A battery combustion detection and automatic extinguishing method based on image recognition

Abstract

The application discloses a battery combustion detection and automatic fire extinguishing method based on image recognition, and relates to the technical fields of fire safety and image processing. The method comprises the following steps: S1, outputting a battery combustion panoramic monitoring image sequence; S2, outputting a fire source two-dimensional pixel coordinate by improving a BiSeNet model; S3, outputting a fire source space positioning coordinate; S4, outputting an obstacle avoidance planning path of a fire extinguishing device; S5, selecting an optimal injection pressure, and outputting a directional fire extinguishing jet along the obstacle avoidance planning path of the fire extinguishing device; S6, outputting real-time evaluation data of fire extinguishing efficiency, and feeding back and correcting the optimal injection pressure; and S7, calculating a permutation entropy degree to measure system dynamics complexity, and if it is determined that the fire source is relighting, the obstacle avoidance planning path of the fire extinguishing device and the optimal injection pressure are recalculated. The application overcomes the defects of visual perception limitations, single control strategy and lack of feedback optimization of traditional methods, and provides an efficient solution for intelligent prevention and control of battery fires.

Description

Technical Field

[0001] This invention relates to the fields of fire safety and image processing technology, and in particular to a method for battery combustion detection and automatic fire extinguishing based on image recognition. Background Technology

[0002] With the large-scale application of electrochemical energy storage systems, battery combustion accidents are frequent. Traditional fire detection methods suffer from significant response lag and false alarm limitations when facing the characteristics of battery thermal runaway. Existing image recognition-based fire extinguishing methods typically employ a single-view monitoring mode. While they can obtain the approximate location of the fire source, they are limited by the field of view and geometric distortion of fixed cameras, making it difficult to eliminate brightness jumps and geometric discontinuities during the stitching process, resulting in a lack of global consistency in the generated monitoring images. This deficiency in visual field construction further restricts the accuracy of feature extraction. Conventional convolutional neural networks often ignore the anisotropic singularities implicit in the frequency domain when processing the high-frequency flicker textures and smoke diffusion patterns unique to battery combustion, making it difficult to separate high-fidelity flame texture details and semantic features in complex backgrounds, thus reducing the robustness of fire source identification. In addition, existing automatic fire extinguishing control logic mainly relies on preset pressure thresholds or simple distance measures, ignoring the information uncertainty in the fire source state evolution process and lacking an adaptive pressure optimization mechanism based on the rate of information entropy decrease, resulting in low utilization of fire extinguishing media. Meanwhile, traditional monitoring methods lack quantitative analysis of the dynamic complexity of the brightness time series in the monitored area after the flame is extinguished, and cannot effectively capture the unique signs of re-ignition in battery thermal runaway, making it difficult for the fire extinguishing system to achieve intelligent closed-loop control from "single extinguishing" to "complete suppression".

[0003] Therefore, how to provide a method for battery combustion detection and automatic fire extinguishing based on image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a method for battery combustion detection and automatic fire extinguishing based on image recognition. It constructs a panoramic visual field by fusing geodesic distance and Laplace's pyramid, extracts flame features using an improved BiSeNet model combined with non-subsampled second-generation curvelet transform and Lipschitz exponent, and constructs a four-dimensional spatiotemporal hypervoxel structure to segment and locate dynamic fire sources. The optimal injection pressure is selected based on the maximization criterion of information entropy gain in the uncertainty of the fire source state, and an efficiency evaluation and reignition monitoring mechanism based on Shannon information entropy and permutation entropy is constructed to achieve closed-loop adaptive control of the fire extinguishing process. By establishing a closed-loop control throughout the entire process from "panoramic visual construction" to "reignition dynamics monitoring," this invention effectively eliminates geometric distortion and brightness jumps during image stitching, achieving precise four-dimensional spatiotemporal positioning and adaptive pressure optimization of dynamic fire sources. It solves the technical problems of low recognition accuracy, positioning drift, and lack of reignition warning in battery combustion detection. This invention overcomes the limitations of traditional methods in visual perception, single control strategies, and lack of feedback optimization, providing an efficient solution for intelligent prevention and control of battery fires.

[0005] A method for battery combustion detection and automatic fire extinguishing based on image recognition according to an embodiment of the present invention specifically includes: S1. Construct a panoramic visual field, calculate the optimal suture line based on geodesic distance, use Laplacian pyramid layered weighted fusion to eliminate geometric faults and brightness jumps, and output a panoramic monitoring image sequence of battery combustion. S2. Input the battery combustion panoramic monitoring image sequence into the improved BiSeNet model to construct a high-fidelity flame texture detail feature tensor. Extract semantic features based on non-subsampled second-generation curvelet transform and Lipschitz exponential adaptive estimation. Utilize frequency domain phase correlation and morphological reconstruction weighted aggregation features to output the two-dimensional pixel coordinates of the fire source. S3. Construct a four-dimensional spatiotemporal supervoxel graph structure based on the two-dimensional pixel coordinates of the fire source, segment the dynamic fire source supervoxel region using graph clustering and calculate the spatiotemporal centroid, and output the spatial positioning coordinates of the fire source. S4. Using the spatial positioning coordinates of the fire source as the target, the path cost is progressively optimized by using a bidirectional fast expansion random tree algorithm to expand bidirectionally and perform rewiring operations, and the obstacle avoidance planning path of the fire extinguishing device is searched and output. S5. Calculate the information entropy gain of the uncertainty of the fire source state under different pressures based on the spatial positioning coordinates of the fire source, dynamically select the optimal spray pressure with the maximization of the information entropy decrease rate as the criterion, and output a directional fire extinguishing jet along the obstacle avoidance planning path of the fire extinguishing device. S6. Based on the panoramic monitoring image sequence of battery combustion and the fire source texture and smoke morphology feature data, calculate the Shannon information entropy of the grayscale histogram and the degree of correlation attenuation between consecutive frames, output real-time evaluation data of fire extinguishing efficiency, and provide feedback to correct the optimal injection pressure. S7. Based on the real-time evaluation data of the fire extinguishing effectiveness, after the flame disappears, enter the dormant monitoring stage, map the brightness time series of the acquired image into a symbol vector, calculate the permutation entropy to measure the dynamic complexity of the system, and if it is determined that the fire source reignites, recalculate the obstacle avoidance planning path of the fire extinguishing device and the optimal spray pressure.

[0006] Optionally, S1 specifically includes: S11. Project the original images captured by multiple cameras onto a unified spherical manifold, and use the optical flow field to perform sub-pixel level registration of feature points in the overlapping areas of adjacent views to construct a geometrically aligned panoramic visual field. S12. Based on the geometrically aligned panoramic visual field, calculate the sum of the absolute value of the color difference and the gradient magnitude difference of the corresponding pixels in the overlapping area as the matching cost, traverse all possible stitching paths and accumulate the total generation value, and select the path with the smallest total generation value as the best stitching line. S13. Perform Laplacian pyramid decomposition on the overlapping region image along the optimal suture line, calculate the Euclidean distance from the pixel to the optimal suture line at different scale levels, construct Gaussian weighted fusion coefficients using the Euclidean distance, and perform weighted fusion on the images at each level to obtain multi-scale fusion sub-bands. S14. Perform a pyramid reconstruction operation on the multi-scale fusion sub-band to eliminate geometric breaks and brightness jumps at the splicing points, and output a panoramic monitoring image sequence of battery combustion.

