Flexible blocking method and system for vehicle thermal runaway in charging station

Abstract

The application discloses a charging station vehicle thermal runaway flexible blocking method and system, relates to the technical field of charging blocking control, and comprises the following steps: S1, collecting thermal runaway blocking monitoring data, and performing pretreatment on the thermal runaway blocking monitoring data; S2, extracting thermal risk features corresponding to each effective parking space, calculating risk potential field values corresponding to each sampling position in the station plane by grid points, determining a priority monitoring area, and generating a thermal anomaly trigger data set; S3, determining a target blocking category based on a real thermal runaway posterior judgment result, a blocking control representation quantity and a flexible blocking strategy database, or writing a corresponding environmental pseudo-thermal spot area into a dynamic mask matrix; and S4, performing hard cutting, flexible blocking or continuous monitoring on the corresponding parking space to form a thermal runaway flexible blocking result set. The problems that in the existing thermal runaway disposal of the charging station, a priority risk area is difficult to accurately determine, a blocking decision does not match a field state, and flexible blocking and hard cutting switching are not timely are solved.

Description

Technical Field

[0001] This invention relates to the field of charging interruption control technology, specifically to a flexible interruption method and system for vehicle thermal runaway at charging stations. Background Technology

[0002] With the continuous growth of new energy vehicle ownership, the construction scale of public charging stations, dedicated charging stations, and integrated energy supply stations is constantly expanding. Charging scenarios are evolving from simple energy replenishment to high-density, continuous, and intelligent solutions. In this process, thermal safety monitoring, anomaly identification, risk warning, and coordinated response during vehicle charging have become important research directions in the field of charging station safety operation and maintenance. Especially with the continuous development of technologies such as infrared thermal imaging sensing, multi-source environmental information acquisition, edge intelligent analysis, visual target localization, and coordinated blocking control, how to achieve dynamic perception and safe handling of thermal runaway risks at charging stations based on multi-source monitoring data has become a key focus of ongoing research in related technologies.

[0003] For example, application CN121417435B discloses a quantum-sensory driven active protection method and system for thermal runaway in charging piles, relating to the field of charging pile protection technology. It includes the following steps: S1, real-time acquisition of quantum thermal runaway data, data preprocessing, quantification of thermal risk sensitivity, and construction of a thermal runaway risk dataset; S2, construction of a thermal runaway spatiotemporal matrix, construction of a thermal runaway risk principal component extraction model based on the thermal runaway spatiotemporal matrix, and construction of a risk trend prediction model by combining the thermal runaway spatiotemporal matrix and the thermal runaway risk principal component extraction model; S3, comprehensive judgment of thermal runaway risk to achieve early detection and active protection; S4, determination of safety thresholds, implementation of graded protection response for thermal runaway, and closed-loop optimization of protection capabilities.

[0004] For example, application CN111817402B provides a method and apparatus for full-state battery thermal runaway safety management. This invention consists of a basic data storage device, a loop detection module, a loop state determiner, a threshold setter, a thermal runaway protector, an early warning module, and a loop disconnection alarm module. The thermal runaway protector consists of a constant current charging protector, a constant voltage charging protector, a discharge state protector, and a static state protector. The implementation of this invention subdivides the battery charging and discharging cycle and distinguishes between alarms and early warnings. The thermal runaway early warning can determine whether there is a trend of thermal runaway as early as when the temperature reaches 60% of the thermal runaway alarm temperature and issue an early warning.

[0005] However, most existing technologies focus on judging thermal runaway based on single-point temperature, single threshold, or fixed state division, lacking continuous characterization of the spatial risk correlation of multiple parking spaces in charging stations, making it difficult to dynamically focus on priority risk areas within the station. At the same time, existing technologies usually lack an effective alignment decision mechanism between thermal anomaly identification results and blocking control strategies, which can easily lead to mismatches between risk identification, blocking category determination, and on-site handling actions. This results in inaccurate timing of switching between flexible blocking and hard disconnection, making it difficult to balance handling efficiency and station operation safety in a timely manner.

[0006] Therefore, in order to address the above problems, there is an urgent need for a flexible method and system for preventing thermal runaway of vehicles at charging stations. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a flexible blocking method and system for vehicle thermal runaway at charging stations. This solves the problems of difficulty in accurately determining priority risk areas and mismatch between blocking decisions and on-site conditions in existing thermal runaway handling at charging stations, which leads to untimely switching between flexible blocking and hard cutoff.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a flexible method for preventing thermal runaway of vehicles in charging stations, comprising: S1, collecting thermal runaway prevention monitoring data, performing time base alignment, anomaly removal, noise suppression, and numerical scale unification processing on the thermal runaway prevention monitoring data, and outputting preprocessed thermal runaway prevention monitoring data; S2, extracting thermal risk characteristics corresponding to each effective parking space based on the preprocessed thermal runaway prevention monitoring data, calculating the risk potential field value corresponding to each sampling position in the station plane on a grid-by-grid basis, determining the priority monitoring area, generating local monitoring frames, and generating thermal anomalies based on the continuous frame change results of the local monitoring frames. Trigger dataset; S3, construct aligned control input vector based on thermal anomaly trigger dataset and input it into low-rank aligned control model, output blocking control characterization quantity and real thermal runaway posterior judgment result; determine target blocking category based on real thermal runaway posterior judgment result, blocking control characterization quantity and flexible blocking strategy database, or write the corresponding environmental pseudo hot spot region into dynamic mask matrix; S4, perform hard cut-off, flexible blocking or continuous monitoring on corresponding parking space based on target blocking category, and switch to hard cut-off when any of the following conditions are met: edge hard cut-off condition, flexible blocking failure condition and model verification timeout condition, forming thermal runaway flexible blocking result set.

[0011] Furthermore, the specific steps for collecting thermal runaway interruption monitoring data and performing time reference alignment, anomaly removal, noise suppression, and numerical scale unification on the thermal runaway interruption monitoring data are as follows: Collect thermal runaway interruption monitoring data generated by the target vehicle during the charging monitoring process within the charging station. This data includes parking space number, infrared image frame, infrared acquisition timestamp, visible light image frame, visible light acquisition timestamp, gimbal yaw angle, gimbal pitch angle, lens zoom ratio, charging output voltage, charging output current, cumulative charging capacity, cumulative charging duration, charging gun connection status, and charging relay status. The system collects thermal runaway interruption monitoring data, including device status values, charging pile operating status values, station ambient temperature values, station relative humidity values, station wind speed values, station solar irradiance values, and station illuminance values. For the collected thermal runaway interruption monitoring data, a nearest-neighbor timestamp matching algorithm is used to perform unified time reference alignment processing on the multi-source acquisition sequences. A Hamper filter algorithm is used to perform anomaly identification and removal on the thermal runaway interruption monitoring data. A Kalman filter algorithm is used to perform noise suppression and temporal smoothing on the thermal runaway interruption monitoring data. A range normalization algorithm is used to perform numerical scale unification processing on the thermal runaway interruption monitoring data, outputting preprocessed thermal runaway interruption monitoring data.

[0012] Furthermore, based on the preprocessed thermal runaway interruption monitoring data, the thermal risk characteristics corresponding to each effective parking space are extracted, and the risk potential field value corresponding to each sampling position in the station plane is calculated grid by grid point. The specific steps to determine the priority monitoring area are as follows: A station plane coordinate system is established using the ground corner point on one side of the station entrance as the origin, the horizontal extension direction of the station as the abscissa, and the vertical extension direction of the station as the ordinate. The center position corresponding to each parking space number is then calibrated. Then, the preprocessed thermal runaway interruption monitoring data is read, and the data corresponding to each parking space is merged according to the parking space number. Only parking space numbers with the charging gun connection status value as ON, the charging relay status value as OFF, and the charging pile operation status value as OPEN are retained, forming monitoring units for each effective parking space. For each effective parking space monitoring unit, the thermal imaging area is extracted according to the connected pixel region in the infrared image frame where the temperature value is higher than the sum of the station ambient temperature value and the temperature difference threshold. Each pixel within the thermal imaging area is then extracted. The maximum temperature value is taken as the highest temperature value. The arithmetic mean of the temperature values ​​of each pixel in the thermal imaging area is calculated as the average temperature value. The difference between the highest temperature value and the average temperature value is calculated as the thermal distribution discrete value. The temperature rise rate is calculated based on the difference between the highest temperature values ​​at adjacent sampling times and the sampling time interval. The charging power value is calculated based on the product of the charging output voltage value and the charging output current value. The charging power change rate is calculated based on the difference between the cumulative charging power values ​​at adjacent sampling times and the sampling time interval. The station ambient temperature value, station solar irradiance value, and station illuminance value are used as environmental constraints. Together with the highest temperature value, temperature rise rate value, thermal distribution discrete value, charging power value, and charging power change rate value, they form the parking space risk feature vector. Discrete sampling is performed on the station plane according to the grid spacing n. The risk potential field value is calculated for each grid point. The grid point with the largest risk potential field value is selected as the priority monitoring center point. The grid where the priority monitoring center point is located and its eight neighboring grids constitute the priority monitoring area.

[0013] Further, the specific steps for calculating the risk potential field value for each grid point are as follows: For any coordinate point within the station plane, take the L2 norm of the risk characteristic vector of each effective parking space to obtain the risk intensity value of each effective parking space; multiply the risk intensity value of each effective parking space by the risk amplification factor and then take the natural exponent to obtain the risk amplification term of each effective parking space; calculate the squared difference between the horizontal coordinate and the squared difference between the vertical coordinate and the center coordinate of each effective parking space, and add a smoothing constant to obtain the distance constraint term of each effective parking space; divide the risk amplification term of each effective parking space by the corresponding distance constraint term to obtain the risk potential field contribution value of each effective parking space to the coordinate point; sum up the risk potential field contribution values ​​of all effective parking spaces to obtain the risk potential field value of the coordinate point.

[0014] Further, the specific steps for generating a thermal anomaly trigger dataset based on the continuous frame change results of the local monitoring frames are as follows: Using the priority monitoring center point as the target position, adjust the gimbal yaw and pitch angle values ​​according to the lateral and longitudinal positional deviations between the target position and the current field of view center of the infrared image frame; then adjust the lens zoom ratio value according to the difference between the imaging size of the priority monitoring area in the infrared image frame and the target imaging size, so that the priority monitoring area is located at the center of the field of view of the adjusted infrared and visible light image frames; then map the priority monitoring area onto the adjusted infrared and visible light image frames to obtain the minimum bounding rectangle of the mapped area; then... After expanding the outer rectangle outward by a pixel width of m, local cropping is performed to form local monitoring frames. For each local monitoring frame, consecutive frame comparisons are performed according to the sampling time sequence. The highest temperature value, average temperature value, and temperature range value within each sampling time are calculated, as well as the difference between the highest temperature value and the average temperature value between adjacent sampling times. When the difference between the highest temperature value, the difference between the average temperature value, and the risk potential field value of the site corresponding to the priority monitoring area all increase sequentially among K consecutive sampling times, the corresponding parking space is marked as a thermal anomaly triggering parking space. The corresponding parking space number, local monitoring frame, parking space risk feature vector, and site risk potential field value are extracted, and the output is a thermal anomaly triggering dataset.

[0015] Further, the specific steps for constructing an alignment control input vector based on the thermal anomaly trigger dataset and inputting it into a low-rank alignment control model to output the blocking control representation quantity and the true thermal runaway posterior judgment result are as follows: Read the thermal anomaly trigger dataset, extract the infrared image frames, visible light image frames, charging output voltage values, charging output current values, cumulative charging power values, cumulative charging duration values, station ambient temperature values, station relative humidity values, station wind speed values, station solar irradiance values, and station illuminance values ​​from the local monitoring frames corresponding to each thermal anomaly trigger parking space at each sampling time, and concatenate them in a fixed order to form the alignment control input vector for each thermal anomaly trigger parking space; Based on a pre-trained multimodal vision-language network fused with convolutional neural networks and Transformers, a low-rank alignment control model is constructed using a low-rank adaptation method, and the low-rank alignment control model is trained using a near-end policy optimization algorithm; Input the alignment control input vector into the trained low-rank alignment control model, and output the original projection weight matrix, low-rank dimensionality reduction matrix, and low-rank dimensionality increase matrix corresponding to the thermal anomaly trigger parking space.

[0016] Furthermore, the specific steps for determining the target blocking category based on the actual thermal runaway posterior judgment results, blocking control characterization quantities, and flexible blocking strategy database, or for writing the corresponding environmental pseudo-hot spot region into the dynamic mask matrix, are as follows: Perform original projection calculation on the aligned control input vector using the original projection weight matrix to obtain the original projection result; perform dimensionality reduction projection calculation on the aligned control input vector using a low-rank dimensionality reduction matrix, and perform dimensionality increase projection calculation on the dimensionality reduction projection result using a low-rank increase matrix to obtain the low-rank additional projection result; superimpose the original projection result and the low-rank additional projection result to obtain the corresponding blocking control characterization quantity; extract the blocking category encoding vectors corresponding to the first-level blocking decision, second-level blocking decision, and maintenance monitoring decision, respectively, from the flexible blocking strategy database; calculate the cosine similarity between the blocking control characterization quantity and each blocking category encoding vector, and determine the blocking category corresponding to the blocking category encoding vector with the largest cosine similarity as the target blocking category.

[0017] Further, the specific steps for performing hard blocking, flexible blocking, or continuous monitoring on the corresponding parking space based on the target blocking category are as follows: Read the target blocking category corresponding to each thermal anomaly-triggered parking space, and extract the corresponding maximum temperature value, temperature rise rate value, thermal distribution discrete value, station risk potential field value, blocking control characterization quantity, and true thermal runaway posterior judgment result; and calculate the regional temperature gradient value based on the average absolute value of the temperature difference between adjacent pixels in the hot spot area of ​​the infrared image frame of the corresponding parking space; when the maximum temperature value is greater than the extreme blocking threshold, the corresponding parking space does not enter the low-rank alignment control model verification process and is directly marked as a hard-blocking parking space; when the maximum temperature value is not greater than the extreme blocking threshold and is within the fuzzy warning temperature range, and the regional temperature gradient value is greater than the primary warning threshold, the corresponding parking space is input into the low-rank alignment control model for true thermal runaway posterior judgment; when the true thermal runaway posterior judgment value is not greater than the posterior judgment threshold... When the value is reached, the corresponding parking space is marked as a continuously monitored parking space; for hard-cut parking spaces, the charging gun disconnection control command and the charging relay disconnection control command are output to cut off the charging circuit of the corresponding parking space; for flexible blocking parking spaces, the corresponding flexible blocking action is generated according to the target blocking category; when the target blocking category is a level one blocking decision, the sound and light alarm control command and the station alarm information push command are generated; when the target blocking category is a level two blocking decision, the parking spaces whose center coordinate distance is less than the distance threshold are determined according to the location of the corresponding parking space to form a flexible handling action area, and the directional spray control command, water curtain isolation control command and evacuation guidance control command are output; for continuously monitored parking spaces, the infrared image frame, visible light image frame and charging output current value of the corresponding parking space are continuously collected, and when the corresponding parking space has been written into the dynamic mask matrix, the temperature gradient characteristics of the area corresponding to the dynamic mask matrix are masked.

