A method for early warning analysis of thermal runaway from battery infrared images

Abstract

This invention discloses a method for early warning analysis of thermal runaway from battery infrared images, specifically relating to the field of infrared thermal imaging visual processing of power battery packs. It addresses the technical problem of the difficulty in consistently and early identifying local thermal anomalies under overall temperature drift and reflection interference. The method involves calibrating continuous infrared thermal image frames into a temperature matrix sequence, extracting low-fluctuation connected reference regions based on the acquisition window to generate a reference temperature field sequence, obtaining a temperature difference matrix sequence through differential analysis, and accumulating these to form a temperature rise trend matrix sequence and trend description. Further, it constructs isothermal boundary sequences for candidate anomaly regions and calculates diffusion consistency evidence. Combining the reference regions to form a trend baseline and evidence baseline generates an early warning criterion, outputting the early warning level and the location of candidate anomaly regions. The early warning criterion reduces threshold dependence based on frame index alignment and gating rules.

Description

Technical Field

[0001] This invention relates to the field of infrared thermal imaging visual processing of power battery packs, and more specifically, to a method for thermal runaway early warning analysis of battery infrared images. Background Technology

[0002] In the operation and maintenance of power battery packs and energy storage battery systems, to detect early signs of abnormal temperature rise in cells or modules, the industry commonly uses infrared thermal imaging for non-contact monitoring of the temperature distribution on the outer surface of the battery, and analyzes and issues warnings on the thermal images at the end-user or industrial control side. Existing technologies, such as the patent "A Method for Detecting Thermal Runaway of Lithium Batteries Based on Infrared Thermal Imager" (publication number CN113379628A), involve preprocessing the acquired thermal images and delineating candidate regions. After selecting the region with the highest temperature, the highest temperature in that region is compared with a preset threshold to trigger an alarm. Another existing patent, "A Battery Pack Temperature Detection System and Method" (application number CN111403836A), uses temperature points collected by the battery management system as a benchmark, combining them with the overall infrared thermal image of the cells within the battery pack. The temperature of each cell is calculated on the overall thermal image to determine if the temperature is abnormal and to provide early warning control.

[0003] In the above applications, although both types of solutions have introduced infrared images into the early warning link, the key basis for early warning judgment is still more inclined to whether there is a sufficiently prominent high temperature point or whether the calculated cell temperature has formed a clear anomaly. In other words, the risk signal is mainly tied to the temperature value or absolute temperature anomaly of a certain local area. This exposes a weakness in the early stages of thermal runaway that is not easily covered by simple optimization: the transition of a battery from normal operating conditions to runaway does not always begin with a clearly visible ultra-high temperature hotspot. In the early stages, it often manifests as a slow shift in the heat distribution pattern, subtle changes in the local heating rhythm and diffusion direction, and these changes are intertwined with the apparent temperature difference caused by charging / discharging switching, heat dissipation strategy adjustment, airflow disturbances in the battery compartment, and obstructions and reflections. When the algorithm focuses the criteria on a single maximum temperature or a single point temperature anomaly, in order to avoid being led by normal fluctuations, the thresholds and rules are often set more stably. As a result, anomalies must accumulate to a more obvious degree before triggering, and the intervention window is passively shortened. Conversely, if the criteria are made more sensitive in order to detect earlier, it is easy to mistake short-term operating condition fluctuations or local pseudo-hotspots for risk signals, resulting in frequent and difficult-to-interpret alarms. It is difficult for the operation and maintenance team to make accurate handling based on alarms, ultimately causing an engineering contradiction of delayed warning triggering and frequent false alarms.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for thermal runaway early warning analysis of battery infrared images. This method involves calibrating continuous infrared thermal images into a temperature matrix sequence, extracting low-fluctuation connected reference regions based on the acquisition window to generate a reference temperature field sequence, obtaining a temperature difference matrix sequence through differential analysis, accumulating these sequences to form a temperature rise trend matrix sequence and a trend description, further constructing isothermal boundary sequences for candidate abnormal regions and calculating diffusion consistency evidence, combining the reference regions to form a trend baseline and evidence baseline to generate an early warning criterion, and outputting the early warning level and the location of candidate abnormal regions. The early warning criterion is based on frame index alignment and gating rules to reduce threshold dependence, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] S1: Acquire continuous infrared thermal image frames within the field of view of a fixedly installed infrared thermal imager and record the acquisition time of each frame. Based on the radiometric calibration relationship of the thermal imager, convert the pixel response values ​​into temperature matrices and form a temperature matrix sequence according to the acquisition time order.

[0008] S2: Within the acquisition window, calculate the fluctuation matrix and temperature deviation matrix based on the temperature matrix sequence and acquisition time. Construct a set of reference candidate pixels based on the median of the fluctuation matrix and the median of the temperature deviation matrix, and determine the reference region under the four-connectivity constraint. The reference region is used to generate reference temperature frame by frame and construct a reference temperature field sequence.

[0009] S3: The temperature matrix subsequence and the reference temperature field sequence are differentially divided according to the frame index to form a temperature difference matrix sequence. Under the constraint of the acquisition time, the temperature difference matrix sequence generates a temperature difference change rate matrix sequence and is converted into a positive increment matrix sequence. The positive increment matrix sequence generates a temperature rise trend matrix sequence and a persistence matrix sequence and combines them with the reference area to generate a trend description.

[0010] S4: Based on the trend description, a set of candidate anomaly regions is obtained, and the core location and core temperature difference are determined on the temperature difference matrix sequence. The core temperature difference generates an isothermal boundary sequence, and the consistency of the outward expansion direction and the continuity of the boundary gradient are calculated to obtain evidence of diffusion consistency. After the candidate anomaly region set is associated across frames through the overlap, the candidate anomaly region trajectory and diffusion consistency evidence are output.

[0011] S5: The reference area generates the temperature difference baseline bandwidth and the trend baseline bandwidth. The candidate anomaly area trajectory generates the diffusion consistency evidence baseline bandwidth and constructs the discriminant. The discriminant is used to generate the early warning criterion through continuous gating and multi-condition consistency rules.

[0012] Furthermore, step S1 includes:

[0013] A fixed-mount infrared thermal imager continuously acquires infrared thermal image frames within the field of view and writes the acquisition time for each frame. If the acquisition time is reversed or repeated, the infrared thermal image frame is discarded. The pixel response value of the infrared thermal image frame is converted into a temperature matrix according to the radiometric calibration relationship. The invalid pixel positions marked in the bad pixel table are filled with the median of the eight-connected neighborhood. The temperature matrix is ​​sorted according to the acquisition time to form a temperature matrix sequence.

[0014] Furthermore, step S2 includes:

[0015] The temperature matrix sequence is truncated into acquisition windows in ascending order of acquisition time to form temperature matrix subsequences and acquisition time subsequences. Based on the temperature matrix subsequences, the fluctuation matrix and temperature deviation matrix are calculated. A set of reference candidate pixels is constructed according to the median limit of the fluctuation and the median limit of the temperature deviation, and the index position of the reference seed pixel is determined.

[0016] Furthermore, step S2 also includes:

[0017] A four-connected breadth-first search is performed within the reference candidate pixel set using the reference seed pixel index position to obtain the reference region. The median temperature of the reference region is taken frame by frame to form a reference temperature sequence and generate a reference temperature field sequence. The reference temperature field sequence, the temperature matrix subsequence, and the acquisition time subsequence are bound and output according to the frame index.

[0018] Furthermore, step S3 includes:

[0019] The temperature matrix subsequence is aligned with the reference temperature field sequence by frame index and the acquisition time subsequence is checked to be consistent. Adjacent items in the acquisition time subsequence are kept increasing. The temperature difference matrix sequence is formed by the difference of pixel index. The temperature difference matrix sequence is used to calculate the temperature difference change rate matrix sequence between adjacent frames based on the acquisition time difference, and the positive temperature difference change rate is converted into a positive increment matrix sequence based on the acquisition time difference.

[0020] Furthermore, step S3 also includes:

[0021] The positive increment matrix sequence is accumulated by frame index within the acquisition window to generate a temperature rise trend matrix sequence. When the positive increment is zero, the temperature rise trend is reset and the persistence matrix sequence is reset simultaneously. The reference region calculates the reference trend level and constructs a set of trend-significant pixels. The set of trend-significant pixels is then labeled with four-connectivity to obtain a set of trend-significant connected regions, which are then summarized and written into the trend description.

[0022] Furthermore, step S4 includes:

[0023] The set of significantly connected regions in the trend description is mapped to a set of candidate abnormal regions by frame index. The core location and core temperature difference are determined within the range of the candidate abnormal regions in the temperature difference matrix sequence. The temperature difference matrix sequence, the temperature rise trend matrix sequence, and the subsequence of the acquisition time are kept consistent with the frame index and a consistency check is completed.