[0007] Optionally, the improved BiSeNet model includes a spatial path feature extraction layer, a contextual path feature extraction layer, an attention-weighted aggregation layer, a segmentation output layer, and a centroid calculation layer: The spatial path feature extraction layer is used to map the battery combustion panoramic monitoring image sequence to a polar coordinate grid, perform one-dimensional processing along the radial direction, and calculate the second central moment of the fractional-order spectrum using the energy difference between the flame flicker frequency and the background noise fractional-order spectrum with the transformation order as the independent variable. The gradient descent algorithm is used to iteratively search for the minimum point of the second central moment to lock the optimal transformation order. Within the optimal fractional-order domain, the energy distribution entropy of the transformation coefficient matrix is ​​calculated, a dynamic threshold is set to remove low-energy diffuse noise components, and high-energy focused texture components are retained. After inverse transformation reconstruction, a high-fidelity flame texture detail feature tensor is output. The context path feature extraction layer is used to perform non-downsampled second-generation curvelet transform on the high-fidelity flame texture detail feature tensor to obtain multi-directional sub-band coefficients; by utilizing the anisotropic singularity of the flame edge in different directional sub-bands, an adaptive estimation operator of the Lipschitz exponent based on logarithmic scale variation is constructed; by setting a regularity threshold, noise components with negative Lipschitz exponents are removed, while geometric structural components with positive exponents are retained and recombined, and a semantic feature tensor containing precise location priors is output; The attention-weighted aggregation layer is used to transform the high-fidelity flame texture detail feature tensor and the semantic feature tensor to the frequency domain. It uses phase spectrum correlation to locate spatial offset and drives the morphological reconstruction operator to perform erosion and dilation with connectivity constraints on the amplitude spectrum, removing isolated noise spikes and retaining continuous energy clusters. After frequency domain reconstruction and residual correction, a spatial weight map is generated, and a fused feature tensor that suppresses background artifacts and is spatially aligned is output. The segmentation output layer is used to input the fusion feature tensor that suppresses background artifacts and is spatially aligned into the convolutional layer for upsampling and prediction. After processing by the Softmax function, it outputs fire source texture and smoke morphology feature data, which includes flame region mask and smoke region mask. The centroid calculation layer is used to perform morphological processing on the flame region mask and the smoke region mask, calculate the weighted center of the mask pixel coordinates, and output the two-dimensional pixel coordinates of the fire source.

[0008] Optionally, S3 specifically includes: S31. Based on the disparity information in the panoramic monitoring image sequence, perform stereo matching calculation to generate an original depth map. Project the original depth map onto the geometrically aligned panoramic visual field to generate a panoramic depth map. Using the two-dimensional pixel coordinates of the fire source as spatial seed points, combine the panoramic depth map to establish pixel adjacency associations between consecutive time frames to construct a four-dimensional spatiotemporal supervoxel map structure containing spatial location and timestamp. S32. Calculate the inter-frame optical flow vector field based on the panoramic monitoring image sequence. According to the optical flow vector field and the texture features of the nodes in the four-dimensional spatiotemporal hypervoxel graph structure, calculate the color similarity, spatial proximity and motion state consistency between the nodes. Traverse the nodes and aggregate the nodes that meet the dynamic combustion characteristics to segment out the dynamic fire source hypervoxel region. S33. Assign weight coefficients based on combustion response intensity to all pixels within the dynamic fire source super-voxel region, calculate the weighted three-dimensional spatial coordinate mean in combination with the panoramic depth map, and output the fire source spatial positioning coordinates.

[0009] Optionally, S4 specifically includes: S41. Using the spatial positioning coordinates of the fire source as the target node and the current position of the fire extinguishing device as the starting node, construct a search space in the preset global dynamic environment semantic map and initialize a bidirectional fast-expanding random tree. S42. Generate new nodes by expanding from the starting node and the target node to the other random sampling point respectively, determine the collision between the new node and the obstacle, and if there is no collision, add the new node to the bidirectional fast expansion random tree and perform rewiring operation to progressively optimize the local path cost until the bidirectional tree nodes meet to generate the initial connection path. S43. Perform smoothing and redundant node removal operations on the initial connection path to eliminate unnecessary turning points in the path and output the obstacle avoidance planning path of the fire extinguishing device.

[0010] Optionally, S5 specifically includes: S51. Based on the spatial positioning coordinates of the fire source and the environmental wind direction data, construct the jet motion differential equation, calculate the liquid fracture length and spray drift distance under the action of gravity and wind resistance, and generate the predicted coverage probability distribution of the target fire source area. S52. Based on the predicted coverage probability distribution, calculate the information entropy of the uncertainty of the fire source state under different pressure levels, construct the mapping relationship between pressure adjustment and information entropy gain, and select the pressure level that maximizes the rate of decrease of information entropy as the optimal injection pressure. S53. Control the fire extinguishing device to adjust to the optimal spray pressure, generate the corresponding spray pressure command, move along the obstacle avoidance planning path of the fire extinguishing device to the range, and output a directional fire extinguishing jet to the target fire source area.

[0011] Optionally, S6 specifically includes: S61. Based on the panoramic monitoring image sequence of battery combustion and the texture and smoke morphology feature data of the fire source, extract the pixel intensity time series of the target area, calculate the mutual information value of pixel intensity between adjacent frames, and measure the degree of correlation decay between consecutive frames. S62. Construct a grayscale histogram of the target area and calculate the Shannon information entropy. Then, weight and fuse the Shannon information entropy with the correlation attenuation degree of the preceding and following frames to generate real-time fire extinguishing performance evaluation data. S63. Construct a fire extinguishing efficiency error transfer function, calculate the pressure compensation gradient based on the real-time fire extinguishing efficiency evaluation data, use the pressure compensation gradient to correct the optimal spray pressure, and output the updated spray pressure command.

[0012] Optionally, S7 specifically includes: S71. After the flame characteristics disappear according to the real-time evaluation data of the fire extinguishing efficiency, enter the dormant monitoring mode, continuously collect the image brightness data of the monitoring area and construct the brightness time series. S72. Map the brightness time series to a symbol vector, calculate the permutation entropy of the symbol vector, and output the system dynamics complexity; S73. If the system dynamics complexity exceeds the preset reignition threshold, the fire source is determined to reignite, the current dormant state is terminated, and the obstacle avoidance planning path and optimal spray pressure of the fire extinguishing device are recalculated based on the updated environmental data until the fire source is completely extinguished.

[0013] The beneficial effects of this invention are: (1) This invention establishes a fire source identification system that deeply integrates frequency domain texture details and spatial domain semantic features by employing an improved BiSeNet model combined with non-subsampled second-generation curvelet transform and Lipschitz exponent estimation. The spatial path uses fractional-order spectral energy differences to lock the optimal transform order and accurately extract high-fidelity flame texture details; the context path is based on the multi-directional subband coefficients of the second-generation curvelet transform and uses the Lipschitz exponent to adaptively remove noise components while retaining geometric structural features with anisotropic singularity. The attention-weighted aggregation layer uses frequency domain phase correlation to drive morphological reconstruction, removes isolated noise spikes, and retains continuous energy clusters. This system achieves accurate decoupling and enhancement of weak texture features and complex semantic information of battery combustion, significantly improving the robustness of fire source identification in complex backgrounds.

[0014] (2) This invention achieves a leap from physical quantity control to information uncertainty elimination in fire extinguishing strategy by introducing the information entropy gain maximization criterion. Based on the spatial location coordinates of the fire source, a jet motion differential equation is constructed, the predicted coverage probability distribution is calculated, and the information entropy of the fire source state uncertainty under different pressure levels is quantified. The optimal injection pressure is dynamically selected with the goal of maximizing the information entropy decrease rate. This model models the fire extinguishing process as a process of reducing system uncertainty, establishes a mathematical and logical relationship between pressure regulation and fire extinguishing effectiveness, and ensures the maximization of fire extinguishing medium utilization under complex environmental interference.