[0018] Furthermore, the specific steps for switching to hard cutoff when any one of the following conditions is met—edge hard cutoff condition, flexible blocking failure condition, or model verification timeout condition—to form a thermal runaway flexible blocking result set are as follows: During the execution of the flexible blocking action, the infrared image frame, charging output current value, charging gun connection status value, and charging relay status value of the corresponding parking space are continuously read, and the highest temperature value, temperature rise rate value, regional temperature gradient value, and temperature gradient change rate value are recalculated based on the infrared image frame; wherein, the regional temperature gradient value is represented by the average of the absolute values ​​of the temperature differences between adjacent pixels in the hot spot region in the infrared image frame of the corresponding parking space, and the temperature gradient change rate value is obtained by subtracting the regional temperature gradient value of the previous sampling time from the regional temperature gradient value of the current sampling time, and then dividing by the sampling time interval; when the highest temperature value and temperature rise rate value corresponding to adjacent sampling times in W consecutive sampling times increase sequentially... If either of the following conditions is met: the forward inference time of the low-rank aligned control model exceeds the response time threshold, or the temperature gradient rate of change exceeds the gradient rate of change threshold, the low-rank aligned control model's processing authority is immediately stripped, and the corresponding parking space is switched to a hard-cut-off parking space from the edge side, outputting a charging gun disconnection control command and a charging relay disconnection control command; the blocking execution results of each parking space are summarized; when the charging gun connection status of the corresponding parking space is disconnected and the charging relay status is disconnected, the corresponding parking space is marked as a blocking completed parking space; when the highest temperature value corresponding to adjacent sampling times in the corresponding parking space no longer increases and the temperature rise rate value no longer increases in the W consecutive sampling times after the flexible blocking action is executed, the corresponding parking space is marked as a flexible blocking completed parking space; then the blocking completed parking spaces, flexible blocking completed parking spaces, and continuously monitored parking spaces are summarized according to the parking space number to form a thermal runaway flexible blocking result set.

[0019] The second aspect of this invention provides a flexible thermal runaway prevention system for vehicles in charging stations, comprising: a data acquisition and processing module, a risk focusing triggering module, a low-rank alignment decision module, and a flexible prevention linkage module. The data acquisition and processing module is used to acquire thermal runaway prevention monitoring data, perform time base alignment, anomaly removal, noise suppression, and numerical scale unification processing on the thermal runaway prevention monitoring data, and output preprocessed thermal runaway prevention monitoring data. The risk focusing triggering module is used to extract thermal risk characteristics corresponding to each effective parking space based on the preprocessed thermal runaway prevention monitoring data, calculate the risk potential field value corresponding to each sampling position in the station plane on a grid-by-grid basis, determine the priority monitoring area, and generate local monitoring frames. The continuous frame change results generate a thermal anomaly trigger dataset; the low-rank alignment decision module is used to construct an alignment control input vector based on the thermal anomaly trigger dataset and input it into the low-rank alignment control model, outputting the blocking control characterization quantity and the true thermal runaway posterior judgment result; the target blocking category is determined based on the true thermal runaway posterior judgment result, the blocking control characterization quantity, and the flexible blocking strategy database, or the corresponding environmental pseudo hot spot region is written into the dynamic mask matrix; the flexible blocking linkage module is used to perform hard cut-off, flexible blocking, or continuous monitoring on the corresponding parking space based on the target blocking category, and switches to hard cut-off when any of the following conditions are met: edge hard cut-off condition, flexible blocking failure condition, and model verification timeout condition, forming a thermal runaway flexible blocking result set.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) A flexible blocking method and system for vehicle thermal runaway in charging stations. By extracting the thermal risk characteristics corresponding to each effective parking space and calculating the risk potential field value corresponding to each sampling position in the station plane by grid point, the spatial characterization of the thermal risk distribution of multiple parking spaces in the station is realized. It can determine the priority monitoring area in the continuous risk distribution, so that the subsequent local monitoring frame acquisition and thermal anomaly triggering are more concentrated in the high-risk position, and improve the pertinence and effectiveness of early thermal anomaly location.

[0023] (2) A flexible method and system for preventing thermal runaway of vehicles in charging stations generates local monitoring frames by adjusting the field of view of the PTZ camera, mapping the region and cropping the local area. Based on the continuous frame change results of the local monitoring frames, a thermal anomaly trigger dataset is generated. This avoids relying solely on the global image or single-point temperature threshold for rough judgment. It allows the thermal anomaly triggering process to combine local temperature changes, regional heat distribution changes and risk potential field changes, thereby improving the stability and accuracy of local thermal anomaly triggering.

[0024] (3) A flexible blocking method and system for vehicle thermal runaway at charging stations is proposed. By constructing the thermal anomaly trigger dataset as an aligned control input vector, and freezing the main parameters, a low-rank aligned control model is introduced with low-rank adaptation parameters in the query projection layer and value projection layer. The output blocking control representation quantity and the posterior judgment result of real thermal runaway are provided. It can complete the thermal runaway blocking scenario adaptation while retaining the pre-trained multimodal representation capability. Furthermore, by combining the temperature evolution residual value, reflection interference representation quantity and dynamic mask matrix, the real thermal runaway and environmental pseudo hot spots are distinguished. This enables the target blocking category judgment to have posterior verification basis and environmental interference suppression capability, thereby improving the accuracy of blocking decision, anti-interference capability and scenario adaptability.

[0025] (4) A flexible blocking method and system for vehicle thermal runaway at charging stations, which links the target blocking category with hard cut-off, flexible blocking and continuous monitoring, and switches to hard cut-off when the edge hard cut-off condition, the flexible blocking failure condition or the model verification timeout condition is met, forming an edge safety fallback mechanism; at the same time, the flexible blocking combines directional spraying, water curtain isolation, evacuation guidance and sound and light guidance control, so that the handling action can take into account vehicle charging circuit blocking, thermal diffusion suppression, personnel evacuation and continuous monitoring, thereby improving the real-time performance, safety and reliability of thermal runaway handling closed loop. Attached Figure Description

[0026] Figure 1 Flowchart of a flexible method for preventing thermal runaway of vehicles at charging stations;

[0027] Figure 2 A structural diagram of a flexible thermal runaway prevention system for vehicles at a charging station.

[0028] Figure 3 A coupled distribution diagram of the risk potential field at the site and the contribution of the source point;

[0029] Figure 4 This is a schematic diagram showing the location of the priority monitoring area and the adjustment of the pan-tilt-zoom (PTZ) field of view. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-4This invention provides a technical solution: a flexible method for preventing thermal runaway of vehicles in charging stations, comprising: S1, collecting thermal runaway prevention monitoring data, performing time base alignment, anomaly removal, noise suppression, and numerical scale unification processing on the thermal runaway prevention monitoring data, and outputting preprocessed thermal runaway prevention monitoring data; S2, extracting thermal risk characteristics corresponding to each effective parking space based on the preprocessed thermal runaway prevention monitoring data, calculating the risk potential field value corresponding to each sampling position in the station plane for each grid point, determining the priority monitoring area, generating local monitoring frames, and generating thermal anomaly triggering data based on the continuous frame change results of the local monitoring frames. S3, construct an aligned control input vector based on the thermal anomaly trigger dataset and input it into the low-rank aligned control model, outputting the blocking control characterization quantity and the actual thermal runaway posterior judgment result; determine the target blocking category based on the actual thermal runaway posterior judgment result, the blocking control characterization quantity, and the flexible blocking strategy database, or write the corresponding environmental pseudo hot spot region into the dynamic mask matrix; S4, perform hard cut-off, flexible blocking, or continuous monitoring on the corresponding parking space based on the target blocking category, and switch to hard cut-off when any of the following conditions are met: edge hard cut-off condition, flexible blocking failure condition, and model verification timeout condition, forming a thermal runaway flexible blocking result set.

[0032] Specifically, the steps for collecting thermal runaway interruption monitoring data and performing time reference alignment, anomaly removal, noise suppression, and numerical scale unification on the thermal runaway interruption monitoring data are as follows: Collect thermal runaway interruption monitoring data generated by the target vehicle during the charging monitoring process within the charging station. This data includes parking space number, infrared image frame, infrared acquisition timestamp, visible light image frame, visible light acquisition timestamp, gimbal yaw angle, gimbal pitch angle, lens zoom ratio, charging output voltage, charging output current, cumulative charging capacity, cumulative charging duration, charging gun connection status, charging relay status, charging pile operating status, station ambient temperature, station relative humidity, and station wind speed. The monitoring system collects data on the following parameters: solar irradiance and illumination at the charging station. Parking space numbers are obtained by matching the charging station's backend coding table with the parking space coding table. During this process, the device address corresponding to each charging gun is bound to a fixed parking space code to ensure accurate merging of subsequent monitoring data according to the parking space number. Infrared image frames are acquired in real-time by an infrared thermal imaging camera mounted on the station's monitoring pole. The infrared thermal imaging camera outputs thermal image data with a pixel temperature matrix, and the infrared acquisition timestamp is automatically written by the camera's internal clock at the moment of image output. Visible light image frames are acquired in real-time by a visible light camera coaxially mounted with the infrared thermal imaging camera. The visible light acquisition timestamp is obtained by the visible light camera... The camera's internal clock is automatically written in at the moment of image output; the gimbal yaw angle is obtained in real time through the angle encoder built into the gimbal servo driver; the gimbal pitch angle is obtained in real time through the gimbal pitch axis angle encoder; the lens zoom ratio is obtained by reading the zoom feedback register of the motorized zoom lens; the charging output voltage is acquired by the voltage sensor at the DC output terminal of the charging pile; the charging output current is acquired by the Hall current sensor in the DC output circuit of the charging pile; the cumulative charging capacity is obtained by the charging pile metering unit through time integration of the charging output voltage and charging output current; and the cumulative charging duration is obtained by the charging pile controller through the continuous running time from the start of charging to the current sampling time. The charging gun connection status value is obtained by jointly reading the charging gun lock switch status signal and the connection confirmation signal; the charging relay status value is obtained by reading the relay auxiliary contact feedback signal; the charging pile operation status value is obtained by reading the status register of the charging pile operation controller; the station ambient temperature value is obtained by the temperature sensor installed on the station monitoring pole; the station relative humidity value is obtained by the humidity sensor installed at the same location; the station wind speed value is obtained by the wind speed sensor installed in the open area above the station; the station solar irradiance value is obtained by the solar irradiance sensor installed in an unobstructed location; and the station illuminance value is obtained by the illuminance sensor installed above the monitoring area.During the data acquisition process, a complete sampling is triggered at a fixed sampling period of 0.5 to 2 seconds. This sampling period is derived from the intersection of the frame rate range of the infrared thermal imaging camera, the output frequency of the visible light camera, the gimbal servo feedback refresh cycle, the charging pile operation data upload cycle, and the site environment sensor update cycle. Within this range, the thermal anomaly evolution process can be covered, and the stable merging of multi-source data under the same time reference can be ensured. Within the same sampling period, the infrared image frame, visible light image frame, charging output voltage value, charging output current value, cumulative charging power value, cumulative charging time value, charging gun connection status value, charging relay status value, and charging pile operation status are read synchronously. The system collects and records the status values, ambient temperature, relative humidity, wind speed, solar irradiance, and illuminance of the thermal runaway interruption monitoring data, and writes each data point with its corresponding timestamp. This ensures that each thermal runaway interruption monitoring data point has a clear time and parking space identifier. For the collected thermal runaway interruption monitoring data, a nearest-neighbor timestamp matching algorithm is used to perform unified time reference alignment processing on the multi-source acquisition sequences. Specifically, the infrared acquisition timestamp is used as the primary time reference, and timestamps are added to the visible light acquisition timestamp sequence, the charging output voltage sampling timestamp sequence, the charging output current sampling timestamp sequence, the cumulative charging power sampling timestamp sequence, the cumulative charging duration sampling timestamp sequence, and the charging gun connection status sampling timestamp sequence. The sampling record with the smallest absolute time difference from each infrared acquisition timestamp is searched within the following time series: charging relay status value sampling time series, charging pile operation status value sampling time series, station ambient temperature value sampling time series, station relative humidity value sampling time series, station wind speed value sampling time series, station solar irradiance value sampling time series, and station illuminance value sampling time series. When the time difference is not greater than a preset matching tolerance, the corresponding sampling record is mapped to that infrared acquisition timestamp to form a unified sampling entry. When the time difference is greater than the preset matching tolerance, the sampling entry is marked as an entry to be compensated. The preset matching tolerance ranges from 20 milliseconds to 100 milliseconds, and the value is derived from the infrared thermal imaging camera and visible light. The data processing method includes: the internal clock drift range of the optical camera, the statistical results of the data upload delay of the charging pile, the statistical results of the transmission jitter of the environmental sensor bus, and the maximum stable synchronization deviation of multi-source devices in the on-site commissioning records. The principle of this processing method is that there are slight deviations in the triggering start points of multi-source acquisition devices and local jitter in the sampling period. By matching the nearest timestamp, the sampling results that are closest in time can be merged into the same time base without changing the source of the original physical quantity values. This ensures that the infrared image frames, visible light image frames, charging status data, and environmental data corresponding to the same moment have a unified temporal meaning. The Hamper filtering algorithm is used to perform anomaly identification and rejection on the thermal runaway blocking monitoring data.Specifically, a sliding window is established on the time series corresponding to each parking space number. The window length ranges from 5 to 11 sampling points. This value is derived from the stability of thermal anomalies during short-term continuous sampling, the normal fluctuation width of charging output voltage and charging output current values, and the statistical results of short-term disturbance cycles from environmental sensors. Within each sliding window, the median and median absolute deviation of the charging output voltage, charging output current, cumulative charging capacity, cumulative charging duration, station ambient temperature, station relative humidity, station wind speed, station solar irradiance, and station illuminance values ​​are calculated. Then, the absolute value of the difference between the center sample value and the corresponding median is multiplied by the median absolute deviation. The comparison is performed using a threshold coefficient. When the absolute value of the difference exceeds the threshold limit, the sampled value is identified as an outlier and removed. The threshold coefficient is between 3 and 5, and the values ​​are derived from the general statistical discrimination criteria of Hamper filtering, the nominal accuracy range of each sensor, and the outlier distribution statistics of the sampled data during the trial operation of the site. For infrared and visible light image frames, the overall average brightness, overall average temperature, and effective pixel ratio of the image are first extracted as image quality discrimination measures. Then, the same Hamper filtering processing approach is used to identify and remove abruptly abnormal frames. The threshold for the effective pixel ratio of the image is between 0.8 and 0.95, and the values ​​are derived from the camera field-of-view occlusion test results, image quality statistics under rain and fog conditions, and on-site installation results. Angle calibration results; the principle of this processing method is that the Hamper filter, with the median as the central statistic, has a strong ability to suppress isolated outliers caused by local pulse interference, instantaneous sampling jumps, and communication jitter, and can avoid outliers directly entering the subsequent thermal risk feature extraction process and causing misjudgment; the Kalman filter algorithm is used to perform noise suppression and time-series smoothing on the thermal runaway interruption monitoring data; specifically, one-dimensional state update equations and observation equations are established for the charging output voltage value, charging output current value, cumulative charging capacity value, cumulative charging duration value, station ambient temperature value, station relative humidity value, station wind speed value, station solar irradiance value, and station illuminance value, respectively, and the state estimate value of the previous sampling time is used as... For the predicted value, the observed value after anomaly removal at the current sampling time is used as the update value. The Kalman gain is calculated based on the prediction error covariance and the observation noise covariance. The predicted value is then corrected using the Kalman gain to obtain the smoothed state estimate. The initial value of the prediction error covariance comes from the factory accuracy parameters of each sensor, and the initial value of the observation noise covariance comes from the variance statistics of the continuous sampling data during the station's no-load phase. Iterative correction is then made in conjunction with the residual sequence during the on-site trial operation phase. For infrared image frames and visible light image frames, instead of directly filtering each pixel of the entire frame, Kalman smoothing is performed on the center temperature trajectory, hot spot area trajectory, and image center position trajectory extracted between frames to keep the image sequence continuous in the time dimension.The center temperature trajectory is obtained by extracting the location of the highest temperature pixel and its corresponding temperature value in the infrared image frame. The hot spot area trajectory is obtained by converting the area of ​​connected pixel regions whose temperature is higher than the ambient reference temperature plus a temperature difference threshold. The image center position trajectory is calculated by the center position of the target vehicle in the visible light image frame. The temperature difference threshold is set between 5 and 15 degrees Celsius, and the values ​​are derived from the statistical results of the temperature rise of the power battery casing surface under normal charging conditions, the results of the solar radiation impact test at the site, and the actual test results of the surface temperature deviating from the background temperature at the initial stage of thermal anomalies. The principle of this processing method is that Kalman filtering uses the state prediction result of the previous moment and the observation result of the current moment to perform weighted fusion, which can preserve the true trend of change. While suppressing random noise, the continuous sampling results are made closer to the actual physical evolution process. The range normalization algorithm is used to perform numerical scaling uniform processing on the thermal runaway interruption monitoring data, and the preprocessed thermal runaway interruption monitoring data is output. Specifically, the minimum and maximum values ​​of charging output voltage, charging output current, cumulative charging amount, cumulative charging time, station ambient temperature, station relative humidity, station wind speed, station solar irradiance, and station illuminance are statistically analyzed within a preset statistical period. The minimum value is subtracted from any sampled value, and then divided by the difference between the maximum and minimum values ​​to obtain the normalized result. The preset statistical period is 24 hours to 168 hours, and the values ​​are derived from the diurnal environmental change cycle of the station. The data includes statistics on the charging load variation cycle during the week and weekdays, as well as the weekly operational fluctuations. For the charging gun connection status, charging relay status, and charging pile operation status, a status mapping method is used to unify the numerical range, where "on," "closed," and "operating" are mapped to 1, and "off," "released," and "stopped" are mapped to 0. For infrared and visible light image frames, pixel grayscale values ​​and pixel temperature values ​​are compressed to a unified numerical range to ensure that image and numerical data have a comparable scale basis in subsequent joint processing. The principle behind this processing method is that different physical quantities have dimensional and numerical range differences; the charging output voltage value is in volts, the charging output current value is in amperes, and the cumulative charging capacity value... The data is expressed in kilowatt-hours (kWh), cumulative charging time in seconds, ambient temperature in degrees Celsius, relative humidity in percentage, wind speed in meters per second, solar irradiance in watts per square meter, and illuminance in lux. Without standardized processing, data with large numerical ranges will dominate subsequent thermal risk feature extraction results, leading to imbalance in the multi-source data fusion process. Therefore, range normalization is used to compress data of different units and orders of magnitude into a unified numerical range, thereby eliminating the adverse effects of dimensional differences on subsequent risk potential field calculation, alignment control input vector construction, and blocking control characterization calculation. The final output is preprocessed thermal runaway blocking monitoring data.