[0024] Furthermore, step S4 also includes:

[0025] The candidate anomaly region set is constructed by extracting the temperature difference level based on the core temperature difference to form an isothermal boundary sequence. The isothermal boundary sequence extracts the consistency of the outward expansion direction and the continuity of the boundary gradient to form evidence of diffusion consistency. The candidate anomaly region set is used to complete cross-frame association with the overlap degree and the distance of the core position to form the candidate anomaly region trajectory and output it.

[0026] Among them, the consistency of the outward expansion direction is obtained by the cosine of the angle between the outward expansion displacement vector and the radial vector. When the magnitude of any vector is zero, the consistency of the outward expansion direction is recorded as zero. The continuity of the boundary gradient is obtained by the relationship between the discrete gradient of the boundary pixel index position and the radial direction. The discrete gradient is composed of the temperature difference difference between adjacent pixels in the row direction and the temperature difference difference between adjacent pixels in the column direction. When the boundary pixel is located at the outer edge of the matrix, resulting in insufficient neighborhood, the one-sided difference is used instead. The median value of the consistency of the outward expansion direction and the continuity of the boundary gradient in the temperature difference level dimension is taken to form the diffusion consistency evidence of the candidate anomaly region in the current frame.

[0027] Furthermore, step S5 includes:

[0028] The temperature difference matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence, trend description, acquisition time subsequence, and candidate abnormal area trajectory are checked for frame index consistency. The regional temperature rise trend summary value and regional persistence summary value in the trend description are mapped to the candidate abnormal area trajectory according to the overlap of the pixel index position set. Candidate abnormal areas that fail to be mapped are removed from the candidate abnormal area trajectory.

[0029] Furthermore, step S5 also includes:

[0030] The reference region generates the upper bound of the temperature difference baseline bandwidth and the upper bound of the trend baseline bandwidth in the temperature difference matrix sequence and the temperature rise trend matrix sequence. The candidate anomaly region trajectory generates the upper bound of the evidence baseline bandwidth in the diffusion consistency evidence and completes the domain truncation. The discriminant quantity, combined with the continuous gating output, focuses on warning, warning and emergency warning, and outputs the collection time, warning level, core location and candidate anomaly region pixel index location set.

[0031] The technical effects and advantages of the thermal runaway early warning analysis method for battery infrared images of this invention are as follows:

[0032] This invention employs a holistic, collaborative approach—establishing a baseline using a reference region, stripping away the background through temperature difference, continuously characterizing trends, expanding morphological verification, and triggering early warnings through baseline discrimination—to fundamentally isolate the overall temperature drift and sporadic bright spots in infrared thermal imaging that most easily interfere with early warning judgments. This eliminates the reliance on fixed thresholds or single-frame extreme values ​​for early warning assessments, instead identifying early evolutionary characteristics of local thermal anomalies based on a self-consistently generated background baseline within the same field of view. The reference region maintains spatial consistency within the acquisition window, allowing the background baseline to adjust synchronously with changes in operating conditions. Temperature difference and trends emphasize continuous accumulation rather than instantaneous fluctuations over time, making it difficult for short-term reflections or local disturbances to form stable evidence of anomalies. Isothermal boundary sequences and diffusion consistency evidence further verify, from a spatial morphological perspective, whether the anomaly possesses continuous characteristics of diffusion from a point to a surface, thus making the early warning more closely aligned with the actual evolutionary process of thermal runaway precursors.

[0033] In the application scenarios of power battery packs and energy storage battery cabinets, the above-mentioned collaborative mechanism enables the early warning output to be both timely and stable: in the face of common interferences such as air-cooling switching, cabin temperature drift, and light reflection, the early warning criteria can automatically converge around the baseline to avoid misjudging background fluctuations as anomalies; in the face of the process of local thermal anomalies from slight temperature rise to gradual spread, the method can continuously track and gradually improve the early warning level, while outputting the location of candidate anomaly areas to facilitate subsequent linkage and verification, thereby improving the interpretability and on-site usability of the early warning results. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for early warning analysis of thermal runaway using infrared images of a battery according to the present invention. Detailed Implementation

[0035] 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.

[0036] Please see Figure 1 This invention provides a method for early warning analysis of thermal runaway using infrared images of batteries, comprising:

[0037] S1: Acquire continuous infrared thermal image frames within the field of view of a fixedly installed infrared thermal imager and record the acquisition time of each frame. Based on the radiometric calibration relationship of the thermal imager, convert the pixel response values ​​into temperature matrices and form a temperature matrix sequence according to the acquisition time order.

[0038] S2: Within the acquisition window, calculate the fluctuation matrix and temperature deviation matrix based on the temperature matrix sequence and acquisition time. Construct a set of reference candidate pixels based on the median of the fluctuation matrix and the median of the temperature deviation matrix, and determine the reference region under the four-connectivity constraint. The reference region is used to generate reference temperature frame by frame and construct a reference temperature field sequence.

[0039] S3: The temperature matrix subsequence and the reference temperature field sequence are differentially divided according to the frame index to form a temperature difference matrix sequence. Under the constraint of the acquisition time, the temperature difference matrix sequence generates a temperature difference change rate matrix sequence and is converted into a positive increment matrix sequence. The positive increment matrix sequence generates a temperature rise trend matrix sequence and a persistence matrix sequence and combines them with the reference area to generate a trend description.

[0040] S4: Based on the trend description, a set of candidate anomaly regions is obtained, and the core location and core temperature difference are determined on the temperature difference matrix sequence. The core temperature difference generates an isothermal boundary sequence, and the consistency of the outward expansion direction and the continuity of the boundary gradient are calculated to obtain evidence of diffusion consistency. After the candidate anomaly region set is associated across frames through the overlap, the candidate anomaly region trajectory and diffusion consistency evidence are output.

[0041] S5: The reference area generates the temperature difference baseline bandwidth and the trend baseline bandwidth. The candidate anomaly area trajectory generates the diffusion consistency evidence baseline bandwidth and constructs the discriminant. The discriminant is used to generate the early warning criterion through continuous gating and multi-condition consistency rules.

[0042] This invention addresses infrared thermal imaging monitoring of power battery packs and energy storage battery cabinets, proposing a thermal runaway early warning analysis method based on infrared images. The core idea is to automatically extract the background temperature benchmark from the thermal image sequence without relying on fixed thresholds and external sensors, and identify the abnormal evolution process of continuous accumulation of local temperature difference with outward expansion on this benchmark, thereby achieving earlier and more stable early warning output.

[0043] The method first converts continuous infrared thermal images into a temperature matrix sequence organized by pixel index, ensuring that the temperature of each frame strictly corresponds to the acquisition time. Then, within the acquisition window, aiming for stable temperature changes, connectivity constraints are introduced to select a spatially continuous reference region with minimal temporal fluctuations within the field of view, forming a reference temperature field sequence frame by frame. This is used to eliminate overall temperature rise or fall caused by factors such as air-cooling switching and cabin temperature drift. Based on the reference temperature field sequence, the temperature matrix of each frame is differentially analyzed to obtain a temperature difference matrix sequence. Then, in the time dimension, only a positively accumulated and persistent temperature rise trend matrix is ​​extracted and summarized to form a trend description, distinguishing between short-term reflective bright spots or random fluctuations and genuine continuous temperature rise.

[0044] Building upon this foundation, this invention further organizes regions with significant trends into a set of candidate anomaly regions. It utilizes a temperature difference field to determine the core location and core temperature difference, and generates a sequence of isothermal boundaries extending outwards from the core layer by layer, based on the core temperature difference. By examining the consistency of the outward movement of the boundaries over time and the continuity of the temperature gradient at the boundaries, it constructs evidence of diffusion consistency to characterize the true evolutionary features of thermal anomalies spreading from points to areas. Finally, this invention uses the temperature difference and trend baseline bandwidth obtained from the reference region statistics as a background benchmark, and combines the diffusion consistency evidence and trend description of the candidate anomaly regions to form a multi-condition consistency early warning criterion. It outputs early warning results aligned with the data acquisition time, while also providing the location and early warning level of the candidate anomaly regions, facilitating subsequent handling and coordinated action.

[0045] Power battery packs and energy storage battery cabinets commonly use fixed-installation infrared thermal imagers for online monitoring. Subsequent steps require calculating the temporal evolution of temperature distribution within the same field of view. This requires that each frame of thermal image corresponds to a specific acquisition time, and that each pixel within each frame can be converted into a temperature value with the same physical meaning. However, the raw output of infrared thermal images is usually pixel response values ​​or radiometric values, which are easily affected by calibration parameters, saturated pixels, and dead pixels when used directly for time-series analysis. Therefore, step S1 focuses on acquiring thermal images, aligning time, and constructing a temperature matrix sequence with the minimum acquisition amount, providing a temperature matrix sequence for step S2 that can directly calculate temperature fluctuations.

