A deep learning-based early warning method for thermal runaway of lithium batteries

Abstract

The application discloses a lithium battery thermal runaway early warning method based on deep learning, comprising: obtaining a time-consistent low signal-to-noise ratio thermal image sequence of a lithium battery; generating a lithium battery thermal sub-zone mapping table; obtaining an original low signal-to-noise ratio infrared sequence, outputting a lithium battery sub-zone reconstructed thermal field sequence; outputting a lithium battery sub-zone rough risk mask; outputting a lithium battery sub-zone accurate positioning mask; generating a lithium battery thermal sub-zone risk parameter table; obtaining a lithium battery thermal runaway risk score list, comparing the lithium battery multi-level thermal runaway risk score list with lithium battery dynamic threshold values set according to perturbation monitoring sections, gradual sections and diffusion risk sections, and generating a lithium battery segmented thermal runaway early warning list. The application improves the distinguishability of real thermal anomaly characteristics in the lithium battery sub-zone reconstructed thermal field sequence, and reduces the false detection probability caused by busbar heating, high-voltage wire harness heating and sensor background noise.

Description

Technical Field

[0001] This invention relates to the field of lithium battery thermal runaway technology, and in particular to an early warning method for lithium battery thermal runaway based on deep learning. Background Technology

[0002] With the rapid development of new energy vehicles, electrochemical energy storage power stations, and distributed energy storage systems, lithium-ion batteries have been widely used due to their high energy density, long cycle life, and excellent power performance. However, lithium-ion batteries are prone to localized micro-short circuits, separator degradation, and exothermic side reactions, which can induce thermal runaway.

[0003] In existing technologies, battery management systems mainly rely on voltage, current, and contact temperature sensors to monitor battery status. However, due to the limited number and fixed placement of contact temperature sensors, only discrete measurement point information can be obtained, making it difficult to reflect the complete temperature field distribution on the surface of the lithium battery pack. When a local micro-short circuit or early side reaction occurs inside the battery, the resulting slight temperature rise often has not yet propagated to the sensor installation location, leading to delayed warnings or even missed detections.

[0004] To obtain richer thermal field information, infrared thermal imaging technology is increasingly being applied to the field of lithium battery thermal runaway monitoring. In practical applications, the background noise generated by the infrared focal plane array sensor itself, background thermal interference from normal heat sources such as internal busbars and high-voltage wiring harnesses of the lithium battery pack, and thermal radiation disturbances in complex environments can lead to a low signal-to-noise ratio in the acquired infrared thermal images, making it difficult to accurately extract the weak temperature rise characteristics of only a few degrees Celsius in the very early stages of thermal runaway.

[0005] Within a lithium-ion battery pack, individual lithium-ion cells are typically arranged in a high-density configuration. When a single cell experiences abnormal heating, the heat diffuses along the heat conduction path to the surrounding area, creating a blurred-edge heat diffusion region in infrared thermal imaging. Existing image segmentation and target detection methods struggle to accurately distinguish between the actual heating cell and neighboring cells affected by heat diffusion, resulting in insufficient positioning accuracy and hindering precise risk identification at the individual lithium-ion cell level. Summary of the Invention

[0006] One objective of this invention is to propose an early warning method for thermal runaway of lithium batteries based on deep learning. This invention improves the identifiability of real thermal anomaly features in the reconstructed thermal field sequence of lithium battery sub-regions and reduces the probability of false detections caused by busbar heating, high-voltage harness heating and sensor background noise.

[0007] According to an embodiment of the present invention, a deep learning-based early warning method for thermal runaway of lithium batteries includes: The original low signal-to-noise ratio thermal image sequence of the lithium battery pack during operation was collected and preprocessed to obtain a time-consistent low signal-to-noise ratio thermal image sequence of the lithium battery. Based on the physical structure, heat dissipation path and isothermal distribution of the lithium battery pack, the surface of the lithium battery pack is divided into several lithium battery thermal sub-regions, and a lithium battery thermal sub-region mapping table is generated. The time-consistent low signal-to-noise ratio thermal image sequence of lithium battery is rasterized and segmented according to the lithium battery thermal sub-region mapping table to obtain the original low signal-to-noise ratio infrared sequence. For each lithium battery thermal sub-region, the corresponding original low signal-to-noise ratio infrared sequence is input into the lithium battery region autoencoder to output the lithium battery sub-region reconstructed thermal field sequence. The reconstructed thermal field sequence of the lithium battery sub-region is input into the Swin-Transformer network of the lithium battery region to extract the weak spatial thermal coupling characteristics of adjacent cells in the lithium battery thermal sub-region caused by early local micro short circuits, and output a rough risk mask of the lithium battery sub-region. The coarse risk mask of the lithium battery sub-region is input into the lithium battery point quadtree partitioning and incoherent region detector. The boundary of the coarse risk mask of the lithium battery sub-region is geometrically refined, and the precise positioning mask of the lithium battery sub-region is output. Based on the precise positioning mask of lithium battery sub-region, the absolute temperature, temperature rise rate, relative temperature difference and thermal runaway risk probability of the target single lithium battery cell are extracted to generate a lithium battery thermal sub-region risk parameter table. Input the risk parameter table of each lithium battery thermal sub-region into the hierarchical risk model, calculate the multi-level thermal runaway risk score according to the lithium battery thermal sub-region, lithium battery module and lithium battery pack level, and obtain the lithium battery thermal runaway risk score list. Compare the multi-level thermal runaway risk score list of lithium batteries with the lithium battery dynamic threshold set according to the perturbation monitoring segment, the gradual segment and the diffusion risk segment to generate the lithium battery segmented thermal runaway early warning list.

[0008] Optionally, the preprocessing includes time synchronization and formatting of the original low signal-to-noise ratio thermal image sequence.

[0009] Optionally, based on the physical structure, heat dissipation path, and isothermal distribution of the lithium battery pack, the surface of the lithium battery pack is divided into several lithium battery thermal sub-regions, including: Obtain the arrangement position of individual lithium cells in the lithium battery pack, the projected size of individual lithium cells, the boundary of the lithium battery module, and the surface contour of the lithium battery pack, and construct the lithium battery pack structure set. The distance between the surface structures of any two individual lithium battery cells is calculated based on the lithium battery pack structure set, and the adjacency matrix of the lithium battery pack structure is constructed. Based on the heat dissipation path of the lithium battery pack, and combined with the adjacency matrix of the lithium battery pack structure and the distance of the lithium battery surface structure, a heat dissipation path correlation matrix of the lithium battery pack is constructed. The average surface temperature of each individual lithium battery cell under static benchmark conditions is obtained, and the isothermal distribution matrix of the lithium battery pack is constructed. The adjacency matrix of the lithium battery pack structure, the correlation matrix of the lithium battery pack heat dissipation path, and the isothermal distribution matrix of the lithium battery pack are weighted and fused to generate the thermal sub-region correlation matrix of the lithium battery. Based on the correlation matrix of lithium battery thermal sub-regions, the surface of the lithium battery pack is divided into spectral clusters and thermal aggregations to obtain a set of lithium battery thermal sub-regions; Each lithium battery thermal sub-region is bound to the infrared thermal image coordinate range of the lithium battery pack surface, the individual lithium cell number, and the lithium battery module number to generate a lithium battery thermal sub-region mapping table. Each lithium battery thermal subregion is assigned a lithium battery thermal subregion identifier based on the lithium battery thermal subregion mapping table.

[0010] Optionally, the step of rasterizing and segmenting the time-consistent low signal-to-noise ratio thermal image sequence of lithium batteries according to the lithium battery thermal sub-region mapping table includes: Based on the lithium battery thermal sub-region mapping table, the coordinate range of each lithium battery thermal sub-region in the infrared thermal image pixel coordinate system is extracted, and a corresponding lithium battery thermal sub-region grid mask is constructed. Each time-consistent low signal-to-noise ratio thermal image of a lithium battery is rasterized according to the thermal sub-region raster mask of the lithium battery to obtain the original low signal-to-noise ratio infrared image. The original low signal-to-noise ratio infrared image is input into the encoder of the corresponding lithium battery region autoencoder to obtain the corresponding lithium battery sub-region thermal field latent features. The decoder of the lithium battery region autoencoder decodes the latent thermal field features of the lithium battery sub-region and outputs the residual thermal interference map of the lithium battery sub-region. The residual temperature response corresponding to the residual thermal interference map of the lithium battery sub-region is extracted from the original low signal-to-noise ratio infrared image, and the reconstructed temperature response value of each pixel coordinate position is calculated. The reconstructed thermal field map of the lithium battery sub-region is composed of the reconstructed temperature response values ​​of all pixel coordinate positions and output. The reconstructed thermal field maps of all lithium battery sub-regions are arranged chronologically according to the acquisition time, forming the reconstructed thermal field sequence of the lithium battery sub-regions corresponding to the lithium battery thermal sub-regions.

[0011] Optionally, inputting the reconstructed thermal field sequence of the lithium battery sub-region into the Swin-Transformer network of the lithium battery region includes: Each frame of the reconstructed thermal field map of the lithium battery sub-region is input into the lithium battery sub-region feature embedding layer of the lithium battery region Swin-Transformer network to obtain the lithium battery sub-region thermal field embedding feature map. The thermal field embedding feature map of the lithium battery sub-region is divided into regular windows to obtain a set of non-overlapping local windows of the lithium battery sub-region. For each non-overlapping local window of the lithium battery sub-region, non-overlapping local window self-attention calculation is performed to obtain the thermal field correlation characteristics within the lithium battery sub-region window; All thermal field correlation features within the lithium battery sub-region window are backfilled into the feature coordinates corresponding to the thermal field embedding feature map of the lithium battery sub-region according to the non-overlapping local window positions of the corresponding lithium battery sub-region, thus obtaining the self-attention feature map of the lithium battery sub-region window. Perform a moving window offset on the self-attention feature map of the lithium battery sub-region window to obtain the moving window feature map of the lithium battery sub-region; The moving window feature map of the lithium battery sub-region is re-divided into regular windows, and self-attention calculation based on the moving window is performed to obtain the cross-window thermal field correlation feature map of the lithium battery sub-region. The cross-window thermal field correlation feature map of the lithium battery sub-region is used as the input of the self-attention block of the next lithium battery region. The self-attention calculation of non-overlapping local windows and the self-attention calculation based on moving windows are stacked layer by layer as a self-attention block of the lithium battery region, and the weak spatial thermal strong coupling feature map of the lithium battery sub-region is output. A coarse risk mapping is performed on the weak spatial thermal coupling feature map of the lithium battery sub-region to obtain a coarse risk probability map of the lithium battery sub-region. Based on the coarse risk probability map of the lithium battery sub-region, a block-level coarse risk mask of the lithium battery sub-region at the Tokens scale is generated.