[0033] In this implementation plan, by refining and defining the acquisition sources, time reference alignment criteria, anomaly discrimination criteria, noise suppression criteria, and numerical scale unification criteria for multi-source heterogeneous data in thermal runaway interruption monitoring data, this ensures that parking space numbers, infrared image frames, visible light image frames, charging output voltage values, charging output current values, cumulative charging power values, cumulative charging duration values, charging gun connection status values, charging relay status values, charging pile operating status values, site ambient temperature values, site relative humidity values, site wind speed values, site solar irradiance values, and site illuminance values ​​have a unified time correspondence, stable data reliability, and a consistent numerical expression basis before entering subsequent processing. This improves the temporal consistency, data comparability, and fusion effectiveness of thermal runaway interruption monitoring data, providing a reliable data foundation for subsequent thermal risk feature extraction, risk potential field value calculation, thermal anomaly trigger determination, and interruption control characterization quantity generation.

[0034] Specifically, based on the preprocessed thermal runaway prevention monitoring data, the thermal risk characteristics corresponding to each effective parking space are extracted, and the risk potential field value corresponding to each sampling position in the station plane is calculated grid by grid point to determine the priority monitoring area. The specific steps are as follows: Using the ground corner point on one side of the station entrance as the origin, the horizontal extension direction of the station as the abscissa, and the vertical extension direction of the station as the ordinate, a station plane coordinate system is established, and the coordinates of the center position corresponding to each parking space number are calibrated. Specifically, a fixed corner point on one side of the entrance is selected as the zero point on the station ground. The horizontal and vertical coordinate values ​​of the center point of each parking space number in the station plane coordinate system are determined by a combination of a total station, a laser rangefinder, and the station's as-built plan. The vertical coordinate value is used to characterize the position of the parking space in the width direction of the depot, and the longitudinal coordinate value is used to characterize the position of the parking space in the depth direction of the depot. By mapping the parking space number to the center coordinate, the spatial position basis required for subsequent risk potential field value calculation is formed. Then, the preprocessed thermal runaway interruption monitoring data is read, and the data corresponding to each parking space is merged according to the parking space number. Only the parking space number with the charging gun connection status value as "on", the charging relay status value as "closed", and the charging pile operation status value as "operating" is retained to form the monitoring unit of each effective parking space. The principle of this processing method is that only when the target vehicle is in the actual charging state, the thermal imaging results, charging output voltage value, charging output current value, cumulative charging amount value, and cumulative charging time value are relevant to thermal runaway. Risk assessment is directly significant. If the charging gun connection status is disconnected, the charging relay status is disconnected, and the charging pile operation status is stopped, then the corresponding parking space lacks a real-time charging thermal load. Excluding it can prevent non-charging parking spaces from interfering with the priority monitoring area determination. For the monitoring unit of each valid parking space, the thermal imaging area is extracted according to the connected pixel regions in the infrared image frame whose temperature value is higher than the sum of the station ambient temperature value and the temperature difference threshold. Specifically, the temperature value of each pixel in the infrared image frame is first compared with the station ambient temperature value at the same sampling time. Then, the comparison result is superimposed with the preset temperature difference threshold for determination. Pixels that meet the condition of having a temperature value higher than the sum of the station ambient temperature value and the temperature difference threshold are retained. Subsequently, eight steps are performed on adjacent pixels that meet the condition. Neighborhood connectivity markers are used to form one, two, three, or even more connected pixel regions. Connected pixel regions with an area greater than a minimum area threshold are then selected as the thermal imaging region. The temperature difference threshold ranges from 5 to 15 degrees Celsius, preferably from 8 to 12 degrees Celsius. These values ​​are derived from statistical results of normal charging temperature rise on the surface of the power battery casing, results of solar-induced temperature rise tests at the site, and measured results of surface temperature deviating from background temperature during the initial stage of thermal anomalies. The minimum area threshold ranges from 20 to 200 pixels, preferably from 50 to 120 pixels. These values ​​are derived from the spatial resolution of the infrared thermal imaging camera, the imaging size of the target vehicle within the monitoring field of view, statistical results of noise and bright spot areas, and the minimum stable display area of ​​the thermal anomaly region in consecutive frames.The principle behind this processing method is that the surface of the target vehicle experiences a mild temperature rise during normal charging. Directly relying on a single high-temperature pixel for judgment is easily affected by reflections, sensor noise, and solar irradiance disturbances. By introducing the ambient temperature of the charging station as a background temperature benchmark, and then introducing a temperature difference threshold to form a secondary discrimination boundary, the truly localized thermal anomaly areas deviating from the background temperature can be separated from the overall image. Through connected pixel region extraction, the spatial integrity of the hotspot area can be maintained. Furthermore, the maximum value among the pixel temperature values ​​within the thermal imaging area is extracted as the highest temperature value, the arithmetic mean of the pixel temperature values ​​within the thermal imaging area is calculated as the average temperature value, and the difference between the highest temperature value and the average temperature value is calculated as the thermal distribution dispersion value. Among these, the highest... Temperature values ​​characterize the temperature level of the strongest thermal anomaly point in the thermal imaging area; average temperature values ​​characterize the overall heating level of the thermal imaging area; and thermal distribution dispersion values ​​characterize the deviation of local high-temperature points from the overall temperature background of the area. When the thermal distribution dispersion value increases, it indicates that the heat distribution within the thermal imaging area tends to be uneven, with heat concentrated in local locations, reflecting a trend of thermal anomalies developing from uniform temperature rise to localized heat accumulation. The temperature rise rate is calculated based on the difference between the highest temperature values ​​at adjacent sampling times and the sampling time interval. The charging power value is calculated based on the product of the charging output voltage and charging output current values. The charging power change rate is calculated based on the difference between the cumulative charging power values ​​at adjacent sampling times and the sampling time interval. The temperature rise rate value is used to characterize the localized temperature distribution. The rate of temperature rise of a thermal anomaly reflects its dynamic evolution. The charging power value characterizes the instantaneous electrical power input to the target vehicle during the current charging process, while the charging capacity change rate characterizes the rate of energy accumulation per unit time, reflecting the continuous input state of the charging load. The principle of using a combined representation of charging power and charging capacity change rate is that the charging power value emphasizes the instantaneous load level, while the charging capacity change rate emphasizes the continuous input state. Their combination can more accurately depict the load background when a thermal anomaly occurs. The station's ambient temperature, solar irradiance, and illuminance are used as environmental constraints, and together with the maximum temperature, temperature rise rate, thermal distribution dispersion, charging power, and charging capacity change rate, they form a comprehensive analysis. A parking space risk feature vector is generated. The ambient temperature value of the parking lot is used to characterize the overall temperature background; the solar irradiance value of the parking lot is used to characterize the intensity of the additional heating effect of direct sunlight on the vehicle surface; and the illuminance value of the parking lot is used to characterize the degree of influence of the parking lot's brightness on image quality and surface reflectivity. By incorporating environmental constraints into the parking space risk feature vector, it can reflect not only the target vehicle's own heating state and charging load state, but also the degree of disturbance of the external environment on the surface thermal performance, thereby improving the parking space risk feature vector's ability to characterize actual thermal runaway risks. The parking lot plane is discretely sampled according to the grid spacing n, and the risk potential field value is calculated for each grid point. The grid point with the largest risk potential field value is selected as the priority monitoring center point.The priority monitoring area is defined by the grid containing the priority monitoring center point and its eight adjacent grids. The specific operation is as follows: using the station's planar coordinate system as a reference, sampling columns are divided along the horizontal coordinate axis at a fixed grid spacing of n, and sampling rows are divided along the vertical coordinate axis at a fixed grid spacing of n, resulting in a regular grid distribution across the entire station plane. At each grid intersection, the corresponding risk potential field value is calculated, and the local high-value area formed by several consecutive grid points is identified as a candidate grid area. All candidate grid areas are compared, and the grid area with the largest risk potential field value is selected as the priority monitoring area. The grid spacing n ranges from 0.5 meters to 2 meters, preferably from 1 meter to 1.5 meters, and the value is derived from the width of a single parking space, the center distance between adjacent parking spaces, the accuracy of the gimbal's field of view adjustment, and the positioning accuracy requirements of the priority monitoring area. The principle behind this approach is that the parking space risk feature vector can only characterize the risk status of a single parking space, while actual thermal anomalies within a charging station have spatial coupling characteristics, with thermal radiation, airflow disturbance, and on-site obstruction effects existing between adjacent parking spaces. By calculating the risk potential field value point by point, the risk intensity of multiple effective parking spaces can be continuously mapped to changes in spatial location. This allows for the identification of areas with the highest risk concentration within the overall space of the facility, providing accurate priority monitoring areas for the generation of subsequent local monitoring frames.

[0035] In this implementation plan, by incorporating the spatial location corresponding to the parking space number, the thermal imaging area in the infrared image frame, the highest temperature value, the average temperature value, the discrete value of heat distribution, the temperature rise rate value, the charging power value, the charging power change rate value, the station ambient temperature value, the station solar irradiance value, and the station illuminance value into the same risk characterization process, and further mapping the discrete risk status of each effective parking space to a continuously distributed risk potential field value within the station plane, the determination of the priority monitoring area is no longer limited to the local anomaly of a single parking space, but can comprehensively reflect changes in charging load, the degree of surface heat accumulation, the impact of environmental disturbances, and spatial proximity propagation relationships. This improves the accuracy, foresight, and spatial targeting of the priority monitoring area location, providing a more stable risk focusing basis for the generation of subsequent thermal anomaly trigger datasets.

