A method for monitoring thermal runaway gases in lithium batteries based on infrared hyperspectral remote sensing

Abstract

This invention discloses a method for monitoring thermal runaway gases in lithium batteries based on infrared hyperspectral remote sensing. The method includes: acquiring time-series hyperspectral data of the entire thermal runaway process of a lithium battery; extracting pixel-by-pixel background spectra before thermal runaway occurs, and identifying gas plume regions during the thermal runaway process based on preset gas characteristic absorption peaks as effective pixels, using their spectral curves as target spectra; subtracting the background spectrum and the contribution of the gas's own thermal radiation to obtain differential spectra; constructing a physical information neural network, using the target spectrum and differential spectrum of each effective pixel as input, and sequentially performing dynamic range compression, spectral self-attention weighted enhancement, graph attention spatial aggregation, and unsupervised training with composite loss based on fixed physical quantities as priors, outputting the gas column concentration of each effective pixel; and evaluating the gas generation and diffusion process based on the spatiotemporal distribution of the column concentration. This method achieves non-contact, real-time, multi-component monitoring and early warning of thermal runaway gases in lithium batteries.

Description

Technical Field

[0001] This invention relates to the field of lithium battery safety monitoring technology, and more specifically to a method for monitoring thermal runaway gases in lithium batteries based on infrared hyperspectral remote sensing. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, energy storage power stations, and portable electronic devices due to their high energy density and long cycle life. However, under conditions such as overheating, overcharging, and mechanical abuse, lithium-ion batteries are prone to thermal runaway, releasing large amounts of flammable and toxic gases, such as dimethyl carbonate (DMC), methane (CH4), carbon monoxide (CO), carbon dioxide (CO2), hydrogen fluoride (HF), and hydrogen chloride (HCl), seriously threatening human safety and the ecological environment. Therefore, early gas monitoring during the thermal runaway process of lithium-ion batteries is of great significance for fire early warning and ensuring public safety.

[0003] Existing technologies for monitoring thermal runaway gases in lithium batteries mainly include: contact sensors (such as electrochemical sensors and semiconductor sensors) are deployed inside the battery module or at the exhaust port, but they have limitations such as slow response, susceptibility to poisoning, and difficulty in covering the entire space; laboratory methods such as gas chromatography-mass spectrometry (GC-MS) require sampling and analysis and cannot monitor in real time; infrared thermal imaging can only sense temperature changes and cannot identify the types of gases.

[0004] Therefore, how to provide a non-contact, real-time, multi-component gas monitoring method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a method for monitoring thermal runaway gases in lithium batteries based on infrared hyperspectral remote sensing, which overcomes or at least partially solves the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: S1. Continuously acquire time-series hyperspectral data of the lithium battery from the beginning of thermal runaway to the end of thermal runaway. S2. From the acquired time-series hyperspectral data, extract the data before thermal runaway as background spectral data. For each spatial location, calculate its average value before thermal runaway as the background spectrum of the pixel at that location. Extract the data during thermal runaway as the spectral data to be processed. Identify the gas plume region in the current frame based on the characteristic absorption peak of the preset gas. Select the pixels in the gas plume region as effective pixels. Use the spectral curve of the effective pixels as the target spectrum of the pixel. S3. For each pixel, subtract the background spectrum corresponding to the pixel and the contribution of gas self-thermal radiation obtained based on the pixel from the target spectrum of the pixel to obtain the corrected differential spectrum. S4. Construct a physical information neural network. The physical information neural network takes the target spectrum of each pixel and its corresponding differential spectrum as input, and combines it with a preset gas absorption cross section to perform the following in sequence: dynamic range compression of the differential spectrum of each pixel, spectral self-attention weighted enhancement of the target spectrum of each pixel, graph attention spatial aggregation of all pixels in the same frame based on neighborhood relationship, and unsupervised training with a composite loss based on a fixed physical quantity as a priori, and outputs the gas column concentration of each pixel. S5. Evaluate the generation and diffusion process of thermal runaway gas based on the spatiotemporal distribution of the gas column concentration.