[0012] Optionally, the step of inputting the coarse risk mask of the lithium battery sub-region into the lithium battery point quadtree partitioning and incoherent region detector includes: The rough risk mask of the lithium battery sub-region is back-projected onto the infrared thermal image pixel coordinate system to obtain the pixel-scale rough risk mask of the lithium battery sub-region. Boundary extraction is performed on the coarse risk mask of the lithium battery sub-region at the pixel scale to obtain the coarse risk boundary band of the lithium battery sub-region. The rough risk boundary zone of the lithium battery sub-region is input into the lithium battery point quadtree partitioning model to form several lithium battery point quadtree nodes. When the lithium battery point quadtree node meets the stopping partitioning condition, it is determined as the leaf node of the lithium battery sub-region boundary quadtree. The temperature gradient within the rough risk boundary band of the lithium battery sub-region is calculated based on the reconstructed thermal field map of the lithium battery sub-region, and the temperature gradient field of the lithium battery sub-region is obtained. The quad-leaf tree nodes of the lithium battery sub-region boundary and the temperature gradient field of the lithium battery sub-region are input into the incoherent region detector to detect the discontinuity of the temperature gradient between the quad-leaf tree nodes of adjacent lithium battery sub-region boundaries, and the incoherent boundary response of the lithium battery sub-region is obtained. When the incoherent boundary response of the lithium battery sub-region is not less than the incoherent boundary response threshold, the common boundary between the four-way leaf nodes of two adjacent lithium battery sub-region boundaries is determined as the incoherent boundary segment of the lithium battery sub-region. Subpixel-level geometric refinement is performed on the incoherent boundary segments of the lithium battery sub-region to generate subpixel fine boundaries of the lithium battery sub-region. Based on the sub-pixel fine boundary of the lithium battery sub-region, the boundary replacement and region closure of the pixel-scale coarse risk mask of the lithium battery sub-region are performed to output the precise positioning mask of the lithium battery sub-region.

[0013] Optionally, the step of extracting the absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability of the target single lithium battery cell based on the precise positioning mask of the lithium battery sub-region includes: The target single lithium cell corresponding to the precise positioning mask of the lithium battery sub-region is determined based on the lithium battery thermal sub-region mapping table; Based on the reconstructed temperature response value of the lithium battery sub-region precisely located mask coverage, the absolute temperature of the target single lithium battery cell is extracted; The temperature rise rate of the target single lithium battery cell is calculated based on the absolute temperature of consecutive frames. The relative temperature difference of the target lithium battery cell is calculated by subtracting the average reference temperature of the non-target single lithium battery cell within the same lithium battery thermal sub-region from the absolute temperature. The absolute temperature, temperature rise rate, and relative temperature difference of the target single lithium battery cell are normalized to calculate the probability of thermal runaway risk of the target single lithium battery cell. The target single lithium battery cell identifier, the absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability of the target single lithium battery cell are bound together to generate a lithium battery thermal sub-region risk parameter table.

[0014] Optionally, inputting the risk parameter table of each lithium battery thermal subregion into the hierarchical risk model includes: Input the risk parameter table of each lithium battery thermal subregion into the hierarchical risk model to calculate the lithium battery thermal subregion risk score at the lithium battery thermal subregion level. Based on the risk score of the lithium battery thermal sub-region, the lithium battery modules to which the lithium battery thermal sub-region belongs are aggregated, and the risk score of the lithium battery module at the lithium battery module level is calculated. Based on the risk score of the lithium battery module, the risk score of the entire lithium battery pack is calculated by aggregating the risk scores of the entire lithium battery pack. By binding the risk scores of the lithium battery thermal sub-region, the lithium battery module, and the lithium battery pack, a multi-level thermal runaway risk score list for lithium batteries is obtained. Based on the multi-level thermal runaway risk score list of lithium batteries, set the dynamic thresholds of lithium batteries corresponding to the perturbation monitoring segment, the gradual segment, and the diffusion risk segment. The list of multi-level thermal runaway risk scores for lithium batteries is compared with the dynamic thresholds of lithium batteries corresponding to the perturbation monitoring segment, the gradual segment, and the diffusion risk segment to determine the segmented thermal runaway warning type of lithium battery corresponding to each thermal sub-region.

[0015] Optionally, the lithium battery segmented thermal runaway warning types include: Early warning of perturbation monitoring segment: The risk score of the lithium battery thermal sub-region is not less than the dynamic threshold of the perturbation monitoring segment of the lithium battery thermal sub-region, and the risk score of the lithium battery thermal sub-region is less than the dynamic threshold of the progressive segment of the lithium battery thermal sub-region. Progressive stage warning: The risk score of the lithium battery thermal sub-region is not less than the dynamic threshold of the progressive stage of the lithium battery thermal sub-region, and the risk score of the lithium battery module is less than the dynamic threshold of the diffusion risk stage of the lithium battery module. Early warning of diffusion risk segment: The risk score of the lithium battery module is not less than the dynamic threshold of the diffusion risk segment of the lithium battery module, or the risk score of the entire lithium battery pack is not less than the dynamic threshold of the diffusion risk segment of the entire lithium battery pack. No warning: The risk score of the lithium battery thermal sub-region is less than the dynamic threshold of the lithium battery thermal sub-region perturbation monitoring segment.

[0016] The beneficial effects of this invention are: This invention utilizes the synergistic processing of lithium battery thermal sub-region division, lithium battery region autoencoder residual denoising, and lithium battery sub-region thermal field reconstruction. After dividing the surface of the lithium battery pack into several lithium battery thermal sub-regions, residual interference is stripped from each sub-region. The lithium battery region autoencoder is used to regionalize the multiple heat source interference, infrared focal plane array background noise, and low-frequency background heat sources within the lithium battery thermal sub-regions. This allows the residual thermal interference map of the lithium battery sub-regions to be directionally stripped from the original low signal-to-noise ratio infrared image, preserving the weak temperature rise response caused by local micro-short circuits. This improves the identifiability of real thermal anomaly features in the reconstructed thermal field sequence of the lithium battery sub-regions and reduces the probability of false detections caused by busbar heating, high-voltage harness heating, and sensor background noise.

[0017] This invention employs a Swin-Transformer network for lithium battery regions to perform non-overlapping local window self-attention calculation and moving window-based self-attention calculation on the reconstructed thermal field sequence of lithium battery sub-regions. This enables the model to not only extract temperature changes within a single pixel or local convolutional neighborhood, but also to establish thermal field feature associations between adjacent cell boundary tokens within the lithium battery thermal sub-region. Through regular window partitioning, in-window self-attention, moving window offset, and cross-window thermal field association feature backfilling, the temperature rise responses of adjacent cells originally located on both sides of different non-overlapping local window boundaries are brought into the same feature interaction range. This enhances the weak, gradual, and localized spatial thermal strong coupling features within the lithium battery thermal sub-region, providing a more stable risk region foundation for single lithium battery cell-level positioning.

[0018] This invention improves the positioning accuracy and segmented early warning reliability in thermal runaway scenarios through continuous processing of lithium battery point quadtree partitioning, incoherent region detectors, and hierarchical risk models. For the boundary ambiguity problem caused by thermal conduction and diffusion in infrared thermal images, a coarse risk mask of block-level lithium battery sub-regions at the token scale is back-projected onto the infrared thermal image pixel coordinate system. The coarse risk boundary band is recursively partitioned using a lithium battery point quadtree, and the temperature gradient discontinuity between leaf nodes of adjacent boundary quadtrees is identified using an incoherent region detector. A precise positioning mask for lithium battery sub-regions is formed through sub-pixel-level geometric refinement, reducing positioning adhesion caused by thermal diffusion of adjacent individual lithium battery cells. Simultaneously, a segmented thermal runaway early warning list corresponding to spatial location and risk stage is formed, improving the interpretability and engineering usability of early warning results. Attached Figure Description

[0019] Figure 1 This is a flowchart of a deep learning-based early warning method for thermal runaway in lithium batteries proposed in this invention. Figure 2 A schematic diagram of the lithium battery pack structure in Embodiment 2 of the proposed method for early warning of thermal runaway of lithium batteries based on deep learning; Figure 3 The image shows the thermal sub-region partitioning results of a lithium battery in Example 2 of a deep learning-based early warning method for thermal runaway in a lithium battery. Figure 4 Example 2 of the proposed method for early warning of thermal runaway in lithium batteries based on deep learning, ID-Z2-04 detail diagram of the lithium battery thermal sub-region partitioning result. Detailed Implementation

[0020] Example 1: Reference Figure 1 A deep learning-based early warning method for thermal runaway in lithium batteries includes: The original low signal-to-noise ratio thermal image sequence of the lithium battery pack during operation was collected and preprocessed to obtain a time-consistent low signal-to-noise ratio thermal image sequence of the lithium battery. In this embodiment, preprocessing includes time synchronization and formatting of the original low signal-to-noise ratio thermal image sequence.

[0021] Based on the physical structure, heat dissipation path and isothermal distribution of the lithium battery pack, the surface of the lithium battery pack is divided into several lithium battery thermal sub-regions, and a lithium battery thermal sub-region mapping table is generated. In this embodiment, based on the physical structure, heat dissipation path, and isothermal distribution of the lithium battery pack, the surface of the lithium battery pack is divided into several lithium battery thermal sub-regions, including: Obtain the arrangement position of individual lithium cells in the lithium battery pack, the projected size of individual lithium cells, the boundary of the lithium battery module, and the surface contour of the lithium battery pack, and construct the lithium battery pack structure set. In Example 1, in constructing the lithium battery pack structure set, an infrared camera intrinsic and extrinsic parameter calibration matrix is ​​introduced to establish a homography perspective transformation mapping relationship between the infrared thermal image pixel coordinate system and the physical coordinate system of the lithium battery pack surface.

[0022] The lithium battery pack structure consists of physical structural elements of several individual lithium battery cells. Each individual lithium battery cell's physical structural elements include the lateral center position of the individual lithium battery cell in the coordinate system of the lithium battery pack surface, the longitudinal center position of the individual lithium battery cell in the coordinate system of the lithium battery pack surface, the projected length of the individual lithium battery cell in the coordinate system of the lithium battery pack surface, the projected width of the individual lithium battery cell in the coordinate system of the lithium battery pack surface, the cell number of the individual lithium battery cell, and the lithium battery module number to which the individual lithium battery cell belongs.

[0023] The distance between the surface structures of any two individual lithium battery cells is calculated based on the lithium battery pack structure set, and the adjacency matrix of the lithium battery pack structure is constructed. In Example 1, the adjacency matrix of the lithium battery pack structure represents the physical structural adjacency relationship between individual lithium battery cells: If the distance between the surface structures of two individual lithium-ion cells is not greater than the threshold for determining the physical adjacency of individual lithium-ion cells, and the two individual lithium-ion cells are not the same individual lithium-ion cell, then the structural adjacency state between the two individual lithium-ion cells is set as adjacent.

[0024] If the structural distance between two individual lithium battery cells is greater than the lithium battery surface distance threshold, or if the two individual lithium battery cells are the same individual lithium battery cell, then the structural adjacency state between the two individual lithium battery cells is set to non-adjacent.

