limit of the material or the duration is shorter than the thermal relaxation time of the material, it is determined to be a hardware failure and eliminated. The second channel implements adversarial generation verification: a generative adversarial network is used to reconstruct a synthetic background signal containing normal system fluctuations and typical noise. The spectrum and spatial distribution characteristics of the real signal and the synthetic signal are compared through the discriminator, focusing on detecting whether there is a specific frequency band energy concentration phenomenon and characteristic heat diffusion topological pattern that are not included in the background signal. When the conclusions of the dual-channel verification conflict, the cross-channel arbitration mechanism is activated, and the warning level is dynamically adjusted according to the degree of deviation of the signal characteristics.

[0113] Fault signals that pass verification enter the classification decision process, extracting physical screening compliance, spatial resonance feature strength, and historical case matching to form a multidimensional classification vector. A hierarchical decision tree is used to first distinguish between hardware faults and battery faults. Battery faults are further subdivided into internal micro-shorts, cooling system anomalies, and poor interface contact caused by aging. A diagnostic report is then generated, including the fault location coordinates.

[0114] Ultimately, a graded response is triggered based on the fault type and confidence level: low-risk signals record characteristic patterns and enhance monitoring sensitivity; medium-risk signals limit battery output power and initiate enhanced monitoring of the abnormal area; and high-risk signals immediately trigger the optically isolated fuse protection circuit, severing the faulty unit's electrical connection in a fraction of a second and uploading the 3D positioning information to a cloud-based diagnostic platform via a secure communication protocol. In the event that multiple adjacent units simultaneously display risk signals, the cluster fault emergency mechanism is activated to directly trigger regional fuse protection.

[0115] It is important to further clarify that, during implementation, a forward heat conduction algorithm is employed to address areas with insufficient sensor coverage during the construction of the dynamic spatiotemporal neural field. The unsteady-state heat conduction equation is solved using adjacent sensor nodes as boundary conditions. When the fitting error between the inferred results and the measured boundary data consistently falls below an acceptable threshold, virtual nodes are activated to participate in the calculation. The weights of transfer paths between graph nodes are updated according to the thermodynamic principle of irreversibility. If an abnormal reverse heat flow is detected, the path is immediately frozen. For conventional paths, the weight coefficients are dynamically adjusted based on the correlation between the temperature changes of adjacent nodes.

[0116] Theoretical Joule thermal power calculation utilizes a dynamic internal resistance model that couples temperature and state of charge. A multi-dimensional parameter mapping table and trilinear interpolation algorithm provide real-time output of the operating condition-adapted internal resistance value. Current acquisition utilizes a dual-source verification mechanism using a primary sensor and voltage inference. Redundant sensor arbitration is activated when significant deviations are detected. Spatial thermal power distribution calculations incorporate cell position weighting to compensate for uneven heat dissipation within the module.

[0117] The hardware system utilizes a heterogeneous computing architecture, with dedicated coprocessors responsible for spatiotemporal neural field calculation, electrothermal delay analysis, and dual-channel verification. The temperature sensing network is deployed in a cellular topology, with increased sensor density in key areas and an anti-interference circuit design. The fuse protection circuit integrates a high-response optical driver module, ensuring electrical isolation within a fraction of a second after receiving a command.

[0118] A strict time series control mechanism is established throughout the diagnostic process to ensure consistent timing from data collection, feature extraction, verification analysis, to response execution. The system incorporates a self-calibration module to continuously monitor the health of sensor devices. Upon detecting a signal anomaly, it automatically freezes the faulty node and activates a backup sensing channel. Upon receiving fault information, the cloud-based diagnostic platform automatically generates repair guidance plans based on a historical case library, forming a complete technical closed loop from real-time diagnosis to maintenance decision-making.

[0119] The coordinated implementation of these technical solutions enables accurate identification and rapid resolution of early-stage faults such as micro-short circuits within power batteries, effectively resolving the technical challenge of subtle fault signals being masked by complex operating conditions. All technical aspects are verified through the dual constraints of physical laws and data feature analysis to ensure the reliability of diagnostic conclusions and the timeliness of early warnings.