[0007] Preferably, step S2 specifically includes: Determine the onset of thermal runaway based on the temperature profile of the lithium battery or manual interpretation. Select The average value of several frames of hyperspectral data within a time period is calculated for each spatial location and used as the background spectrum. Select Each frame of data within a time period is used as spectral data to be processed.

[0008] Preferably, the contribution of the gas's own thermal radiation in step S3 is obtained in the following way: For the effective pixel, the radiance of each wavenumber in the target spectrum of the pixel is converted into brightness temperature. The maximum brightness temperature under all wavenumbers is taken as the gas temperature, and then the gas self-radiation is calculated according to Planck's law.

[0009] Preferably, the dynamic range compression of the differential spectrum includes performing a signed logarithmic transformation on the corrected differential spectrum.

[0010] Preferably, the spectral self-attention weighted enhancement includes: normalizing the target spectrum and inputting it into a multilayer perceptron, wherein the input dimension of the multilayer perceptron is the number of bands of the target spectrum, the intermediate layer dimension is half of the input dimension, the output is activated by a sigmoid function to obtain weights, and the weights are multiplied element-wise with the normalized target spectrum to obtain enhanced spectral features.

[0011] Preferably, the graph attention space aggregation based on pixel neighborhood relationships includes: taking each pixel corresponding to the target spectrum as a node and constructing a graph structure with 8-neighborhood connection edges; the initial feature of each node is composed of three parts: the enhanced spectral feature, the differential spectrum after dynamic range compression, and the physical prior column concentration calculated by matched filtering; inputting the graph structure into two layers of four-head graph attention layers in sequence, aggregating the neighborhood node features through self-attention coefficients, and outputting the aggregated feature of each node; mapping the aggregated feature through a fully connected layer to obtain the gas column concentration.

[0012] Preferably, the physical prior column concentration for each pixel is calculated using a matched filter formula:

[0013] in, Indicates the target spectrum, Indicates the background spectrum. This represents the target gas absorption cross section obtained in advance from the HITRAN database. This corresponds to the self-radiation of the gas in the corresponding pixel.

[0014] Preferably, the unsupervised training of the composite loss with fixed physical quantities as priors includes: fixing the gas temperature corresponding to the gas self-thermal radiation contribution obtained in step S3, the background spectrum obtained in step S2, and the gas absorption cross section as constants, and not participating in gradient updates; The loss function for unsupervised training using the composite loss method includes spectral weighted L1 loss, morphological similarity loss, and prior anchoring loss.

[0015] Preferably, the loss is:

[0016] Where N is the total number of wavelength points, The normalized weights for the target gas absorption cross section; To reconstruct the spectrum, specifically the gas column concentration predicted by the neural network. Calculate transmittance Output reconstructed spectrum ; Morphological similarity loss:

[0017] in, The cosine similarity function; Prior anchoring loss:

[0018] in The gas column concentration predicted by the network; Physical prior column concentrations calculated based on the matched filtering principle; The total loss is: .

[0019] Preferably, step S5 specifically includes: assessing the generation time, diffusion path, and peak concentration of the thermal runaway gas based on the spatiotemporal distribution of the gas column concentration; and comparing the gas column concentration with a preset warning threshold, and issuing a warning message when the gas column concentration exceeds the preset threshold.

[0020] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: 1. Non-contact, long-distance detection: No contact with the battery body is required, the integrity of the battery structure is not affected, and real-time monitoring can be performed from a safe distance.

[0021] Early warning capability: The release of gas molecules (such as dimethyl carbonate (DMC), CH4, and C2H4) often precedes a sharp rise in temperature and visible fumes. Hyperspectral imaging can detect these gases at extremely low concentrations (in the ppm·m range) through characteristic absorption peaks.

[0022] 2. Simultaneous identification of multiple components: Different gases have different absorption peaks, and multiple thermal runaway products can be quantified simultaneously through spectral inversion, achieving "fingerprint-level" identification.

[0023] 3. Spatial distribution information: It can draw a two-dimensional concentration distribution map of the gas plume and locate the cell where thermal runaway begins.