[0025] The surface distance threshold for lithium batteries is determined based on the statistical value of the designed center-to-center distance between adjacent individual lithium battery cells. ,in This represents the median of the nearest neighbor center distances of individual lithium-ion cells in the surface coordinate system of the lithium-ion battery pack. This is the adjacency tolerance coefficient, with a value range of [value range missing]. In Example 1, when the nearest neighbor center distance of the square single-cell lithium battery is 34mm, A thickness of 40mm can be selected, based on the fact that it covers directly adjacent individual lithium battery cells within the same lithium battery module and excludes individual lithium battery cells that are separated by gaps, separated by modules, or not on direct heat transfer paths.

[0026] Based on the heat dissipation path of the lithium battery pack, and combined with the adjacency matrix of the lithium battery pack structure and the distance of the lithium battery surface structure, a heat dissipation path correlation matrix of the lithium battery pack is constructed. In Example 1, when the structural adjacency between two individual lithium-ion cells is adjacent, the equivalent thermal conductivity between the individual lithium-ion cells is divided by the sum of the structural distance on the lithium battery surface and the minimum constant to obtain the initial heat dissipation path correlation strength between the individual lithium-ion cells. When the structural adjacency between two individual lithium-ion cells is not adjacent, the initial heat dissipation path correlation strength between the individual lithium-ion cells is set to zero, and the initial heat dissipation path correlation strength between the individual lithium-ion cells is divided by the sum of the maximum value of the initial heat dissipation path correlation strength between all individual lithium-ion cells and the normalized minimum constant to obtain the heat dissipation path correlation strength between the individual lithium-ion cells.

[0027] The equivalent thermal conductivity between individual lithium battery cells is calculated based on the equivalent series thermal resistance of the heat dissipation path. It is obtained by dividing the total thickness of the heat dissipation path by the sum of the ratios of the thickness of each heat dissipation medium to its thermal conductivity. The specific value is determined based on the actual air, thermally conductive silicone, insulating bracket, metal shell or heat spreader material composition, thickness ratio and assembly contact state between adjacent individual lithium battery cells.

[0028] The average surface temperature of each individual lithium battery cell under static benchmark conditions is obtained, and the isothermal distribution matrix of the lithium battery pack is constructed. In Example 1, the absolute value of the temperature difference between the average surface temperature of two individual lithium battery cells under static reference conditions of rated charge / discharge rate and standard operating temperature is calculated. The absolute value of the temperature difference is divided by the preset maximum allowable design temperature difference in the battery pack as the normalization denominator constant to obtain the normalized value of the isothermal temperature difference. The normalized value of the isothermal temperature difference is negative and then exponentially mapped to obtain the isothermal distribution intensity between individual lithium battery cells.

[0029] The adjacency matrix of the lithium battery pack structure, the correlation matrix of the lithium battery pack heat dissipation path, and the isothermal distribution matrix of the lithium battery pack are weighted and fused to generate the thermal sub-region correlation matrix of the lithium battery. Based on the correlation matrix of lithium battery thermal sub-regions, the surface of the lithium battery pack is divided into spectral clusters and thermal aggregations to obtain a set of lithium battery thermal sub-regions; In Example 1, the lithium battery thermal sub-region set represents all lithium battery thermal sub-regions divided from the surface of the lithium battery pack. Each lithium battery thermal sub-region is composed of individual lithium battery cells with physical structural adjacency, heat dissipation path correlation, and similar isothermal distribution. The lithium battery thermal sub-region number to which each individual lithium battery cell belongs is determined.

[0030] Each lithium battery thermal sub-region is bound to the infrared thermal image coordinate range of the lithium battery pack surface, the individual lithium cell number, and the lithium battery module number to generate a lithium battery thermal sub-region mapping table. In Example 1, based on the homography perspective transformation mapping relationship, each lithium battery thermal sub-region is bound to the infrared thermal image pixel coordinate range of the lithium battery pack surface, the individual lithium battery cell number, and the lithium battery module number.

[0031] Each lithium battery thermal subregion is assigned a lithium battery thermal subregion identifier based on the lithium battery thermal subregion mapping table.

[0032] The time-consistent low signal-to-noise ratio thermal image sequence of lithium battery is rasterized and segmented according to the lithium battery thermal sub-region mapping table to obtain the original low signal-to-noise ratio infrared sequence. For each lithium battery thermal sub-region, the corresponding original low signal-to-noise ratio infrared sequence is input into the lithium battery region autoencoder to output the lithium battery sub-region reconstructed thermal field sequence. In this embodiment, the time-consistent low signal-to-noise ratio thermal image sequence of lithium battery is rasterized and segmented according to the lithium battery thermal sub-region mapping table, including: Based on the lithium battery thermal sub-region mapping table, the coordinate range of each lithium battery thermal sub-region in the infrared thermal image pixel coordinate system is extracted, and a corresponding lithium battery thermal sub-region grid mask is constructed. In Example 1, the grid mask of the lithium battery thermal sub-region indicates whether the pixel coordinates in the infrared thermal image pixel coordinate system belong to the corresponding lithium battery thermal sub-region: when the pixel coordinates are within the coordinate range of the corresponding lithium battery thermal sub-region in the infrared thermal image pixel coordinate system, the mask state of the corresponding pixel coordinates is set to belong to the lithium battery thermal sub-region; otherwise, it does not belong to the corresponding lithium battery thermal sub-region.

[0033] Each time-consistent low signal-to-noise ratio thermal image of a lithium battery is rasterized according to the thermal sub-region raster mask of the lithium battery to obtain the original low signal-to-noise ratio infrared image. The original low signal-to-noise ratio infrared image is input into the encoder of the corresponding lithium battery region autoencoder to obtain the corresponding lithium battery sub-region thermal field latent features. In Example 1, the lithium battery region autoencoder includes a lithium battery sub-region input layer, an encoder, a decoder, and a lithium battery sub-region thermal field reconstruction layer.

[0034] The lithium battery sub-region input layer receives the original low signal-to-noise ratio infrared image. The encoder sequentially performs local temperature response convolution extraction, nonlinear activation, downsampling compression, and channel feature aggregation on each frame of the original low signal-to-noise ratio infrared image to obtain the latent thermal field features of each lithium battery thermal sub-region in each frame. The latent thermal field features of the lithium battery sub-region represent the local temperature gradient, low-frequency background heat source distribution, infrared focal plane array background noise, and weak temperature rise response in the original low signal-to-noise ratio infrared image within the corresponding lithium battery thermal sub-region.

[0035] The decoder of the lithium battery region autoencoder decodes the latent thermal field features of the lithium battery sub-region and outputs the residual thermal interference map of the lithium battery sub-region. In Example 1, the lithium battery region autoencoder is pre-trained based on historical samples of the normally operating passive region that has not experienced thermal runaway. During training, a target function jointly constructed by global reconstruction loss and residual smoothing adversarial regularization constraint term is used to constrain the decoder, causing the decoder to generate a feature decoupling barrier. This allows the decoder to only reconstruct high-frequency random background noise and low-frequency slowly varying background heat sources in a directional manner, without restoring abnormal weak temperature rise responses. This completes the low-level decoupling of interference features and target warning signals in the latent features, and outputs the residual thermal interference map of the lithium battery sub-region.

[0036] The residual temperature response corresponding to the residual thermal interference map of the lithium battery sub-region is extracted from the original low signal-to-noise ratio infrared image, and the reconstructed temperature response value of each pixel coordinate position is calculated. The reconstructed thermal field map of the lithium battery sub-region is composed of the reconstructed temperature response values ​​of all pixel coordinate positions and output. In Example 1, the original low signal-to-noise ratio infrared image and the residual thermal interference of the lithium battery sub-region are compared. Figure 1 The lithium battery sub-region thermal field reconstruction layer of the same input lithium battery region autoencoder is used to extract the residual temperature response corresponding to the residual thermal interference map of the lithium battery sub-region from the original low signal-to-noise ratio infrared image.

[0037] Since the original low signal-to-noise ratio infrared image and the residual thermal interference image of the lithium battery sub-region maintain the same tensor dimension, the thermal field reconstruction layer of the lithium battery sub-region directly subtracts the residual temperature response value in the residual thermal interference image of the lithium battery sub-region from the temperature response value in the original low signal-to-noise ratio infrared image at the same pixel coordinate position inside, to obtain the reconstructed temperature response value at each pixel coordinate position. The reconstructed temperature response values ​​of all pixel coordinate positions are used to form and output the reconstructed thermal field image of the lithium battery sub-region. The reconstructed thermal field image of the lithium battery sub-region is used to remove the true thermal field distribution of the lithium battery sub-region after removing multiple heat source interference and the background noise of the infrared focal plane array.

[0038] The reconstructed thermal field maps of all lithium battery sub-regions are arranged chronologically according to the acquisition time, forming the reconstructed thermal field sequence of the lithium battery sub-regions corresponding to the lithium battery thermal sub-regions.

[0039] The reconstructed thermal field sequence of the lithium battery sub-region is input into the Swin-Transformer network of the lithium battery region to extract the weak spatial thermal coupling characteristics of adjacent cells in the lithium battery thermal sub-region caused by early local micro short circuits, and output a rough risk mask of the lithium battery sub-region. In this embodiment, the reconstructed thermal field sequence of the lithium battery sub-region is input into the Swin-Transformer network of the lithium battery region, including: Each frame of the reconstructed thermal field map of the lithium battery sub-region is input into the lithium battery sub-region feature embedding layer of the lithium battery region Swin-Transformer network to obtain the lithium battery sub-region thermal field embedding feature map. In Example 1, the reconstructed thermal field map of the lithium battery sub-region is divided into local thermal image blocks and linear feature mapping to obtain the thermal field embedding feature map of the lithium battery sub-region. The thermal field embedding feature map of the lithium battery sub-region represents the local temperature response distribution in the reconstructed thermal field map of the lithium battery sub-region.

[0040] The thermal field embedding feature map of the lithium battery sub-region is divided into regular windows to obtain a set of non-overlapping local windows of the lithium battery sub-region. When dividing the thermal field embedding feature map of the lithium battery sub-region into regular windows, the upper left corner Tokens of the thermal field embedding feature map of the lithium battery sub-region is used as the starting position. Local rectangular feature regions are successively extracted along the horizontal direction according to the preset window width as the horizontal step size, and local rectangular feature regions are successively extracted along the vertical direction according to the preset window height as the vertical step size, so that the non-overlapping local windows of any two adjacent lithium battery sub-regions do not overlap in the Tokens coverage area.

[0041] When the edge region of the thermal field embedding feature map of the lithium battery sub-region is less than the size of a complete window, zero-filling or boundary copying filling is performed on the edge region to form a complete non-overlapping local window of the lithium battery sub-region, and a boundary-filling attention mask is generated simultaneously.

