Substation fire warning method and system based on multi-spectrum fusion and anti-interference AI

Abstract

The present application belongs to the technical field of substation fire early warning, and particularly relates to a substation fire early warning method and system based on multispectral fusion and anti-interference AI. The method comprises the following steps: obtaining multispectral, infrared reference and low-resolution hyperspectral data and preprocessing; extracting high-pass components to construct a local double-scale coefficient matrix, combining spectral correction and total variation regularization to solve an anti-interference fusion image; inputting the fusion image and the low-resolution hyperspectral data into a functional data analysis framework, applying kernel norm constraint to reconstruct a hyperspectral feature cube; constructing a double-map structure and introducing a horseshoe prior to dynamically adjust the smoothing strength, combining low-rank constraint to determine an abnormal heat source area, extracting fire-sensitive features to generate a fire-sensitive feature vector set; inputting the fire-sensitive feature vector set into an AI model embedded in an interference suppression module to output fire probability values and development trend grades, and correcting according to real-time working conditions, and warning when the threshold is exceeded. The present application effectively suppresses complex interference and realizes precise early warning of early-stage fire.

Description

Technical Field

[0001] This invention belongs to the field of substation fire early warning technology, specifically involving a substation fire early warning method and system based on multispectral fusion and anti-interference AI, and particularly a substation primary equipment fire early warning method and system based on multispectral fusion and anti-interference AI algorithm. Background Technology

[0002] Under long-term high-load operation, primary equipment in substations is prone to overheating due to factors such as insulation aging and increased contact resistance, which can then develop into fire accidents. Among existing substation fire detection technologies, contact temperature measurement has limited deployment points and slow response; infrared thermal imaging has a high false alarm rate due to changes in ambient temperature and emissivity; and smoke detection is a passive alarm method after a fire has occurred, all of which are insufficient to meet the needs of very early warning.

[0003] In recent years, multispectral sensing and deep learning technologies have been attempted to be introduced into the field of substation fire early warning. However, existing solutions have the following shortcomings: First, substations have strong electromagnetic interference and complex background heat sources, and existing methods are insufficient in suppressing interference such as dynamic lighting changes and static structural occlusion. Second, there is a lack of spatial detail injection and spectral fidelity processing for multispectral images, resulting in low image quality. Third, it is difficult to reconstruct hyperspectral features from limited multispectral data, making it impossible to detect early weak thermal anomalies. Fourth, deep learning models are not designed for strong interference scenarios, and the accuracy of discrimination drops significantly under complex operating conditions. Fifth, there is a lack of dynamic confidence correction mechanisms based on real-time operating conditions, and fixed thresholds are prone to missed or false alarms.

[0004] To address the aforementioned issues, this invention proposes a substation fire early warning method and system based on multispectral fusion and anti-interference AI. Summary of the Invention

[0005] This invention aims to address the technical problems in existing substation fire early warning technologies, such as insufficient interference suppression capabilities, poor multispectral fusion quality, lack of hyperspectral features, weak anti-interference capabilities of deep learning models, and lack of dynamic confidence correction. It provides a substation fire early warning method and system based on multispectral fusion and anti-interference AI.

[0006] According to a first aspect of the present invention, a substation fire early warning method based on multispectral fusion and anti-interference AI is provided, comprising: acquiring substation multispectral image data, registered infrared thermal imaging reference images, and low spatial resolution hyperspectral data; constructing a multispectral data cube and reference image data, and performing radiometric and geometric preprocessing; extracting high-pass components from the preprocessed multispectral images and reference images respectively; constructing a local dual-scale coefficient matrix adaptively fitting the high-pass domain relationship; combining spectral correction and total variational regularization; iteratively solving and outputting an anti-interference fused image; and inputting the anti-interference fused image and the low spatial resolution hyperspectral data into a functional data analysis framework to estimate the regression... The parameters are optimized and nuclear norm constraints are applied to reconstruct the hyperspectral feature cube. A dual-graph structure is constructed based on the hyperspectral feature cube, and a horseshoe prior is introduced to dynamically adjust the smoothing intensity. Aberrant heat source regions are determined by combining low-rank constraints. Multiple fire-sensitive features are extracted within the aberrant heat source regions to generate a fire-sensitive feature vector set. The fire-sensitive feature vector set is input into an AI model embedded in the interference suppression module to output the fire probability value and the fire development trend level. A dynamic correction factor is constructed based on real-time collected environmental and equipment operation data to correct the confidence level of the fire probability value and the fire development trend level. When the corrected output exceeds the preset warning threshold, a warning signal is output.

[0007] In some optional implementations of certain embodiments, the iterative solution to output the anti-interference fusion image includes: upsampling the preprocessed multispectral image to the resolution of the reference image; extracting the high-frequency difference components of the multispectral image and the reference image respectively; calculating the ratio of the local mean and the ratio of the local standard deviation of the multispectral image and the reference image using the neighborhood window of each pixel as a unit, and using their product as the local dual-scale coefficient, combining them pixel by pixel to form a local dual-scale coefficient matrix; constructing a high-pass injection fidelity term with the local dual-scale coefficient matrix as weights to constrain spatial detail injection, constructing a spectral correction term for radiance ratio consistency to maintain spectral fidelity, and introducing a total variational regularization term to maintain edge structure; weighting and combining the high-pass injection fidelity term, the spectral correction term, and the total variational regularization term into a convex energy functional, iteratively solving it using the alternating direction multiplier method, and outputting the anti-interference fusion image.

[0008] In some optional implementations of some embodiments, the size of the neighborhood window is dynamically adjusted according to the local texture complexity, including: calculating the gradient magnitude variance in the neighborhood of each pixel as a texture metric; using a small window in texture-rich areas to maintain detail fitting accuracy; using a large window in flat areas to suppress noise interference; and achieving a continuous and smooth transition of window size in transition areas through linear interpolation.