[0015] (3) This invention achieves real-time assessment of fire extinguishing efficiency and early warning of reignition risk by constructing a multi-dimensional entropy monitoring closed loop based on Shannon information entropy, mutual information, and permutation entropy. The Shannon information entropy of the grayscale histogram and the mutual information values ​​of the preceding and following frames are calculated to measure fire extinguishing efficiency and to provide feedback to correct the injection pressure; the brightness time series during the dormant period is mapped to a symbol vector, and permutation entropy is calculated to measure the dynamic complexity of the system. This mechanism uses nonlinear dynamic indicators to keenly capture the state changes of the battery thermal runaway system, realizes the quantitative judgment of the signs of fire reignition, and constructs a full-process intelligent closed-loop control from "monitoring-fire extinguishing" to "efficiency assessment-reignition warning". Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a battery combustion detection and automatic fire extinguishing method based on image recognition proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the improved BiSeNet model for a battery combustion detection and automatic fire extinguishing method based on image recognition proposed in this invention. Detailed Implementation

[0017] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 and Figure 2 A method for battery combustion detection and automatic fire extinguishing based on image recognition, specifically including: S1. Construct a panoramic visual field, calculate the optimal suture line based on geodesic distance, use Laplacian pyramid layered weighted fusion to eliminate geometric faults and brightness jumps, and output a panoramic monitoring image sequence of battery combustion. S2. Input the panoramic monitoring image sequence of battery combustion into the improved BiSeNet model to construct a high-fidelity flame texture detail feature tensor. Extract semantic features based on non-subsampled second-generation curvelet transform and Lipschitz exponent adaptive estimation. Utilize frequency domain phase correlation and morphological reconstruction weighted aggregation features to output the two-dimensional pixel coordinates of the fire source. S3. Construct a four-dimensional spatiotemporal hypervoxel graph structure based on the two-dimensional pixel coordinates of the fire source, use graph clustering to segment the dynamic fire source hypervoxel region and calculate the spatiotemporal centroid, and output the spatial positioning coordinates of the fire source. S4. Using the spatial location coordinates of the fire source as the target, the path cost is progressively optimized by using the bidirectional fast expansion random tree algorithm to expand bidirectionally and perform rewiring operations, and the obstacle avoidance planning path of the fire extinguishing device is searched and output. S5. Calculate the information entropy gain of the uncertainty of the fire source state under different pressures based on the spatial positioning coordinates of the fire source, dynamically select the optimal jet pressure with the maximization of the information entropy decrease rate as the criterion, and output directional fire extinguishing jet along the obstacle avoidance planning path of the fire extinguishing device. S6. Based on the panoramic monitoring image sequence of battery combustion and the texture and smoke morphology feature data of the fire source, calculate the Shannon information entropy of the grayscale histogram and the degree of correlation attenuation between consecutive frames, output real-time evaluation data of fire extinguishing efficiency, and provide feedback to correct the optimal injection pressure. S7. Based on the real-time evaluation data of fire extinguishing effectiveness, after the flame disappears, the system enters dormant monitoring, maps the brightness time series of the acquired images into a symbol vector, calculates the permutation entropy to measure the dynamic complexity of the system, and if it is determined that the fire source reignites, the obstacle avoidance planning path and optimal spray pressure of the fire extinguishing device are recalculated.

[0019] In this embodiment, S1 specifically includes: S11. Project the original images acquired by multiple cameras onto a unified spherical manifold. Calculate the squared brightness difference of corresponding pixels in the overlapping area of ​​adjacent views as a data term. Simultaneously, calculate the ratio of the dot product to the magnitude of the gradient vectors in the neighborhood of each pixel as a structural similarity term. Set the weight coefficient of the data term to 0.6 and the weight coefficient of the structural similarity term to 0.4. Use the weighted sum as the data fidelity term. Calculate the sum of the squared horizontal and vertical gradients of the optical flow vector as a smoothing term. Set the weight coefficient of the smoothing term to 0.2. Add the data fidelity term and the smoothing term to construct the total energy function. Calculate the partial derivative of the total energy function with respect to the optical flow vector to obtain a system of linear equations. Traverse each pixel in the image and update the optical flow vector of the current pixel using the latest optical flow values ​​of the pixels in the upper and left neighborhoods. Repeat the update operation until the difference between two adjacent iterations is less than 0.001. Obtain the optimal optical flow vector. Based on the optical flow vector, guide the feature points to perform sub-pixel-level displacement adjustments on the spherical manifold to construct a geometrically aligned panoramic visual field.

[0020] S12. Lock the overlapping region in the geometrically aligned panoramic visual field, calculate the sum of the absolute value of the color difference and the gradient magnitude difference of the corresponding pixels in the overlapping region as the matching cost, set the weight coefficient of the color difference to 0.7 and the weight coefficient of the gradient magnitude difference to 0.3, construct a directed cost grid with pixel coordinates as nodes, traverse the grid row by row starting from the first row, calculate the cumulative cost of the three predecessor nodes (top left, top right, top right) of the current pixel node and select the minimum value to add to the cost of the current node, store the result in the cumulative cost matrix, backtrack to the node with the smallest value in the last row of the cumulative cost matrix, search upwards row by row for the predecessor node with the smallest cumulative cost, connect all selected nodes to form the path with the smallest total cost as the best stitching line.

[0021] S13. Perform Laplacian pyramid decomposition on the overlapping region image along the optimal suture line, decomposing the image into a Laplacian residual layer and a Gaussian pyramid layer. Calculate the Euclidean distance from the pixel to the optimal suture line at each level. Set the standard deviation parameter of the Gaussian attenuation model to 5.0. Substitute the Euclidean distance into the Gaussian attenuation model to construct a weighted fusion coefficient that changes nonlinearly with distance. Use this coefficient to perform weighted fusion on the image at each level to obtain a multi-scale fusion subband.

[0022] S14. Perform pyramid reconstruction operation on the multi-scale fusion sub-band, upsample from the top to the bottom level and accumulate the Laplacian residuals of each level to reconstruct the complete image frequency components, fill the frequency missing at the stitching gaps, eliminate geometric discontinuities and brightness jumps at the stitching points, and output the panoramic monitoring image sequence of battery combustion.

[0023] In this embodiment, the improved BiSeNet model includes a spatial path feature extraction layer, a context path feature extraction layer, an attention-weighted aggregation layer, a segmentation output layer, and a centroid calculation layer: The spatial path feature extraction layer is used to read each frame of the panoramic monitoring image sequence of battery combustion, map the image pixel coordinates to a polar coordinate grid, extract the pixel sequence along the radial direction, and convert the pixel sequence into a one-dimensional discrete signal. The angular step size is set to 1 degree and the radial step size to 1 pixel, constructing a set of one-dimensional discrete signals. A fractional Fourier transform is performed on the one-dimensional discrete signal, with an initial transform order of 0.01, a step size of 0.01, and a final value of 1.0. The transform coefficient matrix is ​​calculated iteratively for different transform orders, and the square of the amplitude of each element in the transform coefficient matrix is ​​calculated to obtain the fractional spectral energy distribution. The second central moment of the fractional spectral energy distribution with respect to the geometric center coordinates is calculated, and the Adam algorithm is used to iteratively search for the minimum point of the second central moment and set this point... The corresponding transformation order is set as the optimal transformation order. The transformation coefficient matrix is ​​obtained in the optimal fractional Fourier domain. The sum of the squares of the amplitudes of all elements in the transformation coefficient matrix is ​​calculated as the total energy. The ratio of the square of the amplitude of each element to the total energy is calculated as the probability distribution sequence. The Shannon entropy is calculated using the probability distribution sequence as the energy distribution entropy. The global mean of the transformation coefficient matrix is ​​calculated as the baseline threshold. The product of the baseline threshold and the energy distribution entropy is used as the dynamic segmentation threshold. The transformation coefficient matrix is ​​traversed, and coefficients with moduli less than the dynamic segmentation threshold are set to zero. Low-energy diffuse noise components are removed, and coefficients with moduli greater than or equal to the dynamic segmentation threshold are retained as high-energy focused texture components. An inverse fractional Fourier transform is performed on the processed coefficient matrix to output a high-fidelity flame texture detail feature tensor.