[0042] For each non-overlapping local window of the lithium battery sub-region, non-overlapping local window self-attention calculation is performed to obtain the thermal field correlation characteristics within the lithium battery sub-region window; In Example 1, the self-attention calculation of the non-overlapping local window of the lithium battery sub-region refers to using all Tokens as the calculation object within the non-overlapping local window of each lithium battery sub-region, generating query features, key features, and value features for each Token, and calculating the similarity relationship between the query feature of any Token and the key features of other Tokens within the non-overlapping local window of each lithium battery sub-region to obtain the thermal field feature association score.

[0043] The thermal field feature association scores are scaled and superimposed with relative position bias. The feature association scores corresponding to the filled tokens are assigned negative infinity and normalized exponentially mapped to obtain the thermal field feature association weights. Based on the thermal field feature association weights, the value features of all tokens in the non-overlapping local window of the lithium battery sub-region are weighted and aggregated to obtain the thermal field association features in the lithium battery sub-region window.

[0044] All thermal field correlation features within the lithium battery sub-region window are backfilled into the feature coordinates corresponding to the thermal field embedding feature map of the lithium battery sub-region according to the non-overlapping local window positions of the corresponding lithium battery sub-region, thus obtaining the self-attention feature map of the lithium battery sub-region window. Perform a moving window offset on the self-attention feature map of the lithium battery sub-region window to obtain the moving window feature map of the lithium battery sub-region; Moving window offset refers to shifting the self-attention feature map of the lithium battery sub-region window by half a preset window height in the vertical direction and by half a preset window width in the horizontal direction. This allows tokens that were originally located on both sides of the non-overlapping local window boundaries of different lithium battery sub-regions to enter the same local window in subsequent window division, forming a cross-window thermal field feature interaction relationship.

[0045] The moving window feature map of the lithium battery sub-region is re-divided into regular windows, and self-attention calculation based on the moving window is performed to obtain the cross-window thermal field correlation feature map of the lithium battery sub-region. In Example 1, when the feature map of the moving window of the lithium battery sub-region is re-divided into regular windows, the upper left corner Tokens after the moving window is offset is taken as the starting position. Local rectangular feature areas are successively cut along the horizontal direction according to the preset window width as the horizontal step size, and local rectangular feature areas are successively cut along the vertical direction according to the preset window height as the vertical step size, forming several moving local windows of the lithium battery sub-region. Each lithium battery sub-region moving local window contains multiple tokens that enter the same window region after the moving window offset. These multiple tokens include boundary tokens that originally belonged to non-overlapping local windows of different lithium battery sub-regions.

[0046] Self-attention computation based on moving windows refers to generating query features, key features, and value features for tokens within each moving local window of the lithium battery sub-region. Based on the similarity relationship between the query features and key features, the association weight of the thermal field features within the moving local window of the lithium battery sub-region is calculated. The value features are then weighted and aggregated using these thermal field feature association weights to obtain the thermal field association features of the moving local window of the lithium battery sub-region. All thermal field association features of the moving local window of the lithium battery sub-region are backfilled according to their corresponding positions in the feature map of the moving local window of the lithium battery sub-region. A reverse position restoration is then performed in the opposite direction of the moving window offset, so that the backfilled features are realigned to the original spatial coordinates of the thermal field embedded feature map of the lithium battery sub-region. This results in a cross-window thermal field association feature map of the lithium battery sub-region, which represents the spatial association between the temperature rise responses of adjacent cell boundaries originally distributed in different non-overlapping local windows within the thermal sub-region of the lithium battery.

[0047] The cross-window thermal field correlation feature map of the lithium battery sub-region is used as the input of the self-attention block of the next lithium battery region. The self-attention calculation of non-overlapping local windows and the self-attention calculation based on moving windows are stacked layer by layer as a self-attention block of the lithium battery region, and the weak spatial thermal strong coupling feature map of the lithium battery sub-region is output. A coarse risk mapping is performed on the weak spatial thermal coupling feature map of the lithium battery sub-region to obtain a coarse risk probability map of the lithium battery sub-region. Based on the coarse risk probability map of the lithium battery sub-region, a block-level coarse risk mask of the lithium battery sub-region at the Tokens scale is generated.

[0048] In Example 1, the weak spatial thermal coupling feature map of the lithium battery sub-region is input into the coarse risk mapping layer and Sigmoid mapping is performed to obtain a coarse risk probability map of the lithium battery sub-region, representing the probability that each Tokens spatial scale location within the lithium battery thermal sub-region belongs to the early thermal runaway risk region: When the thermal runaway risk probability corresponding to the coarse risk probability map of the lithium battery sub-region at the Tokens spatial scale position is not less than the coarse risk probability threshold used to generate the coarse risk mask of the lithium battery sub-region, the mask state at the corresponding Tokens spatial scale position is set to belong to the coarse risk region of the lithium battery sub-region; otherwise, it is set to not belong to the coarse risk region of the lithium battery sub-region. The coarse risk probability threshold is determined based on the Tokens-level probability distribution of the coarse risk probability map of the lithium battery sub-region in the normal operation sample.

[0049] In this embodiment, the SwinTransformer network for the lithium battery region uses the reconstructed thermal field map of each lithium battery sub-region as input, and the input tensor dimension is... ,in This refers to the number of samples processed in a batch. and The first The vertical and horizontal pixel counts of each lithium battery thermal sub-region are used, with 1 channel representing the single-channel temperature response. The network includes a lithium battery sub-region feature embedding layer, three lithium battery region self-attention blocks, and a coarse risk mapping layer. The lithium battery sub-region feature embedding layer divides the reconstructed thermal field map of the lithium battery sub-region into... Local thermal images are mapped to 64-dimensional tokens. The number of layers for the self-attention blocks in the three lithium battery regions are respectively... The number of attention heads are respectively Window size is Tokens, the moving window offset is half the window size; the coarse risk mapping layer outputs a coarse risk probability map of the lithium battery sub-region at the Tokens scale, with an output dimension of And generate a rough risk mask for the block-level lithium battery sub-region at the token scale; the network parameter scale is to Each lithium battery thermal sub-region has trainable parameters, and the network weights are shared by each sub-region. Different sub-region sizes are adapted by using a lithium battery thermal sub-region grid mask and a boundary-filling attention mask. The training data sources include normal operating conditions, operating conditions in micro-short circuit embodiment 2, thermal abuse or overcharge experimental conditions, and thermal simulation generated lithium battery sub-region reconstructed thermal field maps. The training label is the Tokens scale early thermal runaway risk area label.

[0050] Compared to existing Swin-Transformer models, the lithium battery region Swin-Transformer network changes the input from general natural images to the thermal field reconstruction of lithium battery thermal sub-regions, replaces the fixed whole-image window partitioning with a regular window partitioning adapted to the boundaries of lithium battery thermal sub-regions, replaces ordinary moving window attention with lithium battery sub-region moving window attention combined with boundary-filling attention masks, and changes the classification output head to a coarse risk mapping layer at the token scale. This allows the network to enhance the weak spatial thermal coupling characteristics between adjacent individual lithium battery cells within the lithium battery thermal sub-regions, reduce the interference of infrared focal plane array background noise and lithium battery sub-region boundary-filling tokens on attention calculation, and output a coarse risk mask for lithium battery sub-regions at a low parameter scale that can be used by lithium battery point quadtree partitioning and incoherent region detectors.

[0051] The coarse risk mask of the lithium battery sub-region is input into the lithium battery point quadtree partitioning and incoherent region detector. The boundary of the coarse risk mask of the lithium battery sub-region is geometrically refined, and the precise positioning mask of the lithium battery sub-region is output. In this embodiment, a coarse risk mask of the lithium battery sub-region is input into the lithium battery point quadtree partitioning and incoherent region detector, including: The rough risk mask of the lithium battery sub-region is back-projected onto the infrared thermal image pixel coordinate system to obtain the pixel-scale rough risk mask of the lithium battery sub-region. Back projection refers to mapping the mask state corresponding to each Token spatial scale position in the rough risk mask of the block-level lithium battery sub-region at the Token scale back to the corresponding pixel coordinate range in the infrared thermal image pixel coordinate system, so that each pixel coordinate in the infrared thermal image pixel coordinate system obtains a mask state consistent with its Token spatial scale position.

[0052] Boundary extraction is performed on the coarse risk mask of the lithium battery sub-region at the pixel scale to obtain the coarse risk boundary band of the lithium battery sub-region. In Example 1, dilation and etch processing are performed on the pixel-scale lithium battery sub-region coarse risk mask, and the difference between the mask area obtained after dilation and the mask area obtained after etch processing is extracted to obtain the lithium battery sub-region coarse risk boundary band, which represents the boundary region in the pixel-scale lithium battery sub-region coarse risk mask that needs to be geometrically refined.

[0053] The rough risk boundary zone of the lithium battery sub-region is input into the lithium battery point quadtree partitioning model to form several lithium battery point quadtree nodes. When the lithium battery point quadtree node meets the stopping partitioning condition, it is determined as the leaf node of the lithium battery sub-region boundary quadtree. In Example 1, the lithium battery point quadtree partitioning model recursively divides the pixel region covered by the rough risk boundary band of the lithium battery sub-region into four sub-regions according to the horizontal and vertical center lines until the corresponding lithium battery point quadtree node meets the stopping partitioning condition, thus obtaining the leaf node of the lithium battery sub-region boundary quadtree.

[0054] The conditions for stopping partitioning are as follows: when the temperature response variance within the corresponding lithium battery point quadtree node is less than the preset lithium battery boundary temperature response variance threshold, or when the size of the corresponding lithium battery point quadtree node is not greater than the preset minimum partitioning size, the partitioning of the corresponding lithium battery point quadtree node is stopped, and the corresponding lithium battery point quadtree node is determined as the leaf node of the lithium battery sub-region boundary quadtree node; when the temperature response variance within the corresponding lithium battery point quadtree node is not less than the preset lithium battery boundary temperature response variance threshold, and the size of the corresponding lithium battery point quadtree node is greater than the preset minimum partitioning size, the corresponding lithium battery point quadtree node is further divided into four child nodes.

[0055] The difference between the reconstructed temperature response value at each pixel coordinate position within the corresponding lithium battery point quadtree node and the average reconstructed temperature response value within the node is squared. The sum of all squared results is then divided by the number of pixels contained within the corresponding lithium battery point quadtree node to obtain the temperature response variance within the node.

[0056] The preset threshold for the boundary temperature response variance of lithium batteries is determined based on the mean and standard deviation of the node temperature response variance within the rough risk boundary band of the lithium battery sub-region in normal operation samples, and its value range is [value missing]. In Example 1, take The preset minimum subdivision size is determined based on the physical resolution of infrared thermal image pixels and the minimum resolvable width of the boundary of a single lithium battery cell. The value ranges from 2×2 pixels to 8×8 pixels. In Example 1, 4×4 pixels is selected. The selection is based on the principle of avoiding excessive subdivision caused by the background noise of the infrared focal plane array while retaining the temperature gradient abrupt change within the rough risk boundary band of the lithium battery sub-region.