[0009] In some optional implementations of certain embodiments, the hyperspectral feature cube is reconstructed by: inputting the anti-interference fused image and the low spatial resolution hyperspectral data into the functional data analysis framework; modeling the spectral response function and spatial point spread function as a combination of smooth basis functions, and estimating the degradation parameters through alternating iterations; reconstructing the multispectral vectors of each pixel in the anti-interference fused image into continuous spectral curves, and establishing a mapping relationship to the hyperspectral image based on the estimated degradation parameters; introducing a low-rank constraint of the nuclear norm to compress the estimation space, and iteratively optimizing the output of the hyperspectral feature cube.

[0010] In some optional implementations of certain embodiments, the estimation of the degradation parameters employs a superpixel-guided partitioning strategy, including: performing superpixel segmentation on the anti-interference fused image to divide it into multiple spectrally homogeneous sub-regions; independently estimating the spectral response function and spatial point spread function of each sub-region, and co-optimizing adjacent sub-regions through smoothness constraints to avoid parameter abrupt changes; applying the nuclear norm constraint to the coefficient matrix of each sub-region respectively, and merging each sub-region after independent singular value thresholding.

[0011] In some optional implementations of certain embodiments, the fire-sensitive feature vector set is generated as follows: Based on the hyperspectral feature cube, a spatial adjacency graph is constructed with each pixel as a vertex and spatial proximity as edges, and a spectral feature graph is constructed with the spectral angular distance between the spectral vectors of each pixel as edge weights, forming the dual-graph structure; a horseshoe prior distribution is introduced into the Bayesian hierarchical model, and an adaptive shrinkage prior is applied to the difference vector of each edge in the graph, so that the smoothing intensity is dynamically adjusted according to the local signal differences; in conjunction with low-rank tensor constraints, the hyperspectral feature cube is decomposed into low-rank background components and sparse anomaly components, and the normal heat distribution field is fitted by the low-rank background components, and the heat source regions deviating from the background are located by the sparse anomaly components as the anomalous heat source regions; the spectral gradient features, spatial texture features, and temporal heat accumulation rate features of each pixel are extracted in the anomalous heat source regions, and the three types of features are concatenated and combined to generate the fire-sensitive feature vector set.

[0012] In some optional implementations of certain embodiments, the construction of the dual-graph structure employs a superpixel-guided reduction strategy, including: performing superpixel segmentation on the spatial dimension of the hyperspectral feature cube, replacing individual pixels with superpixel blocks as graph vertices; the edges of the spatial adjacency graph connect adjacent superpixel blocks, with edge weights determined by the spatial distance between superpixel blocks; the edges of the spectral feature graph connect pairs of superpixel blocks with a spectral angular distance less than a preset threshold, with edge weights determined by the reciprocal of the spectral angular distance; the introduction of the horseshoe prior and the application of the low-rank tensor constraint are both performed at the superpixel level, and the identified abnormal heat source region is output with a superpixel block as the smallest unit.

[0013] In some optional implementations of certain embodiments, the fire probability value and fire development trend level are output in the following manner: the fire-sensitive feature vector set is input into a pre-trained deep learning model, which uses a multi-scale convolutional neural network to extract the spatial local patterns of the feature vectors and uses a Transformer temporal attention mechanism to capture the long-range dependencies of the feature vectors within a continuous time window; the interference suppression module is embedded in the intermediate layer of the deep learning model, which generates an interference frequency band attention mask in a learnable manner and dynamically reduces the weights of the response channels in the feature map related to background thermal fluctuations, ambient light drift, and sensor noise; after multi-scale feature fusion and mapping by a fully connected layer, the deep learning model simultaneously outputs the fire probability value at the current moment and the fire development trend level for multiple future time steps.

[0014] In some optional implementations of certain embodiments, the confidence correction and early warning signal output are achieved through the following methods: real-time acquisition of substation environmental temperature and humidity, wind speed and direction, and current equipment load rate data to construct a multi-dimensional environment-operational state vector; inputting the environment-operational state vector into a pre-established correction factor mapping network to output the dynamic correction factor matching the current operating conditions, wherein the dynamic correction factor is used to multiplicatively adjust the fire probability value and offset the fire development trend level; when the corrected fire probability value exceeds a preset early warning threshold and the corrected fire development trend level is not lower than a preset risk level, it is determined that the early warning triggering condition is met; the early warning signal includes the spatial location coordinates of the abnormal heat source, the fire type prediction label, and the handling priority level information, and is pushed to the local audible and visual alarm device, the centralized monitoring platform, and the mobile terminal respectively through different communication protocols.