[0046] The specific implementation method of step S1 is as follows:

[0047] S101: The acquisition window is created and bound to the acquisition time.

[0048] When a fixed-installation infrared thermal imager is used for monitoring power battery packs or energy storage battery cabinets, the temporal evolution of temperature distribution needs to rely on a stable time axis; otherwise, the subsequent calculation of temperature fluctuations and temperature rise trends will treat uneven sampling as thermal changes.

[0049] The acquisition window is started using a continuous frame acquisition method, with infrared thermal image frames entering the acquisition window in arrival order. Simultaneously with each infrared thermal image frame entering the acquisition window, the acquisition time is obtained from the frame timestamp output by the thermal imager. The frame timestamp comes from either the timestamp field in the thermal imager's communication protocol message or the timestamp field in the thermal imager's output file header; either entry point can be selected. The acquisition time record is bound one-to-one with each infrared thermal image frame and written into the acquisition window. The range of acquisition time values ​​is limited to the range that a monotonically increasing timestamp can represent; adjacent acquisition times are not allowed to be equal or in reverse order. If equal or in reverse order occurs, the corresponding infrared thermal image frame is removed from the acquisition window, and acquisition continues.

[0050] Example: An infrared thermal imager is installed inside the door of the energy storage battery cabinet. After the acquisition window is started, infrared thermal image frames arrive continuously. The acquisition time is directly read from the timestamp field of each frame message. If a frame timestamp is found to be the same as the previous frame, the infrared thermal image frame does not enter the acquisition window, and the next frame continues to be bound to the acquisition time normally.

[0051] S102: Pixel response value analysis and valid pixel determination.

[0052] Infrared thermal images typically contain a pixel response value plane, which may contain saturated pixels, bad pixels, or missing pixels. If the temperature is directly retrieved, non-physical temperatures or inversion may occur, and subsequent matrix operations will also be disrupted.

[0053] Each infrared thermal image frame within the acquisition window is parsed to obtain a pixel response value plane. The pixel response values, as defined by the thermal imager's output format, fall between the minimum and maximum output response values. Valid pixel determination employs a combined two-path approach: the first path reads the pixel state plane output by the thermal imager, which is presented in the output as bad pixel or saturation markers. Locations marked as bad pixels or saturated are determined as invalid pixel locations. The second path, when a pixel state plane is missing, uses the bad pixel table output by the thermal imager. This bad pixel table is derived from factory calibration or maintenance and written to a file; locations recorded in the bad pixel table are determined as invalid pixel locations. The invalid pixel locations obtained from the two paths are combined to form a set of invalid pixel locations.

[0054] The range of values ​​for the invalid cell location set is limited to the row and column positions within the range that the cell index can represent. Positions within the set are not repeated, and positions outside the set are considered valid cell locations.

[0055] S103: Acquisition of radiation calibration relationship parameters and consistency constraints within the window.

[0056] The same model of thermal imager may switch calibration parameters under different temperature measurement ranges or different output modes. If different calibration parameters are mixed within the acquisition window for temperature inversion, the parameter switching will be regarded as a temperature change, and the temperature matrix sequence will lose comparability.

[0057] The set of radiometric calibration parameters is obtained from the metadata output by the thermal imager. The metadata source is limited to one of two types: calibration fields in the thermal imager communication protocol message or calibration fields in the header of the infrared thermal image frame. The set of radiometric calibration parameters includes calibration scale coefficients, calibration temperature scale coefficients, logarithmic correction constants, and calibration bias coefficients; all four parameters are provided by the thermal imager calibration. Within the acquisition window, a set of radiometric calibration parameters is extracted for each frame, and the parameter sets of adjacent frames are compared. When a change in the parameter set occurs, the acquisition window is split according to the position of the change to ensure that infrared thermal image frames entering the same temperature matrix sequence use the same set of radiometric calibration parameters.

[0058] The range of values ​​for the radiation calibration relationship parameter set is limited to the range that the thermal imager calibration field can express, and the subsequent temperature inversion logarithmic input must be positive. Frames that do not meet the requirements are removed from the acquisition window.

[0059] S104: Temperature inversion calculation and invalid pixel repair.

[0060] The temperature matrix needs to be fully covered in the cell index dimension; otherwise, the subsequent calculation of cell-level temperature fluctuations cannot be uniformly traversed. At the same time, temperature inversion must strictly follow the scaling inversion order to avoid inversion failure due to non-positive logarithmic input.

[0061] For each infrared thermal image frame within the acquisition window, temperature inversion calculation is performed for each valid pixel location, with a fixed calculation sequence of four steps. First, the pixel response value is subtracted from the calibration bias coefficient to obtain the bias correction response value. Second, the calibration scaling factor is divided by the bias correction response value to obtain the scaling inversion term. Third, the scaling inversion term is added to the logarithmic correction constant to obtain the logarithmic input term, which must be positive. Fourth, the natural logarithm of the logarithmic input term is taken, and the calibration temperature scaling factor is divided by the natural logarithm to obtain the temperature value.

[0062] Positions within the invalid pixel location set are not subjected to inversion calculations; instead, repair calculations are performed. The repair algorithm uses 8-connected neighborhood median filling. An 8-connected neighborhood is defined as the set of adjacent row and column positions around the invalid pixel location, including the positions above, below, left, right, and four diagonal positions. The repair calculation sequence is fixed in three steps: First, valid pixel locations are selected from the 8-connected neighborhood to obtain a set of valid neighborhood temperature values; second, the set of valid neighborhood temperature values ​​is sorted by size; third, the temperature value located at the middle position after sorting is taken as the repair temperature value and written to the invalid pixel location. If the set of valid neighborhood temperature values ​​is empty, a larger neighborhood is used to continue searching until at least one valid temperature value is obtained.

[0063] Example: An infrared thermal imager is fixedly installed inside the battery pack housing of a vehicle. When a saturated pixel appears in a frame within the acquisition window, the saturated pixel position enters the invalid pixel position set. Temperature inversion skips the saturated pixel position and instead reads the temperature value in the eight connected neighborhood around the saturated pixel. After sorting, the middle temperature value is taken and filled into the saturated pixel position to obtain a continuous temperature matrix without breaks.

[0064] S105: Temperature matrix sequence organization and output.

[0065] Step S2 requires directly traversing the temperature matrix sequence to calculate the temperature fluctuation. If the temperature matrix and the acquisition time are not bound in a unified structure, misalignment will occur during subsequent traversal.

[0066] The infrared thermal images within the acquisition window are sorted in ascending order of acquisition time, and each frame forms a record. The record consists of the acquisition time and the temperature matrix of the same frame. The temperature matrix sequence is output in the order of the records. The value range of the temperature matrix sequence is limited to the effective temperature range of the thermal imager's temperature measurement calibration. Positions that exceed the effective temperature range have been checked by the domain definition and invalid pixel repair path in step S104, and are not allowed to enter the temperature matrix sequence in an uninverted or unrepaired state.

[0067] Step S1 forms a temperature matrix sequence within the field of view of the fixed-installation infrared thermal imager, arranged in strict ascending order according to the acquisition time. Each frame of the temperature matrix sequence is obtained by inversion from the radiometric calibration relationship and invalid pixel repair is completed. The temperature matrix sequence and the acquisition time record together constitute the sole input basis for step S2 to measure pixel stability within the acquisition window and extract the reference area.

[0068] During the operation of the power battery pack and energy storage battery cabinet, the switching of air cooling strategy and the temperature drift of the chamber will cause the temperature matrix sequence to shift upward or downward as a whole. Directly entering the temperature difference matrix calculation will mistake the overall drift as a local abnormal evolution. The reference temperature field sequence needs to be obtained from the temperature matrix sequence and maintain spatial stability. However, there are local heating, bright reflection spots and bad pixel repair residues within the temperature matrix sequence. The extraction of the reference region must simultaneously constrain the amount of temperature fluctuation, temperature level deviation and pixel connectivity to ensure that the reference temperature field sequence has consistent temperature dimensions and consistent time alignment.

[0069] The specific implementation method of step S2 is as follows:

[0070] S201: Capture of the acquisition window and confirmation of input.

[0071] During the monitoring of power battery packs and energy storage battery cabinets, the temperature matrix sequence covers the entire operation process. The extraction of the reference area needs to be completed within a limited time span; otherwise, the reference area will be affected by the switching of operating conditions.