[0057] The temperature gradient within the rough risk boundary band of the lithium battery sub-region is calculated based on the reconstructed thermal field map of the lithium battery sub-region, and the temperature gradient field of the lithium battery sub-region is obtained. In Example 1, the temperature gradient field of the lithium battery sub-region is calculated using the lateral temperature gradient component and the longitudinal temperature gradient component.

[0058] The quad-leaf tree nodes of the lithium battery sub-region boundary and the temperature gradient field of the lithium battery sub-region are input into the incoherent region detector to detect the discontinuity of the temperature gradient between the quad-leaf tree nodes of adjacent lithium battery sub-region boundaries, and the incoherent boundary response of the lithium battery sub-region is obtained. In Example 1, the average temperature gradient components along the normal direction of the four-leaf tree nodes at the boundary of two adjacent lithium battery sub-regions are subtracted and their absolute values ​​are divided by the gradient scale parameter to obtain the incoherent boundary response of the lithium battery sub-region.

[0059] The average temperature gradient component along the boundary normal direction represents the average temperature gradient component of all boundary pixel coordinates within the quad-leaf leaf node of the corresponding lithium battery sub-region in the boundary normal direction.

[0060] When the incoherent boundary response of the lithium battery sub-region is not less than the incoherent boundary response threshold, the common boundary between the four-way leaf nodes of two adjacent lithium battery sub-region boundaries is determined as the incoherent boundary segment of the lithium battery sub-region. The incoherent boundary response threshold is determined based on the statistical distribution of the incoherent boundary responses of the lithium battery sub-regions in the normal operation samples. The initial threshold is the sum of the mean and three standard deviations of the incoherent boundary responses of the corresponding lithium battery sub-regions in the normal operation samples, and is further adjusted to 0.25 based on false alarm verification results. The value is adjusted within the range of 0.50. In Example 1, it is 0.35. The basis for this selection is that it can eliminate the continuous temperature gradient change caused by normal thermal diffusion, while retaining the abrupt boundary response caused by the discontinuity of thermal conduction at the physical boundary of a single lithium battery cell.

[0061] Subpixel-level geometric refinement is performed on the incoherent boundary segments of the lithium battery sub-region to generate subpixel fine boundaries of the lithium battery sub-region. In Example 1, the temperature gradient magnitude of three consecutive pixel positions is selected in the normal direction of the incoherent boundary segment of the lithium battery sub-region. The current pixel position on the incoherent boundary segment of the lithium battery sub-region is moved along its normal direction by a sub-pixel offset to obtain the sub-pixel fine boundary point of the lithium battery sub-region. All the sub-pixel fine boundary points of the lithium battery sub-region are connected in sequence to generate the sub-pixel fine boundary of the lithium battery sub-region.

[0062] The sub-pixel offset satisfies: ; in, This represents the sub-pixel offset of the incoherent boundary segment of the lithium battery sub-region in the normal direction. Indicates the serial number of the thermal sub-region of the lithium battery. This indicates the frame number in the low signal-to-noise ratio thermal image sequence of a lithium battery. This represents the magnitude of the temperature gradient at adjacent pixel locations along the negative normal direction of the incoherent boundary segment of the lithium battery sub-region. This represents the magnitude of the temperature gradient at adjacent pixel locations along the positive normal direction of the incoherent boundary segment of the lithium battery sub-region. This represents the temperature gradient magnitude at the current pixel location on the incoherent boundary segment of the lithium battery sub-region. This represents the subpixel correction constant used to avoid the subpixel offset denominator being zero. This represents the equivalent physical distance between adjacent pixels in the infrared thermal image pixel coordinate system and in the physical coordinate system of the lithium battery pack surface.

[0063] Based on the sub-pixel fine boundary of the lithium battery sub-region, the boundary replacement and region closure of the pixel-scale coarse risk mask of the lithium battery sub-region are performed to output the precise positioning mask of the lithium battery sub-region.

[0064] In Example 1, the sub-pixel fine boundary of the lithium battery sub-region is enclosed to form the internal region of the sub-pixel fine boundary of the lithium battery sub-region: when the pixel coordinates in the infrared thermal image pixel coordinate system are located in the internal region of the sub-pixel fine boundary of the lithium battery sub-region, the mask state of the corresponding pixel coordinates is set to belong to the precise positioning mask region of the lithium battery sub-region; otherwise, the mask state of the corresponding pixel coordinates is set to not belong to the precise positioning mask region of the lithium battery sub-region, thus obtaining the precise positioning mask of the lithium battery sub-region.

[0065] Based on the precise positioning mask of lithium battery sub-region, the absolute temperature, temperature rise rate, relative temperature difference and thermal runaway risk probability of the target single lithium battery cell are extracted to generate a lithium battery thermal sub-region risk parameter table. In this embodiment, based on a precise positioning mask for lithium battery sub-regions, the absolute temperature, temperature rise rate, relative temperature difference, and probability of thermal runaway risk of the target single lithium battery cell are extracted, including: The target single lithium cell corresponding to the precise positioning mask of the lithium battery sub-region is determined based on the lithium battery thermal sub-region mapping table; In Example 1, the number of pixels that are determined to belong to the precise positioning mask area of ​​the lithium battery sub-region within the pixel coordinate range of the corresponding single lithium battery cell in the infrared thermal image pixel coordinate system is divided by the number of pixels of the corresponding single lithium battery cell in the infrared thermal image pixel coordinate system to obtain the coverage area ratio of the precise positioning mask of the lithium battery sub-region.

[0066] The target single lithium battery cell is the single lithium battery cell with the largest proportion of the mask coverage area for precise positioning of the lithium battery sub-region.

[0067] Based on the reconstructed temperature response value of the lithium battery sub-region precisely located mask coverage, the absolute temperature of the target single lithium battery cell is extracted. In Example 1, the reconstructed temperature response values ​​corresponding to all pixel coordinate positions within the precise positioning mask of the target single lithium battery cell in the lithium battery sub-region are summed and divided by the number of pixels of the target single lithium battery cell within the precise positioning mask of the lithium battery sub-region to obtain the absolute temperature of the target single lithium battery cell.

[0068] The temperature rise rate of the target single lithium battery cell is calculated based on the absolute temperature of consecutive frames. The relative temperature difference of the target lithium battery cell is calculated by subtracting the average reference temperature of the non-target single lithium battery cell within the same lithium battery thermal sub-region from the absolute temperature. In Example 1, non-target single lithium battery cells are single lithium battery cells other than the target single lithium battery cells within the same lithium battery thermal sub-region. The average reference temperature of non-target single lithium battery cells within the same lithium battery thermal sub-region represents the overall control thermal field temperature level of non-target single lithium battery cells within the same lithium battery thermal sub-region.

[0069] The absolute temperature, temperature rise rate, and relative temperature difference of the target single lithium battery cell are normalized to calculate the probability of thermal runaway risk of the target single lithium battery cell. In Example 1, the absolute temperature, temperature rise rate, and relative temperature difference of the target single lithium battery cell after normalization are weighted and summed with bias terms. The summation result is then subjected to Sigmoid mapping to obtain the probability of thermal runaway risk.

[0070] The target single lithium battery cell identifier, the absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability of the target single lithium battery cell are bound together to generate a lithium battery thermal sub-region risk parameter table.

[0071] Input the risk parameter table of each lithium battery thermal sub-region into the hierarchical risk model, calculate the multi-level thermal runaway risk score according to the lithium battery thermal sub-region, lithium battery module and lithium battery pack level, and obtain the lithium battery thermal runaway risk score list. Compare the multi-level thermal runaway risk score list of lithium batteries with the lithium battery dynamic threshold set according to the perturbation monitoring segment, the gradual segment and the diffusion risk segment to generate the lithium battery segmented thermal runaway early warning list.

[0072] In this embodiment, the risk parameter table of each lithium battery thermal subregion is input into the hierarchical risk model, including: Input the risk parameter table of each lithium battery thermal subregion into the hierarchical risk model to calculate the lithium battery thermal subregion risk score at the lithium battery thermal subregion level. In Example 1, the hierarchical risk model is a hierarchical weighted fusion model based on the physical assembly level of the lithium battery pack. The hierarchical risk model aggregates the risk parameters of each lithium battery thermal sub-region step by step according to the hierarchical relationship between the lithium battery thermal sub-region, the lithium battery module, and the entire lithium battery pack.

[0073] The thermal runaway risk score of the lithium battery is obtained by weighting and summing the target single lithium battery cell identifier, absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability.

[0074] Based on the risk score of the lithium battery thermal sub-region, the lithium battery modules to which the lithium battery thermal sub-region belongs are aggregated, and the risk score of the lithium battery module at the lithium battery module level is calculated. ; in, Indicates the first The lithium battery thermal sub-region is in the first The corresponding lithium battery thermal sub-region risk score for the frame. This indicates the serial number of the lithium battery module. Indicates the first The set of lithium battery thermal sub-region indices contained in each lithium battery module. Indicates the first The number of lithium battery thermal sub-regions contained in a lithium battery module. Indicates the first The lithium battery module is in the first The risk score of the lithium battery module corresponding to the frame. This represents the aggregation weight corresponding to the highest lithium battery thermal sub-region risk score in the lithium battery module level. This represents the aggregate weight corresponding to the average lithium battery thermal sub-region risk score in the lithium battery module level.

[0075] Based on the risk score of the lithium battery module, the risk score of the entire lithium battery pack is calculated by aggregating the risk scores of the entire lithium battery pack. ; in, This represents the set of serial numbers of the lithium battery modules contained in a complete lithium battery pack. This indicates the number of lithium battery modules contained in the complete lithium battery pack. Indicates the first The risk score of the entire lithium battery pack corresponding to the frame. This represents the aggregate weight corresponding to the highest risk score of the lithium battery module within the entire lithium battery pack. This represents the aggregate weight corresponding to the average risk score of lithium battery modules in the entire lithium battery pack.

[0076] By binding the risk scores of the lithium battery thermal sub-region, the lithium battery module, and the lithium battery pack, a multi-level thermal runaway risk score list for lithium batteries is obtained. Based on the multi-level thermal runaway risk score list of lithium batteries, set the dynamic thresholds of lithium batteries corresponding to the perturbation monitoring segment, the gradual segment, and the diffusion risk segment. In Example 1, the dynamic threshold of the lithium battery includes the dynamic threshold of the perturbation monitoring segment of the thermal sub-region of the lithium battery, the dynamic threshold of the progressive segment of the thermal sub-region of the lithium battery, the dynamic threshold of the diffusion risk segment of the lithium battery module, and the dynamic threshold of the diffusion risk segment of the entire lithium battery pack. The dynamic threshold of the lithium battery thermal sub-region perturbation monitoring segment is obtained by summing the product of the average risk score of the lithium battery thermal sub-region in the frames before the current frame that did not trigger thermal runaway warning and the threshold adjustment coefficient of the perturbation monitoring segment multiplied by the standard deviation of the risk score of the lithium battery thermal sub-region in the frames before the current frame that did not trigger thermal runaway warning.