[0024] The context path feature extraction layer is used to lock the high-fidelity flame texture detail feature tensor. A non-downsampling second-generation curvelet transform is performed on the high-fidelity flame texture detail feature tensor. The decomposition scale is set to 4 layers, with 8 sub-bands per layer, to obtain a multi-scale, multi-directional sub-band coefficient matrix set. For each coefficient point in the sub-band coefficient matrix set, coefficient points with the same spatial coordinates in two adjacent scale levels are located. The difference between the logarithm of the modulus at the current scale and the logarithm of the modulus at the previous scale is calculated. This difference is divided by the scale level change factor of 2 to obtain the estimated Lipschitz exponent for that coefficient point. A regularization threshold of 0 is set, and the estimated Lipschitz exponent for each coefficient point is compared to the regularization threshold. Coefficient points with Lipschitz exponent estimates less than 0 are identified as noise components and their values ​​are adjusted accordingly. Set the value to 0, identify coefficients with Lipschitz exponent estimates greater than or equal to 0 as geometric structural components and retain their original values; perform inverse curvelet transform reorganization on the processed sub-band coefficients of each level, set the coefficient weight of the coarsest scale layer to 0.5 and the coefficient weight of the fine scale layer to 1.0, and perform inverse Radon transform on the sub-band coefficients of each level. The specific steps are as follows: perform one-dimensional inverse Fourier transform on the sub-band coefficients to transform the frequency domain coefficients to the projection domain, calculate the one-dimensional Hilbert transform of the projection data, convolve the transformation result with the original projection data in the angular dimension, perform back-projection integration on the convolution result in the radial direction to obtain the spatial components of each level, calculate the weighted sum of the spatial components of each level, perform bilinear interpolation to adjust the weighted sum image to the same size as the original input image, and output a semantic feature tensor containing precise location priors.

[0025] The attention-weighted aggregation layer transforms the high-fidelity flame texture detail feature tensor and semantic feature tensor to the frequency domain. It calculates the phase spectrum correlation value of the two frequency domain tensors to locate the spatial offset. Based on the spatial offset, the morphological reconstruction operator performs erosion and dilation operations with connectivity constraints on the amplitude spectrum, removing isolated noise spikes and retaining continuous energy clusters. After frequency domain reconstruction and residual correction, a spatial weight map is generated. The spatial weight map is used to perform weighted fusion of the high-fidelity flame texture detail feature tensor and semantic feature tensor, and outputs a fused feature tensor that suppresses background artifacts and is spatially aligned. The specific steps are as follows: Calculate the two-dimensional Fourier transform of the two tensors respectively, extract the phase spectrum and calculate the normalized cross-power spectrum, perform an inverse Fourier transform on the normalized cross-power spectrum to obtain the impulse response function, and search for the peak coordinates of the impulse response function as the spatial offset; perform translation correction on the amplitude spectrum of the semantic feature tensor based on the spatial offset, perform erosion operation on the corrected amplitude spectrum to remove isolated noise points, set the erosion kernel size to 3×3 pixels, perform dilation operation to restore the target size, and retain continuous energy clusters with a connected region area greater than 10 pixels; combine the processed amplitude spectrum with the original phase spectrum and perform an inverse Fourier transform to obtain the reconstructed features, calculate the difference between the reconstructed features and the original features as the residual, and superimpose the residual onto the reconstructed features to generate a spatial weight map; use the spatial weight map to perform element-wise multiplication weighted summation on the high-fidelity flame texture detail feature tensor and the semantic feature tensor, and output a fused feature tensor that suppresses background artifacts and is spatially aligned.

[0026] The segmentation output layer is used to input the spatially aligned and background artifact-suppressed fusion feature tensor into the convolutional layer. The bilinear interpolation algorithm is used to perform upsampling to restore the original image resolution. The Softmax function is used to calculate the probability value of each pixel belonging to the flame category and the probability value of each pixel belonging to the smoke category. The category with the highest probability value is selected as the predicted label of the pixel. The fire source texture and smoke morphology feature data are output, which includes the flame region mask and the smoke region mask. The centroid calculation layer is used to read the flame region mask and the smoke region mask, perform morphological opening operations on the flame region mask and the smoke region mask to remove edge burrs, calculate the average horizontal coordinate of all pixels in the mask area as the horizontal coordinate of the fire source center, and the average vertical coordinate as the vertical coordinate of the fire source center. Combine the horizontal coordinate and vertical coordinate of the fire source center to output the two-dimensional pixel coordinates of the fire source.

[0027] The improved BiSeNet model feature extraction and fusion process proposed in this step is similar to that of the traditional BiSeNet model in that it follows the dual-branch architecture design idea, that is, the spatial path is used to capture spatial detail information and the context path is used to extract semantic context information. Both adopt a feature fusion strategy to combine the two complementary features to achieve the final pixel-level classification, and both rely on the backpropagation algorithm to optimize network parameters.

[0028] The difference lies in that this invention overcomes the limitations of traditional BiSeNet models, which suffer from the loss of high-frequency texture details due to downsampling when using only standard convolution in the spatial path and the difficulty in effectively distinguishing geometric structures from random noise when using dilated convolution in the context path. At the spatial path feature extraction level, this invention introduces fractional-order transform and energy distribution entropy calculation, using the energy difference in the fractional-order spectrum between the flame flicker frequency and background noise to lock the optimal transform order, and uses dynamic thresholding to remove low-energy diffuse noise instead of simple convolution filtering. At the context path feature extraction level, this invention uses non-downsampled second-generation curvelet transform to replace dilated convolution in the traditional model, constructing a Lipschitz exponential adaptive estimation operator based on logarithmic scale changes, and using anisotropic singularity to distinguish noise from structure. At the feature aggregation level, this invention abandons conventional convolution summation or concatenation operations, transforming features to the frequency domain, using phase spectrum correlation to locate spatial offsets, and driving morphological reconstruction operators to perform erosion and dilation with connectivity constraints on the amplitude spectrum, rather than directly applying the attention weight matrix.

[0029] The beneficial effects of the improvements are that this invention, through fractional-order spectral energy difference screening and Lipschitz exponent regularity estimation, can effectively separate high-frequency texture details and random noise interference in battery combustion scenarios from the underlying mechanism of signal processing, solving the problem of insufficient perception of subtle flame textures in the traditional BiSeNet model; by using the aggregation strategy of frequency domain phase correlation-driven morphological reconstruction, it not only eliminates spatial misalignment and background artifacts during multi-path feature fusion, but also preserves the continuous energy clumping structure, significantly improving the signal-to-noise ratio and geometric fidelity of fire source feature extraction under complex lighting and smoke-covered environments, providing high-quality prior data for subsequent accurate positioning.