[0024] 4. Resistance to environmental interference: Through background subtraction and differential spectroscopy technology, interference from sunlight, environmental heat radiation, etc. can be eliminated.

[0025] 5. Strong early warning capability: Gas release usually occurs before a sharp rise in temperature and visible smoke. This invention can detect characteristic gases at extremely low concentrations (ppm·m level), buying valuable time for emergency response. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0027] Figure 1 This is a flowchart of a lithium battery thermal runaway gas monitoring method based on infrared hyperspectral remote sensing provided in an embodiment of the present invention; Figure 2 This is a graph showing the methane (CH4) column concentration inversion effect in an embodiment of the present invention; Figure 3 This is a graph showing the effect of ethylene (C2H4) column concentration inversion in the embodiments of the invention; Figure 4 This is a graph showing the effect of dimethyl carbonate (DMC) column concentration inversion in the embodiments of the invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Infrared hyperspectral imaging technology simultaneously acquires spatial dimensions and spectral dimensions ( The three-dimensional data of the image contains radiation or reflection information in hundreds of consecutive bands for each pixel. Gas molecules have unique "fingerprint" absorption lines in the mid-infrared band, giving hyperspectral technology the following irreplaceable advantages. Addressing the problems of response lag and poor specificity in existing lithium battery thermal runaway monitoring methods, this invention proposes a new monitoring method utilizing the high-resolution identification capability of infrared hyperspectral imaging technology for gas absorption characteristics.

[0030] like Figure 1 As shown in the figure, this invention discloses a method for monitoring thermal runaway gases in lithium batteries based on infrared hyperspectral remote sensing, comprising: S1. Continuously acquire time-series hyperspectral data of the lithium battery from the beginning of thermal runaway to the end of thermal runaway. S2. From the acquired time-series hyperspectral data, extract the data before thermal runaway as background spectral data. For each spatial location, calculate its average value before thermal runaway as the background spectrum of the pixel at that location. Extract the data during thermal runaway as the spectral data to be processed. Identify the gas plume region in the current frame based on the characteristic absorption peak of the preset gas. Select the pixels in the gas plume region as effective pixels. Use the spectral curve of the effective pixels as the target spectrum of the pixel. S3. For each pixel, subtract the background spectrum corresponding to the pixel and the contribution of gas self-thermal radiation obtained based on the pixel from the target spectrum of the pixel to obtain the corrected differential spectrum. S4. Construct a physical information neural network. The physical information neural network takes the target spectrum and its corresponding differential spectrum of each pixel as input, and combines them with a preset gas absorption cross section to perform the following in sequence: dynamic range compression of the differential spectrum of each pixel, spectral self-attention weighted enhancement of the target spectrum of each pixel, graph attention spatial aggregation of all pixels in the same frame based on neighborhood relationship, and unsupervised training with a composite loss based on a fixed physical quantity as a priori, and outputs the gas column concentration of each pixel. S5. Based on the spatiotemporal distribution of gas column concentration, assess the generation and diffusion process of thermal runaway gas.

[0031] The following is a detailed explanation of steps S1-S5: S101: Data Acquisition In one embodiment, the lithium battery under test (such as an 18650 type or a square pouch cell) is placed on a heating plate, which is heated at a constant rate (e.g., 5°C / min) until the battery experiences thermal runaway. An infrared hyperspectral imager is deployed approximately 1 meter above the battery, with the lens adjusted to cover the battery and the surrounding approximately 1 meter × 1 meter area. The imager operates in the mid-infrared band (e.g., 850–1350 cm⁻¹). - ¹), with a spectral resolution better than 4 cm⁻¹ - ¹, with a spatial resolution better than 1 mrad and a sampling frequency not less than 1 Hz. Data is continuously acquired from the start of heating until the end of thermal runaway (e.g., flame extinguishing or gas dissipation), for a total of N frames of hyperspectral data. .