[0077] The average risk score of the lithium battery thermal sub-region in the frames before the current frame that did not trigger thermal runaway warning is summed with the product of the progressive segment threshold adjustment coefficient multiplied by the standard deviation of the risk score of the lithium battery thermal sub-region in the frames before the current frame that did not trigger thermal runaway warning. This product is used to obtain the progressive segment dynamic threshold of the lithium battery thermal sub-region.

[0078] The dynamic threshold of the lithium battery module diffusion risk segment is obtained by summing the product of the average risk score of the lithium battery module in the frames before the current frame where no thermal runaway warning was triggered, multiplied by the threshold adjustment coefficient of the lithium battery module diffusion risk segment by the standard deviation of the risk score of the lithium battery module in the frames before the current frame where no thermal runaway warning was triggered.

[0079] The dynamic threshold of the lithium battery pack diffusion risk segment is obtained by summing the product of the average risk score of the entire lithium battery pack in the frames before the current frame where no thermal runaway warning was triggered, multiplied by the threshold adjustment coefficient of the lithium battery pack diffusion risk segment by the standard deviation of the risk score of the entire lithium battery pack in the frames before the current frame where no thermal runaway warning was triggered.

[0080] The list of multi-level thermal runaway risk scores for lithium batteries is compared with the dynamic thresholds of lithium batteries corresponding to the perturbation monitoring segment, the gradual segment, and the diffusion risk segment to determine the segmented thermal runaway warning type of lithium battery corresponding to each thermal sub-region.

[0081] In this embodiment, the types of lithium battery segmented thermal runaway warnings include: Early warning of perturbation monitoring segment: The risk score of the lithium battery thermal sub-region is not less than the dynamic threshold of the perturbation monitoring segment of the lithium battery thermal sub-region, and the risk score of the lithium battery thermal sub-region is less than the dynamic threshold of the progressive segment of the lithium battery thermal sub-region. Progressive stage warning: The risk score of the lithium battery thermal sub-region is not less than the dynamic threshold of the progressive stage of the lithium battery thermal sub-region, and the risk score of the lithium battery module is less than the dynamic threshold of the diffusion risk stage of the lithium battery module. Early warning of diffusion risk segment: The risk score of the lithium battery module is not less than the dynamic threshold of the diffusion risk segment of the lithium battery module, or the risk score of the entire lithium battery pack is not less than the dynamic threshold of the diffusion risk segment of the entire lithium battery pack. No warning: The risk score of the lithium battery thermal sub-region is less than the dynamic threshold of the lithium battery thermal sub-region perturbation monitoring segment.

[0082] Example 2: Reference Figures 2-4During the operation of a lithium battery pack in an energy storage system, the implementers selected a lithium battery pack consisting of four lithium battery modules as the verification object. The structure of the lithium battery pack is shown in the figure. The four lithium battery modules are arranged in a 2×2 configuration and are designated as module M1, module M2, module M3, and module M4. Each lithium battery module contains 96 individual lithium battery cells in 8 columns × 12 rows, and the entire pack contains a total of 384 individual lithium battery cells. Each individual lithium battery cell is a prismatic lithium battery cell, regularly arranged along the surface of the lithium battery pack. Thermally conductive silicone pads, insulating supports, and metal pressure plates are placed between adjacent individual lithium battery cells. Busbars and high-voltage harnesses are arranged above the modules. The busbars extend laterally along the upper side of the modules, and the high-voltage harnesses are connected to the busbars along the edge of the lithium battery pack. An infrared camera is fixedly installed at the observation window above the lithium battery pack to acquire raw low signal-to-noise ratio thermal image sequences of the lithium battery pack from a top-down perspective. The infrared thermal image resolution is 640×512, and the sampling frequency is 2 frames / second. The system establishes a mapping relationship between the infrared thermal image pixel coordinate system and the physical coordinate system of the lithium battery pack surface through the homography mapping matrix H. The infrared thermal image pixel coordinate system is represented by the horizontal pixel coordinate u and the vertical pixel coordinate v, while the physical coordinate system of the lithium battery pack surface is represented by x and y. The unit of the physical coordinates is mm.

[0083] During normal operation of the lithium battery pack, the system recorded that the busbar surface temperature was typically between 38.5℃ and 42.8℃, the temperature near the high-voltage harness was typically between 36.9℃ and 40.6℃, while the surface temperature of a normal individual lithium battery cell was typically between 30.8℃ and 34.7℃. The standard deviation of the background noise generated by the infrared focal plane array after continuous operation was approximately 0.42℃. When a single lithium battery cell experiences an early micro-short circuit, its initial temperature rise is typically only 1.5℃ to 3.5℃, significantly lower than the normal operating temperature of the busbar and high-voltage harness, and easily mixed with the background noise of the infrared focal plane array, normal heat source diffusion, and the heat transfer response of adjacent cells.

[0084] Before system deployment, implementers collected normal operation samples as training samples. The training samples included 280,000 frames of normal charging samples, 310,000 frames of normal discharging samples, 180,000 frames of standby temperature equalization samples, 62,000 frames of samples showing high busbar temperature but normal individual lithium-ion cells, 41,000 frames of samples showing high high voltage harness temperature but normal individual lithium-ion cells, and 26,000 frames of infrared focal plane array background noise enhancement samples, totaling 899,000 frames. Validation samples included 120,000 frames of normal samples under different charging states and 48,000 frames of localized micro-short circuit samples from Example 2. In the localized micro-short circuit samples of Example 2, the early temperature rise range of abnormal individual lithium-ion cells was 1.2℃ to 5.0℃, with samples below 3℃ accounting for 57.4%. Traditional infrared fixed threshold methods use a 5℃ relative temperature difference as the alarm condition, traditional contact-type BMS temperature threshold methods use 45℃ as the alarm condition, and traditional CNN segmentation methods use whole-pack infrared images for end-to-end risk area segmentation.

[0085] During a verification process, the system acquired raw low signal-to-noise ratio (SNR) thermal imaging sequences of the lithium battery pack during operation. In these sequences, a single lithium-ion cell near the lower right of module M2 began to exhibit localized abnormal heat leakage. This cell was located within the key lithium battery thermal sub-region ID-Z2-04 in the image. Three minutes after the anomaly began, the temperature in the central region of this single lithium-ion cell rose from 32.1℃ to 33.0℃, only 0.8℃ higher than the average temperature of non-target single lithium-ion cells within the same thermal sub-region. The traditional infrared fixed threshold method did not trigger an alarm, and the reading from the traditional contact-type BMS temperature sensor rose from 31.9℃ to 32.1℃, also without triggering an alarm. At this point, the highest temperature in the busbar area of ​​the infrared thermal image reached 41.6℃, and the highest temperature in the high-voltage harness area reached 39.8℃, both significantly higher than the early temperature rise area of ​​the abnormal single lithium-ion cell.

[0086] The system establishes a lithium battery pack structure based on the arrangement of individual lithium-ion cells, their projected dimensions, the boundaries of the lithium battery modules, and the surface contour of the lithium battery pack. It then divides the surface of the lithium battery pack according to the physical adjacency of the cells, the correlation of heat dissipation paths, and the similarity of isothermal distributions. In this lithium battery pack, the median nearest neighbor center distance of individual lithium-ion cells is 34 mm, and a surface distance threshold of 40 mm is used as the threshold for determining the physical adjacency of individual lithium-ion cells. Based on this threshold, the system excludes individual lithium-ion cells that cross module gaps, cross large-distance indirect heat transfer paths, or are spatially discontinuous. The surface of the lithium battery pack is divided into 16 lithium battery thermal sub-regions, with each lithium battery module divided into 4 lithium battery thermal sub-regions. Each lithium battery thermal sub-region contains 24 individual lithium-ion cells (4×6). Among them, module M1 corresponds to ID-Z1-01, ID-Z1-02, ID-Z1-03 and ID-Z1-04, module M2 corresponds to ID-Z2-01, ID-Z2-02, ID-Z2-03 and ID-Z2-04, module M3 corresponds to ID-Z3-01, ID-Z3-02, ID-Z3-03 and ID-Z3-04, and module M4 corresponds to ID-Z4-01, ID-Z4-02, ID-Z4-03 and ID-Z4-04. The individual lithium battery cell exhibiting abnormal heat leakage is located in the lithium battery thermal sub-region ID-Z2-04. This lithium battery thermal sub-region contains 24 individual lithium battery cells in 4 columns × 6 rows. Its coordinate range in the infrared thermal image pixel coordinate system is u=210–365, v=150–292, corresponding to physical dimensions of approximately 4 columns ≈ 272 mm and 6 rows ≈ 408 mm.

[0087] The system rasterizes and segments the time-consistent low signal-to-noise ratio (SNR) thermal image sequence of lithium batteries according to the lithium battery thermal sub-region mapping table. For lithium battery thermal sub-region ID-Z2-04, the system generates a corresponding lithium battery thermal sub-region raster mask and extracts the original low SNR infrared image of this sub-region according to the pixel range of u=210–365 and v=150–292. Five minutes after the anomaly started, the average temperature of the target single lithium battery cell area in the original low SNR infrared image of the sub-region was 33.5℃, and the average temperature of the non-target single lithium battery cells in the same sub-region was 32.4℃, with a relative temperature difference of 1.1℃. Since ID-Z2-04 is located in the lower right local area of ​​module M2, close to the thermal diffusion edge of the upper busbar and high-voltage harness of the module, the standard deviation of the local background temperature fluctuation in this sub-region in the infrared thermal image reached 0.44℃. The weak temperature rise boundary presents a discontinuous and unstable patchy response in the original low SNR infrared image.

[0088] The system inputs the original low signal-to-noise ratio infrared image corresponding to ID-Z2-04 into the lithium battery region autoencoder. The encoder of the lithium battery region autoencoder extracts latent features of the local temperature gradient, low-frequency background heat source distribution, infrared focal plane array background noise, and weak temperature rise response within this sub-region, and the decoder outputs the residual thermal interference map of the lithium battery sub-region. After residual stripping, the standard deviation of temperature fluctuation in the non-target region within ID-Z2-04 decreases from 0.44℃ to 0.16℃, the slowly varying background response formed by the heat diffusion from the busbars at the edge of this sub-region is weakened by approximately 63.2%, and the intensity of the infrared focal plane array stripe background noise decreases from the original mean square response of 0.178 to 0.052. At the same time, the reconstructed temperature response of the central region of the abnormal single lithium battery cell remains at 33.4℃ and has not been smoothed out as background noise, indicating that the lithium battery region autoencoder retains the weak temperature rise response corresponding to the very early abnormal heat leakage within the sub-region.