[0030] In this embodiment, S3 specifically includes: S31. Read the left and right views from the panoramic monitoring image sequence of battery combustion, calculate the sum of the absolute values ​​of the grayscale differences of corresponding pixels in the left and right views as the matching cost, set the search range to 0 to 64 pixels, traverse each disparity value in the search range, select the disparity value with the smallest matching cost as the coarse disparity, use the quadratic curve fitting method to calculate the sub-pixel offset of the point with the smallest matching cost, add the coarse disparity and the sub-pixel offset to obtain the accurate disparity map, divide the value of the accurate disparity map by the baseline length and multiply by the focal length value, take the reciprocal of the calculation result to obtain the original depth map, use the homography matrix to project the original depth map onto the geometrically aligned panoramic visual field, use the bilinear interpolation method to fill the pixel gaps to generate the panoramic depth map; read the two-dimensional pixel coordinates of the fire source, use them as spatial seed points to index the corresponding depth values ​​in the panoramic depth map, set the time radius of the spatiotemporal voxel to 3 frames and the spatial radius to 5 pixels, search for pixels with a depth value difference of less than 0.1 meters between consecutive time frames to establish adjacency associations, and construct a four-dimensional spatiotemporal supervoxel map structure containing spatial location coordinates and timestamps.

[0031] S32. Read the panoramic monitoring image sequence and calculate the inter-frame optical flow vector field between the previous frame and the current frame using the polynomial expansion method. The specific steps are as follows: Gaussian blur the image, set the Gaussian kernel size to 15 pixels, calculate the neighborhood matrix of each pixel in the image, fit the polynomial coefficients of the neighborhood matrix using the least squares method, compare the polynomial coefficient matrices of the two frames to calculate the displacement vector, and output the inter-frame optical flow vector field; based on the coordinate position of the nodes in the four-dimensional spatiotemporal supervoxel image structure, index the optical flow vector field and image texture features, calculate the Euclidean distance between nodes in the RGB color space as color similarity, calculate the Euclidean distance between nodes in spatial coordinates as spatial proximity, and calculate the cosine similarity of the node optical flow vector as motion state consistency; set the color similarity threshold to 30, the spatial proximity threshold to 5 pixels, and the motion state consistency threshold to 0.8, traverse the nodes and filter the nodes that meet the three metrics simultaneously, perform a region-growing-based aggregation operation on the nodes that meet the conditions, and segment the dynamic fire source supervoxel region.

[0032] S33. Read the dynamic fire source super-voxel region, calculate the product of the brightness value and saturation value of each pixel in the region as the combustion response intensity, search for the maximum value of the combustion response intensity of all pixels in the region, divide the combustion response intensity of each pixel by the maximum value to obtain the normalized weight coefficient with a value between 0 and 1; read the panoramic depth map, extract the corresponding depth value according to the pixel coordinate, subtract the horizontal coordinate of the image center from the horizontal coordinate of the pixel to obtain the horizontal coordinate offset, subtract the vertical coordinate of the image center from the vertical coordinate of the pixel to obtain the vertical coordinate offset, multiply the horizontal coordinate offset by the depth value and divide by the focal length value to obtain the spatial X-axis coordinate, multiply the vertical coordinate offset by the depth value and divide by the focal length value to obtain the spatial Y-axis coordinate, use the depth value as the spatial Z-axis coordinate, combine the three coordinate values ​​to obtain the three-dimensional spatial coordinate, calculate the weighted average of the three-dimensional spatial coordinate and the normalized weight coefficient, and output the spatial positioning coordinates of the fire source.

[0033] The four-dimensional spatiotemporal super-voxel fire source localization process proposed in this step is similar to the traditional three-dimensional spatial localization method in that it relies on depth information and two-dimensional pixel coordinates obtained by a visual sensor, and derives the position of the target in physical space through coordinate system transformation and geometric operations. It also uses image segmentation technology to separate the target area from the background to improve the localization accuracy.

[0034] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on single-frame static images for 3D coordinate inversion or isolated temporal processing. Building upon traditional models that only construct spatial voxels, this invention adds a four-dimensional spatiotemporal map construction step, integrating timestamps as a fourth dimension into the voxel structure. It utilizes disparity information to establish pixel adjacency relationships between consecutive frames, rather than simply performing independent voxelization between frames. In the region segmentation step, it uses optical flow vector fields to calculate the consistency of node motion states, combining color similarity and spatial proximity for multi-dimensional feature aggregation, rather than relying solely on single manifold cutting based on color or spatial distance. Finally, in the coordinate output step, it introduces combustion response intensity as a weighting coefficient, combining it with the panoramic depth map to calculate the weighted 3D spatial coordinate mean, rather than simply calculating the geometric centroid.

[0035] The beneficial effects of the improvements are that, by constructing a four-dimensional spatiotemporal hypervoxel structure, this invention can model the dynamic evolution of flame combustion as a unified spatiotemporal geometric whole, breaking the limitation of low utilization of dynamic features due to the inter-frame independence of traditional methods, and realizing a leap from static spatial positioning to dynamic spatiotemporal tracking. This design significantly enhances the algorithm's robustness to irregular changes in flame morphology and smoke obstruction, enabling more accurate capture of the core area of ​​the fire source in the spatiotemporal continuous domain. The weighted centroid calculation based on combustion response intensity effectively suppresses the weights of edge noise and interference areas, achieving high-precision positioning of the spatial center of the fire source, and providing a reliable spatial reference for subsequent path planning and injection control.

[0036] In this embodiment, S4 specifically includes: S41. Lock the spatial positioning coordinates of the fire source output from the previous step and set them as the target node. Read the current GPS coordinates of the fire extinguishing device as the starting node. Load the preset global dynamic environment semantic map and discretize the three-dimensional space into a voxel grid with a side length of 0.1 meters. Construct a search space based on the voxel grid. Initialize a bidirectional fast-expanding random tree in the search space and construct a starting tree with the starting node as the root node and a target tree with the target node as the root node. Set the initial node set of the two trees to contain only the coordinates of the root node.

[0037] S42. Generate a uniformly distributed random 3D coordinate point in the search space as a sampling point. Calculate the node in the starting tree and the target tree that is closest to the sampling point in Euclidean distance as the nearest node. Set the step size to 0.2 meters, calculate the unit vector pointing from the nearest node to the sampling point, multiply the unit vector by the step size to obtain the offset, and add the coordinates of the nearest node and the offset to generate a new node. Read the obstacle object pixel information in the global dynamic environment semantic map, calculate the distance between the new node and the center point of the obstacle object pixel. If the distance is greater than the safety threshold of 0.5 meters, it is determined to be a no-collision, and the new node is added to the corresponding tree structure. Calculate the connection cost between the new node and the neighboring nodes in the tree, select the connection method with the lowest cost, perform a rewiring operation to optimize the local path, and repeat the above expansion steps until the distance between the new node in the starting tree and the node in the target tree is less than 0.2 meters. Connect the nodes of the two trees to generate the initial connection path.

[0038] S43. Read the initial connection path, traverse the coordinates of each node in the path, calculate the connection vector between the current node and the second successor node, check whether there are obstacle objects on the connection vector, if there are no obstacles, remove the first successor node in the middle, and perform redundant node removal operation; traverse the turning points of the path, calculate the sum of the vector pointing from the previous node to the turning point and the vector pointing from the previous node to the next node as the smoothing tangent vector, multiply the smoothing tangent vector by the scaling factor 0.3 to obtain the control offset, add the coordinates of the turning point to the control offset to obtain the smoothing control point, insert smoothing control points before and after the turning point, use cubic polynomial interpolation to calculate the trajectory coordinates between the inserted points, eliminate unnecessary turning points in the path, and output the obstacle avoidance planning path of the fire extinguishing device.