[0036] Specifically, the steps for calculating the risk potential field value point by point are as follows: For any coordinate point within the station plane, take the L2 norm of the risk feature vector of each effective parking space to obtain the risk intensity value of each effective parking space; firstly, read the risk feature vector corresponding to each effective parking space. The risk feature vector consists of the highest temperature value, temperature rise rate value, heat distribution dispersion value, charging power value, charging power change rate value, station ambient temperature value, station solar irradiance value, and station illuminance value. Then, according to the L2 norm calculation rules, calculate each component of the risk feature vector of the parking space separately. The risk intensity value of the corresponding effective parking space is obtained by squaring the square, summing the results, and then taking the square root of the sum. The principle behind this method is that the L2 norm can compress a multidimensional thermal risk representation into a single scalar, allowing for comparison of risk levels across different effective parking spaces on a unified scale while maintaining the positive impact of high-risk components on the overall risk intensity value. The risk intensity value of each effective parking space is then multiplied by a risk amplification factor and the natural exponent is taken to obtain the risk amplification term for each effective parking space. The risk amplification factor is a pre-defined dimensionless positive real number used to adjust the risk intensity value within the natural exponent. The degree of amplification in the index mapping; the risk amplification coefficient ranges from 0.1 to 1.5, with a preferred range of 0.3 to 0.8; when the risk amplification coefficient increases, the difference between the risk amplification terms corresponding to each effective parking space is further widened, and the focus on high-risk areas within the station plane is improved; when the risk amplification coefficient decreases, the difference between the risk amplification terms corresponding to each effective parking space decreases, and the risk potential field distribution tends to be flatter; the value of the risk amplification coefficient is based on the L2 norm distribution range of the parking space risk feature vector in the station's historical thermal anomaly samples, and the location sensitivity of the priority monitoring area. The requirements for suppressing the peak value of the risk potential field and the spatial density of parking spaces in the station are determined by comparing the risk potential field map generated under different risk amplification coefficients with the actual occurrence areas of historical thermal anomalies. The natural index calculation adopts an exponential mapping method with the natural constant e as the base, so that effective parking spaces with higher risk intensity values ​​receive greater contribution weight in the subsequent risk potential field calculation. To prevent the natural index mapping from causing the risk amplification term to be too large, leading to calculation overflow or difficulties in engineering implementation, an upper limit cutoff value of 1.0 to 10 is set for the risk amplification term of each effective parking space.Dimensionless values ​​between 0 and 1, or risk intensity values ​​normalized to the range of 0 to 1 before exponential mapping, ensure that the risk potential field contribution value is stable and controllable during the calculation process. The principle behind this approach is that if only linear mapping is used, the differences between risk intensity values ​​are easily weakened by distance factors in spatial diffusion calculations. However, after natural exponential amplification, the impact of high-risk parking spaces on the surrounding area within the station plane can be highlighted, thereby enhancing the risk focusing effect. The squared difference between the horizontal and vertical coordinates between the coordinate point and the center coordinates of each effective parking space is calculated, and a smoothing constant is added to obtain the distance constraint term for each effective parking space. For any coordinate point, the horizontal coordinates of that coordinate point are read. The system first reads the longitudinal and lateral coordinates of the center coordinates of each valid parking space, then reads the longitudinal and lateral coordinates of the center coordinates of each valid parking space. It then calculates the lateral coordinate difference, followed by the longitudinal coordinate difference, and then squares both. Finally, it sums the squared lateral and longitudinal coordinate differences and adds a smoothing constant to obtain the distance constraint term for the corresponding valid parking space. The smoothing constant is between 0.1 and 2, preferably between 0.3 and 1, and its value is derived from the parking space spacing distribution, the range of grid spacing n, and the near-field peak suppression requirement of the risk potential field. The principle behind this processing method is that the sum of the squared lateral and longitudinal coordinate differences characterizes the influence of the planar distance from the coordinate point to the center of the valid parking space, and the smoothing constant is used to avoid the influence of the lateral and longitudinal coordinate differences on the center coordinates of the valid parking space. When the coordinates coincide, the denominator approaches zero to prevent abnormal divergence of the risk potential field value at local locations. The risk amplification term for each valid parking space is divided by its corresponding distance constraint term to obtain the risk potential field contribution value of each valid parking space to the coordinate point. Each valid parking space corresponds to an independent risk potential field contribution value. The larger the risk amplification term, the stronger the thermal risk of the valid parking space itself, and the more significant its impact on the external spatial location. The smaller the distance constraint term, the closer the coordinate point is to the center of the valid parking space, and the stronger the impact of the valid parking space on that coordinate point. Dividing these two terms forms a decaying spatial contribution relationship, ensuring that high-risk parking spaces generate high contribution values ​​in close-range areas and low contribution values ​​in distant areas. This applies to all valid parking spaces. The risk potential field contribution values ​​are accumulated to obtain the risk potential field value of the coordinate point. Specifically, for any coordinate point within the charging station plane, the risk potential field contribution values ​​corresponding to all valid parking spaces are accumulated one by one to form the risk potential field value of that coordinate point. The larger the risk potential field value, the stronger the combined impact of multiple high-risk parking spaces on that coordinate point. The principle behind this processing method is that thermal anomalies within the charging station do not exist in isolation. The thermal risk status of a single valid parking space can spread its influence to the surrounding area through spatial proximity. By accumulating the risk potential field contribution values ​​of all valid parking spaces, a continuously distributed risk potential field value map can be constructed within the charging station plane, thereby providing quantifiable spatial risk basis for subsequent priority monitoring area location.

[0037] The specific formula for calculating the risk potential field value is as follows:

[0038] ;

[0039] In the formula, Represents the coordinates of the station in the plane at time t. , The risk potential value at that location, This indicates the total number of available parking spaces. , () represents the center coordinates of the vth valid parking space. This represents the risk amplification factor. This represents the parking space risk feature vector of the v-th valid parking space. The L2 norm of the risk feature vector of the v-th valid parking space is represented. Represents the smoothing constant, ( , ) represents the coordinates of any point within the station plane.

[0040] In this implementation plan, by uniformly mapping the risk feature vectors of each valid parking space to a risk intensity value, and then combining the risk amplification coefficient, distance constraint term, and risk potential field contribution value to construct a continuously distributed risk potential field value within the station plane, the thermal anomaly information originally scattered across each parking space can be transformed into a spatially continuous and comparable risk distribution result. This allows the determination of priority monitoring areas to not only reflect the thermal risk level of a single parking space, but also to reflect the strength of the superimposed influence formed by multiple valid parking spaces under spatial proximity conditions. This improves the accuracy, stability, and advance warning of risk cluster area identification in the station, and provides a more reliable spatial positioning basis for subsequent local monitoring frame generation and thermal anomaly trigger dataset extraction.

[0041] In this embodiment, Table 1 is a risk potential field value data table, listing the total number of valid parking spaces, risk amplification factor, smoothing constant, grid coordinates, and risk potential field value for each of the five sampling grids. The center coordinates of the three valid parking spaces are as follows: valid parking space 1 is (2.0, 3.0), valid parking space 2 is (6.0, 4.0), and valid parking space 3 is (8.0, 7.0). Specifically: in grid 1, the total number of valid parking spaces is 3, the risk amplification factor is 0.6, the smoothing constant is 0.5, the grid coordinates are (2.0, 4.0), and the corresponding risk potential field value is 2.8227; in grid 2, the total number of valid parking spaces is 3, the risk amplification factor is 0.6, the smoothing constant is 0.5, the grid coordinates are (4.0, 4.0), and the corresponding risk potential field value is 1.4955; in grid 3, the total number of valid parking spaces is 3, the risk amplification factor is 0.6, and the smoothing constant is... 0.5, grid coordinates (6.0, 6.0), corresponding risk potential field value is 1.7175; in grid 4, the total number of effective parking spaces is 3, the risk amplification factor is 0.6, the smoothing constant is 0.5, grid coordinates (7.0, 3.0), corresponding risk potential field value is 1.5208; in grid 5, the total number of effective parking spaces is 3, the risk amplification factor is 0.6, the smoothing constant is 0.5, grid coordinates (9.0, 8.0), corresponding risk potential field value is 2.3041.

[0042] Table 1 Risk Potential Value Data Table

[0043]

[0044] like Figure 3 As shown in the figure, based on the station's planar coordinate system, the center position of each effective parking space is taken as the risk source point. Effective parking spaces 1, 2, and 3 correspond to different parking space risk feature vectors. Based on the intensity of the risk feature vector corresponding to each effective parking space and the distance relationship between any point in the station plane and the center coordinates of each effective parking space, the risk potential field value U is calculated point by point, forming a continuous risk potential field distribution for the station. The colored areas in the figure represent the magnitude of the risk potential field value; the brighter the color, the higher the risk potential field value at the corresponding location. Grids 1 to 5 represent discrete sampling grid points in the station plane, and the corresponding risk potential field value is marked next to each grid point. As can be seen from the figure, the risk potential field value is the largest at grid 1. Therefore, the area where grid 1 is located is determined as the priority monitoring area, and the pan-tilt unit and lens are further driven to focus and image this area to generate a local monitoring frame. This figure intuitively reflects that the invention does not rely solely on the temperature anomaly of a single parking space for judgment, but rather constructs a continuous risk potential field by comprehensively considering the thermal risk characteristics and spatial propagation relationships of multiple effective parking spaces, thereby achieving dynamic determination of the priority monitoring area.

[0045] like Figure 4The diagram illustrates the relationship between determining the priority monitoring area and focusing the pan-tilt-zoom (PTZ) field of view based on the risk potential field value in this invention. Multiple parking spaces are set up within the site plan, and the target vehicle is parked in its corresponding space. After extracting thermal risk features from each valid parking space and calculating the risk potential field value corresponding to each sampling location within the site plan on a grid-by-grid basis, the area with the highest risk potential field value is determined as the priority monitoring area. The multi-layered closed curves surrounding the vehicle distribution in the figure represent the spatial distribution of the risk potential field value from high to low. The central region represents a concentrated area with a high risk potential field value, which corresponds to the priority monitoring area in this invention. The PTZ camera is deployed on the side of the station, and its field of view covers the parking area. After determining the priority monitoring area, the PTZ yaw angle and PTZ pitch angle are adjusted according to the lateral and longitudinal positional deviations between the center coordinates of the priority monitoring area and the current field of view of the infrared image frame. Then, the lens zoom ratio is adjusted according to the difference between the imaging size of the priority monitoring area in the infrared image frame and the imaging size of the target, so that the priority monitoring area is located at the center of the field of view of the adjusted infrared image frame and visible light image frame. This provides a spatial pointing basis for subsequent mapping of the priority monitoring area, obtaining the minimum bounding rectangle, and forming local monitoring frames.

[0046] Specifically, the steps for generating local monitoring frames and generating a thermal anomaly trigger dataset based on the continuous frame change results of the local monitoring frames are as follows: Using the priority monitoring center point as the target position, adjust the gimbal yaw angle and gimbal pitch angle values ​​according to the lateral and longitudinal positional deviations between the target position and the current field of view center of the infrared image frame. First, determine the target position based on the lateral and longitudinal coordinates of the center of the priority monitoring area in the station's plane coordinate system. Then, read the field of view center position corresponding to the infrared image frame at the same sampling time, calculate the lateral and longitudinal offset directions of the target position in the field of view, and then convert the lateral positional deviation into a gimbal yaw angle correction and the longitudinal positional deviation into a gimbal pitch angle correction, and drive the gimbal. The process involves adjusting the viewing angle. The principle behind this method is that the priority monitoring area is determined by the region with the highest risk potential field value. If the original monitoring orientation is maintained, high-risk areas in the infrared image frame may deviate from the center of the field of view, resulting in subsequent local cropping areas containing too much irrelevant background. By adjusting the gimbal orientation based on the lateral and longitudinal positional deviations, high-risk areas can be brought into the main monitoring field of view as quickly as possible. Then, the lens zoom ratio is adjusted according to the difference between the imaging size of the priority monitoring area in the infrared image frame and the target imaging size, ensuring that the priority monitoring area is located at the center of the adjusted infrared and visible light image frames. The imaging size is represented by the outer width and outer height pixel values ​​of the priority monitoring area in the infrared image frame, while the target... The imaging size is characterized by a pre-defined pixel range, ensuring that the priority monitoring area in the infrared image frame is neither lower than the minimum recognition size nor exceeds the field of view. The lens zoom ratio ranges from 1x to 20x, preferably from 3x to 10x, and is determined by the focal length coverage of the pan-tilt camera, the parking space monitoring distance range, and the minimum recognition size requirement for the thermal imaging area. The principle behind this processing method is that when the imaging size of the priority monitoring area in the infrared image frame is too small, it is difficult to stably extract thermal anomaly details; when the imaging size is too large, local areas exceed the effective field of view. By adjusting the lens zoom ratio, the priority monitoring area can maintain appropriate resolution in subsequent local monitoring frames. The priority monitoring area is then mapped onto the adjusted infrared image. In the infrared and visible light image frames, the minimum bounding rectangle of the mapped area is obtained. First, based on the imaging relationship adjusted according to the gimbal yaw angle, gimbal pitch angle, and lens zoom ratio, the priority monitoring area in the station's planar coordinate system is projected onto the pixel coordinate plane of the infrared and visible light image frames. Then, the bounding box is extracted from the projected area boundary to obtain the minimum bounding rectangle that completely encloses the priority monitoring area. The principle behind this processing method is that the priority monitoring area is initially defined in the station's planar coordinate system, and subsequent continuous frame analysis depends on the image pixel area. Therefore, it is necessary to establish the correspondence between spatial location and image location through coordinate mapping. The minimum bounding rectangle is used to compress the background area as much as possible while ensuring the complete preservation of the target area.After expanding the minimum bounding rectangle outwards by a pixel width *m*, a local cropping is performed to form a local monitoring frame. The pixel width *m* ranges from 10 to 80 pixels, preferably 20 to 50 pixels, and its value is derived from the gimbal tracking error range, the edge jitter amplitude of the infrared image frame, the registration deviation range of the visible light image frame, and the required boundary expansion margin of the thermal imaging area. The principle behind this processing method is that if the local monitoring frame is directly cropped using the minimum bounding rectangle, it is highly sensitive to slight offsets, and the target area can easily be truncated due to gimbal jitter between consecutive sampling moments. By expanding the pixel width *m* outwards, necessary boundary margins can be preserved for the thermal imaging area, improving the stability of continuous frame comparison. The frame comparison is performed sequentially according to the sampling time. The highest temperature, average temperature, and temperature range within each local monitoring frame at each sampling time are calculated, along with the difference between the highest and average temperature values ​​between adjacent sampling times. The temperature range is obtained by subtracting the lowest pixel temperature value from the highest pixel temperature value within the local monitoring frame, representing the dispersion of temperature distribution within a local area. During continuous frame comparison, the local monitoring frames corresponding to adjacent sampling times are used as the comparison objects. Within a uniform cropping range, the highest temperature, average temperature, and temperature range for each time point are extracted, and the difference between the highest and average temperature values ​​between adjacent sampling times is calculated. The principle behind this processing method is that a single... The high temperature result at a single sampling moment may be affected by instantaneous disturbances, while continuous frame comparison can characterize the continuous evolution trend of local thermal anomalies in the time dimension, making the thermal anomaly trigger determination based on continuous changes. When the difference between the highest temperature value, the difference between the average temperature value, and the risk potential field value of the station corresponding to the priority monitoring area all increase sequentially among K consecutive sampling moments, the corresponding parking space is marked as a thermal anomaly triggering parking space, and the corresponding parking space number, local monitoring frame, parking space risk feature vector, and station risk potential field value are extracted, and the output is a thermal anomaly triggering dataset. Among them, 3 to 10 consecutive sampling moments are taken, preferably 4 to 6, and the values ​​are derived from the initial holding period of thermal anomalies in the charging station. The statistical results of the continuous time, the completion time of the PTZ redirection, and the stability requirements of continuous frame analysis; "all increasing sequentially" specifically refers to the fact that in the K-1 sets of adjacent difference sequences formed by K consecutive sampling times, the difference of the later set is greater than the difference of the previous set, so that the difference sequence of the highest temperature value, the difference sequence of the average temperature value, and the station risk potential field value sequence all show an increasing state simultaneously; the principle of adopting this processing method is that only when the local thermal anomaly shows an increasing trend at the temperature peak level, the overall heating level, and the spatial risk accumulation level, it indicates that the corresponding parking space has entered a stage of continuous deterioration from general thermal fluctuation. At this time, outputting the thermal anomaly trigger dataset can provide more reliable anomaly starting point data for the subsequent construction of low-rank aligned control input vectors.