[0015] According to a second aspect of the present invention, a substation fire early warning system based on multispectral fusion and anti-interference AI is provided, for implementing the substation fire early warning method based on multispectral fusion and anti-interference AI as described above, comprising: a data acquisition and preprocessing module, used to acquire substation multispectral image data, registered infrared thermal imaging reference image, and low spatial resolution hyperspectral data, construct a multispectral data cube and reference image data, and perform radiometric and geometric preprocessing; an anti-interference fusion image generation module, used to extract high-pass components from the preprocessed multispectral image and reference image respectively, construct a local dual-scale coefficient matrix adaptively fitting high-pass domain relationship, combine spectral correction and total variational regularization, and iteratively solve to output an anti-interference fusion image; and a hyperspectral feature reconstruction module, used to reconstruct the anti-interference fusion image and the low spatial resolution hyperspectral data. Using the same input function-based data analysis framework, the degradation parameters are estimated and nuclear norm constraints are applied to reconstruct the hyperspectral feature cube. A feature vector set generation module is used to construct a dual-graph structure based on the hyperspectral feature cube, and introduces a horseshoe prior to dynamically adjust the smoothing intensity. Combined with low-rank constraints, abnormal heat source regions are identified, and multiple fire-sensitive features are extracted within these regions to generate a fire-sensitive feature vector set. A fire prediction module is used to input the fire-sensitive feature vector set into an AI model embedded in the interference suppression module, outputting the fire probability value and fire development trend level. A confidence correction and early warning output module is used to construct a dynamic correction factor based on real-time collected environmental and equipment operation data, correct the fire probability value and fire development trend level with confidence, and output an early warning signal when the corrected output exceeds a preset early warning threshold.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) By extracting high-pass components and constructing a local dual-scale coefficient matrix to adaptively fit the high-pass domain relationship, combined with spectral correction and total variational regularization iterative solution, the interference of complex illumination and structural occlusion in substations is effectively suppressed, and the multispectral image fusion quality and spatial detail restoration accuracy are significantly improved; (2) The anti-interference fused image and low-resolution hyperspectral data are input into a functional data analysis framework, and the hyperspectral feature cube is reconstructed through degradation parameter estimation and low-rank constraint of nuclear norm, so as to mine fine spectral information from limited multispectral data and realize the effective perception of early weak thermal anomalies; (3) Constructing The dual-graph structure introduces horseshoe priors to dynamically adjust the smoothing intensity, and combines low-rank constraints to accurately separate background heat distribution and abnormal heat sources. It extracts multi-dimensional sensitive features such as spectral gradient, spatial texture and heat accumulation rate, and combines AI models with embedded interference suppression modules to dynamically reduce interference frequency band response, which significantly improves the detection accuracy and robustness under complex working conditions. (4) Based on real-time environment and equipment operation data, dynamic correction factors are constructed to perform confidence adaptive correction, which overcomes the defects of fixed thresholds that are easy to miss or false alarm, making the early warning output more in line with the actual working conditions of substations. The computational complexity is greatly reduced by the superpixel-guided graph reduction strategy, and it has good real-time deployment capabilities. Attached Figure Description

[0017] Figure 1 A schematic diagram of the substation fire early warning method based on multispectral fusion and anti-interference AI provided in this application.

[0018] Figure 2 The architecture diagram of the substation fire early warning system with multispectral fusion and anti-interference AI provided for this application.

[0019] Figure 3 A schematic diagram of the electronic device provided in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0021] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0022] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1 like Figure 1 As shown, this application provides a substation fire early warning method based on multispectral fusion and anti-interference AI, including the following steps.

[0024] S1: Acquire multispectral image data of the substation, registered infrared thermal imaging reference image and low spatial resolution hyperspectral data, construct multispectral data cube and reference image data, and perform radiometric and geometric preprocessing.

[0025] In this step, a multispectral sensor array is deployed around the primary equipment of the substation to simultaneously collect multi-band data such as ultraviolet, infrared thermal imaging, and visible light. Data alignment is completed according to the time stamp and spatial coordinates of each sensor to eliminate spatiotemporal misalignment caused by differences in acquisition time and installation position deviation. Then, the original digital signal is converted into physical quantity value through radiometric calibration, and the lens distortion effect is eliminated through geometric correction to form a standard data format that is easy to process later.

[0026] S2: Extract high-pass components from the preprocessed multispectral image and the reference image respectively, construct a local dual-scale coefficient matrix to adaptively fit the high-pass domain relationship, and combine spectral correction and total variation regularization to iteratively solve and output an anti-interference fused image.

[0027] In this step, the low-resolution multispectral image is first upsampled to the same size as the infrared thermal imaging reference image using an interpolation algorithm. The high-frequency difference between the two images after subtracting their own low-frequency components is calculated to extract their respective detail information. Then, taking each pixel and its surrounding neighborhood as the object of study, the similarity between the multispectral and reference images in spectral intensity and spatial texture is calculated. The product of the two is used as the weight coefficient of the pixel. After traversing pixel by pixel, a weight matrix with the same size as the image is formed. Subsequently, three constraints are constructed. The first is based on the weight matrix to control the high-frequency details of the reference image to be injected into the multispectral image in a pixel-by-pixel differentiated manner. The second constrains the ratio of radiation intensity of the fused image and the original image to remain consistent at different resolutions to protect spectral characteristics. The third applies an overall smoothing constraint to the image while allowing numerical jumps at edge positions to preserve the device outline. The three constraints are weighted and summed to form the solution objective. Finally, the optimal solution is approximated by iteratively using the alternating direction multiplier method, outputting a fused image that retains both multispectral color information and infrared high-resolution details.

[0028] Specifically, in this embodiment, the iterative solution for outputting the anti-interference fusion image includes the following steps.

[0029] S2.1 Upsample the preprocessed multispectral image to the resolution of the reference image, and extract the high-frequency difference components of the multispectral image and the reference image respectively. In this step, the low-resolution multispectral image is first enlarged to the same size as the infrared thermal imaging reference image using an interpolation algorithm. Then, the high-frequency difference component of the reference image is obtained by subtracting it from the low-frequency version obtained by downsampling and then upsampling. This component mainly reflects the spatial detail information such as the edge and texture of the device. Similarly, the high-frequency difference component of the multispectral image is obtained by subtracting it from its own low-frequency version. The degree of difference between the two high-frequency components reflects the inconsistency between them in terms of detail information.

[0030] S2.2. Using the neighborhood window of each pixel as a unit, calculate the ratio of the local mean of the multispectral image to that of the reference image and the ratio of the local standard deviation. The product of the two forms the local dual-scale coefficients, which are then combined pixel by pixel into a coefficient matrix.