[0072] A data acquisition window is used to extract frames from the temperature matrix sequence while maintaining frame continuity. The acquisition window extracts consecutive frames from the temperature matrix sequence in ascending order of acquisition time. The starting position of the acquisition window is determined by the early warning analysis trigger point, which is limited to the starting frame index provided by an external call. The ending position of the acquisition window is determined by a fixed time span, which is given by a preset configuration and remains unchanged. The ending position is determined by the frame index when the accumulated acquisition time reaches the fixed time span. Any adjacent acquisition times within the acquisition window maintain strict ascending order. If a frame is reversed or repeated, the corresponding frame is removed from the acquisition window, and the process continues to fill in the gaps. The output objects are the temperature matrix subsequence and the acquisition time subsequence corresponding to the acquisition window.

[0073] S202: Calculation of temperature fluctuation and generation of fluctuation matrix.

[0074] The reference area needs to reflect stable temperature changes. Stable characteristics cannot be judged by the temperature level of a single frame; the temperature fluctuation needs to be obtained from the time series changes within the acquisition window.

[0075] The same set of operations is performed on each cell index position within the acquisition window.

[0076] The first step is to traverse the temperature matrix of two adjacent frames in the time sequence of the acquisition window and read the temperature value at the same pixel index position.

[0077] The second step is to calculate the difference between the temperature value of the next frame and the temperature value of the previous frame, and take the absolute value.

[0078] The third step is to read the difference between the acquisition time of the next frame and the acquisition time of the previous frame, and keep the difference positive.

[0079] The fourth step is to divide the absolute value of the temperature difference by the time difference between the acquisition times to obtain the absolute value of the temperature change rate between adjacent frames.

[0080] The fifth step involves summing the absolute values ​​of the temperature change rates of all adjacent frames within the acquisition window and dividing by the number of adjacent frames to obtain the temperature fluctuation. After the temperature fluctuation is calculated at all pixel index positions, a fluctuation matrix is ​​formed. The elements of the fluctuation matrix take non-negative values, and the fluctuation matrix and the temperature matrix maintain consistency in the pixel index dimension.

[0081] S203: Temperature deviation matrix generation and reference region determination.

[0082] The reference area needs to simultaneously satisfy the conditions of stable temperature change and temperature level close to the normal field of view. Relying solely on the minimum fluctuation will include fixed heat sources or fixed cold areas in the reference area.

[0083] The processing logic is divided into four actions: generating a temperature deviation matrix, constructing a set of reference candidate pixels, determining the index position of the reference seed pixel, and searching for four-connected components, while keeping the reference area unchanged within the acquisition window.

[0084] The temperature deviation matrix is ​​generated in the following order: for each pixel index position, all frame temperature values ​​within the acquisition window are collected and the median value is taken to obtain the pixel temporal median temperature; the median value of all pixel temporal median temperatures is taken to obtain the global median temperature; the absolute difference between the pixel temporal median temperature and the global median temperature is calculated for each pixel index position to obtain the temperature deviation matrix.

[0085] The reference candidate pixel set is constructed in the following order: the median of all elements of the fluctuation matrix is ​​taken to obtain the fluctuation median; the median of all elements of the temperature deviation matrix is ​​taken to obtain the temperature deviation median; and the pixel index positions that have fluctuations not exceeding the fluctuation median and temperature deviations not exceeding the temperature deviation median are collected as the reference candidate pixel set.

[0086] The reference seed cell index position is determined using a deterministic convergence rule: the cell index position with the smallest fluctuation is selected from the reference candidate cell set; when fluctuations are equal, the cell index position with the smallest temperature deviation is selected; when temperature deviations are equal, the first cell index position is selected in the order of row index priority followed by column index, thus obtaining the reference seed cell index position.

[0087] The four-connected connected component search uses a breadth-first search: starting from the reference seed cell index position, the cell index positions in the four-connected neighborhood that belong to the reference candidate cell set and have not been visited are added to the access queue layer by layer until the access queue is empty, and the access set forms the reference region.

[0088] For example, the field of view of the energy storage battery cabinet includes both the surface of the battery module and the busbar area. The busbar area shows a significant deviation in the temperature deviation matrix. The reference candidate pixel set excludes the busbar area. The reference area obtained by breadth-first search falls on the continuous insulating plate area at the edge of the battery module. The reference area remains unchanged within the acquisition window.

[0089] S204: Reference temperature calculation and reference temperature field sequence generation.

[0090] The reference temperature needs to reflect the temperature level of the reference area in each frame and remain robust to a small number of abnormal pixels. The reference temperature uses the median temperature of the reference area, and the reference temperature field uses a constant matrix to satisfy the pixel-by-pixel differential input form.

[0091] For each frame of the temperature matrix within the acquisition window, extract all pixel temperature values ​​from the reference area and take the median to obtain the reference temperature; write the reference temperature into a reference temperature field of the same size as the temperature matrix, and take the same reference temperature for each pixel index position in the reference temperature field; arrange all reference temperature fields in the time order of the acquisition window to obtain a reference temperature field sequence, and form a reference temperature sequence in the same order; keep the reference temperature and the reference temperature field element value range within the effective temperature range of the infrared thermal imager temperature measurement calibration.

[0092] S205: Summary of Output Packaging and Processing.

[0093] The calculation of the temperature difference matrix requires that the temperature matrix subsequence and the reference temperature field sequence be strictly aligned at the same acquisition time, and the output object needs to carry the reference region for verification of the reference temperature source.

[0094] The temperature matrix subsequence, reference temperature field sequence, reference temperature sequence, reference region, and acquisition time subsequence are bound and output one by one according to the frame index. The frame index alignment rule is fixed so that the same frame index corresponds to the same acquisition time, the same temperature matrix, the same reference temperature field, and the same reference temperature.

[0095] Step S2 calculates the fluctuation matrix and temperature deviation matrix within the acquisition window, constructs a set of reference candidate pixels using double median constraints, and determines the reference region using a four-connected breadth-first search. The reference region is used to generate reference temperatures frame by frame and construct a reference temperature field sequence. The output object meets the input requirements of the alignment difference in step S3.

[0096] The infrared monitoring field of view of the power battery pack and energy storage battery cabinet includes superimposed signals of overall temperature drift and local thermal anomalies. Step S2 has already determined the reference area and formed a reference temperature field sequence within the acquisition window. The reference temperature field sequence is used to separate the overall temperature drift from the temperature matrix subsequence. However, local thermal anomalies often manifest as the continuous accumulation of temperature difference structure over time rather than a single-frame extreme value change in the early stage. Step S3 needs to align and differ the temperature matrix subsequence with the reference temperature field sequence within the acquisition window to obtain the temperature difference matrix sequence, and perform time dimension processing on the temperature difference matrix sequence to obtain the temperature rise trend matrix and trend description, which are used by step S4 to directly extract the candidate anomaly region set.

[0097] The specific implementation method of step S3 is as follows:

[0098] S301: Temperature difference matrix sequence generation and alignment check.

[0099] The temperature difference matrix sequence is used to separate the overall temperature drift effect from the temperature matrix subsequence. The temperature difference matrix sequence must satisfy the requirements of frame index consistency and cell index consistency.

[0100] Read the temperature matrix subsequence, reference temperature field sequence, and acquisition time subsequence output in step S2. Check that the number of frames in the temperature matrix subsequence and the reference temperature field sequence are consistent, that the acquisition time corresponding to the same frame index is consistent, and that the temperature matrix corresponding to the same frame index and the reference temperature field have the same cell index size. After the alignment check passes, perform a difference operation according to the frame index and cell index. The difference operation uses the temperature value at the same cell index position to subtract the temperature value of the reference temperature field at the same cell index position to obtain the temperature difference matrix sequence. The value range of the elements in the temperature difference matrix sequence is allowed to be positive or negative, and the value range is constrained by the effective temperature range of the temperature measurement calibration of the temperature matrix subsequence and the reference temperature field sequence.

[0101] S302: Generation of positive increment matrix sequence.

[0102] The temperature rise trend needs to highlight the information of continuous temperature rise. Positive increments are used to retain the increase in temperature difference between adjacent frames as an accumulative amount, so as to avoid diluting the evidence of continuous temperature rise during the fall process.

[0103] The temperature difference matrix of adjacent frames is traversed in the order of the acquisition time subsequence. The temperature difference value of the next frame and the temperature difference value of the previous frame are read pixel by pixel and the temperature difference value is calculated. At the same time, the acquisition time of the next frame and the acquisition time of the previous frame are read to obtain the acquisition time interval. The temperature difference difference is divided by the acquisition time interval to obtain the temperature difference change rate and written into the temperature difference change rate matrix sequence. When the temperature difference change rate is positive, the temperature difference change rate is multiplied by the acquisition time interval to obtain the positive increment and written into the positive increment matrix sequence. When the temperature difference change rate is zero or negative, zero is written into the positive increment matrix sequence. If the acquisition time interval is found to be not positive, the adjacent frame pair is removed from the temperature difference change rate matrix sequence and the positive increment matrix sequence and the next pair of adjacent frames is used to continue generation. The dimensions of the elements of the temperature difference change rate matrix sequence are maintained as the ratio of temperature units to time units, and the dimensions of the elements of the positive increment matrix sequence are maintained as temperature units.