[0047] In this implementation scheme, by continuously mapping the spatial location of the priority monitoring area to infrared image frames and visible light image frames, and continuously discriminating the highest temperature value, average temperature value, and temperature range value within a unified local monitoring frame range, the identification of thermal anomaly-triggered parking spaces can be based on the common constraints of spatial focusing results and temporal evolution results. This avoids false triggering caused by high temperature disturbances in a single frame, improves the stability, relevance, and reliability of the thermal anomaly trigger dataset extraction, and provides more accurate anomaly input basis for subsequent calculation of blocking control characterization quantities and determination of target blocking categories.

[0048] Specifically, the specific steps for constructing an aligned control input vector based on the thermal anomaly trigger dataset and inputting it into a low-rank aligned control model, outputting the blocking control representation quantity and the actual thermal runaway posterior judgment result are as follows: Read the thermal anomaly trigger dataset, extract the infrared image frames, visible light image frames, charging output voltage values, charging output current values, cumulative charging capacity values, cumulative charging duration values, station ambient temperature values, station relative humidity values, station wind speed values, station solar irradiance values, and station illuminance values ​​from the local monitoring frames corresponding to each thermal anomaly trigger parking space at each sampling time, and concatenate them in a fixed order to form the aligned control input vector for each thermal anomaly trigger parking space; among which, first according to the thermal anomaly trigger... The parking space number corresponding to the thermal anomaly trigger dataset is read, and the data records corresponding to each sampling time are sorted in ascending order by sampling timestamp. Then, the local monitoring frames, infrared image frames, visible light image frames, charging output voltage values, charging output current values, cumulative charging power values, cumulative charging time values, station ambient temperature values, station relative humidity values, station wind speed values, station solar irradiance values, and station illuminance values ​​corresponding to the same thermal anomaly trigger parking space at continuous sampling times are extracted. For infrared image frames and visible light image frames, they are first resampled according to a unified input size and then converted into fixed-length image feature sequences. For charging output voltage values, charging output current values, and cumulative charging power values... The values ​​of infrared image frame features, visible light image frame features, charging output voltage value sequence, charging output current value sequence, cumulative charging capacity value sequence, cumulative charging time value sequence, station ambient temperature value sequence, station relative humidity value sequence, station wind speed value sequence, station solar irradiance value sequence, and station illuminance value sequence are constructed into a numerical feature sequence according to a time index consistent with the image sampling time. Then, the fixed-length image feature sequence and the numerical feature sequence are concatenated end-to-end according to a preset splicing order to form an alignment control input vector. The preset splicing order is: infrared image frame features, visible light image frame features, charging output voltage value sequence, charging output current value sequence, cumulative charging capacity value sequence, cumulative charging time value sequence, station ambient temperature value sequence, station relative humidity value sequence, station wind speed value sequence, station solar irradiance value sequence, and station illuminance value sequence. The principle behind this processing method is that, during the local thermal evolution of a parking space triggered by a thermal anomaly, it is simultaneously affected by image representation information, charging condition information, and environmental disturbance information. By concatenating the multi-source features into a single aligned control input vector under a unified time sequence, the subsequent low-rank aligned control model can form a consistent representation of the same thermal anomaly process. A low-rank aligned control model is constructed based on a pre-trained multimodal vision-language network that integrates convolutional neural networks and Transformers. The backbone parameters of the pre-trained multimodal vision-language network are frozen, and low-rank dimensionality reduction matrices and low-rank dimensionality increase matrices are inserted into the query projection layer and value projection layer of the self-attention mechanism, respectively, to form low-rank adaptation parameters.The pre-trained multimodal vision-language network is used to uniformly map infrared image frame features, visible light image frame features, and numerical feature sequences. Frozen backbone parameters are used to maintain the pre-trained multimodal vision-language network's general representational ability in basic cross-modal mapping. Low-rank dimensionality reduction and increase matrices are used to handle incremental alignment under thermal runaway flexible tampering tasks. Low-rank adaptation parameters only serve as trainable parameters within the low-rank alignment control model during model construction and training, and are not used as outputs during the inference phase. The low-rank dimension ranges from 4 to 64, preferably from 8 to 32, with the value determined by the thermal anomaly sample size, model response time requirements, and edge deployment computing power constraints. A near-end strategy optimization algorithm is used to optimize the low-rank alignment control model. The low-rank adaptation parameters are trained with training objectives including the accuracy of blocking category determination, the false alarm suppression rate of environmental hot spots, and the timeliness of blocking response. During training, a training set is constructed using historical samples of parking spaces triggered by thermal anomalies. Infrared image frame features, visible light image frame features, numerical feature sequences corresponding to charging conditions, and numerical feature sequences corresponding to environmental disturbances are used as training inputs. Real thermal runaway annotation results, environmental false hot spot annotation results, and blocking category annotation results are used as training supervision information. The low-rank adaptation parameters in the query projection layer and the value projection layer are iteratively updated to enable the low-rank aligned control model to adapt to the thermal anomaly blocking decision task of the charging station while keeping the core parameters frozen. The principle behind this approach is that pre-training multimodal vision-language... While the language network possesses general cross-modal representation capabilities, it lacks specific alignment relationships for flexible thermal runaway containment scenarios. By introducing a small number of incremental parameters through low-rank adaptation, task-specific alignment can be achieved without significantly increasing computational burden. Training with a proximal policy optimization algorithm ensures that the update direction of the low-rank adaptation parameters aligns with the actual containment decision objective. The alignment control input vector is input into the trained low-rank alignment control model, which outputs the containment control representation quantity for the corresponding thermally triggered parking space and the actual thermal runaway posterior judgment result. The containment control representation quantity characterizes the matching relationship between the thermally triggered parking space and different containment categories, while the actual thermal runaway posterior judgment result includes the actual thermal runaway posterior judgment value and the actual thermal runaway... The posterior judgment value is used to characterize the probability that the hotspot region is caused by a heat source inside the vehicle. The principle behind this approach is that the low-rank dimensionality reduction matrix and the low-rank dimensionality increase matrix serve as low-rank adaptation parameters within the model, adjusting the mapping relationship of multimodal features in the query projection layer and the value projection layer. During the inference phase, the original projection weight matrix, low-rank dimensionality reduction matrix, and low-rank dimensionality increase matrix are no longer output. Instead, the model directly outputs blocking control representation quantities that can be used for subsequent flexible blocking strategy database matching, as well as the true thermal runaway posterior judgment results that can be used for environmental pseudo-hotspot removal and dynamic mask matrix generation. This ensures that the output of the low-rank aligned control model can directly connect with the target blocking category determination, environmental pseudo-hotspot shielding, and subsequent blocking linkage control processes.

[0049] In this implementation, by uniformly constructing the local monitoring frames, multi-source operating condition data, and environmental data from the thermal anomaly trigger dataset into an aligned control input vector, and freezing the alignment control input vector into the backbone parameters, a low-rank aligned control model is introduced with low-rank adaptation parameters in the query projection layer and value projection layer. This model can achieve task adaptation in the thermal runaway flexible blocking scenario while retaining the general cross-modal representation capability of the pre-trained multimodal vision-language network, using a small number of low-rank incremental parameters. The low-rank aligned control model directly outputs the blocking control representation quantity and the actual thermal runaway posterior judgment result, enabling the image representation, charging condition representation, and environmental disturbance representation of the thermal anomaly triggering parking space to simultaneously participate in the blocking category determination and environmental pseudo-hot spot removal. This avoids the problem of unclear output objects caused by using the low-rank parameters inside the model as inference output, thereby improving the scenario adaptability of the target blocking category determination, the reliability of posterior verification, and the real-time performance of blocking linkage control.

[0050] Specifically, the steps for determining the target blocking category based on the actual thermal runaway posterior judgment results, blocking control characterization quantities, and flexible blocking strategy database, or for writing the corresponding environmental pseudo hotspot region into the dynamic mask matrix, are as follows: Extract the hotspot contour, hotspot area value, and hotspot area change value corresponding to the infrared image frame in the local monitoring frame; First, perform hotspot region identification on the infrared image frame in the local monitoring frame, identifying connected pixel regions whose temperature values ​​are greater than the sum of the station's ambient temperature value and the temperature difference threshold as hotspot regions. Then, extract the hotspot contour along the hotspot region boundary, and compare the number of pixels within the hotspot region with the single... Multiplying the actual area values ​​of each pixel yields the hotspot area value. Subtracting the hotspot area value from the previous sampling time from the current sampling time's hotspot area value and then dividing by the sampling time interval yields the hotspot area change value. The mean visible light brightness, variance of visible light brightness, percentage of pixels exceeding the brightness threshold, and mean texture gradient of the corresponding hotspot region are extracted from the visible light image frame. These values ​​are used as reflection interference characterization quantities. Specifically, the hotspot region in the infrared image frame is first classified according to the infrared image frame and the visible light image frame... The field-of-view mapping relationship between the pixels is projected onto the visible light image frame to obtain the visible light pixel set corresponding to the hot spot region. The visible light brightness mean is obtained by calculating the arithmetic mean of the brightness values ​​of each pixel in the visible light pixel set, and the visible light brightness variance is obtained by the mean of the squared deviations of the brightness values ​​of each pixel in the visible light pixel set relative to the visible light brightness mean. The brightness threshold is obtained by adding the brightness deviation threshold to the mean of the neighborhood background brightness after expanding the hot spot region by a preset pixel width. The percentage of pixels exceeding the brightness threshold is the sum of the number of pixels in the visible light pixel set whose brightness values ​​are greater than the brightness threshold and the total number of pixels in the visible light pixel set. The ratio; the mean texture gradient is obtained by calculating the arithmetic mean of the gradient magnitudes of each pixel in the visible light pixel set using the Sobel gradient; the brightness deviation threshold is set to 15 to 45 gray levels, preferably 20 to 35 gray levels, and the value is derived from the statistical results of the brightness difference between the station's solar reflection samples, shadow samples, and normal reflective samples on the vehicle surface; the preset pixel width is set to 5 to 30 pixels, preferably 8 to 20 pixels, and the value is derived from the statistical results of the registration error between the infrared image frame and the visible light image frame, the hot spot boundary expansion error, and the boundary expansion of the local reflection area of ​​the vehicle. The principle of this processing method is that environmental pseudo hot spots are usually accompanied by local reflection enhancement, brightness abrupt changes, and texture gradient anomalies. Using the mean visible light brightness, visible light brightness variance, the proportion of pixels exceeding the brightness threshold, and the mean texture gradient as reflection interference characterization quantities can provide a calculable basis for distinguishing between real thermal runaway caused by heat sources inside the vehicle and environmental pseudo hot spots caused by solar reflection and ground thermal radiation; the real-time power value is calculated based on the charging output voltage value and the charging output current value;Based on the real-time power value, the station ambient temperature value, and the sampling time interval corresponding to the local monitoring frame, a solid-state heat diffusion calculation model based on Fourier's law of heat transfer is used to calculate the theoretical temperature evolution value corresponding to each sampling moment. The real-time power value is obtained by multiplying the charging output voltage value and the charging output current value at the same sampling moment. The real-time power value serves as a characterization of the internal heat source caused by the electrical energy input during vehicle charging. The station ambient temperature value serves as the external boundary temperature value, and the sampling time interval serves as the time step for heat diffusion calculation. The solid-state heat diffusion calculation model describes the heat diffusion process within the local area of ​​the vehicle according to Fourier's law of heat transfer, calculating the theoretical temperature evolution value that should be formed under the action of the internal heat source at each sampling moment. The temperature evolution residual is calculated between the theoretical temperature evolution value and the measured temperature evolution value in the local monitoring frame. The measured temperature evolution value consists of the temperature statistics at each sampling time within the hotspot region of the local monitoring frame. The temperature evolution residual is represented by the absolute value or root mean square error of the difference between the measured temperature evolution value and the theoretical temperature evolution value at the same sampling time. The principle behind this processing method is that if the hotspot is mainly caused by internal heat sources within the vehicle, the measured temperature evolution trend should correspond to the theoretical temperature evolution trend driven by the real-time power value. If the hotspot is mainly caused by environmental reflection or external thermal radiation, a significant temperature evolution will occur between the measured temperature evolution trend and the theoretical temperature evolution trend driven by internal heat sources. The residuals are used as the posterior judgment criteria for the low-rank aligned control model, including temperature evolution residuals, hot spot area change values, reflection interference characterization quantities, station solar irradiance values, and station illuminance values, to obtain the true thermal runaway posterior judgment value. The true thermal runaway posterior judgment value is output by the trained low-rank aligned control model based on these residuals, hot spot area change values, reflection interference characterization quantities, station solar irradiance values, and station illuminance values, and is used to characterize the possibility that the hot spot area is caused by a heat source inside the vehicle. In an optional implementation, the true thermal runaway posterior judgment value is generated through a posterior judgment branch in the low-rank aligned control model. This posterior judgment branch uses a Bayesian directed acyclic graph to determine the true thermal runaway value of the parking space triggered by the thermal anomaly. Probabilistic inference is performed on the thermal runaway state. Specifically, the temperature evolution residual, hot spot area change, and blocking control characterization are used as internal heat source observation evidence nodes; the reflection interference characterization, station solar irradiance, and station illuminance are used as environmental pseudo-hot spot observation evidence nodes; and the actual thermal runaway state is used as a latent variable node to be determined. The node conditional probability relationship is determined based on the correspondence between the actual thermal runaway samples, environmental pseudo-hot spot samples, and each observation evidence node in the historical thermal anomaly samples. The posterior distribution corresponding to the actual thermal runaway state is updated based on the temperature evolution residual, hot spot area change, reflection interference characterization, station solar irradiance, and station illuminance at the current sampling time, resulting in the posterior determination value for the actual thermal runaway.When using the logarithmic dominance ratio, the natural logarithm of the ratio of the posterior probability corresponding to the true thermal runaway state to the posterior probability corresponding to the environmental pseudo-hotspot state is taken to obtain the posterior judgment value of the true thermal runaway. The principle behind this approach is that the Bayesian directed acyclic graph can simultaneously incorporate evidence of internal heat sources and environmental interference into the posterior judgment process. This ensures that the posterior judgment value of the true thermal runaway depends not only on infrared temperature changes but also on visible light reflection characteristics and ambient lighting conditions, thereby reducing the probability of misjudging environmental pseudo-hotspots caused by solar reflection and ground thermal radiation. Furthermore, the principle behind this posterior judgment approach is that the temperature evolution residual value reflects the consistency between the measured thermal evolution and the theoretical thermal evolution of the internal heat source, while the hotspot area change value reflects the thermal anomaly region. The expansion trend, the reflection interference characterization quantity, the station solar irradiance value, and the station illuminance value reflect the intensity of environmental light and reflection disturbance. Using this information together as the posterior judgment basis can reduce the probability of environmental false hot spots being misjudged as real thermal runaway. When the posterior judgment value of real thermal runaway is greater than the posterior judgment threshold, the blocking category coding vectors corresponding to the first-level blocking decision, the second-level blocking decision, and the maintenance monitoring decision are extracted from the flexible blocking strategy database. The flexible blocking strategy database is formed by summarizing historical thermal anomaly handling records, handling action category labels, and action execution effect evaluation results. The flexible blocking strategy database pre-stores multiple blocking category coding vectors. The first-level blocking decision corresponds to the early warning prompt handling category, and the second-level blocking... The decision corresponds to the linkage isolation and disposal category, while the monitoring decision corresponds to the continuous observation category. The encoding vector for each blocking category maintains the same dimension as the blocking control representation quantity, used to represent the standard reference position of different blocking decisions in the control representation space. The posterior judgment threshold is set between 0.60 and 0.85, preferably between 0.70 and 0.80. This value is derived from the differentiation results between real thermal runaway samples and environmental pseudo-hot spot samples in historical thermal anomaly samples, and is selected based on the minimum weighted sum of the false positive rate of real thermal runaway and the false positive rate of environmental pseudo-hot spots. The principle behind this approach is that if categories are directly divided based solely on a single value of the blocking control representation quantity, the category boundaries are prone to instability due to differences in the numerical distribution of different feature dimensions. Pre-constructing the blocking... The category encoding vector maps different blocking decisions to standard category centers in the control representation space, facilitating similarity matching. The cosine similarity between the blocking control representation and each blocking category encoding vector is calculated, and the blocking category corresponding to the blocking category encoding vector with the highest cosine similarity is determined as the target blocking category. Specifically, the cosine value of the angle between the blocking control representation and each blocking category encoding vector is calculated; a higher cosine similarity indicates that the blocking control representation and the corresponding blocking category encoding vector are closer in direction. Then, all cosine similarity results are compared, and the blocking category encoding vector corresponding to the highest cosine similarity value is selected as the target blocking category.The principle behind this processing method is that cosine similarity reflects the proximity of the blocking control representation quantity and the blocking category encoding vector in the feature distribution direction. It is not directly governed by absolute amplitude differences and is suitable for category matching in multidimensional control representation spaces, thus making the determination of the target blocking category more stable, clear, and calculable. When the true posterior judgment value of thermal runaway is not greater than the posterior judgment threshold, the pixel coordinate range of the corresponding hotspot region, the writing timestamp, and the mask duration period number are written into the dynamic mask matrix, and the temperature gradient features of the corresponding region in the dynamic mask matrix are masked during subsequent monitoring periods. The pixel coordinate range is used to record the location of the hotspot region determined to be an environmental pseudo-hotspot, and is written into... The timestamp records the generation time of the dynamic mask matrix, and the mask duration period number limits the effective period for which the hotspot area is shielded in subsequent monitoring cycles. The principle behind this approach is that when the actual thermal runaway posterior judgment value does not exceed the posterior judgment threshold, the corresponding hotspot area is more likely formed by environmental reflection, ground thermal radiation, or light disturbance. If the temperature gradient characteristics of this area continue to directly participate in subsequent thermal anomaly triggering, it can easily cause repeated false triggers. By recording and shielding the temperature gradient characteristics of the corresponding area through the dynamic mask matrix, the interference of environmental pseudo-hotspots on subsequent inspection focusing and blocking decisions can be reduced, while simultaneously preserving the continuous acquisition status of infrared image frames, visible light image frames, and charging output current values.