[0031] For each pixel location, a neighborhood window is defined. Within this window, the ratio of the average brightness of the multispectral image to that of the reference image in the local region and the ratio of the intensity of local fluctuations are calculated. The former reflects the difference in spectral response, and the latter reflects the difference in texture structure. The two are multiplied to obtain a comprehensive coefficient. The closer this coefficient is to 1, the better the consistency of the spectrum and texture between the multispectral image and the reference image in the local region, and the higher the fidelity during detail injection. After traversing the entire image pixel by pixel, all coefficients are arranged according to pixel position to form a coefficient matrix with the same size as the image.

[0032] S2.3. Using the coefficient matrix as weights, construct a high-pass injection fidelity term to constrain spatial detail injection, and at the same time construct a spectral correction term for radiation intensity ratio consistency to maintain spectral fidelity, and introduce a total variational regularization term to maintain edge structure.

[0033] This step constructs three constraint terms. The first term uses the coefficient matrix as pixel-wise weights to constrain the high-frequency details of the reference image to be injected into the multispectral image according to their degree of differentiation. The higher the coefficient, the greater the injection weight. This constraint is called the high-pass injection fidelity term. The second term constrains the pixel-wise radiance ratio between the fused image and the reference image to remain consistent at different resolutions, avoiding changes to the spectral characteristics of the original multispectral image when injecting details. This constraint is called the spectral correction term. The third term constrains the sum of the absolute values ​​of the differences between adjacent pixels in the horizontal and vertical directions of the fused image to be as small as possible, but allows for larger jumps at edge positions, so as to preserve the device contour boundary while suppressing noise. This constraint is called the total variational regularization term.

[0034] S2.4. Combine the three weighted terms into a convex energy functional and solve iteratively using the alternating direction multiplier method.

[0035] This step adds the three constraints mentioned above according to their respective weights to form a convex energy function with a unique optimal solution; the alternating direction multiplier method is used to solve the problem. This method decomposes the complex problem into multiple subproblems. In each subproblem, only one variable is updated while the other variables are fixed. The global optimal solution is gradually approached by repeatedly updating the variables alternately; each iteration checks whether the difference between the current result and the previous result is less than the preset precision. If the precision is satisfied, the iteration is terminated, and the final high-resolution anti-interference fusion image is output.

[0036] Furthermore, in this embodiment, the size of the neighborhood window is dynamically adjusted according to the local texture complexity: the variance of the gradient magnitude within the neighborhood of each pixel is calculated as a texture metric; a small window is used in texture-rich areas to maintain detail fitting accuracy, a large window is used in flat areas to suppress noise interference, and linear interpolation is used in transition areas to achieve a smooth transition in window size. The choice of window size directly affects the accuracy of coefficient estimation; a single fixed size cannot simultaneously meet the different needs of texture-rich and flat areas. In this step, an initial neighborhood is defined at each pixel location, the variation amplitude of the grayscale value of each pixel in the horizontal and vertical directions within the neighborhood is calculated, and its variance is statistically analyzed as a quantitative indicator of texture complexity; a larger variance indicates that the region contains richer details such as edges and corners, while a smaller variance indicates that the grayscale distribution in the region is more uniform. The window size is dynamically determined based on the texture measurement results: In texture-rich areas, a smaller neighborhood window is used for coefficient calculation to accurately capture local detail changes and avoid excessive smoothing of edges due to an overly large window; in flat areas, a larger neighborhood window is used to increase the number of pixels involved in the calculation, improve statistical stability, and effectively suppress the interference of isolated noise points on coefficient estimation; for transitional areas between the two, the window size is linearly interpolated between two preset values ​​to achieve a continuous and smooth transition of the window size, avoiding discontinuous jumps in calculation results between adjacent pixels due to abrupt changes in window size, and ultimately ensuring that the local dual-scale coefficients of each pixel maintain both detail accuracy and noise suppression capability across the entire image.

[0037] Specifically, this invention constructs the following convex energy functional and defines a method for calculating the local dual-scale coefficient matrix to achieve pixel-wise adaptive detail injection: (I) Local dual-scale coefficient matrix , (1) in, This represents the spatial row and column position of a pixel in the image; respectively with Within the neighborhood window centered on the image, the local gray-scale mean of the multispectral image and the local gray-scale mean of the infrared reference image are calculated. These represent the local grayscale standard deviations of the multispectral image and the infrared reference image within the corresponding window, respectively. This is a local dual-scale coefficient; the closer the value is to 1, the higher the consistency of the local spectrum and texture between the two images.

[0038] The size of the neighborhood window mentioned above is not a fixed value, but is dynamically adjusted according to the local texture complexity: , (2) in, For pixels The actual size (side length) of the neighboring window used at that location. These are the preset maximum and minimum window sizes; The measure of local texture complexity is defined as the variance of gradient magnitude within the neighborhood. Within the entire map area The maximum and minimum values; The gradient magnitude variance, These are the gradient operators for the horizontal and vertical directions.

[0039] (II) Anti-interference fusion of convex energy functionals , (3) in, Let be a matrix, representing the fused image to be solved. This represents the final output anti-interference fused image; represents the local dual-scale coefficient matrix calculated pixel by pixel using formula (1); ⊙ represents the Hadamard product, i.e., element-wise multiplication; This represents the high-frequency difference components of the infrared reference image. ; For the high-frequency difference components of a multispectral image, ); This is a preprocessed infrared thermal imaging reference image; This is a multispectral image upsampled to the resolution of the reference image; This is a low-pass filter operator (using Gaussian filtering); The downsampling operator (using bicubic interpolation); This is a gradient operator along the spectral dimension, used to calculate the radiance ratio; For anisotropic total variational regularization; These are the weighting coefficients for the three constraints.