[0039] In this embodiment, S5 specifically includes: S51. Lock the spatial positioning coordinates of the fire source output from the previous step, read the wind speed and wind direction data transmitted by the environmental sensor, and construct the differential equation of jet motion. The specific steps are as follows: calculate the vector sum of the gravitational acceleration component of the ejected droplet and the air resistance component generated by the wind speed as the resultant acceleration, perform a double integral operation on the resultant acceleration with respect to time to obtain the displacement offset of the droplet in three-dimensional space, superimpose the displacement offset onto the spatial positioning coordinates of the fire source, set the droplet diameter distribution range to 0.1 mm to 1 mm, traverse and calculate the landing position of droplets with different diameters, statistically analyze the probability density distribution of the landing position, and generate the predicted coverage probability distribution of the target fire source area.

[0040] S52. Read the predicted coverage probability distribution, set the pressure level search range to 0.5 MPa to 1.5 MPa, and the step size to 0.1 MPa. Iterate through and calculate the predicted coverage probability distribution corresponding to each pressure level. Calculate the logarithm of each probability value in the predicted coverage probability distribution, multiply it by the negative of the probability value itself, and sum them up to obtain the information entropy of the uncertainty of the fire source state. Calculate the absolute value of the difference between the information entropy of the current pressure level and the information entropy of the previous pressure level. Divide the absolute value of the difference by the pressure step size to obtain the information entropy decrease rate. Select the pressure value corresponding to the maximum information entropy decrease rate as the optimal injection pressure.

[0041] S53. Lock the optimal spray pressure output from the previous step to generate the corresponding spray pressure command, read the obstacle avoidance planning path of the fire extinguishing device, calculate the Euclidean distance between the real-time coordinates of the fire extinguishing device moving along the path and the spatial positioning coordinates of the fire source, compare the Euclidean distance with the maximum range value of the fire extinguishing device, lock the position of the fire extinguishing device when the Euclidean distance is less than the maximum range value, adjust the output power of the pump body according to the spray pressure command, control the fire extinguishing device to adjust to the optimal spray pressure, and output a directional fire extinguishing jet to the target fire source area.

[0042] The fire extinguishing jet motion modeling and pressure optimization control process proposed in this step is similar to the traditional fire extinguishing device spray control method in that both are based on the principles of fluid mechanics and kinematics to establish a droplet motion model, use wind speed and direction data collected by environmental sensors to correct the trajectory, and use the fire source coordinates as the target to guide the movement and spray of the fire extinguishing device to achieve physical coverage and suppression of the fire.

[0043] The difference lies in that this invention breaks away from traditional methods that rely solely on experience to set a fixed injection pressure or adjust the pressure simply by adjusting the distance ratio, neglecting the uncertainty of the probability distribution of the extinguishing agent's landing point under complex environmental wind fields. In the landing point prediction step, this invention constructs a jet motion differential equation, introduces a droplet diameter distribution variable, and iteratively calculates the landing points of droplets of different sizes under the coupling effect of gravity and wind resistance, generating a predicted coverage probability distribution of the target fire source area, rather than simply calculating a single ideal trajectory. In the pressure decision step, the information entropy characterizing the uncertainty of the fire source state is calculated using the predicted coverage probability distribution. The optimal injection pressure is dynamically selected based on maximizing the rate of decrease of information entropy, rather than using a constant pressure value or manually adjusting the pressure. In the execution control step, the Euclidean distance between the real-time coordinates of the fire extinguishing device moving along the obstacle avoidance path and the coordinates of the fire source is used as the trigger criterion to lock the optimal operating position within the range.

[0044] The beneficial effects of this improvement are as follows: By constructing a jet motion differential equation that includes droplet size distribution, this invention can simulate the drift and diffusion behavior of multi-diameter droplets in real spraying scenarios. This breaks through the limitations of traditional methods that assume a single trajectory, resulting in poor environmental adaptability, and achieves probabilistic and accurate prediction of the coverage range of extinguishing agents under complex wind fields. By using the maximization criterion of information entropy decrease rate to select the optimal spray pressure, it can quantitatively assess the degree of reduction of fire source uncertainty by different pressure levels, fundamentally solving the problem of lack of theoretical basis for pressure regulation, significantly improving fire extinguishing efficiency and reducing extinguishing agent waste. The path locking mechanism combined with the range criterion ensures that the fire extinguishing device outputs the jet at the optimal pressure in the optimal geometric position, enhancing the system's autonomous decision-making and precise fire extinguishing capabilities in complex fire environments.

[0045] In this embodiment, S6 specifically includes: S61. Read the panoramic monitoring image sequence of battery combustion, lock the dynamic fire source super voxel region segmented in the previous step as the target region, extract the gray value of all pixels in the target region and calculate the average value to obtain the pixel intensity time series, calculate the joint probability distribution of pixel intensity of two adjacent frames, calculate the mutual information value of pixel intensity based on the joint probability distribution, subtract the mutual information value of the current frame from the mutual information value of the previous frame to obtain the mutual information difference, divide the mutual information difference by the mutual information value of the previous frame to obtain the degree of correlation attenuation between the two frames.

[0046] S62. Read the grayscale image of the target region in the current frame, count the frequency of each grayscale level in the grayscale image as the grayscale distribution probability, calculate the natural logarithm of the grayscale distribution probability, multiply it by the negative of the grayscale distribution probability itself, and sum them up to obtain the Shannon information entropy; set the weight coefficient of the correlation decay degree to 0.6 and the weight coefficient of the Shannon information entropy to 0.4, calculate the product of the correlation decay degree and the weight coefficient 0.6 as the first weighted component, calculate the product of the Shannon information entropy and the weight coefficient 0.4 as the second weighted component, add the first weighted component and the second weighted component to generate real-time fire extinguishing efficiency assessment data.

[0047] S63. Read the real-time fire extinguishing performance evaluation data, set the performance target value to 0, calculate the difference between the real-time fire extinguishing performance evaluation data and the performance target value to obtain the performance error, construct the error transfer function, multiply the performance error by the proportional coefficient 0.5 to obtain the pressure compensation gradient; read the optimal spray pressure output by the previous step, add the pressure compensation gradient to the optimal spray pressure to obtain the updated spray pressure value, and generate the updated spray pressure command based on the updated spray pressure value.

[0048] In this embodiment, S7 specifically includes: S71. Read the real-time fire extinguishing efficiency evaluation data generated in the previous steps. When the data value is lower than the preset extinguishing threshold of 0.05 for 10 consecutive frames, it is determined that the flame characteristics have disappeared, and the system status is switched to sleep monitoring mode. In sleep monitoring mode, continuously collect RGB images of the monitoring area, calculate the average value of the R channel, G channel and B channel of all pixels in each frame image, take the maximum value of the three channel averages as the brightness feature value of the current frame, and arrange the brightness feature values ​​in time order to construct a brightness time series.

[0049] S72. Read the brightness time series, set the embedding dimension to 5 and the time delay to 1, extract 5 consecutive data points from the brightness time series to construct a subsequence, sort the data points in the subsequence in ascending order of numerical value, and map the sorted index order to symbol vectors; traverse the brightness time series to count the occurrence frequency of each symbol vector, calculate the natural logarithm of the occurrence frequency of each symbol vector multiplied by the negative of the occurrence frequency itself, and sum them up to obtain the system dynamic complexity.