[0104] S303: Generation of temperature rise trend matrix sequence and persistence matrix sequence.

[0105] The temperature rise trend matrix sequence is used to characterize the cumulative intensity of continuous temperature rise, and the persistence matrix sequence is used to characterize the number of frames of continuous temperature rise. The two constraints together can distinguish the short fluctuations of the reflected bright spot from the slow and continuous temperature rise of the cell surface.

[0106] The temperature rise trend matrix and the persistence matrix corresponding to the first frame of the acquisition window are initialized to zero. The temperature rise trend matrix and persistence matrix are updated in frame index order, with a fixed two-branch update rule:

[0107] When the positive increment matrix sequence has a positive value at the corresponding cell index position, the temperature rise trend matrix writes the cumulative result of the temperature rise trend value and the positive increment value of the previous frame at the corresponding cell index position, and the persistence matrix writes the result of the persistence count of the previous frame plus one at the corresponding cell index position; when the positive increment matrix sequence has a zero value at the corresponding cell index position, the temperature rise trend matrix writes zero at the corresponding cell index position, and the persistence matrix writes zero at the corresponding cell index position; the value range of the elements of the temperature rise trend matrix sequence is non-negative, the value range of the elements of the persistence matrix sequence is non-negative integer, and the number of elements of the persistence matrix sequence does not exceed the number of frames in the acquisition window.

[0108] The example illustrates that in the thermal imaging field of the energy storage battery cabinet, reflected bright spots appear. The location of the reflected bright spots shows a positive temperature difference value followed by a negative value in adjacent frames. The persistence matrix frequently returns to zero at the location of the reflected bright spots, and the temperature rise trend matrix is ​​difficult to accumulate continuously at the location of the reflected bright spots. In contrast, the localized heating area on the surface of the battery cell shows a continuous positive temperature difference value in adjacent frames, and the persistence matrix continuously increases. The temperature rise trend matrix continuously accumulates to form a distinct area.

[0109] S304: Trend description generation.

[0110] The trend description is used to convert the pixel-level temperature rise trend matrix sequence into regional structured information. The trend description needs to include a set of significantly trend-significant connected regions and a summary value of the connected regions, so that step S4 can extract a set of candidate abnormal regions and perform isothermal boundary sequence construction.

[0111] Traverse the temperature rise trend matrix sequence by frame index. The reference region, determined in step S2, remains unchanged within the acquisition window. Extract all temperature rise trend values ​​within the reference region from each frame's temperature rise trend matrix and calculate the median to obtain the reference trend level. Collect the cell index positions where the temperature rise trend value is greater than the reference trend level into a set of trend-significant cells. If the set of trend-significant cells is empty, write an empty set of trend-significant connected regions in the corresponding frame for trend description. If the set of trend-significant cells is not empty, use a four-connectivity breadth-first search to label connected components, obtaining a set of trend-significant connected regions. For each connected region in the set of trend-significant connected regions, extract the set of temperature rise trend values ​​and the set of persistence counts within the connected region, calculate the median for each, and obtain the regional temperature rise trend summary value and the regional persistence summary value. Write the set of connected region cell index positions, the regional temperature rise trend summary value, and the regional persistence summary value into the trend description and store them by frame index. The reference trend level, the regional temperature rise trend summary value, and the regional persistence summary value are all non-negative integers.

[0112] S305: Summary of Output Packaging and Processing.

[0113] Step S4 requires reading the temperature difference matrix sequence, the temperature rise trend matrix sequence, and the trend description simultaneously. The output encapsulation must maintain frame index alignment and be consistent with the cell index.

[0114] The temperature difference matrix sequence, positive increment matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence, trend description, and acquisition time subsequence are bound and output one by one according to the frame index. The binding rule is fixed so that the same frame index corresponds to the same acquisition time, the same temperature difference matrix, the same temperature rise trend matrix, the same persistence matrix, and the same trend description.

[0115] Step S3 involves frame-by-frame difference analysis of the temperature matrix subsequence and the reference temperature field sequence within the acquisition window to obtain the temperature difference matrix sequence. The temperature difference matrix sequence calculates the temperature difference change rate matrix sequence based on the acquisition time subsequence and converts the positive temperature difference change rate into a positive increment matrix sequence. The positive increment matrix sequence generates a temperature rise trend matrix sequence and a persistence matrix sequence through continuous accumulation and reset rules. The reference area is used to construct a reference trend level and generate a trend description. The output object satisfies the direct call requirement of step S4 in both the frame index dimension and the cell index dimension.

[0116] In the infrared monitoring scenario of power battery packs and energy storage battery cabinets, step S3 has already output the temperature difference matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence and trend description. The trend description provides a set of significantly connected regions and a region summary value. However, the set of significantly connected regions may still contain reflective bright spots and short-term areas caused by local thermal disturbances of connectors. Step S4 needs to convert the set of significantly connected regions into a set of candidate abnormal regions, construct an isothermal boundary sequence on the temperature difference matrix sequence, and then extract diffusion consistency evidence from the boundary evolution for step S5 to generate early warning criteria.

[0117] The specific implementation method of step S4 is as follows:

[0118] S401: Generation and alignment verification of candidate abnormal region sets.

[0119] The temperature rise trend matrix sequence and trend description have separated the continuously rising areas from the recurring fluctuations. The candidate abnormal region set needs to be consistent with the temperature difference matrix sequence within the same frame index; otherwise, the core location will fall into the wrong frame temperature difference field.

[0120] Read the temperature difference matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence, trend description, and acquisition time subsequence output in step S3, and perform three consistency checks: check for consistent frame count, check for consistent acquisition times corresponding to the same frame index, and check for consistent matrix size corresponding to the same frame index. After the consistency checks pass, the set of significantly connected regions of the trend in each frame in the trend description is directly used as the set of candidate abnormal regions for the same frame. Any candidate abnormal region in the set of candidate abnormal regions is represented by a set of pixel index positions. The pixel index positions are limited to the range of pixel indexes in the temperature difference matrix sequence, and the candidate abnormal regions do not share pixel index positions.

[0121] S402: Determination of core location and core temperature difference.

[0122] The isothermal boundary sequence needs to be observed from the peak of the temperature difference field outwards to determine its expansion pattern. Fixing the core position inside the candidate anomaly region can prevent external heat sources from pulling the boundary off course.

[0123] For each candidate anomaly region in the candidate anomaly region set of each frame, traverse all the cell index positions contained in the candidate anomaly region within the temperature difference matrix of the corresponding frame in the temperature difference matrix sequence, and record the cell index position with the largest temperature difference as the core position. The core temperature difference is defined as the temperature difference value corresponding to the core position. When the maximum temperature difference values ​​are tied, select the first cell index position as the core position in the order of row index priority followed by column index. The core position and core temperature difference are written into the attribute field of the candidate anomaly region. The value range of the core temperature difference is constrained by the value range of the temperature difference matrix sequence.

[0124] S403: Construction of isothermal boundary sequences.

[0125] Consistent evidence of diffusion requires observing the boundary expansion process at multiple temperature difference levels. The selection of temperature difference levels needs to balance determinism and computational burden, avoiding the use of all temperature difference values ​​within the candidate anomaly region for boundary construction.

[0126] For each candidate anomaly region in each frame, temperature difference values ​​are first collected and sorted within the candidate anomaly region. After sorting, temperature difference levels are uniformly extracted according to a preset number of layers. The preset number of layers is derived from a preset configuration and is an integer not less than two. The temperature difference level includes the core temperature difference and covers the temperature difference range within the candidate anomaly region. Each temperature difference level corresponds to a core connected hyperlevel set, which consists of pixel index positions that satisfy the condition that the temperature difference is not lower than the temperature difference level and that are 4-connected to the core location. 4-connectivity is achieved through breadth-first search. Each temperature difference level corresponds to an isothermal boundary, which consists of pixel index positions in the core connected hyperlevel set that satisfy the condition that at least one pixel index position in the 4-connected neighborhood falls outside the set. The isothermal boundaries obtained for the same candidate anomaly region at each temperature difference level are sorted from high to low according to the temperature difference level to form a frame of isothermal boundary sequence. The isothermal boundary sequences of each frame within the acquisition window are stored according to the frame index to form an isothermal boundary sequence.

[0127] Example: In the field of view of the energy storage battery cabinet, the surface of the battery cell is locally heated. The core position is located in the center of the area. After the temperature difference level is extracted from high to low, the core connected super-level collection expands layer by layer as the temperature difference level decreases. The isothermal boundary moves outward layer by layer from the vicinity of the core, and the boundary shape remains continuous sheet-like.