[0040] (III) Iterative solution using the alternating direction multiplier method Equation (3) is solved iteratively using the alternating direction multiplier method (ADMM), and its iterative format is as follows: (4) , (5) , (6) in, This represents the current iteration step number; This represents the sum of the data fidelity term and the spectral correction term in formula (3), excluding the total variation term. This represents the penalty parameter, which controls the iteration convergence speed. As an auxiliary variable, it is used to decouple the total variation terms; Represents the dual variable (scaled form) used for Lagrange multiplier updates; This represents the spatial gradient operator, which includes both horizontal and vertical directions; This represents the soft-threshold shrinkage operator; the iteration termination condition is... Or, the maximum number of iterations can be reached as 100.

[0041] Input preprocessing and ; Calculate the dynamic window size pixel by pixel according to formula (2), and calculate pixel by pixel according to formula (1). Combining ; calculate separately and ;initialization , , ; Repeat formulas (4)-(6) until convergence; Output .

[0042] S3: Input the anti-interference fused image and the low spatial resolution hyperspectral data into a functional data analysis framework, estimate the degradation parameters and apply nuclear norm constraints, and reconstruct the hyperspectral feature cube.

[0043] In this step, the fused image generated in the previous step is input into a functional data analysis framework along with the pre-acquired low-resolution hyperspectral data. The core idea of ​​this framework is to fit the discrete spectral sampling values ​​of each pixel in each band into a continuous and smooth spectral curve, thereby mathematically restoring the physical continuity of the spectrum. First, the framework is used to iteratively estimate the spectral response function and the spatial point spread function of the sensor. The former describes the sensitivity of each band of the sensor to incident light, and the latter describes the blurring and diffusion effect of the imaging system on point light sources. Then, a nuclear norm constraint is added during the reconstruction process, which forces the sum of singular values ​​of the data to be reconstructed in the spectral dimension to be as small as possible. This utilizes the high correlation between hyperspectral bands to compress the solution space. Through repeated iterations, the final output is a hyperspectral feature cube that contains both rich spatial details and complete spectral information.

[0044] Specifically, in this embodiment, the hyperspectral feature cube is reconstructed in the following manner.

[0045] S3.1. Input the anti-interference fused image and the low spatial resolution hyperspectral data into a functional data analysis framework. This step uses the previously generated anti-interference fused image and the pre-acquired low spatial resolution hyperspectral data as input into a functional data analysis framework. This framework differs from the traditional approach of treating the spectrum as a discrete vector. Its basic idea is to treat the discrete sampled values ​​of each pixel in each band as observation points on a continuous curve, thereby maintaining the physical continuity and smoothness of the spectrum in subsequent processing.

[0046] S3.2. The spectral response function and spatial point spread function are modeled as combinations of smooth basis functions, and the degradation parameters are estimated through alternating iterations. In this step, the spectral response function and spatial point spread function of the sensor are respectively represented as weighted combinations of several smooth basis functions. The spectral response function describes the sensitivity of the sensor to incident light in each band, and the spatial point spread function describes the diffusion effect of the imaging system on an ideal point light source. An alternating iterative strategy is used to solve the problem: first, the spatial point spread function is fixed to estimate the spectral response function, then the spectral response function is fixed to update the spatial point spread function. The two are repeated alternately to gradually approximate the true degradation parameters, making the degradation model highly consistent with the actual imaging system.

[0047] S3.3 Reconstruct the multispectral vectors of each pixel in the fused image into a continuous spectral curve, establishing a mapping relationship to the hyperspectral image. In this step, the multispectral vector at each pixel location in the anti-interference fused image, i.e., the discrete values ​​of that pixel in each multispectral band, is expanded and fitted into a continuous and smooth spectral curve through basis function expansion, so that it is no longer limited by the finite number of discrete bands; then, using the spectral response function estimated in S3.2, a mathematical mapping relationship between this continuous curve and the hyperspectral image is established, projecting the multispectral information of each pixel onto the hyperspectral space to obtain an initial estimate of each pixel in the entire band range.

[0048] S3.4 Introducing a low-rank constraint based on the nuclear norm to compress the estimation space and iteratively optimizing the output hyperspectral feature cube. This step introduces a nuclear norm constraint during the reconstruction process, which forces the sum of singular values ​​in the spectral dimension of the hyperspectral data to be reconstructed to be as small as possible. Since there is a high degree of correlation between continuous bands in the hyperspectral spectrum, its inherent dimensionality is much lower than the actual number of bands; this constraint effectively compresses the solution space. The mapping relationship established in S3.3 is used as a data fidelity term, which, together with the nuclear norm constraint term, constitutes the optimization objective. Through iterative adjustments, the spectral curve shape of each pixel is gradually adjusted until both data consistency and low-rank characteristics are simultaneously satisfied, ultimately outputting a feature cube with high spatial and spectral resolution.

[0049] Specifically, this invention models the degradation parameters as a combination of smooth basis functions, and reconstructs the hyperspectral feature cube by applying sub-regional nuclear norm constraints after alternating iterative estimation. (I) Smooth basis function representation of degenerate parameters The spectral response function and the spatial point spread function are respectively expressed as a weighted combination of B-spline basis functions: , , (7) in, For spectral wavelength variables; Spatial coordinates (continuous values); For the first One B-spline spectral basis function; For the first One B-spline spatial basis function (two-dimensional separable B-spline); The spectral response function is the first One expansion coefficient; The first point spread function of space One expansion coefficient; The number of basis functions is automatically selected by the Akaike Information Criterion (AIC), typically M=8 and N=6.

[0050] (ii) Independent estimation of superpixel partitions Superpixel segmentation (using the SLIC algorithm) is performed on the anti-interference fused image, dividing it into... Each sub-region has spectrally homoproton regions. Degradation parameters are estimated independently for each sub-region, while neighboring regions are collaboratively optimized through smoothness constraints. , (8) in, For superpixel sub-region indexing, ; For the first The expanded coefficient vector of the spectral response of each sub-region; For the first The spatial point diffusion expansion coefficient vector of each sub-region; For the first Observational data for each sub-region (a combination of fused images and low-resolution hyperspectral data); For the first Hyperspectral data to be reconstructed for each sub-region; For the reason A defined spectral response function matrix; For the reason A defined spatial point spread function matrix; In order to be with the first A set of indexes of spatially adjacent subregions; Weights for smoothness constraints in adjacent regions; Let be the Euclidean norm of the vector.