[0051] In this implementation scheme, the true thermal runaway state is determined by combining changes in infrared hotspots, visible light reflection interference characterization, and temperature evolution residual values. This can distinguish between abnormal heat sources inside the vehicle and false hotspots in the environment. When the posterior judgment value of the true thermal runaway meets the blocking conditions, the target blocking category is matched based on the blocking control characterization quantity. When the blocking conditions are not met, the hotspot region is written into the dynamic mask matrix to reduce repeated triggering of the same false hotspot, thereby improving the accuracy, anti-interference ability, and continuous monitoring stability of the flexible blocking decision for thermal runaway.

[0052] Specifically, the steps for performing hard blocking, flexible blocking, and continuous monitoring on corresponding parking spaces based on the target blocking category are as follows: First, read the target blocking category corresponding to each thermal anomaly-triggered parking space, and extract the corresponding maximum temperature value, temperature rise rate value, thermal distribution discrete value, station risk potential field value, blocking control characterization quantity, and actual thermal runaway posterior judgment result. Specifically, first, read the target blocking category according to the parking space number, then extract the maximum temperature value, temperature rise rate value, thermal distribution discrete value, station risk potential field value, blocking control characterization quantity, and actual thermal runaway posterior judgment value corresponding to the parking space number from the thermal anomaly trigger dataset, the blocking control characterization quantity calculation result, and the actual thermal runaway posterior judgment result, forming the basis for blocking execution judgment of the thermal anomaly-triggered parking space; after actual thermal runaway... The verification judgment results include the true thermal runaway posterior judgment value, which characterizes the probability that the hot spot region is caused by a heat source inside the vehicle; and the regional temperature gradient value is calculated based on the average absolute value of the temperature difference between adjacent pixels in the hot spot region in the infrared image frame of the corresponding parking space; the regional temperature gradient value is used to characterize the degree of temperature spatial change inside and near the boundary of the hot spot region; the principle of this processing method is that the target blocking category reflects the category matching result in the control representation space, the highest temperature value, the temperature rise rate value, the heat distribution discrete value, and the station risk potential field value reflect the actual risk state of the parking space triggered by thermal anomalies, the blocking control representation quantity reflects the comprehensive control representation result after multi-source input fusion, and the true thermal runaway posterior judgment value reflects the thermal The possibility that the heat source in the affected area is caused by a heat source inside the vehicle, and the temperature gradient value of the area reflects the spatial expansion state of the thermal anomaly. Using these data together as the basis for execution can avoid the result of a single category being out of touch with the actual risk situation on site. When the highest temperature value exceeds the extreme value blocking threshold, the corresponding parking space does not enter the low-rank alignment control model verification process, but is directly marked as a hard-cut-off parking space. The extreme value blocking threshold is set between 60 degrees Celsius and 120 degrees Celsius, preferably between 80 degrees Celsius and 100 degrees Celsius. The value is derived from the test data of the temperature rise characteristics before thermal runaway of the power battery, the safety management and control specifications of the charging station, and the advance requirement for fire-fighting response at the station. The principle of this processing method is that when the highest temperature value has exceeded the extreme value blocking threshold, the parking space that triggered the thermal anomaly has already entered the fast-track phase. In the case of a rapid temperature rise risk, if we continue to wait for the model verification results, the electrical energy input in the charging circuit may continue to accumulate. Therefore, we directly enter the hard cut-off process from the edge side. When the highest temperature value is not greater than the extreme value blocking threshold and is within the fuzzy warning temperature range, and the regional temperature gradient value is greater than the primary warning threshold, the corresponding parking space will be input into the low-rank aligned control model for a post-hoc judgment of the actual thermal runaway. The fuzzy warning temperature range is defined by the environmental baseline temperature value and the extreme value blocking threshold. The environmental baseline temperature value is the sum of the station's environmental temperature value and the temperature difference threshold. The primary warning threshold is the judgment threshold for the regional temperature gradient value. The primary warning threshold is determined by the statistical results of the boundary gradient of the normal charging hot spot, the statistical results of the boundary gradient of the environmental pseudo hot spot, and the initial gradient mutation samples of the thermal anomaly.The principle behind this processing method is that if the highest temperature value does not reach the extreme value blocking threshold, direct hard cutting off may cause over-processing. However, when the highest temperature value is within the fuzzy warning temperature range and the regional temperature gradient value is greater than the primary warning threshold, the thermal anomaly triggering parking space has already shown a local thermal anomaly expansion trend. Therefore, the low-rank alignment control model is triggered to perform a post-hoc judgment of the true thermal runaway to distinguish between the true thermal runaway caused by heat sources inside the vehicle and environmental pseudo-hot spots formed by solar reflection and ground thermal radiation. When the post-hoc judgment value of the true thermal runaway is not greater than the post-hoc judgment threshold, the corresponding parking space is marked as a continuously monitored parking space. The post-hoc judgment threshold is set between 0.60 and 0.85, preferably between 0.70 and 0.80, and the value is derived from the true thermal anomaly samples in the historical data. The distinction between thermal runaway samples and environmental pseudo-hotspot samples was determined, with the selection criterion being the minimum weighted sum of the false negative rate of actual thermal runaway and the false negative rate of environmental hotspots. The principle behind this approach is that when the posterior judgment value of actual thermal runaway is not greater than the posterior judgment threshold, the corresponding hotspot area is more likely to be an environmental pseudo-hotspot, and direct triggering of hard cutoff or flexible blocking is not advisable. Marking the corresponding parking space as a continuously monitored parking space can reduce false blocking. When the highest temperature value is not greater than the extreme value blocking threshold and the target blocking category is either a first-level blocking decision or a second-level blocking decision, the corresponding parking space is marked as a flexible blocking parking space. The first-level blocking decision corresponds to the early warning and prompt handling category, and the second-level blocking decision corresponds to the linkage isolation handling category. Both are applicable when the charging circuit is still short-lived. Under the premise of maintaining a controlled state, flexible measures are initiated. The principle behind this approach is that if the highest temperature value has not yet reached the extreme value blocking threshold, it indicates that there is still a buffer window for handling the on-site risk. At this time, by initiating tiered flexible blocking in conjunction with the target blocking category, time can be gained for on-site coordinated handling while maintaining local control. When the target blocking category is "maintain monitoring decision," the corresponding parking space is marked as a continuously monitored parking space. "Maintain monitoring decision" corresponds to a state where the current thermal anomaly-triggered parking space still needs continuous monitoring but active blocking actions are not yet executed. The principle behind this approach is that although some thermal anomaly-triggered parking spaces have entered the anomaly triggering process, the matching result between the blocking control representation quantity and the blocking category encoding vector points to the continuously observed category. In this case, continued monitoring is initiated. Continuous monitoring helps avoid excessive intervention. For hard-block parking spaces, output charging gun disconnect control commands and charging relay disconnect control commands to cut off the charging circuit of the corresponding parking space. The charging gun disconnect control command drives the charging gun to perform the disconnect action, and the charging relay disconnect control command drives the charging relay to switch to the disconnect state. After both are executed, the circuits corresponding to the charging output voltage value and the charging output current value are disconnected. The principle of this processing method is that the risk of thermal runaway is directly related to continuous power input. By simultaneously cutting off the charging gun connection path and the charging relay conduction path, the injection of external power into abnormal parking spaces can be reduced. For soft-block parking spaces, corresponding soft-block actions are generated according to the target blocking category.When the target blocking category is a Level 1 blocking decision, an audible and visual alarm control command and an in-station alarm information push command are generated. The audible and visual alarm control command triggers the alarm device near the corresponding parking space, while the in-station alarm information push command sends the parking space number, risk status, and handling prompts to the in-station monitoring terminal. The principle behind this approach is that Level 1 blocking decisions correspond to a state requiring on-site attention but not yet requiring coordinated isolation measures; audible and visual prompts and information pushes are prioritized to enable rapid manual intervention. When the target blocking category is a Level 2 blocking decision, parking spaces whose center coordinates are less than a distance threshold are determined based on their location to form a flexible handling area, and a directional spray control command is output. Water curtain isolation control commands and evacuation guidance control commands are generated. The distance threshold ranges from 1 to 5 meters, preferably from 1.5 to 3 meters, and is derived from the parking space spacing, the range of heat radiation diffusion, and the coverage radius of the fire sprinkler system. The directional sprinkler control command activates the local sprinkler system facing the corresponding parking space, the water curtain isolation control command activates the water curtain around the corresponding parking space, and the evacuation guidance control command drives the on-site audio-visual indicators to guide personnel away from the flexible response area. When generating the evacuation guidance control command, the center coordinates of the parking space corresponding to the secondary blocking decision are used as the risk source coordinates. The heat radiation safety radius is calculated based on the highest temperature value, temperature rise rate value, station wind speed value, and the station passage extension direction markings. A hazard exclusion zone is constructed covering the entire radius. The thermal radiation safety radius is set at 2 to 12 meters, preferably 3 to 8 meters, derived from estimated heat release from abnormal power battery temperatures, parking space spacing, safe personnel distance, and advance response requirements for fire suppression at the facility. When a wind direction sensor is installed at the facility, the wind direction value is further read to correct the direction of the thermal radiation safety radius. When no wind direction sensor is installed, the main extension direction of the facility's passageway is used as the diffusion direction marker. A safety distance matrix is ​​then constructed based on the risk source coordinates, adjacent parking space center coordinates, personnel positioning coordinates, exit coordinates, and passageway boundary coordinates. Elements in the safety distance matrix represent the personnel positioning coordinates and the coordinates of the risk source and adjacent parking space centers. The distance values ​​between the marker, channel boundary coordinates, and outlet coordinates are calculated. Then, based on the station wind speed, diffusion direction markers, and channel grid connectivity, a hot flue gas convection diffusion model is established. The diffusion impact value of each channel grid is calculated, and channel grids with diffusion impact values ​​greater than the diffusion impact threshold are incorporated into the hazard rejection zone. The diffusion impact threshold is set between 0.30 and 0.70, preferably between 0.40 and 0.60, and the values ​​are derived from hot flue gas diffusion simulation samples, channel width, and personnel passage safety margin. An evacuation potential energy field is constructed based on the hazard rejection zone and outlet coordinates. The risk source coordinates, the thermal radiation safety radius coverage area, and channel grids with diffusion impact values ​​greater than the diffusion impact threshold form a repulsive potential energy field, while the outlet coordinates form an attractive potential energy field.Among them, the repulsive potential energy is negatively correlated with the distance from the corresponding channel grid to the risk source coordinates, while the attractive potential energy is positively correlated with the distance from the corresponding channel grid to the exit coordinates. For each person's location coordinates, an evacuation path from the current location to the exit coordinates is searched in the evacuation potential energy field, and the time slice occupancy of the evacuation paths corresponding to different people is checked. When two evacuation paths occupy the same channel grid in the same time slice, the waiting time slice of the path corresponding to the person farther from the exit coordinates is extended, or the path corresponding to the person farther from the exit coordinates is switched to an adjacent alternative channel grid to generate a multi-person non-intersecting evacuation path. Based on the multi-person non-intersecting evacuation path, an audio-visual guidance control sequence is output, which includes an audio-visual alarm device. The system uses a numbered system, including the indicator light number, direction marker, activation time stamp, and duration value, to drive the on-site audible and visual alarm devices and indicator lights to guide personnel away from the hazardous area along a multi-person, non-intersecting evacuation route. The principle behind this approach is that the risk corresponding to the secondary blocking decision has the potential for spatial diffusion. The flexible response area is no longer limited to a single parking space but covers the influence range adjacent to the abnormal parking space. Through the coordinated execution of directional spraying, water curtain isolation, and evacuation guidance, heat transmission can be suppressed, the heat impact on adjacent parking spaces can be blocked, and the risk of personnel exposure can be reduced. Simultaneously, a safety distance matrix is ​​used to clarify the spatial relationship between personnel, risk sources, passage boundaries, and exits, utilizing thermal... Radiation safety radius and hot flue gas convection diffusion model determine the danger repulsion zone. Repulsion and attraction potential energy guide personnel to avoid the heat radiation and flue gas diffusion areas and move towards the exit. Time slice occupancy verification generates multi-person, non-intersecting evacuation paths, reducing congestion in narrow passages and improving the matching of evacuation guidance control commands with the on-site thermal diffusion situation. For continuously monitored parking spaces, continuous acquisition of infrared image frames, visible light image frames, and charging output current values ​​is maintained. When the corresponding parking space has been written into the dynamic mask matrix, the temperature gradient characteristics of the corresponding area are masked. Furthermore, the continuously monitored parking spaces continue to retain local thermal state tracking and charging status during subsequent sampling. Tracking is used to reassess changes in the highest temperature value, temperature rise rate value, heat distribution discrete value, site risk potential field value, and regional temperature gradient value. The area corresponding to the dynamic mask matrix is ​​the environmental pseudo-hotspot area written when the actual thermal runaway posterior judgment value is not greater than the posterior judgment threshold. Temperature gradient feature masking is used to reduce the repeated interference of the same environmental pseudo-hotspot area on thermal anomaly triggering and blocking decisions in subsequent monitoring cycles. The principle of this processing method is that although the parking space has not yet entered the hard cut-off state or the flexible blocking state, the evolution of thermal anomalies is time-varying. Maintaining the continuous acquisition of infrared image frames, visible light image frames, and charging output current values ​​is beneficial for timely switching of blocking strategies when subsequent risks change.Simultaneously, masking the temperature gradient characteristics of the region corresponding to the dynamic mask matrix can reduce repeated false triggering caused by environmental pseudo-hotspots while preserving continuous data acquisition.