[0032] S102: Background and Target Spectrum Extraction In one embodiment, the onset time of thermal runaway is determined by synchronously recorded temperature curves. (Usually marked by a sharp rise in battery surface temperature). Select Five frames of data from when the battery had not yet produced gas during the time period were used, and the average value was taken as the background spectrum for each spatial location. Select All frames or keyframes within the time period are used as spectral data to be processed; gas plume regions in the current frame are identified based on the characteristic absorption peaks of preset gases, and pixels within the gas plume regions are selected as valid pixels. The spectral curve of the valid pixels is defined as the target spectrum of that pixel. Subsequent processing is performed only on the valid pixels.

[0033] In this embodiment, the preset gas (i.e., the gas to be monitored) can be pre-set according to actual safety requirements, such as dimethyl carbonate (DMC), methane (CH4), ethylene (C2H4), and other typical gaseous products of lithium battery thermal runaway. Different target gases correspond to different characteristic absorption peaks, and their absorption cross-section data can be obtained in advance from the HITRAN database.

[0034] It should be noted that, to clarify the data hierarchy in this method, the following definitions are used: "A frame" refers to hyperspectral data, which is a complete two-dimensional spatial image acquired at a certain time t, in which each spatial coordinate (x, y) corresponds to a pixel.

[0035] "A pixel" refers to a sampling point located at a fixed spatial coordinate (x, y) in a frame, which contains a complete spectral curve covering all wavenumbers.

[0036] "Effective pixels" refer to pixels in a frame of an image that are selected after the gas plume region is identified and located in the region of the preset gas characteristic absorption peak. The spectral curves of these pixels contain significant gas absorption information.

[0037] "Target spectrum" refers to the spectral vector at a specific pixel location within a frame acquired during thermal runaway. In other words, each frame contains multiple pixels, and each pixel corresponds to a target spectrum.

[0038] In subsequent processing, spectral dimension operations (such as dynamic range compression and self-attention weighting) are performed independently for each pixel; spatial dimension operations (such as attention aggregation) use all pixels in the same frame as nodes and utilize the neighborhood relationships between pixels to pass information.

[0039] S103: Subtracting background and gas intrinsic radiation, the corrected differential spectrum is obtained. In one embodiment, for each frame, each pixel within that frame is traversed, and each pixel is read. In each wave number Radiance value at location / target spectrum and the background spectrum at the corresponding location. Calculate using the following steps: (1) Calculate the brightness temperature:

[0040] Wherein, the first radiation constant is: Second radiation constant: ; (2) Take the maximum value of the brightness temperature of all wavenumbers of the pixel as the gas temperature estimate: ; It should be noted that the radiation spectrum of high-temperature solids exhibits a broad and smooth distribution, lacking selective characteristic peaks within the gas's characteristic absorption band, and only providing slowly varying background radiation. In contrast, the radiation of the target gas exhibits significant wavelength selectivity within its characteristic absorption band, determined by molecular energy level transitions. Therefore, even if the effective pixels include solid surface radiation, the gas contribution can still be distinguished from the solid background radiation by its characteristic spectral shape, and the band corresponding to the maximum brightness temperature across the entire spectrum reflects the gas's own thermal state.

[0041] In one embodiment, for cases where extreme high temperatures cause detector signal saturation, a brightness temperature threshold Tmax = 450 K is set. When the brightness temperature of a pixel exceeds Tmax for all wavenumbers in the characteristic band, the pixel does not perform gas column concentration inversion and is filled by interpolation of effective neighboring pixels.

[0042] (3) Calculate the gas's own radiation contribution:

[0043] (4) Subtract the background spectrum and gas self-radiation to obtain the corrected difference spectrum. :

[0044] S104: Constructing and Training Physical Information Neural Networks This embodiment constructs a physical information neural network, which uses the target spectrum... Corrected differential spectrum and the preset target gas absorption cross section For input, perform the following operations in sequence.

[0045] (1) Dynamic range compression of differential spectroscopy: A signed logarithmic transform is performed on the corrected differential spectrum to compress the high dynamic range:

[0046] Simultaneously, the target spectrum is normalized using the min-max method to obtain the normalized target spectrum. .