[0051] (III) Reconstruction of Low-Rank Constraints on Nuclear Norm Apply nuclear norm constraints at the sub-region level, and then stitch together the results after independently performing singular value threshold shrinkage on each sub-region: , (9) in, The hyperspectral feature cube to be reconstructed (expanded into matrix form). The final reconstructed output is a hyperspectral feature cube; For the first Each sub-region corresponds to submatrices; For global observation data ( (in conjunction with low-resolution hyperspectral data). For nuclear norm, for The One singular value; The low-rank constraint regularization coefficient is determined adaptively using the L-curve method.

[0052] exist Perform SLIC superpixel segmentation to obtain Each sub-region; for each sub-region Estimate alternately according to formula (8) and (Fix one side and update the other side, repeat until convergence); Solve the optimization problem of formula (9) using the singular value threshold shrinkage algorithm: In each iteration, for each sub-region Perform singular value decomposition to determine values ​​less than the threshold. The singular values ​​are set to zero; the sub-regions are shrunk and then stitched together to form a complete system. It is reshaped into a three-dimensional hyperspectral feature cube.

[0053] Furthermore, in this embodiment, the estimation of degradation parameters employs a superpixel-guided partitioning strategy: S3.2.1 Perform superpixel segmentation on the anti-interference fused image, dividing it into spectrally homogeneous sub-regions. This step performs superpixel segmentation on the anti-interference fused image, dividing the image into several spatially continuous and spectrally homogeneous sub-regions. Superpixel segmentation aggregates adjacent pixels with similar colors, textures, and other characteristics into a whole, making the pixels within each sub-region tend to be consistent in spectral characteristics, providing reasonable partitioning for subsequent independent processing of each region.

[0054] S3.2.2. The spectral response function and spatial point spread function are independently estimated for each sub-region, while adjacent regions are collaboratively optimized through smoothness constraints to avoid abrupt parameter changes. Within each sub-region, the spectral response function and spatial point spread function are independently estimated to adapt the degradation parameters to the local imaging characteristics of different land cover types or equipment materials. Simultaneously, the parameter estimation results between adjacent sub-regions are collaboratively optimized through smoothness constraints, forcing the degradation parameters of adjacent regions to maintain a continuous transition at the boundaries, avoiding abrupt parameter changes caused by independent estimation of sub-regions, and ensuring the spectral consistency and spatial continuity of the fused image at the boundaries of sub-regions.

[0055] S3.2.3 The nuclear norm constraint is applied to the coefficient matrix of each sub-region, and then merged after threshold shrinkage of independent singular values. The nuclear norm constraint is not applied to the entire hyperspectral data, but is applied to the coefficient matrix within each sub-region. That is, the spectral singular values ​​of each sub-region are calculated and threshold shrinkage is performed, weak components below the threshold are removed and the main components are retained, thereby strengthening the low-rank characteristics within each sub-region. After the shrinkage of each sub-region is completed, they are spliced ​​and merged according to their original spatial positions to form a complete hyperspectral feature cube, so that the low-rank constraint can exert the optimal constraint effect in different land cover types.

[0056] S4: Construct a dual-graph structure based on the hyperspectral feature cube, and introduce a horseshoe prior to dynamically adjust the smoothing intensity. Combine low-rank constraints to determine abnormal heat source regions, extract multiple fire-sensitive features within the abnormal heat source regions, and generate a fire-sensitive feature vector set.

[0057] In this step, two graph structures are constructed on the hyperspectral feature cube: a spatial graph formed by connecting pixels based on spatial proximity, and a spectral graph formed by connecting pixels based on spectral curve similarity, forming a dual-graph structure. A horseshoe prior is introduced into the model solution, allowing the smoothing intensity of each connecting edge to autonomously adjust based on the actual difference between the pixel values ​​at both ends. Edges with large differences have reduced constraints to preserve abrupt edges, while edges with small differences have increased constraints to smooth noise. Combined with low-rank constraints, the data is decomposed into a background subject and abnormal residuals. The background subject represents the normal equipment temperature distribution, while areas in the abnormal residuals that significantly deviate from zero are suspected heat source locations. Finally, within the suspected heat source regions, the steepness of the spectral change with wavelength, the spatial texture roughness, and the cumulative rate of temperature increase over time are calculated. These indicators are then concatenated and combined into a numerical vector that fully describes the early characteristics of a fire.

[0058] Specifically, in this embodiment, the fire-sensitive feature vector set is generated in the following way.

[0059] S4.1 Based on the hyperspectral feature cube, a spatial adjacency graph is constructed with each pixel as a vertex and spatial proximity as an edge, and a spectral feature graph is constructed with the spectral angular distance between the spectral vectors of each pixel as the edge weight, forming a dual-graph structure.

[0060] This step constructs two graph structures based on a hyperspectral feature cube. The spatial adjacency graph uses each pixel as a vertex, connecting spatially adjacent pixels with edges to characterize the spatial continuity and spread trend of the heat source region. The spectral feature graph uses the spectral vector of each pixel as a vertex, calculating the spectral angular distance between any two pixel spectral curves. The smaller the angle, the closer the spectral characteristics of the two pixels. Pixel pairs with a distance less than a preset threshold are connected with edges, the edge weight being the reciprocal of the spectral angular distance. These two types of graphs collaboratively describe the geometric structure of the data from two dimensions: spatial adjacency and spectral similarity.