[0053] In this implementation plan, by using the target blocking category together with the highest temperature value, temperature rise rate value, heat distribution dispersion value, site risk potential field value, blocking control characterization quantity, actual thermal runaway posterior judgment value, and regional temperature gradient value for blocking execution determination, it is possible to classify and handle hard cut-off, flexible blocking, and continuous monitoring based on the distinction between actual thermal runaway and environmental pseudo-hot spots. At the same time, under the secondary blocking decision, the combination of directional spraying, water curtain isolation, evacuation guidance, and dynamic mask matrix shielding mechanism can make the blocking action adapt to the intensity of on-site thermal anomalies, spatial diffusion trends, and personnel evacuation needs, thereby improving the targeting of the flexible blocking process for thermal runaway, on-site safety, and the anti-interference capability of continuous monitoring.

[0054] Specifically, the steps to switch to hard cutoff and form a thermal runaway flexible blocking result set when any one of the following conditions is met—edge hard cutoff condition, flexible blocking failure condition, and model verification timeout condition—are as follows: During the execution of the flexible blocking action, the infrared image frame, charging output current value, charging gun connection status value, and charging relay status value of the corresponding parking space are continuously read; among them, according to the sampling period corresponding to the preprocessed thermal runaway blocking monitoring data, the parking spaces that have entered the flexible blocking state are cyclically read, and the reading objects are limited to the infrared image frame, charging output current value, charging gun connection status value, and charging relay status value corresponding to the parking space; at each sampling moment, the thermal runaway blocking data is re-extracted based on the infrared image frame. In the imaging area, the highest temperature value, regional temperature gradient value, and temperature gradient change rate value are recalculated. Then, the temperature rise rate value is recalculated based on the difference between the highest temperature values ​​at adjacent sampling times and the sampling time interval. The regional temperature gradient value is represented by the average of the absolute values ​​of the temperature differences between adjacent pixels within the hotspot region in the infrared image frame of the corresponding parking space. The temperature gradient change rate value is obtained by subtracting the regional temperature gradient value at the previous sampling time from the regional temperature gradient value at the current sampling time, and then dividing by the sampling time interval. The principle behind this processing method is that after the flexible blocking action is executed, the thermal anomaly state may continue to evolve. Relying solely on the judgment result at the start time of the flexible blocking action cannot reflect subsequent changes; therefore, it is necessary to consider the changes in the flexible blocking action. During the execution of the flexible blocking action, the system continuously tracks the highest temperature value, temperature rise rate, regional temperature gradient value, temperature gradient change rate value, charging output current value, charging gun connection status value, and charging relay status value to promptly identify whether the flexible blocking has failed and whether the edge side needs to take over the operation. When any of the following conditions are met: the highest temperature value and temperature rise rate value corresponding to adjacent sampling times increase sequentially within W consecutive sampling times; the forward inference time of the low-rank aligned control model exceeds the response time threshold; or the temperature gradient change rate value exceeds the gradient change rate threshold, the low-rank aligned control model's handling authority is immediately stripped, and the edge side switches the corresponding parking space to a hard-cut-off parking space, outputting a charging gun disconnection control command and charging... The relay disconnection control command; where W ranges from 3 to 8, preferably from 3 to 5, and the values ​​are derived from the requirements for identifying short-term continuous evolution of thermal anomalies, the range of sampling period settings, and the response time of flexible blocking actions; "the highest temperature value increases sequentially" means that the highest temperature value of the later sampling time is greater than the highest temperature value of the previous sampling time in a series of W consecutive sampling times; "the temperature rise rate value increases sequentially" means that the temperature rise rate value of the later sampling time is greater than the temperature rise rate value of the previous sampling time in a series of W consecutive sampling times; the response time threshold ranges from 50 milliseconds to 500 milliseconds, preferably from 100 milliseconds to 300 milliseconds, and the values ​​are derived from the real-time requirements of edge inference, the linkage time limit of charging station blocking actions, and the development speed of the thermal runaway front."The forward inference time of the low-rank alignment control model is greater than the response time threshold" refers to the time from the start of the alignment control input vector being fed into the low-rank alignment control model to the completion of the output blocking control representation and the actual thermal runaway posterior judgment result being greater than the response time threshold; the gradient change rate threshold is set at 0.5 degrees Celsius per pixel per second to 5 degrees Celsius per pixel per second, preferably 1 degree Celsius per pixel per second to 3 degrees Celsius per pixel per second, and the values ​​are derived from normal charging temperature gradient change samples, rapid thermal anomaly expansion samples, and infrared sampling periods; the principle behind this processing method is that when the maximum temperature value and the temperature rise rate value continue to increase, it indicates that the flexible blocking action has failed to effectively suppress the risk development; when the forward inference time of the low-rank alignment control model is greater than the response time threshold... When the temperature gradient change rate exceeds the threshold, it indicates that the spatial temperature difference in the hotspot area is rapidly expanding, posing a risk of thermal anomalies spreading from the local area to the periphery. Therefore, when any of these conditions are met, the edge side directly takes over and switches to hard-block parking spaces to reduce the risk accumulation caused by model waiting and delays in flexible handling. The blocking execution results for each parking space are summarized. Specifically, the execution status data of hard-blocked parking spaces, flexible-blocked parking spaces, and continuously monitored parking spaces after this round of handling are summarized according to the parking space number, forming a blocking execution result record for each parking space. The blocking execution result record includes at least the parking space number, charging gun connection status value, charging relay status value, and highest temperature. The data includes the following parameters: temperature rise rate, temperature gradient, temperature gradient change rate, and current status. The principle behind this processing method is that both the completion and flexible blocking determination rely on post-execution status data. First, the blocking execution results for each parking space are summarized, providing a consistent data foundation for subsequent unified determination. When the charging gun connection status for the corresponding parking space is "disconnected" and the charging relay status is "disconnected," the corresponding parking space is marked as a completed blocking parking space. "Charging gun connection status is disconnected" indicates that the physical connection of the charging interface has been released, and "charging relay status is disconnected" indicates that the electrical connection of the charging main circuit has been cut off. Only when both conditions are met simultaneously is the corresponding parking space marked as a completed blocking parking space. The principle behind this processing method is... Since the disconnection of the charging gun connection status alone cannot fully indicate that the electrical circuit has been isolated, and the disconnection of the charging relay status alone cannot fully indicate that the external connection has been released, the simultaneous fulfillment of both states is used as the blocking completion criterion to improve the accuracy of the blocking completion determination. When the highest temperature value of adjacent sampling times no longer increases and the temperature rise rate no longer increases in the W consecutive sampling times after the flexible blocking action is executed, the corresponding parking space is marked as a flexible blocking completed parking space. Among them, "the highest temperature value no longer increases" means that the highest temperature value of the later sampling time is less than the highest temperature value of the previous sampling time in the W consecutive sampling times, or the difference between the highest temperature value of the later sampling time and the highest temperature value of the previous sampling time is less than the preset temperature fluctuation tolerance."The temperature rise rate no longer increases" means that in a series of W consecutive sampling times, the temperature rise rate at the next sampling time is less than the temperature rise rate at the previous sampling time, or the difference between the temperature rise rate at the next sampling time and the temperature rise rate at the previous sampling time is less than the preset rate fluctuation tolerance. The preset temperature fluctuation tolerance is 0.1°C to 1°C, preferably 0.2°C to 0.5°C, and is derived from the infrared thermometry resolution and the amplitude of the station background thermal disturbance. The preset rate fluctuation tolerance is 0.01°C per second to 0.5°C per second, preferably 0.05°C per second to 0.2°C per second, and is derived from the fluctuation range of the temperature rise rate values ​​calculated from adjacent sampling times and the thermal state recovery characteristics after the flexible blocking action. The principle of this processing method is that the completion of flexible blocking does not require a significant drop in temperature in a short period of time, but rather requires the suppression of the thermal anomaly evolution trend. Therefore, the maximum temperature value no longer increases and the temperature rise rate no longer increases are determined by the following conditions. Increasing the threshold as a criterion for completing flexible blocking better aligns with the actual thermal state changes after flexible handling. Furthermore, parking spaces with completed blocking, those with completed flexible blocking, and those under continuous monitoring are aggregated according to their parking space numbers to form a thermal runaway flexible blocking result set. This set records the final state category corresponding to each parking space number, including parking spaces with completed blocking, those with completed flexible blocking, and those under continuous monitoring. The thermal runaway flexible blocking result set also records the final charging gun connection status, final charging relay status, final maximum temperature, final temperature rise rate, final area temperature gradient, and final temperature gradient change rate for each parking space. The principle behind this approach is that by uniformly aggregating parking spaces with completed blocking, those with completed flexible blocking, and those under continuous monitoring, a complete closed-loop thermal runaway handling result can be formed, providing consistent data output for subsequent on-site recording and storage, alarm verification, and strategy optimization.

[0055] In this implementation scheme, by continuously tracking infrared image frames, charging output current values, charging gun connection status values, and charging relay status values ​​during the execution of the flexible blocking action, and combining the highest temperature value, temperature rise rate value, regional temperature gradient value, temperature gradient change rate value, and forward inference time of the low-rank aligned control model to determine whether to switch to hard cutoff, the edge side can take over the handling in a timely manner when the flexible blocking fails, the model verification timeout occurs, or the thermal anomaly spatial expansion intensifies. At the same time, the blocking completion status is determined by jointly judging the charging gun connection status value and the charging relay status value, and the flexible blocking completion status is determined by the continuous change of the highest temperature value and temperature rise rate value. This enables the thermal runaway flexible blocking result set to accurately reflect the final handling status of each parking space, thereby improving the real-time performance, fallback reliability, and status determination accuracy of the thermal runaway handling closed loop.

[0056] like Figure 2As shown, the second aspect of the present invention provides a flexible thermal runaway prevention system for vehicles in charging stations, comprising: a data acquisition and processing module, a risk focusing triggering module, a low-rank alignment decision module, and a flexible prevention linkage module, wherein: the data acquisition and processing module is used to acquire thermal runaway prevention monitoring data, perform time base alignment, anomaly removal, noise suppression, and numerical scale unification processing on the thermal runaway prevention monitoring data, and output preprocessed thermal runaway prevention monitoring data; the risk focusing triggering module is used to extract the thermal risk characteristics corresponding to each effective parking space based on the preprocessed thermal runaway prevention monitoring data, calculate the risk potential field value corresponding to each sampling position in the station plane on a grid-by-grid basis, determine the priority monitoring area, and generate local monitoring frames based on local monitoring The continuous frame change results generate a thermal anomaly trigger dataset; the low-rank alignment decision module is used to construct an alignment control input vector based on the thermal anomaly trigger dataset and input it into the low-rank alignment control model, outputting the blocking control characterization quantity and the actual thermal runaway posterior judgment result; the target blocking category is determined based on the actual thermal runaway posterior judgment result, the blocking control characterization quantity, and the flexible blocking strategy database, or the corresponding environmental pseudo hot spot region is written into the dynamic mask matrix; the flexible blocking linkage module is used to perform hard cut-off, flexible blocking, or continuous monitoring on the corresponding parking space based on the target blocking category, and switches to hard cut-off when any of the following conditions are met: edge hard cut-off condition, flexible blocking failure condition, and model verification timeout condition, forming a thermal runaway flexible blocking result set.

[0057] In this implementation plan, by sequentially connecting the preprocessing of thermal runaway containment monitoring data, extraction of thermal risk features, calculation of risk potential field value, generation of thermal anomaly triggering dataset, calculation of containment control characterization quantity, determination of target containment category, and output of thermal runaway flexible containment result set, a continuous handling chain can be formed from risk discovery to containment execution and result confirmation. This ensures that the identification basis, decision-making basis, and execution basis of thermal anomaly triggered parking spaces remain consistent throughout, improving the overall coordination, decision-making accuracy, and completeness of the thermal runaway flexible containment process.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0059] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A flexible method for preventing thermal runaway of vehicles at charging stations, characterized in that, Includes the following steps: S1 collects thermal runaway arrest monitoring data, performs time base alignment, anomaly removal, noise suppression and numerical scale unification on the thermal runaway arrest monitoring data, and outputs the preprocessed thermal runaway arrest monitoring data. S2, based on the preprocessed thermal runaway interruption monitoring data, extract the thermal risk characteristics corresponding to each effective parking space, calculate the risk potential field value corresponding to each sampling position in the station plane for each grid point, determine the priority monitoring area, and generate local monitoring frames. Based on the continuous frame change results of the local monitoring frames, generate a thermal anomaly trigger dataset. S3: Construct an aligned control input vector based on the thermal anomaly triggering dataset and input it into the low-rank aligned control model. Output the blocking control characterization quantity and the true thermal runaway posterior judgment result. Determine the target blocking category based on the true thermal runaway posterior judgment result, the blocking control characterization quantity, and the flexible blocking strategy database, or write the corresponding environmental pseudo hot spot region into the dynamic mask matrix. S4 performs hard cut-off, flexible cut-off, or continuous monitoring on the corresponding parking space based on the target blocking category, and switches to hard cut-off when any of the following conditions are met: edge hard cut-off condition, flexible cut-off failure condition, and model verification timeout condition, forming a set of thermal runaway flexible cut-off results.

2. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 1, characterized in that: The specific steps for collecting thermal runaway arrest monitoring data and performing time-base alignment, anomaly removal, noise suppression, and numerical scale unification on the thermal runaway arrest monitoring data are as follows: The system collects thermal runaway interruption monitoring data generated by target vehicles during the charging process within the charging station. The thermal runaway interruption monitoring data includes parking space number, infrared image frame, infrared acquisition timestamp, visible light image frame, visible light acquisition timestamp, gimbal yaw angle value, gimbal pitch angle value, lens zoom ratio value, charging output voltage value, charging output current value, cumulative charging amount value, cumulative charging duration value, charging gun connection status value, charging relay status value, charging pile operation status value, station ambient temperature value, station relative humidity value, station wind speed value, station solar irradiance value, and station illuminance value. For the collected thermal runaway arrest monitoring data, the nearest neighbor timestamp matching algorithm is used to perform unified time reference alignment processing on the multi-source acquisition sequences; the Hamper filter algorithm is used to perform anomaly identification and removal on the thermal runaway arrest monitoring data; the Kalman filter algorithm is used to perform noise suppression and time series smoothing on the thermal runaway arrest monitoring data; and the range normalization algorithm is used to perform numerical scale unification processing on the thermal runaway arrest monitoring data, outputting the preprocessed thermal runaway arrest monitoring data.

3. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 2, characterized in that: The specific steps for extracting thermal risk characteristics corresponding to each effective parking space based on the preprocessed thermal runaway prevention monitoring data, and calculating the risk potential field value corresponding to each sampling position in the station plane for each grid point to determine the priority monitoring area are as follows: Using the ground corner point on one side of the station entrance as the origin and the horizontal extension direction of the station as the horizontal axis, a station plane coordinate system is established, and the coordinates of the center position corresponding to each parking space number are calibrated. Then, the pre-processed thermal runaway interruption monitoring data is read, and the data corresponding to each parking space is merged according to the parking space number. Only the parking space number with the charging gun connection status value as ON, the charging relay status value as OFF, and the charging pile operation status value as OPEN is retained, forming a monitoring unit for each valid parking space. For each valid parking space monitoring unit, the thermal imaging area is extracted from the connected pixel regions in the infrared image frame whose temperature values ​​are higher than the sum of the ambient temperature and the temperature difference threshold. The maximum temperature value among the pixel temperatures within the thermal imaging area is extracted as the highest temperature value, and the arithmetic mean of the pixel temperatures within the thermal imaging area is calculated as the average temperature value. The difference between the highest temperature value and the average temperature value is calculated as the thermal distribution discrete value. The temperature rise rate is calculated based on the difference between the highest temperature values ​​at adjacent sampling times and the sampling time interval. The charging power value is calculated based on the product of the charging output voltage and the charging output current. The charging power change rate is calculated based on the difference between the cumulative charging power values ​​at adjacent sampling times and the sampling time interval. The ambient temperature, solar irradiance, and illuminance of the parking space are used as environmental constraints, and together with the highest temperature value, temperature rise rate, thermal distribution discrete value, charging power value, and charging power change rate, they form a parking space risk feature vector. Discrete sampling is performed on the station plane according to the grid spacing n, and the risk potential field value is calculated for each grid point. The grid point with the largest risk potential field value is selected as the priority monitoring center point. The priority monitoring area is formed by the grid where the priority monitoring center point is located and the eight neighboring grids.

4. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 3, characterized in that: The specific steps for calculating the risk potential field value point by grid are as follows: For any coordinate point within the station plane, take the L2 norm of the risk feature vector of each effective parking space to obtain the risk intensity value of each effective parking space; multiply the risk intensity value of each effective parking space by the risk amplification factor and then take the natural exponent to obtain the risk amplification term of each effective parking space. Calculate the squared difference between the horizontal coordinate and the squared difference between the coordinate point and the center coordinates of each effective parking space, and add a smoothing constant to obtain the distance constraint term for each effective parking space; divide the risk amplification term of each effective parking space by the corresponding distance constraint term to obtain the risk potential field contribution value of each effective parking space to the coordinate point. The risk potential field contribution values ​​of all valid parking spaces are summed to obtain the risk potential field value of the coordinate point.

5. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 3, characterized in that: The specific steps for generating local monitoring frames and generating a thermal anomaly trigger dataset based on the continuous frame change results of the local monitoring frames are as follows: Using the priority monitoring center point as the target position, the gimbal yaw and pitch angles are adjusted based on the lateral and longitudinal positional deviations between the target position and the current field of view center of the infrared image frame. Then, the lens zoom ratio is adjusted based on the difference between the imaging size of the priority monitoring area in the infrared image frame and the target imaging size, ensuring that the priority monitoring area is located at the center of the field of view of both the adjusted infrared and visible light image frames. The priority monitoring area is then mapped onto the adjusted infrared and visible light image frames to obtain the minimum bounding rectangle of the mapped area. Finally, the area is locally cropped after expanding outwards by a pixel width m along the perimeter of the minimum bounding rectangle to form a local monitoring frame. For local monitoring frames, perform continuous frame comparison according to the sampling time sequence, calculate the highest temperature value, average temperature value, and temperature range value within the local monitoring frame at each sampling time, and calculate the difference between the highest temperature value and the difference between the average temperature value at adjacent sampling times; when the difference between the highest temperature value, the difference between the average temperature value, and the risk potential field value of the station corresponding to the priority monitoring area all increase sequentially among K consecutive sampling times, mark the corresponding parking space as a thermal anomaly triggering parking space, and extract the corresponding parking space number, local monitoring frame, parking space risk feature vector, and station risk potential field value, and output as a thermal anomaly triggering dataset.

6. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 5, characterized in that: The specific steps for constructing an aligned control input vector based on the thermal anomaly triggering dataset, inputting it into a low-rank aligned control model, and outputting the blocking control representation quantity and the actual thermal runaway posterior judgment result are as follows: Read the thermal anomaly trigger dataset, extract the infrared image frame, visible light image frame, charging output voltage value, charging output current value, cumulative charging power value, cumulative charging time value, station ambient temperature value, station relative humidity value, station wind speed value, station solar irradiance value and station illuminance value from the local monitoring frame corresponding to each thermal anomaly trigger parking space at each sampling time, and splice them in a fixed order to form the alignment control input vector of each thermal anomaly trigger parking space; A low-rank alignment control model is constructed based on a pre-trained multimodal vision-language network fused with convolutional neural networks and Transformers. The backbone parameters of the pre-trained multimodal vision-language network are frozen, and low-rank dimensionality reduction matrices and low-rank dimensionality increase matrices are inserted into the query projection layer and value projection layer of the self-attention mechanism, respectively, to form low-rank adaptation parameters. The low-rank adaptation parameters are trained using a proximal policy optimization algorithm, with training objectives including the accuracy of blocking category judgment, the false alarm suppression rate of environmental hotspots, and the timeliness of blocking response. The alignment control input vector is input into the trained low-rank alignment control model, which outputs the blocking control representation quantity corresponding to the thermal anomaly-triggered parking space and the posterior judgment result of the actual thermal runaway. The posterior judgment result of the actual thermal runaway includes the actual thermal runaway posterior judgment value, which is used to characterize the probability that the hotspot region is caused by a heat source inside the vehicle.

7. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 6, characterized in that: The specific steps for determining the target blocking category based on the actual thermal runaway posterior judgment results, blocking control characterization quantities, and flexible blocking strategy database, or for writing the corresponding environmental pseudo-hotspot region into the dynamic mask matrix, are as follows: Extract the hotspot contour, hotspot area value, and hotspot area change value corresponding to the infrared image frame in the local monitoring frame; extract the mean visible light brightness, visible light brightness variance, percentage of pixels exceeding the brightness threshold, and mean texture gradient value of the corresponding hotspot region from the visible light image frame, and use the mean visible light brightness, visible light brightness variance, percentage of pixels exceeding the brightness threshold, and mean texture gradient value as reflection interference characterization quantities; among them, the percentage of pixels exceeding the brightness threshold is the ratio of the number of pixels with brightness values ​​greater than the brightness threshold in the hotspot region to the total number of pixels in the hotspot region. The real-time power value is calculated based on the charging output voltage and charging output current. Based on the real-time power value, the station ambient temperature value, and the sampling time interval corresponding to the local monitoring frame, a solid heat diffusion calculation model based on Fourier's heat transfer law is used to calculate the theoretical temperature evolution value corresponding to each sampling time. The temperature evolution residual value between the theoretical temperature evolution value and the measured temperature evolution value in the local monitoring frame is calculated. Temperature evolution residuals, hotspot area changes, reflection interference characteristics, station solar irradiance, and station illumination are used as the posterior judgment criteria for the low-rank aligned control model to obtain the true thermal runaway posterior judgment value. When the true thermal runaway posterior judgment value is greater than the posterior judgment threshold, the blocking category coding vectors corresponding to the first-level blocking decision, second-level blocking decision, and maintenance monitoring decision are extracted from the flexible blocking strategy database. The cosine similarity between the blocking control characteristics and each blocking category coding vector is calculated, and the blocking category corresponding to the blocking category coding vector with the largest cosine similarity is determined as the target blocking category. When the true thermal runaway posterior judgment value is not greater than the posterior judgment threshold, the pixel coordinate range, writing timestamp, and mask duration period number of the corresponding hotspot region are written into the dynamic mask matrix, and the temperature gradient features of the region corresponding to the dynamic mask matrix are masked in subsequent monitoring periods.

8. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 7, characterized in that: The specific steps for performing hard blocking, flexible blocking, or continuous monitoring on the corresponding parking space based on the target blocking category are as follows: Read the target blocking category corresponding to each thermal anomaly-triggered parking space, and extract the corresponding maximum temperature value, temperature rise rate value, thermal distribution discrete value, station risk potential field value, blocking control characterization quantity and actual thermal runaway posterior judgment result; and calculate the regional temperature gradient value based on the average absolute value of the temperature difference between adjacent pixels in the hot spot area in the infrared image frame of the corresponding parking space. When the highest temperature value exceeds the extreme value blocking threshold, the corresponding parking space will not enter the low-rank alignment control model verification process, and will be directly marked as a hard-cut-off parking space. When the highest temperature value is not greater than the extreme value blocking threshold and is within the fuzzy warning temperature range, and the regional temperature gradient value is greater than the primary warning threshold, the corresponding parking space will be input into the low-rank aligned control model for a post-hoc judgment of the actual thermal runaway; when the post-hoc judgment value of the actual thermal runaway is not greater than the post-hoc judgment threshold, the corresponding parking space will be marked as a continuously monitored parking space. For parking spaces that are hard-cut off, output the charging gun disconnect control command and the charging relay disconnect control command to cut off the charging circuit of the corresponding parking space. For flexible barrier parking spaces, corresponding flexible barrier actions are generated based on the target barrier category. Specifically, when the target barrier category is a Level 1 barrier decision, an audible and visual alarm control command and an in-station alarm information push command are generated. When the target barrier category is a Level 2 barrier decision, parking spaces whose center coordinates are less than a distance threshold are determined according to the corresponding parking space location to form a flexible treatment action area, and directional spray control command, water curtain isolation control command, and evacuation guidance control command are output. For continuously monitored parking spaces, the infrared image frames, visible light image frames, and charging output current values ​​of the corresponding parking spaces are continuously acquired. When the corresponding parking space has been written into the dynamic mask matrix, the temperature gradient characteristics of the area corresponding to the dynamic mask matrix are masked.

9. The flexible method for preventing thermal runaway of vehicles in charging stations according to claim 8, characterized in that: The specific steps for switching to hard cut when any one of the following conditions is met—edge hard cut condition, flexible blocking failure condition, or model verification timeout condition—to form a thermal runaway flexible blocking result set are as follows: During the execution of the flexible blocking action, the infrared image frame, charging output current value, charging gun connection status value, and charging relay status value of the corresponding parking space are continuously read. Based on the infrared image frame, the highest temperature value, temperature rise rate value, regional temperature gradient value, and temperature gradient change rate value are recalculated. Among them, the regional temperature gradient value is represented by the average of the absolute values ​​of the temperature difference between adjacent pixels in the hot spot area in the infrared image frame of the corresponding parking space. The temperature gradient change rate value is obtained by subtracting the regional temperature gradient value of the previous sampling time from the regional temperature gradient value of the current sampling time and then dividing by the sampling time interval. When any of the following conditions are met: the highest temperature value and temperature rise rate value corresponding to adjacent sampling times increase sequentially in W consecutive sampling times, the forward inference time of the low-rank alignment control model is greater than the response time threshold, and the temperature gradient change rate value is greater than the gradient change rate threshold, the processing authority of the low-rank alignment control model is immediately stripped, and the corresponding parking space is switched to a hard-cut-off parking space by the edge side, and the charging gun disconnection control command and the charging relay disconnection control command are output. The results of the blocking actions for each parking space are summarized. When the charging gun connection status of the corresponding parking space is disconnected and the charging relay status is disconnected, the corresponding parking space is marked as a parking space where blocking is completed. When the highest temperature value corresponding to adjacent sampling times no longer increases and the temperature rise rate no longer increases for W consecutive sampling times after the execution of the flexible blocking action, the corresponding parking space is marked as a parking space where flexible blocking is completed. Then, the parking spaces where blocking is completed, the parking spaces where flexible blocking is completed, and the parking spaces under continuous monitoring are summarized according to the parking space number to form a set of results for flexible blocking of thermal runaway.

10. A flexible thermal runaway prevention system for vehicles in charging stations, characterized in that, include: The module comprises a data acquisition and processing module, a risk-focused triggering module, a low-rank alignment decision-making module, and a flexible blocking linkage module, among which: The data acquisition and processing module is used to acquire thermal runaway arrest monitoring data, perform time base alignment, anomaly removal, noise suppression and numerical scale unification processing on the thermal runaway arrest monitoring data, and output preprocessed thermal runaway arrest monitoring data. The risk focusing triggering module is used to extract the thermal risk characteristics corresponding to each effective parking space based on the preprocessed thermal runaway interruption monitoring data, calculate the risk potential field value corresponding to each sampling position in the station plane for each grid point, determine the priority monitoring area, generate local monitoring frames, and generate a thermal anomaly triggering dataset based on the continuous frame change results of the local monitoring frames. The low-rank alignment decision module is used to construct an alignment control input vector based on the thermal anomaly triggering dataset and input it into the low-rank alignment control model, outputting the blocking control characterization quantity and the actual thermal runaway posterior judgment result; and to determine the target blocking category based on the actual thermal runaway posterior judgment result, the blocking control characterization quantity and the flexible blocking strategy database, or to write the corresponding environmental pseudo hot spot region into the dynamic mask matrix. The flexible blocking linkage module is used to perform hard blocking, flexible blocking, or continuous monitoring on the corresponding parking space based on the target blocking category, and switches to hard blocking when any of the following conditions are met: edge hard blocking condition, flexible blocking failure condition, and model verification timeout condition, forming a set of thermal runaway flexible blocking results.

Images