[0047] (2) Spectral self-attention weighted enhancement: Normalized target spectrum The input spectrum is a self-attention layer, which consists of a multilayer sensor with an input dimension equal to the number of target spectral bands, intermediate layers with halved dimensions, and an output activated by Sigmoid. The output weights are related to... Element-wise multiplication yields enhanced spectral features. ; This mechanism enables the network to automatically focus on the characteristic absorption peak band of the target gas, suppressing irrelevant bands and noise.

[0048] (3) Physical prior calculation for matched filtering: Using the classic matched filter formula, combined with a preset gas absorption cross section Using the gas self-radiation obtained in step S3, calculate the physical prior column concentration for each pixel:

[0049] Among them, gas absorption cross section Pre-fetched from the HITRAN database, Divide by 1000 and use the resulting prior feature unit to convert to ppm·m (conversion factor: 1 ppm·m = 2.6867 × 10¹). 5 The standardized prior concentration is obtained by using molecules / cm².

[0050] (4) Graph structure construction and node feature splicing: For the current frame of data being processed, all pixels within that frame are treated as graph nodes (limited to pixels covering the lithium battery body and the gas diffusion space above, depending on the experimental field of view). A graph is constructed by connecting edges in 8-neighborhoods. Each node's initial features are composed of three parts: enhanced spectral features. Difference spectrum after logarithmic transformation Standardized prior concentration .

[0051] It should be noted that dynamic range compression and spectral self-attention weighted enhancement are spectral dimension operations performed independently for each pixel; graph attention spatial aggregation is a spatial dimension operation performed simultaneously for all pixels in the same frame.

[0052] (5) Spatial feature aggregation: The graph structure of the current frame is sequentially input into two layers of four-head graph attention layers (GATConv). Each layer dynamically aggregates the features of the neighboring nodes in the frame through self-attention coefficients and outputs the aggregated features of each node.

[0053] (6) Column concentration output: The aggregated features are input into a fully connected layer to map the gas column concentration of that pixel. (Unit: ppm·m)

[0054] (7) Unsupervised training: The gas temperature obtained in step S3 is fixed. The background spectrum obtained in step S2 and gas absorption cross section Since the variable is constant (not involved in gradient updates), the Adam optimizer (learning rate 0.03, iterations 120) is used to minimize the composite loss function. .

[0055] in, , namely spectrally weighted L1 loss, is used to force the network to fit gas characteristic peaks; Where N is the total number of wavelength points, The normalized weights for the target gas absorption cross section; To reconstruct the spectrum, specifically the gas column concentration predicted by the neural network. Calculate transmittance Output reconstructed spectrum ; The calculation is based on the Lambert-Beer law and the infrared radiative transfer equation, and the output is constrained to conform to the radiative transfer equation to ensure that the network has physical interpretability.

[0056] That is, morphological similarity loss, constraining the waveform of the reconstructed spectrum to be consistent with that of the target spectrum; in, The cosine similarity function; This refers to the prior anchoring loss, which guides the network output to approximate the physical prior column concentration.

[0057] After training converges, the network directly outputs the column concentration for each pixel. Repeat the above process for all target frames to obtain the spatiotemporal distribution of each gas column concentration. For different target gases (such as CH4, C2H4, etc.), only the absorption cross-section file needs to be changed and the characteristic bands adjusted accordingly; the rest of the process remains the same.

[0058] S105: Spatiotemporal Distribution Assessment and Early Warning Repeat steps S103-S104 for each frame of data during the thermal runaway process to obtain a two-dimensional distribution map of the gas column concentration at each time point. Based on this: Evaluation of generation time: Plot the average column concentration over time in the region of interest (e.g., the 10×10 pixel area directly above the battery), and take the moment when the concentration exceeds 3 times the noise base as the starting point of gas release.

[0059] Evaluation of diffusion path: The column concentration distribution of each frame is superimposed on the original radiance image in the form of a pseudo-color image. Red indicates high concentration areas and blue indicates low concentration areas. The process of gas plume diffusing outward from the battery surface can be observed by playing the image dynamically.