[0120] A fault diagnosis and early warning method for a power battery intelligent thermal management system includes the following steps:

[0121] Step S1: The distributed temperature sensor array collects the battery module surface temperature data stream in real time, and the current and voltage monitoring module synchronously obtains the charge and discharge current and cell voltage signals, which are input into the processing system after eliminating electromagnetic crosstalk through the anti-interference circuit;

[0122] Step S2: Construct a dynamic spatiotemporal neural field model: During the training phase, transient thermal disturbance data simulating micro-short circuit characteristics is injected to enable the adaptive graph neural network to learn the topological differences between normal and abnormal heat propagation. During online operation, the sparse sensor data is mapped to a virtual 3D grid, and dynamic graph convolution is used to extract spatial correlation strength and temporal decay characteristics to generate a thermal field feature fingerprint.

[0123] Step S3: Perform electrothermal delay coupling analysis: Calculate the theoretical Joule heating power distribution based on the internal resistance parameters coupled with the real-time current and temperature-state-of-charge, align the thermal field signature with the theoretical value at the millisecond level, and identify systematic advances in actual temperature rise relative to the theoretical value through cross-correlation detection;

[0124] Step S4: Perform dual-channel verification on the marked suspected signal:

[0125] Physical constraint screening: Solve the theoretical solution boundary of the heat diffusion equation in the abnormal area and eliminate signals that exceed the thermodynamic limit of the material;

[0126] Adversarial generation verification: Using a generative adversarial network to reconstruct background noise signals, the discriminator detects the unique spatial resonant frequencies and thermal diffusion topological patterns in the real signal;

[0127] Step S5: Initiate cross-channel arbitration: When the conclusions of physical screening and adversarial verification conflict, dynamically adjust the warning level based on the degree of deviation of signal characteristics;

[0128] Step S6: Fault classification decision: Integrating physical conformity, resonance characteristic strength, and historical case matching, a hierarchical decision tree is used to distinguish sensor failure, cooling system abnormality, battery micro-short circuit, and aging poor contact;

[0129] Step S7: triggering a hierarchical response:

[0130] Low-risk signals record characteristic patterns and improve monitoring sensitivity;

[0131] A medium risk signal limits output power and initiates local enhanced monitoring;

[0132] The high-risk signal immediately drives the optical fuse circuit to cut off the faulty unit and outputs the three-dimensional positioning coordinates at the same time;

[0133] Step S8: Cluster fault emergency processing: When risk signals appear synchronously in adjacent units, skip the hierarchical response and directly execute regional fuse protection;

[0134] Step S9: Upload the encrypted positioning data to the cloud platform through a secure communication protocol, and associate it with the historical case library to generate a maintenance decision plan.

[0135] By constructing a dynamic spatiotemporal neural field model, we can accurately capture sub-℃ weak thermal anomaly signals, and combine it with electrothermal delay analysis to achieve early identification of micro-short-circuit faults, solving the problem of fault precursor signals being overwhelmed by complex working conditions; the dual-channel verification mechanism relies on the dual mutual verification of the rigid constraints of physical laws and intelligent residual feature recognition to eliminate the risk of misjudgment caused by cooling system fluctuations and sensor noise, and improve the accuracy of early warning; the coordinated implementation of the hierarchical response strategy and the cluster fault direct fuse mechanism builds a closed-loop safety protection system from precise intervention in the risk budding stage to rapid isolation of the critical point of thermal runaway, effectively extending the safety margin of the battery system.

[0136] Precise thermal modeling technology based on the three-dimensional coupling of temperature, state of charge, and aging ensures diagnostic reliability under all life cycle conditions and reduces maintenance costs. The collaborative design of heterogeneous hardware architecture and anti-interference sensor networks overcomes real-time bottlenecks, enabling complex analysis processes to meet the stringent latency requirements of on-board systems. The intelligent closed-loop fault location and maintenance decision-making not only enhances safety assurance levels, but also provides high-value data support for battery health management.