[0128] S404: Calculation of evidence for diffusion consistency.

[0129] Reflected bright spots and local thermal disturbances of connectors often exhibit boundary jumps and outward drift. Diffusion consistency evidence solidifies the consistency of the outward diffusion direction and the continuity of the boundary gradient into computable fields for judgment in step S5.

[0130] For each isothermal boundary sequence of each candidate anomaly region in each frame, firstly, calculate the boundary center position for each isothermal boundary. The boundary center position is obtained by the mean of the row coordinates and the mean of the column coordinates of the boundary cell index position. When the boundary is an empty set, the boundary center position is recorded as the core position and marked as an empty boundary. Then, calculate the outward displacement vector for the boundary center positions at the same temperature difference level in adjacent frames. The outward displacement vector is obtained by subtracting the boundary center position of the previous frame from the boundary center position of the next frame. Next, construct a radial vector for the same temperature difference level in the same frame. The radial vector is obtained by the reverse vector pointing from the boundary center position to the core position. The consistency of the outward expansion direction is obtained by the cosine of the angle between the outward displacement vector and the radial vector. When the magnitude of any vector is zero, the consistency of the outward expansion direction is recorded as zero. Boundary gradient connection The continuity is obtained by relating the discrete gradient at the boundary cell index position to the radial direction. The discrete gradient is constructed using the temperature difference difference between adjacent cells in the row direction and the temperature difference difference between adjacent cells in the column direction. When the boundary cell is located at the outer edge of the matrix, resulting in insufficient neighborhood, a one-sided difference is used instead. The radial direction is obtained by pointing from the core position to the boundary cell index position. If the boundary cell index position coincides with the core position, that cell index position is skipped. The radial gradient sign at the boundary cell index position is statistically represented as a proportional quantity. The more concentrated the proportional quantity, the more stable the boundary gradient continuity. The median value of the consistency of the outward expansion direction and the continuity of the boundary gradient is taken in the temperature difference level dimension to form the diffusion consistency evidence of the candidate anomaly region in the current frame. The diffusion consistency evidence is stored with the frame index and bound to the candidate anomaly region.

[0131] S405: Cross-frame correlation and output encapsulation of candidate abnormal regions.

[0132] Candidate abnormal regions will experience shape drift within the acquisition window. Cross-frame association requires connecting candidate abnormal regions corresponding to the same physical heat source into a trajectory to avoid step S5 splitting the same heat source into multiple short segments.

[0133] The overlap degree is calculated for the candidate anomaly region sets of adjacent frames. The overlap degree is obtained by dividing the number of intersection cells of the two candidate anomaly region cell index position sets by the number of union cells. For each candidate anomaly region in each frame, the candidate anomaly region with the largest overlap degree is selected as the association target in the candidate anomaly region set of the next frame. When the overlap degrees are equal, the candidate anomaly region with the smallest core position distance is selected first. The distance is represented by the sum of the absolute values ​​of the row coordinate difference and the column coordinate difference. When the same candidate anomaly region in the next frame is selected by multiple candidate anomaly regions in the previous frame, the association relationship with the larger overlap degree is retained, and the trajectory of other candidate anomaly regions in the previous frame terminates in the current frame and a new trajectory is created. When all overlap degrees are zero, the candidate anomaly region with the smallest core position distance is selected as the association target. When the core position distances are equal, the row index takes precedence and the column index takes precedence. The association result forms the candidate anomaly region trajectory. The candidate anomaly region trajectory is bound to the candidate anomaly region attribute fields one by one and output. The output fields include the candidate anomaly region set, core position, core temperature difference, isothermal boundary sequence, diffusion consistency evidence, and acquisition time subsequence. The field naming maintains the same frame index alignment rule as the output of step S3.

[0134] Step S4 solidifies the set of significantly connected regions in the trend description into a set of candidate anomaly regions. Within the temperature difference matrix sequence, the core location and core temperature difference of each candidate anomaly region are determined. An isothermal boundary sequence is constructed using the extracted temperature difference levels. Evidence of diffusion consistency is formed based on the evolution of the boundary center location and the statistical analysis of the boundary discrete gradient. Cross-frame association is completed using the overlap and the distance to the core location, and the trajectory of the candidate anomaly region is output.

[0135] In infrared monitoring of power battery packs and energy storage battery cabinets, step S4 has already output the candidate abnormal area trajectory, core location, core temperature difference, and diffusion consistency evidence. Step S3 has already output the temperature difference matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence, and trend description. However, the same field of view has natural fluctuations in temperature difference background and trend background under different operating conditions. If the warning criterion relies on a fixed threshold, the warning trigger will be pulled by the difference in operating conditions. Step S5 forms a trend baseline around the reference area and maps the diffusion consistency evidence and trend description of the candidate abnormal area trajectory to the trend baseline to form an adaptive warning criterion and output the warning result.

[0136] The specific implementation method of step S5 is as follows:

[0137] S501: Input object alignment and mapping rules are determined.

[0138] Early warning criteria need to simultaneously cite temperature difference matrix sequence, temperature rise trend matrix sequence, trend description, candidate anomaly area trajectory and diffusion consistency evidence. Any frame misalignment will mix information from different times into a pseudo-anomaly.

[0139] Read the temperature difference matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence, trend description, and acquisition time subsequence output in step S3. Read the candidate anomaly area trajectory, core location, core temperature difference, and diffusion consistency evidence output in step S4. Verify that the number of frames in the temperature difference matrix sequence, temperature rise trend matrix sequence, and persistence matrix sequence are consistent. Verify that the length of the acquisition time subsequence is consistent with the number of frames. Verify that the frame index covered by the candidate anomaly area trajectory is within the frame index range of the acquisition window.

[0140] The mapping rules are as follows: The trend description in each frame includes the set of pixel index locations of candidate anomaly regions, along with the regional temperature rise trend summary value and the regional persistence summary value. The candidate anomaly region trajectory in the same frame includes the set of pixel index locations, core location, core temperature difference, and diffusion consistency evidence for the candidate anomaly region. Mapping is performed one-to-one matching based on the maximum overlap of the pixel index location sets; when overlap is equal, the minimum core location distance is used to determine the match. After a successful match, the regional temperature rise trend summary value and the regional persistence summary value from the trend description are written into the same candidate anomaly region entry in the candidate anomaly region trajectory. Candidate anomaly region entries that fail to match are removed from the candidate anomaly region trajectory.

[0141] Adjacent terms in the time-series data acquisition remain increasing; elements in the temperature difference matrix sequence and the temperature rise trend matrix sequence maintain the temperature dimension; elements in the persistence matrix sequence remain non-negative integers.

[0142] S502: Trend baseline generation.

[0143] The field of view of the power battery pack and energy storage battery cabinet exhibits overall temperature difference fluctuations and trend fluctuations. The baseline uses reference area statistics to compress background fluctuations into a bandwidth that changes over time, reducing the influence of operating condition switching on early warning criteria.

[0144] Using the reference area output in step S2, for each frame of the acquisition window, extract the temperature difference values ​​of the temperature difference matrix sequence in the reference area. First, take the absolute value of the temperature difference value to obtain the absolute temperature difference set, and then take the median value of the absolute temperature difference set to obtain the median value of the temperature difference baseline. Then, calculate the absolute difference between each absolute temperature difference and the median value of the temperature difference baseline for each absolute temperature difference set to form a temperature difference deviation set. Take the median value of the temperature difference deviation set to obtain the median value of the temperature difference baseline deviation. The upper bound of the temperature difference baseline bandwidth is synthesized by the median value of the temperature difference baseline and the median value of the temperature difference baseline deviation.

[0145] Within the same frame, a trend baseline is calculated for the reference region. The set of temperature rise trend values ​​in the reference region is extracted from the temperature rise trend matrix sequence. The median of the trend baseline is obtained by taking the midpoint of this set. Then, the absolute difference between each temperature rise trend value and the median of the trend baseline is calculated to form a trend deviation set. The median of the trend deviation set is obtained by taking the midpoint of this set. The upper bound of the trend baseline bandwidth is synthesized from the median of the trend baseline and the median of the trend baseline deviation.

[0146] The baseline median, baseline deviation, and upper bound of the baseline bandwidth are non-negative; the baseline median, baseline deviation, and upper bound of the trend bandwidth are also non-negative.

[0147] Example: The switching of the air-cooling strategy of the energy storage battery cabinet causes the overall temperature difference in the field of view to increase. The median of the absolute temperature difference set in the reference area increases synchronously. The upper bound of the temperature difference baseline bandwidth changes with the frame. The core temperature difference discrimination of the candidate abnormal area will not be directly triggered by the overall increase.