[0089] Seven minutes after the anomaly began, the system input the reconstructed thermal field sequence of the lithium battery sub-region formed by ID-Z2-04 into the lithium battery region SwinTransformer network. The lithium battery region SwinTransformer network divided the reconstructed thermal field map of the lithium battery sub-region into local thermal image blocks at the Tokens scale, and then performed non-overlapping local window self-attention calculation. The boundary Tokens between the anomalous single lithium battery cell and its left and right adjacent single lithium battery cells showed a continuously increasing thermal field correlation weight, and the average thermal field feature correlation weight within the window increased from 0.18 in normal samples to 0.37. After the moving window offset, the adjacent cell boundary tokens that were originally distributed on both sides of the boundary of different local windows enter the same moving local window. The cross-window thermal field correlation weight increases to 0.46, while in normal high-load samples, this index is usually below 0.25. The system generates a coarse risk mask of the block-level lithium battery sub-region at the token scale. The position with the highest coarse risk probability falls in the upper part of the center of the target single lithium battery cell in ID-Z2-04, with a probability of 0.64, which exceeds the coarse risk probability threshold of 0.52 set according to the token-level probability distribution of normal samples.

[0090] Due to the presence of thermally conductive silicone pads, insulating supports, and metal pressure plates inside the lithium battery pack, abnormal heat began to spread to the edges of adjacent lithium battery cells within ID-Z2-04. Eleven minutes after the anomaly began, an elliptical thermal halo region formed in the original infrared thermal image, covering the target lithium battery cell and the edges of two adjacent lithium battery cells. Traditional infrared fixed threshold methods classified this elliptical region as a high-temperature area, resulting in a localization result covering all three lithium battery cells with a maximum localization deviation of approximately 17.8 mm. While traditional CNN segmentation methods could output the risk area, their mask boundaries were dragged by the thermal halo, resulting in the target lithium battery cell covering 56.3% of the area, while adjacent lithium battery cells covered 24.1% and 19.6% respectively, leading to unclear target boundaries.

[0091] This method backprojects a block-level coarse risk mask of the lithium battery sub-region at the token scale onto the infrared thermal image pixel coordinate system to obtain a pixel-level coarse risk mask of the lithium battery sub-region. Differences in dilation and corrosion are then extracted to form a coarse risk boundary band for the lithium battery sub-region. A lithium battery point quadtree partitioning model performs recursive partitioning within this boundary band, with a preset threshold of 0.04℃² for the lithium battery boundary temperature response variance and a preset minimum partitioning size of 4×4 pixels. When the temperature response variance within a given lithium battery point quadtree node is 0.087℃² and the node size is 16×16 pixels, the system further divides this node into four child nodes. When the temperature response variance within another node drops to 0.026℃² or the node size reaches 4×4 pixels, the system identifies it as a leaf node of the lithium battery sub-region boundary quadtree. The coarse risk boundary band within ID-Z2-04 is divided into 186 leaf nodes of the boundary quadtree.

[0092] The system calculates the temperature gradient field of the lithium battery sub-region based on the reconstructed thermal field map of the ID-Z2-04. In the continuous thermal halo region between the target lithium battery cell and the adjacent thermally diffused lithium battery cells, the lateral temperature gradient amplitude is mostly between 0.18℃ / mm and 0.29℃ / mm. However, near the actual physical boundary of the target lithium battery cell, due to abrupt changes in the heat conduction path and material interface, the normal temperature gradient amplitude exhibits local peaks ranging from 0.53℃ / mm to 0.68℃ / mm. The incoherent region detector detects the discontinuity of the temperature gradient between the leaf nodes of the quadrilateral leaf nodes at the boundary of adjacent lithium battery sub-regions. The incoherent boundary response between two adjacent leaf nodes at the target boundary reaches 0.47, which is higher than the incoherent boundary response threshold of 0.35, and is therefore identified as an incoherent boundary segment of the lithium battery sub-region. The incoherent boundary response of the ordinary thermally diffused smooth region is 0.18 to 0.27, and is not identified as a valid boundary segment.

[0093] During subpixel-level geometric refinement, the system selects the temperature gradient magnitudes of three consecutive pixel positions along the normal direction of the incoherent boundary segment. In Example 2, the temperature gradient magnitude of adjacent pixels in the negative normal direction of a certain boundary point is 0.49℃ / mm, the current pixel temperature gradient magnitude is 0.66℃ / mm, and the temperature gradient magnitude of adjacent pixels in the positive normal direction is 0.58℃ / mm. Combined with the pixel equivalent physical spacing of 0.72mm, the subpixel offset of this boundary point is calculated to be approximately 0.21mm. The system sequentially connects all subpixel fine boundary points and performs boundary replacement and region closure on the coarse risk mask of the lithium battery sub-region at the pixel scale, outputting a precise positioning mask for the lithium battery sub-region. In the precise positioning mask, the coverage area of ​​the target single lithium battery cell accounts for 84.7%, and the coverage area of ​​two adjacent single lithium battery cells decreases to 8.9% and 6.4%, respectively. Finally, the system determines the target single lithium battery cell as the single lithium battery cell with the largest coverage area within ID-Z2-04.

[0094] During the risk parameter extraction process, the system calculates the absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability of the target single lithium battery cell based on the reconstructed temperature response value of the precise positioning mask covering the lithium battery sub-region. At the 7-minute mark after the anomaly began, the absolute temperature of the target single lithium battery cell was 34.1℃, the temperature rise rate was 0.18℃ / min, the relative temperature difference was 1.4℃, and the thermal runaway risk probability was 0.43. At the 11-minute mark after the anomaly began, the absolute temperature of the target single lithium battery cell was 35.6℃, the temperature rise rate was 0.31℃ / min, the relative temperature difference was 2.7℃, and the thermal runaway risk probability was 0.71. At the 16-minute mark after the anomaly began, the absolute temperature of the target single lithium battery cell was 38.2℃, the temperature rise rate was 0.47℃ / min, the relative temperature difference was 4.6℃, and the thermal runaway risk probability was 0.91. The contact temperature sensor reading only increased from 32.0℃ to 33.6℃, ​​still below the traditional BMS temperature alarm threshold.

[0095] The system inputs the risk parameter tables of each lithium battery thermal sub-region into the hierarchical risk model to obtain the risk scores of the lithium battery thermal sub-region, the lithium battery module, and the entire lithium battery pack. Seven minutes after the anomaly began, the lithium battery thermal sub-region risk score for ID-Z2-04 was 0.44, the dynamic threshold for the perturbation monitoring segment was 0.39, and the dynamic threshold for the gradual transition segment was 0.62. Therefore, the system marked ID-Z2-04 as a warning for the perturbation monitoring segment. Eleven minutes after the anomaly began, the lithium battery thermal sub-region risk score for ID-Z2-04 rose to 0.73, exceeding the dynamic threshold for the gradual transition segment of 0.63, while the risk score for module M2 was 0.57, lower than the dynamic threshold for the lithium battery module diffusion risk segment of 0.70. Therefore, the system marked ID-Z2-04 as a warning for the gradual transition segment. Sixteen minutes after the anomaly began, the risk score of module M2 rose to 0.74, exceeding the dynamic threshold of 0.71 for the diffusion risk segment of the lithium battery module. At the same time, the risk score of the entire lithium battery pack was 0.48, which was lower than the dynamic threshold of 0.66 for the diffusion risk segment of the entire lithium battery pack. Therefore, the system marked ID-Z2-04 as a diffusion risk segment warning and recorded the lithium battery thermal sub-region identifier ID-Z2-04, lithium battery module serial number M2, target single lithium cell identifier, lithium battery segment thermal runaway warning type, lithium battery thermal sub-region risk score, lithium battery module risk score, and lithium battery pack risk score in the lithium battery segment thermal runaway warning list.

[0096] In Example 2, the early warning time for the perturbation monitoring segment is 7 minutes after the anomaly begins. The traditional CNN segmentation method first forms an effective risk area at 13 minutes after the anomaly begins, the traditional infrared fixed threshold method first triggers an alarm at 18 minutes after the anomaly begins, and the traditional BMS temperature threshold method first triggers an alarm at 31 minutes after the anomaly begins. This method is approximately 6 minutes faster than the traditional CNN segmentation method, approximately 11 minutes faster than the traditional infrared fixed threshold method, and approximately 24 minutes faster than the traditional BMS temperature threshold method.

[0097] In terms of positioning, the target single lithium battery cell positioning error of this method is 1.2mm, while that of the traditional CNN segmentation method is 7.4mm, the traditional infrared fixed threshold method is 17.8mm, and the traditional BMS temperature threshold method cannot locate the single lithium battery cell. Regarding mask accuracy, the intersection-over-union (IoU) ratio between the precise positioning mask of the lithium battery sub-region in this method and the manually calibrated target boundary is 0.85, while that of the traditional CNN segmentation method is 0.62, and that of the traditional infrared fixed threshold method is 0.38.

[0098] In the test samples containing multiple sets of normal high-load interference and micro-short circuits in Example 2, the recall rate of this method for samples with slight temperature rises from 1.5℃ to 3.0℃ was 92.8%, compared to 71.6% for the traditional CNN segmentation method, 38.4% for the traditional infrared fixed threshold method, and 0% for the traditional BMS temperature threshold method. In the normal high-load interference samples, when the busbar temperature reached a maximum of 43.5℃ and there were no abnormalities in the individual lithium-ion cells, the false alarm rate of the traditional infrared fixed threshold method was 26.7%, the false alarm rate of the traditional CNN segmentation method was 13.9%, and the false alarm rate of this method was 2.1%. In samples with significant thermal halo diffusion, the average positioning error of the individual lithium-ion cells using this method was 1.5mm, compared to 6.8mm for the traditional CNN segmentation method and 16.2mm for the traditional infrared fixed threshold method.

[0099] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of thermal runaway in lithium batteries based on deep learning, characterized in that, include: The original low signal-to-noise ratio thermal image sequence of the lithium battery pack during operation was collected and preprocessed to obtain a time-consistent low signal-to-noise ratio thermal image sequence of the lithium battery. Based on the physical structure, heat dissipation path and isothermal distribution of the lithium battery pack, the surface of the lithium battery pack is divided into several lithium battery thermal sub-regions, and a lithium battery thermal sub-region mapping table is generated. The time-consistent low signal-to-noise ratio thermal image sequence of lithium battery is rasterized and segmented according to the lithium battery thermal sub-region mapping table to obtain the original low signal-to-noise ratio infrared sequence. For each lithium battery thermal sub-region, the corresponding original low signal-to-noise ratio infrared sequence is input into the lithium battery region autoencoder to output the lithium battery sub-region reconstructed thermal field sequence. The reconstructed thermal field sequence of the lithium battery sub-region is input into the Swin-Transformer network of the lithium battery region to extract the weak spatial thermal coupling characteristics of adjacent cells in the lithium battery thermal sub-region caused by early local micro short circuits, and output a rough risk mask of the lithium battery sub-region. The coarse risk mask of the lithium battery sub-region is input into the lithium battery point quadtree partitioning and incoherent region detector. The boundary of the coarse risk mask of the lithium battery sub-region is geometrically refined, and the precise positioning mask of the lithium battery sub-region is output. Based on the precise positioning mask of lithium battery sub-region, the absolute temperature, temperature rise rate, relative temperature difference and thermal runaway risk probability of the target single lithium battery cell are extracted to generate a lithium battery thermal sub-region risk parameter table. Input the risk parameter table of each lithium battery thermal sub-region into the hierarchical risk model, calculate the multi-level thermal runaway risk score according to the lithium battery thermal sub-region, lithium battery module and lithium battery pack level, and obtain the lithium battery thermal runaway risk score list. Compare the multi-level thermal runaway risk score list of lithium batteries with the lithium battery dynamic threshold to generate the lithium battery segmented thermal runaway early warning list.

2. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 1, characterized in that, The preprocessing includes time synchronization and formatting of the original low signal-to-noise ratio thermal image sequence.

3. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 2, characterized in that, Based on the physical structure, heat dissipation path, and isothermal distribution of the lithium battery pack, the surface of the lithium battery pack is divided into several lithium battery thermal sub-regions, including: Obtain the arrangement position of individual lithium cells in the lithium battery pack, the projected size of individual lithium cells, the boundary of the lithium battery module, and the surface contour of the lithium battery pack, and construct the lithium battery pack structure set. The distance between the surface structures of any two individual lithium battery cells is calculated based on the lithium battery pack structure set, and the adjacency matrix of the lithium battery pack structure is constructed. Based on the heat dissipation path of the lithium battery pack, and combined with the adjacency matrix of the lithium battery pack structure and the distance of the lithium battery surface structure, a heat dissipation path correlation matrix of the lithium battery pack is constructed. The average surface temperature of each individual lithium battery cell under static benchmark conditions is obtained, and the isothermal distribution matrix of the lithium battery pack is constructed. The adjacency matrix of the lithium battery pack structure, the correlation matrix of the lithium battery pack heat dissipation path, and the isothermal distribution matrix of the lithium battery pack are weighted and fused to generate the thermal sub-region correlation matrix of the lithium battery. Based on the correlation matrix of lithium battery thermal sub-regions, the surface of the lithium battery pack is divided into spectral clusters and thermal aggregations to obtain a set of lithium battery thermal sub-regions; Each lithium battery thermal sub-region is bound to the infrared thermal image coordinate range of the lithium battery pack surface, the individual lithium cell number, and the lithium battery module number to generate a lithium battery thermal sub-region mapping table. Each lithium battery thermal subregion is assigned a lithium battery thermal subregion identifier based on the lithium battery thermal subregion mapping table.

4. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 3, characterized in that, The step of rasterizing and segmenting the time-consistent low signal-to-noise ratio thermal image sequence of lithium batteries according to the lithium battery thermal sub-region mapping table includes: Based on the lithium battery thermal sub-region mapping table, the coordinate range of each lithium battery thermal sub-region in the infrared thermal image pixel coordinate system is extracted, and a corresponding lithium battery thermal sub-region grid mask is constructed. Each time-consistent low signal-to-noise ratio thermal image of a lithium battery is rasterized according to the thermal sub-region raster mask of the lithium battery to obtain the original low signal-to-noise ratio infrared image. The original low signal-to-noise ratio infrared image is input into the encoder of the corresponding lithium battery region autoencoder to obtain the corresponding lithium battery sub-region thermal field latent features. The decoder of the lithium battery region autoencoder decodes the latent thermal field features of the lithium battery sub-region and outputs the residual thermal interference map of the lithium battery sub-region. The residual temperature response corresponding to the residual thermal interference map of the lithium battery sub-region is extracted from the original low signal-to-noise ratio infrared image, and the reconstructed temperature response value of each pixel coordinate position is calculated. The reconstructed thermal field map of the lithium battery sub-region is composed of the reconstructed temperature response values ​​of all pixel coordinate positions and output. The reconstructed thermal field maps of all lithium battery sub-regions are arranged chronologically according to the acquisition time, forming the reconstructed thermal field sequence of the lithium battery sub-regions corresponding to the lithium battery thermal sub-regions.

5. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 4, characterized in that, The step of inputting the reconstructed thermal field sequence of the lithium battery sub-region into the Swin-Transformer network of the lithium battery region includes: Each frame of the reconstructed thermal field map of the lithium battery sub-region is input into the lithium battery sub-region feature embedding layer of the lithium battery region Swin-Transformer network to obtain the lithium battery sub-region thermal field embedding feature map. The thermal field embedding feature map of the lithium battery sub-region is divided into regular windows to obtain a set of non-overlapping local windows of the lithium battery sub-region. For each non-overlapping local window of the lithium battery sub-region, non-overlapping local window self-attention calculation is performed to obtain the thermal field correlation characteristics within the lithium battery sub-region window; All thermal field correlation features within the lithium battery sub-region window are backfilled into the feature coordinates corresponding to the thermal field embedding feature map of the lithium battery sub-region according to the non-overlapping local window positions of the corresponding lithium battery sub-region, thus obtaining the self-attention feature map of the lithium battery sub-region window. Perform a moving window offset on the self-attention feature map of the lithium battery sub-region window to obtain the moving window feature map of the lithium battery sub-region; The moving window feature map of the lithium battery sub-region is re-divided into regular windows, and self-attention calculation based on the moving window is performed to obtain the cross-window thermal field correlation feature map of the lithium battery sub-region. The cross-window thermal field correlation feature map of the lithium battery sub-region is used as the input of the self-attention block of the next lithium battery region. The self-attention calculation of non-overlapping local windows and the self-attention calculation based on moving windows are stacked layer by layer as a self-attention block of the lithium battery region, and the weak spatial thermal strong coupling feature map of the lithium battery sub-region is output. A coarse risk mapping is performed on the weak spatial thermal coupling feature map of the lithium battery sub-region to obtain a coarse risk probability map of the lithium battery sub-region. Based on the coarse risk probability map of the lithium battery sub-region, a block-level coarse risk mask of the lithium battery sub-region at the Tokens scale is generated.

6. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 5, characterized in that, The step of inputting a coarse risk mask of the lithium battery sub-region into the lithium battery point quadtree partitioning and incoherent region detector includes: The rough risk mask of the lithium battery sub-region is back-projected onto the infrared thermal image pixel coordinate system to obtain the pixel-scale rough risk mask of the lithium battery sub-region. Boundary extraction is performed on the coarse risk mask of the lithium battery sub-region at the pixel scale to obtain the coarse risk boundary band of the lithium battery sub-region. The rough risk boundary zone of the lithium battery sub-region is input into the lithium battery point quadtree partitioning model to form several lithium battery point quadtree nodes. When the lithium battery point quadtree node meets the stopping partitioning condition, it is determined as the leaf node of the lithium battery sub-region boundary quadtree. The temperature gradient within the rough risk boundary band of the lithium battery sub-region is calculated based on the reconstructed thermal field map of the lithium battery sub-region, and the temperature gradient field of the lithium battery sub-region is obtained. The quad-leaf tree nodes of the lithium battery sub-region boundary and the temperature gradient field of the lithium battery sub-region are input into the incoherent region detector to detect the discontinuity of the temperature gradient between the quad-leaf tree nodes of adjacent lithium battery sub-region boundaries, and the incoherent boundary response of the lithium battery sub-region is obtained. When the incoherent boundary response of the lithium battery sub-region is not less than the incoherent boundary response threshold, the common boundary between the four-way leaf nodes of two adjacent lithium battery sub-region boundaries is determined as the incoherent boundary segment of the lithium battery sub-region. Subpixel-level geometric refinement is performed on the incoherent boundary segments of the lithium battery sub-region to generate subpixel fine boundaries of the lithium battery sub-region. Based on the sub-pixel fine boundary of the lithium battery sub-region, the boundary replacement and region closure of the pixel-scale coarse risk mask of the lithium battery sub-region are performed to output the precise positioning mask of the lithium battery sub-region.

7. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 6, characterized in that, The method of extracting the absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability of a target single lithium battery cell using a precise positioning mask for lithium battery sub-regions includes: The target single lithium cell corresponding to the precise positioning mask of the lithium battery sub-region is determined based on the lithium battery thermal sub-region mapping table; Based on the reconstructed temperature response value of the lithium battery sub-region precisely located mask coverage, the absolute temperature of the target single lithium battery cell is extracted; The temperature rise rate of the target single lithium battery cell is calculated based on the absolute temperature of consecutive frames. The relative temperature difference of the target lithium battery cell is calculated by subtracting the average reference temperature of the non-target single lithium battery cell within the same lithium battery thermal sub-region from the absolute temperature. The absolute temperature, temperature rise rate, and relative temperature difference of the target single lithium battery cell are normalized to calculate the probability of thermal runaway risk of the target single lithium battery cell. The target single lithium battery cell identifier, the absolute temperature, temperature rise rate, relative temperature difference, and thermal runaway risk probability of the target single lithium battery cell are bound together to generate a lithium battery thermal sub-region risk parameter table.

8. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 7, characterized in that, The step of inputting the risk parameter table of each lithium battery thermal subregion into the hierarchical risk model includes: Input the risk parameter table of each lithium battery thermal subregion into the hierarchical risk model to calculate the lithium battery thermal subregion risk score at the lithium battery thermal subregion level. Based on the risk score of the lithium battery thermal sub-region, the lithium battery modules to which the lithium battery thermal sub-region belongs are aggregated, and the risk score of the lithium battery module at the lithium battery module level is calculated. Based on the risk score of the lithium battery module, the risk score of the entire lithium battery pack is calculated by aggregating the risk scores of the entire lithium battery pack. By binding the risk scores of the lithium battery thermal sub-region, the lithium battery module, and the lithium battery pack, a multi-level thermal runaway risk score list for lithium batteries is obtained. Based on the multi-level thermal runaway risk score list of lithium batteries, set the dynamic thresholds of lithium batteries corresponding to the perturbation monitoring segment, the gradual segment, and the diffusion risk segment. The list of multi-level thermal runaway risk scores for lithium batteries is compared with the dynamic thresholds of lithium batteries corresponding to the perturbation monitoring segment, the gradual segment, and the diffusion risk segment to determine the segmented thermal runaway warning type of lithium battery corresponding to each thermal sub-region.

9. The method for early warning of thermal runaway in lithium batteries based on deep learning according to claim 8, characterized in that, The types of segmented thermal runaway warnings for lithium batteries include: Early warning of perturbation monitoring segment: The risk score of the lithium battery thermal sub-region is not less than the dynamic threshold of the perturbation monitoring segment of the lithium battery thermal sub-region, and the risk score of the lithium battery thermal sub-region is less than the dynamic threshold of the progressive segment of the lithium battery thermal sub-region. Progressive stage warning: The risk score of the lithium battery thermal sub-region is not less than the dynamic threshold of the progressive stage of the lithium battery thermal sub-region, and the risk score of the lithium battery module is less than the dynamic threshold of the diffusion risk stage of the lithium battery module. Early warning of diffusion risk segment: The risk score of the lithium battery module is not less than the dynamic threshold of the diffusion risk segment of the lithium battery module, or the risk score of the entire lithium battery pack is not less than the dynamic threshold of the diffusion risk segment of the entire lithium battery pack. No warning: The risk score of the lithium battery thermal sub-region is less than the dynamic threshold of the lithium battery thermal sub-region perturbation monitoring segment.

Images