[0060] Peak concentration: Extracts the maximum column concentration value and its location and time of occurrence across the entire spatiotemporal range.

[0061] In addition, warning thresholds for each gas are preset based on safety standards or experimental experience (e.g., 50 ppm·m for DMC, 100 ppm·m for CH4). During real-time monitoring, if the column concentration of any pixel exceeds this threshold, the system immediately triggers an audible and visual alarm, records the time, location, and concentration value of exceeding the threshold, and sends this information to the monitoring platform via the network. To verify the effectiveness and gas identification capability of this invention, column concentration inversion experiments were conducted on three typical gases during the thermal runaway process of lithium cobalt oxide batteries—methane (CH4), ethylene (C2H4), and dimethyl carbonate (DMC). The comparative analysis of the three figures is as follows: Figure 2 (Methane, CH4): With the smallest molecular weight, it has the strongest diffusion ability. The inversion results show that CH4 has the widest concentration distribution, almost covering the entire field of view, and the high concentration region is diffuse with no obvious local aggregation peaks. This indicates that the present invention can track the long-distance diffusion path of light gases, providing a basis for determining whether the gas has spread outside the battery region.

[0062] Figure 3 (Ethylene, C2H4): With a medium molecular weight, it is a key characteristic product of SEI film decomposition and separator melting during thermal runaway in lithium-ion batteries. (And...) Figure 2 different, Figure 3 The concentration of C2H4 in the column is significantly concentrated near the surface of the battery, with a much smaller outward diffusion range than that of CH4. This spatial distribution characteristic has important diagnostic significance: the localized high concentration of C2H4 indicates that the thermal reaction is in a violent decomposition stage, and the heat source is located in the near-surface region inside the battery. Figure 2 If CH4 has already spread while C2H4 remains concentrated in the source region, it indicates that thermal runaway is about to occur but has not yet spread fully, which is the optimal window for intervention.

[0063] Figure 4 Dimethyl carbonate (DMC): With the largest molecular weight, it is the main solvent in the electrolyte. Inversion results show that DMC exhibits the highest column concentration release and the most complete plume morphology, forming a clear laminar plume rising from the battery surface. The early appearance of DMC (concentration increasing before CH4 / C2H4) can serve as an early warning signal of electrolyte leakage or evaporation.

[0064] Figures 2 to 4 The concentration distribution presented is highly consistent with the chemical mechanism of thermal runaway, which verifies the effectiveness of the present invention in simultaneous detection of multiple components, precise spatial localization, and physical consistency inversion. The spatial distribution differences in the concentrations of the three gas columns can be used for graded early warning of thermal runaway: a single increase in DMC indicates early electrolyte evaporation; simultaneous increases in CH4 and C2H4, with C2H4 concentrated in the near-source region, indicate severe SEI film decomposition and impending thermal runaway; when all three gases reach their peak values, it indicates that thermal runaway has fully occurred. These results fully demonstrate the practical value of the method of this invention in early warning of thermal runaway in lithium batteries.