[0148] S503: Evidence baseline generation and discriminant quantity generation.

[0149] Evidence of diffusion consistency is a morphological statistic. Direct comparison with the upper bound of the temperature difference baseline bandwidth or the upper bound of the trend baseline bandwidth lacks dimensional consistency. The evidence distribution of the set of candidate anomaly region entries in the same frame forms the evidence baseline bandwidth, which can reflect the level of morphological noise.

[0150] For each frame of the acquisition window, extract the set of evidence for consistency in the outward expansion direction and the set of evidence for continuity of the boundary gradient from the trajectory of all candidate anomaly regions in the same frame. Take the median of the set of evidence for consistency in the outward expansion direction to obtain the median of the baseline for consistency in the outward expansion direction. Calculate the absolute difference between each evidence value in the set of evidence for consistency in the outward expansion direction and the median of the baseline for consistency in the outward expansion direction, forming a set of deviations from the baseline for consistency in the outward expansion direction. Take the median of the deviations from the set of deviations from the baseline for consistency in the outward expansion direction to obtain the median deviation of the baseline for consistency in the outward expansion direction. The upper bound of the bandwidth of the baseline for consistency in the outward expansion direction is synthesized from the median of the baseline for consistency in the outward expansion direction and the median deviation of the baseline for consistency in the outward expansion direction, and is truncated to the upper bound of the domain of consistency in the outward expansion direction.

[0151] Perform the same structural operation on the set of evidence for boundary gradient continuity to obtain the baseline median, the baseline deviation median, and the upper bound of the bandwidth of the boundary gradient continuity evidence, and truncate it to the upper bound of the domain of the boundary gradient continuity evidence.

[0152] When the set of candidate anomaly region entries in the same frame is empty, the upper bound of the baseline bandwidth of the outward expansion direction consistency evidence is taken as the upper bound of the domain of the outward expansion direction consistency evidence, and the upper bound of the baseline bandwidth of the boundary gradient continuity evidence is taken as the upper bound of the domain of the boundary gradient continuity evidence. The trend discriminant, core temperature difference discriminant, outward expansion direction consistency discriminant and boundary gradient continuity discriminant are all negative.

[0153] The discriminant generation rules are as follows: The trend discriminant is true when the regional temperature rise trend summary value is greater than the upper bound of the trend baseline bandwidth, and false otherwise; the core temperature difference discriminant is true when the absolute value of the core temperature difference is greater than the upper bound of the temperature difference baseline bandwidth, and false otherwise; the consistency discriminant of the outward expansion direction is true when the evidence of consistency in the outward expansion direction is greater than the upper bound of the baseline bandwidth of the evidence of consistency in the outward expansion direction, and false otherwise; the boundary gradient continuity discriminant is true when the evidence of boundary gradient continuity is greater than the upper bound of the baseline bandwidth of the evidence of boundary gradient continuity, and false otherwise.

[0154] Evidence of consistency in the outward expansion direction is located within the domain of evidence of consistency in the outward expansion direction; evidence of continuity of the boundary gradient is located within the domain of evidence of continuity of the boundary gradient; the discriminant value is either yes or no.

[0155] S504: Early warning criteria generation and early warning level determination.

[0156] The warning level needs to cover the cumulative trend, the consistency of the outward expansion pattern, and the intensity of the temperature difference. Single-condition triggering will introduce the reflected bright spot and the thermal disturbance of the connector into the warning output. The multi-condition consistency rule is more in line with the spatiotemporal evolution characteristics of the precursor to thermal runaway.

[0157] The persistence gate discriminant is generated from the region persistence summary value. The persistence gate discriminant is set to no when the region persistence summary value is zero, and to yes when the region persistence summary value is non-zero.

[0158] The attention warning criterion is jointly determined by the persistence gating discrimination quantity, the trend discrimination quantity, and the consistency discrimination quantity of the outward expansion direction. When all three are true, the attention warning criterion is true; otherwise, the attention warning criterion is false. The alert warning criterion is based on the attention warning criterion being true, plus the boundary gradient continuity discrimination quantity. When the boundary gradient continuity discrimination quantity is true, the alert warning criterion is true; otherwise, the alert warning criterion is false. The emergency warning criterion is based on the alert warning criterion being true, plus the core temperature difference discrimination quantity. When the core temperature difference discrimination quantity is true, the emergency warning criterion is true; otherwise, the emergency warning criterion is false.

[0159] Example: A bright reflective spot appears on the inner side of the power battery pack casing, the trajectory of the candidate abnormal region is repeatedly broken in adjacent frames, the region persistence summary value frequently returns to zero, the persistence gating discriminant is negative for a long time, and the attention warning criterion remains negative; the region persistence summary value of the local heating area on the cell surface remains non-zero, the trend discriminant and the consistency discriminant with the outward expansion direction are stable and positive, the attention warning criterion is positive, and after the boundary gradient continuity discriminant turns positive, the warning warning criterion is positive.

[0160] S505: Early warning result output and location mapping.

[0161] On-site handling requires mapping the warning level and thermal anomaly location to the field of view pixel index position, and the output structure needs to ensure traceability and frame index consistency.

[0162] In each frame of the acquisition window, all candidate anomaly regions within the same frame are traversed, and warning results are generated according to the priority order of emergency warning criteria, warning criteria, and attention warning criteria. The warning result fields include acquisition time, warning level, core location, set of candidate anomaly region pixel index locations, core temperature difference, evidence of consistency in outward expansion direction, and evidence of boundary gradient continuity. When multiple candidate anomaly regions meet the warning criteria in the same frame, multiple warning results are output while maintaining the candidate anomaly region indices unchanged.

[0163] The core location is within the cell index range; the elements of the candidate anomaly region cell index location set are within the cell index range; the warning level is one of three: attention warning, alert warning, or emergency warning.

[0164] Example: At the acquisition time corresponding to a certain frame in the acquisition window, the early warning generation stage begins. Multiple candidate anomaly region entries are provided for this frame. Each entry includes a set of candidate anomaly region pixel index locations, core location, core temperature difference, evidence of consistency in outward expansion direction, and evidence of boundary gradient continuity. Simultaneously, the judgment results for emergency warning criteria, warning criteria, and attention warning criteria are obtained in the same frame. When traversing all candidate anomaly region entries in this frame, the emergency warning criterion result for a specific entry is read first. If the emergency warning criterion is true, the warning level is written as an emergency warning, and the acquisition time, warning level, core location, set of candidate anomaly region pixel index locations, core temperature difference, evidence of consistency in outward expansion direction, and evidence of boundary gradient continuity are encapsulated into a single warning result. If the emergency warning criterion is false, the alert criterion results for the same entry are read. If the alert criterion is true, the warning level is written as "alert warning" and the warning result is encapsulated with the same field and added to the warning result set. If the alert criterion is false, the attention warning criterion results for the same entry are read. If the attention warning criterion is true, the warning level is written as "attention warning" and the warning result is encapsulated with the same field and added to the warning result set. If all three types of warning criteria are false, no warning result is generated. After traversal, the warning result set is sorted according to the priority of the warning level, with emergency warnings before alert warnings, and alert warnings before attention warnings. Within the same warning level, the order of the candidate abnormal area trajectory index remains unchanged. The sorted warning result set is output as the warning result for this frame.

[0165] Step S5 generates the upper bound of the temperature difference baseline bandwidth and the upper bound of the trend baseline bandwidth in the reference area. It also generates the upper bound of the baseline bandwidth for consistency evidence of outward expansion direction and the upper bound of the baseline bandwidth for continuity evidence of boundary gradient in the candidate anomaly area entry set in the same frame. The regional temperature rise trend summary value, regional persistence summary value, core temperature difference, and diffusion consistency evidence are mapped into binary discriminant values. The binary discriminant values ​​are used to generate an early warning level through persistence gating and multi-condition consistency rules. The early warning result outputs the acquisition time and the location elements of the candidate anomaly area.

[0166] Specifically, the above are merely preferred embodiments of this application and are not intended to limit this application.

[0167] Other preset parameters whose acquisition logic and value range are not explicitly explained in this invention can be pre-calibrated through offline simulation testing, or set as fixed values ​​according to on-site operating procedures.

[0168] In the description of this specification, references to terms such as "an embodiment," "an example," and "a specific example" indicate that a particular feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0169] 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 any specific implementation. 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.