[0061] S4.2 Introduce a horseshoe prior distribution in the Bayesian hierarchical model, apply an adaptive shrinkage prior to the difference vector of each edge in the graph, and dynamically adjust the smoothing intensity according to the local signal differences.

[0062] This step introduces a horseshoe prior distribution into the Bayesian hierarchical model, applying an adaptive shrinkage prior to the difference vectors of pixel values ​​at both ends of each edge in the graph. The horseshoe prior is a heavy-tailed distribution, its core characteristic being that it allows each difference term to have an independent shrinkage strength: when the difference value is large, the horseshoe prior assigns a larger shrinkage parameter value to edges with larger difference values, resulting in a weaker shrinkage effect and thus preserving the abrupt changes in the heat source edges; conversely, it assigns a smaller shrinkage parameter value to edges with difference values ​​close to zero, resulting in a stronger shrinkage effect and effectively smoothing background thermal noise. This mechanism eliminates the need for manually setting a global smoothing coefficient; instead, the data itself drives each edge to independently adjust its smoothing strength.

[0063] Specifically, this invention constructs a dual-graph structure at the superpixel level and establishes the following Bayesian hierarchical model to introduce a horseshoe prior: (I) Construction of Superpixel-level Dual-Graph Structure Superpixel segmentation is performed in the spatial dimension of the hyperspectral feature cube, replacing individual pixels with superpixel blocks as graph vertices. Two graph structures are defined: , (10) Spectral feature map edge weights: , (11) in, Superpixel block index, , The spatial coordinates (row, column) of the centroid of the superpixel block; This refers to the spatial bandwidth parameter; It is the mean of the spectral vectors of all pixels within the superpixel block; Spectral angular distance, ; Spectral connection threshold (radians); This represents the edge weights in a spatial adjacency graph; a larger value indicates a stronger spatial association. This represents the edge weights of the spectral feature map; a larger value indicates higher spectral similarity. (II) Horseshoe Prior Adaptive Contraction Model Apply a horseshoe prior to the difference vector of each edge in the graph to adaptively adjust the smoothing intensity based on local signal differences: , (12) , (13) Equivalently, the horseshoe a priori can be represented in the following hierarchical form: , , (14) in, For superpixel blocks and The spectral difference vector between them; For the first The adaptive shrinkage parameter of the strip controls the smoothing intensity of the strip; It is the identity matrix; It follows a normal distribution; It follows a standard semi-Cauchy distribution (only positive values ​​are considered); It exhibits an inverse gamma distribution; As an intermediate auxiliary variable used for hierarchical representation; the core characteristic of the horseshoe prior: when When it is large, The posterior distribution is concentrated at large values, with weak contraction (preserving abrupt changes at the edge of the heat source); when When it is close to 0, The posterior distribution is concentrated at small values, and the contraction is strong (smoothing background thermal noise).

[0064] (III) Low-rank tensor decomposition and anomaly region identification The hyperspectral feature cube is decomposed into a low-rank background component and a sparse anomaly component: , (15) in, The reconstructed hyperspectral feature cube (superpixel level, listed as the spectral mean of each superpixel); The low-rank background component is used to fit the normal thermal distribution field. For sparse anomalous components, locate the anomalous heat source region; To enforce the low-rank nature of the background component as the nuclear norm; For superpixel-level group sparse constraints For the first The coefficient vector corresponding to each superpixel; This is the Laplacian matrix of the spatial adjacency graph and the spectral feature graph; Expand the horseshoe-shaped prior contraction parameter matrix into vector form; These are the three weighting coefficients; It is a vector consisting entirely of 1s.

[0065] The criteria for determining abnormal heat source areas are: if Formula (16) then determines the first One superpixel represents an abnormal heat source region. To set a preset multiple threshold, take =3.0, superpixel blocks above this threshold are marked as abnormal heat source regions. This represents the median function, which takes the middle value after sorting all values ​​in the set in ascending order; when the number of elements in the set is even, it takes the arithmetic mean of the two middle values.

[0066] exist Perform superpixel segmentation in the spatial dimension to obtain A superpixel block; construct a spatial adjacency graph and a spectral feature graph according to formulas (10) and (11) respectively to form a dual-graph structure; use Gibbs sampling or variational inference to solve the horseshoe prior and posterior distributions of formulas (12)-(14) to estimate the sides. The alternating direction multiplier method (ADMM) is used to solve formula (15), and L and S are updated alternately. The abnormal heat source region is determined according to formula (16), and the superpixel block is output as the smallest unit.

[0067] S4.3. Combine low-rank tensor constraints to decompose the hyperspectral feature cube into low-rank background components and sparse anomaly components. Fit the normal thermal distribution field through the low-rank background components and locate the heat source region that deviates from the background through the sparse anomaly components.

[0068] This step treats the hyperspectral feature cube as a three-dimensional tensor and decomposes it into a low-rank background component and a sparse anomaly component through low-rank tensor decomposition. The low-rank background component captures the stable thermal distribution field during normal operation of the device, which exhibits strong correlation in both spatial and spectral dimensions. The sparse anomaly component captures local thermal anomalies that deviate from the background, with only a few pixel locations having non-zero values. By separating the background and the anomaly, the heat source region deviating from the background can be accurately located, achieving effective detection of anomalous heating points.

[0069] S4.4 Extract the spectral gradient features, spatial texture features and temporal thermal accumulation rate features of each pixel within the heat source area, and combine the three types of features in series to generate a fire-sensitive feature vector set.

[0070] Within the located heat source area, three types of features are extracted for each pixel: spectral gradient features measure the steepness of the spectral curve change at abnormal bands, used to distinguish between thermal anomalies and differences in material reflection; spatial texture features describe the roughness and shape regularity of the heat source area's edges, used to distinguish between equipment heating and external heat source interference; and temporal heat accumulation rate features calculate the slope of temperature rise within a continuous time window, used to distinguish between sudden fires and slow heat accumulation. These three types of features are concatenated sequentially to form a high-dimensional vector, which serves as the input to the subsequent AI model.