[0137] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0138] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A fault diagnosis and early warning method for a power battery intelligent thermal management system, characterized in that: It includes the following steps, performed in order: S1) Based on real-time monitoring data from distributed temperature sensors within the battery module, a dynamic spatiotemporal neural field model is constructed through an adaptive graph neural network to generate a thermal field fingerprint that characterizes the topological pattern of heat propagation. S2) aligning the thermal field fingerprint with a theoretical Joule heat power distribution calculated based on the real-time current and internal resistance parameters of the battery in time and space, and analyzing the difference in delay characteristics between the actual temperature rise and the theoretical value through cross-correlation delay detection; S3) performs dual-channel verification on the abnormal signal identified in step S2): Channel 1: Calculate the theoretical solution boundary of the heat diffusion equation in the abnormal area to verify whether the signal meets the material thermodynamic parameter constraints; Channel 2: Generate a background noise signal using a generative adversarial network, and use the discriminator to identify the residual characteristic patterns in the real signal; S4) Trigger a graded warning response based on the verification results.

2. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 1, characterized in that: The step S1) includes: During the training phase, transient thermal disturbance data within a preset temperature disturbance range simulating micro-short circuit characteristics is injected into the adaptive graph neural network; During online operation, dynamic graph convolution is used to extract the spatial correlation strength matrix and time propagation attenuation features.

3. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 1, characterized in that: In the step S2): When it is detected that the actual local temperature rise is ahead of the theoretical Joule heat power by milliseconds, it is marked as a suspected micro-short circuit signal.

4. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 1, characterized in that: The dual-channel verification in step S3) includes: 3a) In channel 1, the theoretical heat conduction boundary is calculated by partitioning the non-uniform material parameters of each battery region, and only the signal within the theoretical solution boundary is retained; 3b) In channel 2, the generator reconstructs a synthetic background signal containing system fluctuations based on normal operating data, and the discriminator identifies the spatial resonant frequency characteristics in the real signal through comparison.

5. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 4, characterized in that: The step 3b) further comprises: The residual feature pattern output by the discriminator is matched with the historical micro-short circuit case library for similarity, and the fault confidence score is output.

6. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 1, characterized in that: The hierarchical warning response in step S4) includes: Level 1 response: Record abnormal pattern characteristics and increase the priority of similar signal identification; Secondary response: Reduce battery output power and initiate enhanced temperature monitoring in abnormal areas; Level 3 response: triggers the fuse protection circuit and outputs the fault location coordinates.

7. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 2, characterized in that: The construction process of the dynamic spatiotemporal neural field model includes: Mapping sparse temperature sensor data to virtual three-dimensional heat conduction grid nodes; The transfer path weights between graph nodes are dynamically updated according to real-time heat flow changes.

8. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 1, characterized in that: The fault diagnosis and early warning method of the power battery intelligent thermal management system further includes, after step S3): The fault signals verified through dual channels are classified into different types to distinguish between micro short circuit, sensor failure or cooling system failure.

9. The fault diagnosis and early warning method for a power battery intelligent thermal management system according to claim 1, characterized in that: The calculation of the theoretical Joule heating power distribution is based on: Dynamic variation model of battery cell internal resistance with temperature and state of charge; Real-time charge and discharge current data collected by the current sensor.

10. A fault diagnosis device for a power battery intelligent thermal management system, characterized in that: include: Temperature sensor arrays, current and voltage monitoring modules, and embedded processors are arranged in key areas of the battery module; The processor is configured to execute the steps of any one of the methods of claims 1 to 9; Wherein, the embedded processor includes: a spatiotemporal neural field calculation unit, configured to implement step S1 of claim 1); an electrothermal delay analysis unit, configured to implement step S2 of claim 1); A dual-channel verification engine, used to implement step S3 of claim 1).

Images