[0065] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0066] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for monitoring thermal runaway gases in lithium batteries based on infrared hyperspectral remote sensing, characterized in that, Includes the following steps: S1. Continuously acquire time-series hyperspectral data of the lithium battery from the beginning of thermal runaway to the end of thermal runaway. S2. From the acquired time-series hyperspectral data, extract the data before thermal runaway as background spectral data. For each spatial location, calculate its average value before thermal runaway as the background spectrum of the pixel at that location. Extract the data during thermal runaway as the spectral data to be processed. Identify the gas plume region in the current frame based on the characteristic absorption peak of the preset gas. Select the pixels in the gas plume region as effective pixels. Use the spectral curve of the effective pixels as the target spectrum of the pixel. S3. For each pixel, subtract the background spectrum corresponding to the pixel and the contribution of gas self-thermal radiation obtained based on the pixel from the target spectrum of the pixel to obtain the corrected differential spectrum. S4. Construct a physical information neural network. The physical information neural network takes the target spectrum of each pixel and its corresponding differential spectrum as input, and combines it with a preset gas absorption cross section to perform the following in sequence: dynamic range compression of the differential spectrum of each pixel, spectral self-attention weighted enhancement of the target spectrum of each pixel, graph attention spatial aggregation of all pixels in the same frame based on neighborhood relationship, and unsupervised training with a composite loss based on a fixed physical quantity as a priori, and outputs the gas column concentration of each pixel. The graph attention space aggregation based on neighborhood relations includes: taking each pixel corresponding to the target spectrum as a node and constructing a graph structure with 8-neighborhood connection edges; the initial feature of each node is composed of three parts: enhanced spectral features, differential spectrum after dynamic range compression, and physical prior column concentration calculated by matched filtering; the graph structure is sequentially input into two layers of four-head graph attention layers, and the neighborhood node features are aggregated through self-attention coefficients to output the aggregated features of each node; the aggregated features are mapped through a fully connected layer to obtain the gas column concentration; The unsupervised training of the composite loss with fixed physical quantities as priors includes: fixing the gas temperature corresponding to the gas self-thermal radiation contribution obtained in step S3, the background spectrum obtained in step S2, and the gas absorption cross section as constants, and not participating in gradient update; The loss function for unsupervised training of the composite loss includes spectral weighted L1 loss, morphological similarity loss, and prior anchoring loss. The spectral weighted L1 loss is: in, Indicates the target spectrum, This represents the background spectrum, where N is the total number of wavelength points. The normalized weights for the target gas absorption cross section; To reconstruct the spectrum, specifically the gas column concentration predicted by the neural network. Calculate transmittance Output reconstructed spectrum ; This represents the target gas absorption cross section obtained in advance from the HITRAN database. This corresponds to the gas self-radiation of the corresponding pixel; Morphological similarity loss: in, The cosine similarity function; Prior anchoring loss: in The gas column concentration predicted by the network; Physical prior column concentrations calculated based on the matched filtering principle; The total loss is: ； S5. Evaluate the generation and diffusion process of thermal runaway gas based on the spatiotemporal distribution of the gas column concentration.

2. The method according to claim 1, characterized in that, Step S2 specifically includes: Determine the onset of thermal runaway based on the temperature profile of the lithium battery or manual interpretation. Select The average value of several frames of hyperspectral data within a time period is calculated for each spatial location and used as the background spectrum. Select Each frame of data within a time period is used as spectral data to be processed.

3. The method according to claim 1, characterized in that, The gas's own thermal radiation contribution in step S3 is obtained in the following way: For the effective pixel, the radiance of each wavenumber in the target spectrum of the pixel is converted into brightness temperature. The maximum brightness temperature under all wavenumbers is taken as the gas temperature, and then the gas self-radiation is calculated according to Planck's law.

4. The method according to claim 1, characterized in that, The dynamic range compression of the differential spectrum includes performing a signed logarithmic transformation on the corrected differential spectrum.

5. The method according to claim 4, characterized in that, The spectral self-attention weighted enhancement includes: normalizing the target spectrum and inputting it into a multilayer perceptron. The input dimension of the multilayer perceptron is the number of bands in the target spectrum, the intermediate layer dimension is half of the input dimension, and the output is activated by a sigmoid function to obtain weights. The weights are then multiplied element-wise with the normalized target spectrum to obtain the enhanced spectral features.

6. The method according to claim 1, characterized in that, The physical prior column concentration for each pixel is calculated using the matched filtering formula: in, Indicates the target spectrum, Indicates the background spectrum. This represents the target gas absorption cross section obtained in advance from the HITRAN database. This corresponds to the self-radiation of the gas in the corresponding pixel.

7. The method according to claim 1, characterized in that, Step S5 specifically includes: assessing the generation time, diffusion path, and peak concentration of the thermal runaway gas based on the spatiotemporal distribution of the gas column concentration; and comparing the gas column concentration with a preset warning threshold, issuing a warning message when the gas column concentration exceeds the preset threshold.

Images