Claims

1. A method for early warning analysis of thermal runaway using infrared images of a battery, characterized in that, Including the following steps: S1: Acquire continuous infrared thermal image frames within the field of view of a fixedly installed infrared thermal imager and record the acquisition time of each frame. Based on the radiometric calibration relationship of the thermal imager, convert the pixel response values ​​into temperature matrices and form a temperature matrix sequence according to the acquisition time order. S2: Within the acquisition window, for each pixel index position, traverse the temperature matrices of adjacent two frames in the time order of the acquisition window, calculate the absolute value of the difference between the temperature value of the later frame and the temperature value of the previous frame, and divide it by the corresponding acquisition time difference to obtain the absolute value of the temperature change rate of adjacent frames; sum up the absolute values ​​of the temperature change rates of all adjacent frames and divide by the logarithm of adjacent frames to obtain the temperature fluctuation of the pixel index position; the temperature fluctuations of all pixel index positions form a fluctuation matrix; calculate the temperature deviation matrix within the acquisition window, construct a reference candidate pixel set based on the median of the fluctuation matrix and the median of the temperature deviation matrix, and determine the reference region under the four-connectivity constraint. The reference region is used to generate reference temperatures frame by frame and construct a reference temperature field sequence. S3: The temperature matrix subsequence and the reference temperature field sequence are differentially divided by frame index to form a temperature difference matrix sequence; the temperature difference matrix of adjacent frames is traversed in the order of the acquisition time subsequence, and the difference between the temperature difference value of the next frame and the temperature difference value of the previous frame is calculated for each pixel, and divided by the acquisition time interval to obtain the temperature difference change rate. The temperature difference change rates of all obtained pixels are combined to form a temperature difference change rate matrix sequence; when the temperature difference change rate of the corresponding pixel is positive, the temperature difference change rate is multiplied by the acquisition time interval as the positive increment of the pixel; when the temperature difference change rate of the corresponding pixel is zero or negative, the positive increment of the pixel is set to zero, and the positive increments of all pixels are combined to form a positive increment matrix sequence; the positive increment matrix sequence generates a temperature rise trend matrix sequence and a persistence matrix sequence, and combines with the reference area to generate a trend description. The trend description is structured information stored by frame index, including at least the set of trend significantly connected regions of the corresponding frame, and the set of pixel index positions, regional temperature rise trend summary value and regional persistence summary value of each trend significantly connected region. S4: Based on the trend description, a set of candidate anomaly regions is obtained, and the core location and core temperature difference are determined on the temperature difference matrix sequence. The core temperature difference generates an isothermal boundary sequence, and the consistency of the outward expansion direction and the continuity of the boundary gradient are calculated to obtain evidence of diffusion consistency. After the candidate anomaly region set is associated across frames through the overlap, the candidate anomaly region trajectory and diffusion consistency evidence are output. S5: The reference area generates the temperature difference baseline bandwidth and the trend baseline bandwidth. The candidate anomaly area trajectory generates the diffusion consistency evidence baseline bandwidth and constructs the discriminant. The discriminant is used to generate the early warning criterion through continuous gating and multi-condition consistency rules.

2. The method for thermal runaway early warning analysis of battery infrared images according to claim 1, characterized in that, Step S1 includes: fixing the infrared thermal imager to continuously acquire infrared thermal image frames within the field of view and writing the acquisition time for each frame; discarding infrared thermal image frames when the acquisition times are reversed or repeated; converting the pixel response values ​​of the infrared thermal image frames into temperature matrices according to the radiometric calibration relationship; filling the invalid pixel positions marked in the bad pixel table with the median of the eight-connected neighborhood; and sorting the temperature matrices according to the acquisition times to form a temperature matrix sequence.

3. The method for early warning analysis of thermal runaway using infrared images of a battery according to claim 2, characterized in that, Step S2 includes: extracting acquisition windows from the temperature matrix sequence in ascending order of acquisition time to form a temperature matrix subsequence and an acquisition time subsequence; calculating the fluctuation matrix and temperature deviation matrix based on the temperature matrix subsequence; constructing a set of reference candidate pixels based on the median limit of the fluctuation and the median limit of the temperature deviation, and determining the index position of the reference seed pixel.

4. The method for early warning analysis of thermal runaway using infrared images of a battery according to claim 3, characterized in that, Step S2 further includes: performing a four-connected breadth-first search within the reference candidate pixel set using the reference seed pixel index position to obtain a reference region; taking the median temperature of the reference region frame by frame to form a reference temperature sequence and generating a reference temperature field sequence; and binding and outputting the reference temperature field sequence, the temperature matrix subsequence, and the acquisition time subsequence according to the frame index.

5. The method for thermal runaway early warning analysis of battery infrared images according to claim 4, characterized in that, Step S3 includes: aligning the temperature matrix subsequence with the reference temperature field sequence according to the frame index and verifying that the acquisition time subsequence is consistent; keeping adjacent items of the acquisition time subsequence increasing; forming a temperature difference matrix sequence by pixel index difference; calculating the temperature difference change rate matrix sequence between adjacent frames based on the acquisition time difference; and converting the positive temperature difference change rate into a positive increment matrix sequence based on the acquisition time difference.

6. The method for thermal runaway early warning analysis of battery infrared images according to claim 5, characterized in that, Step S3 also includes: the positive incremental matrix sequence is accumulated by frame index within the acquisition window to generate a temperature rise trend matrix sequence; the position where the positive increment is zero triggers the temperature rise trend reset and simultaneously resets the persistence matrix sequence; the reference region calculates the reference trend level and constructs a set of trend significant pixels; the set of trend significant pixels is labeled with four-connectivity to obtain a set of trend significant connected regions and is summarized and written into the trend description.

7. The method for thermal runaway early warning analysis of battery infrared images according to claim 6, characterized in that, Step S4 includes: mapping the set of significantly connected regions in the trend description to a set of candidate abnormal regions according to the frame index; determining the core location and core temperature difference within the candidate abnormal region range of the temperature difference matrix sequence; and ensuring that the frame index of the temperature difference matrix sequence, the temperature rise trend matrix sequence, and the subsequence of the acquisition time are consistent and completing the consistency check.

8. The method for thermal runaway early warning analysis of battery infrared images according to claim 7, characterized in that, Step S4 further includes: constructing an isothermal boundary sequence by extracting temperature difference levels from the core temperature difference in the candidate anomaly region set; extracting the consistency of the outward expansion direction and the continuity of the boundary gradient from the isothermal boundary sequence to form diffusion consistency evidence; using the overlap and core position distance to complete cross-frame association of the candidate anomaly region set to form the candidate anomaly region trajectory and output it; wherein, the consistency of the outward expansion direction is obtained by the cosine of the angle between the outward expansion displacement vector and the radial vector, and the consistency of the outward expansion direction is recorded as zero when the magnitude of any vector is zero; the continuity of the boundary gradient is obtained by the relationship between the discrete gradient of the boundary pixel index position and the radial direction, and the discrete gradient is constructed by the temperature difference difference between adjacent pixels in the row direction and the temperature difference difference between adjacent pixels in the column direction. When the boundary pixel is located at the outer edge of the matrix, resulting in insufficient neighborhood, a one-sided difference is used instead; taking the median value of the consistency of the outward expansion direction and the continuity of the boundary gradient in the temperature difference level dimension to form diffusion consistency evidence of the candidate anomaly region in the current frame.

9. The method for thermal runaway early warning analysis of battery infrared images according to claim 8, characterized in that, Step S5 includes: verifying the consistency of the frame index of the temperature difference matrix sequence, temperature rise trend matrix sequence, persistence matrix sequence, trend description, acquisition time subsequence and candidate abnormal area trajectory; mapping the regional temperature rise trend summary value and regional persistence summary value in the trend description to the candidate abnormal area trajectory according to the overlap of the pixel index position set; and removing candidate abnormal areas that fail to be mapped from the candidate abnormal area trajectory.

10. The method for thermal runaway early warning analysis of battery infrared images according to claim 9, characterized in that, Step S5 further includes: generating upper bounds of the temperature difference baseline bandwidth and trend baseline bandwidth in the temperature difference matrix sequence and temperature rise trend matrix sequence for the reference region; for each frame of the acquisition window, extracting the set of evidence for consistency in the outward expansion direction and the set of evidence for continuity of the boundary gradient from the trajectory of the candidate anomaly region, and generating upper bounds of the baseline bandwidth for consistency in the outward expansion direction and the baseline bandwidth for continuity of the boundary gradient respectively; when the upper bound of the baseline bandwidth for consistency in the outward expansion direction is greater than the upper bound of the domain of consistency in the outward expansion direction, setting the upper bound of the baseline bandwidth for consistency in the outward expansion direction to the upper bound of the domain of consistency in the outward expansion direction; when the upper bound of the baseline bandwidth for continuity of the boundary gradient is greater than the upper bound of the domain of continuity of the boundary gradient, setting the upper bound of the baseline bandwidth for continuity of the boundary gradient to the upper bound of the domain of continuity of the boundary gradient; the discriminant quantity, combined with continuous gating, outputs attention warning, alert warning and emergency warning, and outputs the acquisition time, warning level, core location and set of pixel index positions of candidate anomaly regions.

Images