[0050] S73. Read the system dynamics complexity, set the reignition threshold to 0.6, compare the system dynamics complexity with the reignition threshold. If the system dynamics complexity exceeds the reignition threshold, it is determined that the fire source has reignited, and the current dormant state is terminated. Call the environmental sensor to obtain updated wind speed and wind direction data, recalculate the obstacle avoidance planning path and optimal spray pressure of the fire extinguishing device based on the updated environmental data, and repeatedly execute the fire extinguishing operation until the system dynamics complexity is below the reignition threshold for 30 consecutive frames, and it is determined that the fire source is completely extinguished.

[0051] Example 1: To verify the feasibility of this invention in the safe operation and maintenance of electrochemical energy storage power stations, the method of this invention was applied to the intelligent fire protection system of the battery compartment of a provincial power company's energy storage demonstration power station (hereinafter referred to as "Power Station E"). In traditional fire monitoring systems for energy storage power stations, alarm mechanisms based on a single temperature sensor or ordinary smoke detector are typically used. These methods not only struggle to accurately identify weak flames and smoke characteristics in the early stages of battery thermal runaway, but also fail to accurately obtain the spatial three-dimensional coordinates and dynamic evolution trend of the fire source, easily leading to deviations in the placement of extinguishing agents or improper pressure configuration, thus missing the optimal extinguishing opportunity. To solve the above problems, Power Station E decided to adopt the image recognition-based battery combustion detection and automatic fire extinguishing method proposed in this invention.

[0052] During implementation, Power Plant E first acquired raw image streams of the monitoring area using multiple industrial cameras deployed on the top and side walls of the battery compartment. After preprocessing operations such as spherical manifold projection, sub-pixel registration of the optical flow field, and Laplacian pyramid layered weighted fusion, a panoramic monitoring image sequence of battery combustion was constructed, eliminating geometric faults and brightness abrupt changes. Simultaneously, Power Plant E's fire protection and maintenance team precisely labeled the collected multi-source data with fire source types and calibrated pressure thresholds, serving as a benchmark for model training and control strategies.

[0053] Power Plant E improved the BiSeNet model by using a spatial path feature extraction layer to preserve high-energy focused texture components in the optimal fractional order domain. A contextual path feature extraction layer performed a non-downsampled second-generation curvelet transform and eliminated noise components based on the Lipschitz exponent, generating a semantic feature tensor containing precise location priors. Next, an attention-weighted aggregation layer performed phase spectrum correlation and morphological reconstruction in the frequency domain, outputting a fused feature tensor that suppressed background artifacts and was spatially aligned, effectively eliminating invalid interference from ambient lighting and equipment reflections. Subsequently, based on the segmented output of the fire source's two-dimensional pixel coordinates, a four-dimensional spatiotemporal supervoxel map structure was constructed using a panoramic depth map. Graph clustering was used to segment the dynamic fire source supervoxel region and calculate the spatiotemporal centroid, accurately outputting the fire source's spatial positioning coordinates.

[0054] During the core control and fire suppression phase, this invention constructs a jet motion differential equation to calculate the predicted coverage probability distribution of the target fire source area. It then dynamically selects the optimal jet pressure based on maximizing the rate of information entropy decrease, and outputs a directional fire suppression jet along an obstacle avoidance planning path generated by a bidirectionally expanding random tree. Subsequently, the system continuously calculates the Shannon information entropy of the grayscale histogram and the degree of correlation decay between consecutive frames, generating real-time fire suppression efficiency evaluation data and providing feedback to correct the jet pressure. After the flames are extinguished, the system automatically enters a dormant monitoring mode. By calculating the permutation entropy of the brightness time series to measure the system's dynamic complexity, it successfully detects a potential battery reignition sign and triggers a secondary fire suppression command.

[0055] During implementation, the technical team at Power Plant E discovered that, compared to traditional temperature sensing and constant pressure injection methods, the method of this invention significantly improves the accuracy of battery combustion detection and the level of intelligence in fire suppression control. Traditional methods cannot quantify the spatial location and state uncertainty of the fire source and lack the ability to perceive the risk of reignition. In contrast, the method of this invention, through panoramic vision construction, fractional-order spectral feature enhancement, and information entropy closed-loop control, effectively achieves very early warning and adaptive, precise fire suppression for battery fires.

[0056] To further verify the actual performance of the method of the present invention, power plant E conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Performance Comparison of Battery Combustion Detection and Automatic Fire Extinguishing Methods in Power Plant E

[0057] As shown in Table 1, the performance of the fire suppression system of the energy storage power station was comprehensively improved after applying the method of this invention. The accuracy of fire source identification increased from 79.5% with traditional methods to 97.2%, and the average spatial positioning error decreased from 0.85 meters to 0.12 meters, significantly improving the accuracy of fire perception and providing a reliable basis for precise fire suppression. The response time for reignition warning was drastically shortened from 120 seconds to 8 seconds, significantly enhancing the system's timeliness and effectively curbing the spread of battery thermal runaway. In addition, the utilization rate of extinguishing media increased from 65.0% to 92.5%, and the false alarm rate decreased from 8.5% to 0.9%, significantly reducing operation and maintenance costs and the risk of secondary disasters. The equipment property loss rate decreased from 12.0% to 2.5%, effectively ensuring the safety of the core assets of the energy storage power station.

[0058] Through the method of this invention, power station E has successfully achieved accurate perception and adaptive intelligent fire suppression of potential battery combustion hazards, effectively reducing the risk of safety accidents caused by thermal runaway, ensuring the safe and stable operation of the energy storage system, significantly improving the intelligence and digitalization level of the fire protection system, significantly reducing the monitoring pressure on operation and maintenance personnel, enhancing the power station's response capability to sudden fires, and providing strong technical support for the safe construction of electrochemical energy storage power stations.

[0059] 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 battery combustion detection and automatic fire extinguishing based on image recognition, characterized in that, Includes the following steps: S1. Construct a panoramic visual field, calculate the optimal suture line based on geodesic distance, use Laplacian pyramid layered weighted fusion to eliminate geometric faults and brightness jumps, and output a panoramic monitoring image sequence of battery combustion. S2. Input the battery combustion panoramic monitoring image sequence into the improved BiSeNet model to construct a high-fidelity flame texture detail feature tensor. Extract semantic features based on non-subsampled second-generation curvelet transform and Lipschitz exponential adaptive estimation. Utilize frequency domain phase correlation and morphological reconstruction weighted aggregation features to output the two-dimensional pixel coordinates of the fire source. S3. Construct a four-dimensional spatiotemporal hypervoxel graph structure based on the two-dimensional pixel coordinates of the fire source, segment the dynamic fire source hypervoxel region using graph clustering and calculate the spatiotemporal centroid, and output the spatial positioning coordinates of the fire source. S4. Using the spatial positioning coordinates of the fire source as the target, the path cost is progressively optimized by using a bidirectional fast expansion random tree algorithm to expand bidirectionally and perform rewiring operations, and the obstacle avoidance planning path of the fire extinguishing device is searched and output. S5. Calculate the information entropy gain of the uncertainty of the fire source state under different pressures based on the spatial positioning coordinates of the fire source, dynamically select the optimal spray pressure with the maximization of the information entropy decrease rate as the criterion, and output a directional fire extinguishing jet along the obstacle avoidance planning path of the fire extinguishing device. S6. Based on the panoramic monitoring image sequence of battery combustion and the fire source texture and smoke morphology feature data, calculate the Shannon information entropy of the grayscale histogram and the degree of correlation attenuation between consecutive frames, output real-time evaluation data of fire extinguishing efficiency, and provide feedback to correct the optimal injection pressure. S7. Based on the real-time evaluation data of the fire extinguishing effectiveness, after the flame disappears, enter the dormant monitoring stage, map the brightness time series of the acquired image into a symbol vector, calculate the permutation entropy to measure the dynamic complexity of the system, and if it is determined that the fire source reignites, recalculate the obstacle avoidance planning path of the fire extinguishing device and the optimal spray pressure.

2. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, S1 specifically includes: S11. Project the original images captured by multiple cameras onto a unified spherical manifold, and use the optical flow field to perform sub-pixel level registration of feature points in the overlapping areas of adjacent views to construct a geometrically aligned panoramic visual field. S12. Based on the geometrically aligned panoramic visual field, calculate the sum of the absolute value of the color difference and the gradient magnitude difference of the corresponding pixels in the overlapping area as the matching cost, traverse all possible stitching paths and accumulate the total generation value, and select the path with the smallest total generation value as the best stitching line. S13. Perform Laplacian pyramid decomposition on the overlapping region image along the optimal suture line, calculate the Euclidean distance from the pixel to the optimal suture line at different scale levels, construct Gaussian weighted fusion coefficients using the Euclidean distance, and perform weighted fusion on the images at each level to obtain multi-scale fusion sub-bands. S14. Perform a pyramid reconstruction operation on the multi-scale fusion sub-band to eliminate geometric breaks and brightness jumps at the splicing points, and output a panoramic monitoring image sequence of battery combustion.

3. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, The improved BiSeNet model includes a spatial path feature extraction layer, a context path feature extraction layer, an attention-weighted aggregation layer, a segmentation output layer, and a centroid calculation layer. The spatial path feature extraction layer is used to map the battery combustion panoramic monitoring image sequence to a polar coordinate grid, perform one-dimensional processing along the radial direction, and calculate the second central moment of the fractional-order spectrum using the energy difference between the flame flicker frequency and the background noise fractional-order spectrum with the transformation order as the independent variable. The gradient descent algorithm is used to iteratively search for the minimum point of the second central moment to lock the optimal transformation order. Within the optimal fractional-order domain, the energy distribution entropy of the transformation coefficient matrix is ​​calculated, a dynamic threshold is set to remove low-energy diffuse noise components, and high-energy focused texture components are retained. After inverse transformation reconstruction, a high-fidelity flame texture detail feature tensor is output. The context path feature extraction layer is used to perform non-downsampled second-generation curvelet transform on the high-fidelity flame texture detail feature tensor to obtain multi-directional sub-band coefficients; by utilizing the anisotropic singularity of the flame edge in different directional sub-bands, an adaptive estimation operator of the Lipschitz exponent based on logarithmic scale variation is constructed; by setting a regularity threshold, noise components with negative Lipschitz exponents are removed, while geometric structural components with positive exponents are retained and recombined, and a semantic feature tensor containing precise location priors is output; The attention-weighted aggregation layer is used to transform the high-fidelity flame texture detail feature tensor and the semantic feature tensor to the frequency domain. It uses phase spectrum correlation to locate spatial offset and drives the morphological reconstruction operator to perform erosion and dilation with connectivity constraints on the amplitude spectrum, removing isolated noise spikes and retaining continuous energy clusters. After frequency domain reconstruction and residual correction, a spatial weight map is generated, and a fused feature tensor that suppresses background artifacts and is spatially aligned is output. The segmentation output layer is used to input the fusion feature tensor that suppresses background artifacts and is spatially aligned into the convolutional layer for upsampling and prediction. After processing by the Softmax function, it outputs fire source texture and smoke morphology feature data, which includes flame region mask and smoke region mask. The centroid calculation layer is used to perform morphological processing on the flame region mask and the smoke region mask, calculate the weighted center of the mask pixel coordinates, and output the two-dimensional pixel coordinates of the fire source.

4. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, S3 specifically includes: S31. Based on the disparity information in the panoramic monitoring image sequence, perform stereo matching calculation to generate an original depth map. Project the original depth map onto the geometrically aligned panoramic visual field to generate a panoramic depth map. Using the two-dimensional pixel coordinates of the fire source as spatial seed points, combine the panoramic depth map to establish pixel adjacency associations between consecutive time frames to construct a four-dimensional spatiotemporal supervoxel map structure containing spatial location and timestamp. S32. Calculate the inter-frame optical flow vector field based on the panoramic monitoring image sequence. According to the optical flow vector field and the texture features of the nodes in the four-dimensional spatiotemporal hypervoxel graph structure, calculate the color similarity, spatial proximity and motion state consistency between the nodes. Traverse the nodes and aggregate the nodes that meet the dynamic combustion characteristics to segment out the dynamic fire source hypervoxel region. S33. Assign weight coefficients based on combustion response intensity to all pixels within the dynamic fire source super-voxel region, calculate the weighted three-dimensional spatial coordinate mean in combination with the panoramic depth map, and output the fire source spatial positioning coordinates.

5. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, S4 specifically includes: S41. Using the spatial positioning coordinates of the fire source as the target node and the current position of the fire extinguishing device as the starting node, construct a search space in the preset global dynamic environment semantic map and initialize a bidirectional fast-expanding random tree. S42. Generate new nodes by expanding from the starting node and the target node to the other random sampling point respectively, determine the collision between the new node and the obstacle, and if there is no collision, add the new node to the bidirectional fast expansion random tree and perform rewiring operation to progressively optimize the local path cost until the bidirectional tree nodes meet to generate the initial connection path. S43. Perform smoothing and redundant node removal operations on the initial connection path to eliminate unnecessary turning points in the path and output the obstacle avoidance planning path of the fire extinguishing device.

6. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, S5 specifically includes: S51. Based on the spatial positioning coordinates of the fire source and the environmental wind direction data, construct the jet motion differential equation, calculate the liquid fracture length and spray drift distance under the action of gravity and wind resistance, and generate the predicted coverage probability distribution of the target fire source area. S52. Based on the predicted coverage probability distribution, calculate the information entropy of the uncertainty of the fire source state under different pressure levels, construct the mapping relationship between pressure adjustment and information entropy gain, and select the pressure level that maximizes the rate of decrease of information entropy as the optimal injection pressure. S53. Control the fire extinguishing device to adjust to the optimal spray pressure, generate the corresponding spray pressure command, move along the obstacle avoidance planning path of the fire extinguishing device to the range, and output a directional fire extinguishing jet to the target fire source area.

7. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, S6 specifically includes: S61. Based on the panoramic monitoring image sequence of battery combustion and the texture and smoke morphology feature data of the fire source, extract the pixel intensity time series of the target area, calculate the mutual information value of pixel intensity between adjacent frames, and measure the degree of correlation decay between consecutive frames. S62. Construct a grayscale histogram of the target area and calculate the Shannon information entropy. Then, weight and fuse the Shannon information entropy with the correlation attenuation degree of the preceding and following frames to generate real-time fire extinguishing performance evaluation data. S63. Construct a fire extinguishing efficiency error transfer function, calculate the pressure compensation gradient based on the real-time fire extinguishing efficiency evaluation data, use the pressure compensation gradient to correct the optimal spray pressure, and output the updated spray pressure command.

8. The method for battery combustion detection and automatic fire extinguishing based on image recognition according to claim 1, characterized in that, Specifically, S7 includes: S71. After the flame characteristics disappear according to the real-time evaluation data of the fire extinguishing efficiency, enter the dormant monitoring mode, continuously collect the image brightness data of the monitoring area and construct the brightness time series. S72. Map the brightness time series to a symbol vector, calculate the permutation entropy of the symbol vector, and output the system dynamics complexity; S73. If the system dynamics complexity exceeds the preset reignition threshold, the fire source is determined to reignite, the current dormant state is terminated, and the obstacle avoidance planning path and optimal spray pressure of the fire extinguishing device are recalculated based on the updated environmental data until the fire source is completely extinguished.

Images