[0071] Furthermore, in this embodiment, the construction of the dual-graph structure adopts a superpixel-guided reduction strategy: S4.1.1 Perform superpixel segmentation on the spatial dimension of the hyperspectral feature cube, and replace individual pixels with superpixel blocks as graph vertices.

[0072] This step performs superpixel segmentation in the spatial dimension of the hyperspectral feature cube, aggregating spatially continuous pixels with similar spectral characteristics into superpixel blocks, and replacing individual pixels with these superpixel blocks as graph vertices. For a hyperspectral image containing hundreds of thousands of pixels, superpixel segmentation can reduce the vertex size by two to three orders of magnitude, significantly reducing the storage overhead and computation time of subsequent graph processing.

[0073] S4.1.2. Edges in the spatial adjacency graph connect adjacent superpixel blocks, and the edge weights are determined by the spatial distance between the superpixel blocks; edges in the spectral feature graph connect pairs of superpixel blocks whose spectral angular distance is less than a preset threshold, and the edge weights are determined by the reciprocal of the spectral angular distance.

[0074] Two graph structures are constructed at the superpixel level. The spatial adjacency graph connects spatially adjacent superpixel blocks, with edge weights calculated as the spatial distance between the centroids of the superpixel blocks; smaller distances indicate stronger spatial associations. The spectral feature graph calculates the spectral angular distance between the spectral mean vectors of any two superpixel blocks, connecting only pairs of superpixel blocks with a distance less than a preset threshold. Edge weights are the reciprocal of this distance, with higher weights for spectrally similar superpixel pairs, tending to maintain numerical similarity during smoothing, thus effectively capturing long-distance associations that are nonlocal but spectrally homogeneous.

[0075] S4.1.3 The introduction of the horseshoe prior and the application of low-rank tensor constraints are both performed at the superpixel level, and the identified anomalous heat source regions are output as superpixel blocks as the smallest unit. The adjustment of the horseshoe prior's contraction intensity and the sparse anomaly extraction from low-rank tensor decomposition are both performed at the superpixel level; that is, each superpixel block participates as a whole in the dynamic adjustment of smoothing intensity and the background-anomaly separation calculation. The finally identified anomalous heat source regions are output as superpixel blocks as the smallest spatial unit, no longer reverting to the single pixel level, thus preserving the boundary morphology of the anomalous regions while significantly improving computational efficiency.

[0076] S5: Input the fire-sensitive feature vector set into the AI ​​model embedded in the interference suppression module, and output the fire probability value and the fire development trend level.

[0077] In this step, the feature vectors generated in the previous step are input into a pre-trained deep learning network in chronological order. The front end of this network uses a multi-scale convolutional structure to extract local spatial patterns, while the back end uses a Transformer self-attention mechanism to capture dependencies across time windows. An interference suppression module is added to the middle layer of the network. This module automatically learns and generates suppression masks for various types of background interference, weakening channels in the feature map associated with interference factors such as thermal fluctuations, illumination changes, and sensor noise, thereby reducing the impact of interference on the final judgment. After multiple layers of abstraction, the network outputs two results simultaneously through a fully connected layer: the probability of a fire occurring at the current moment and the trend level of the fire intensity over multiple future time steps.

[0078] Specifically, in this embodiment, the fire probability value and development trend level are output in the following manner: S5.1 Input the fire-sensitive feature vector set into a pre-trained deep learning model. The model uses a multi-scale convolutional neural network to extract the spatial local patterns of the feature vectors and uses the Transformer temporal attention mechanism to capture the long-range dependencies of the feature vectors within a continuous time window.

[0079] This step involves arranging the fire-sensitive feature vector set generated in the previous step in chronological order and inputting it into a pre-trained deep learning model. The model's front end employs a multi-scale convolutional neural network, using convolutional kernels of different sizes in parallel to extract local patterns in the spatial dimension of the feature vectors. Small-sized kernels capture fine texture changes, while large-sized kernels capture coarse-grained regional distributions. This combination allows the model to balance detail sensitivity and region awareness. Simultaneously, the model uses a Transformer temporal attention mechanism to calculate attention weights between feature vectors at different times within a continuous time window, thereby capturing the long-term dependencies and evolutionary patterns exhibited during fire development.

[0080] S5.2. An interference suppression module is embedded in the middle layer of the model. This module generates an attention mask for the interference frequency band in a learnable manner and dynamically reduces the weight of the response channels in the feature map that are related to background thermal fluctuations, ambient light drift and sensor noise.

[0081] An interference suppression module is embedded in the middle layer of the model. This module automatically generates a set of attention mask coefficients that correspond one-to-one with the feature map channels in a learnable manner. During model training, these coefficients are continuously optimized through backpropagation, gradually identifying which feature channels are sensitive to interference factors such as background thermal fluctuations, ambient light drift, and sensor noise. During actual inference, the interference suppression module dynamically reduces the weight of these interference-sensitive channels, weakening the impact of interference factors on high-level semantic features, so that subsequent decisions are more focused on effective features related to the fire situation.

[0082] S5.3 After multi-scale feature fusion and fully connected layer mapping, the model simultaneously outputs the probability value of fire occurrence at the current moment and the fire development trend level at multiple future time steps.

[0083] The model's backend fuses features at different scales and inputs them into a fully connected layer. The output layer contains two branches: one branch outputs the probability of a fire occurring at the current moment, reflecting the immediate risk of the current situation; the other branch outputs the fire development trend level for multiple future time steps, reflecting the possible changes in the fire's trajectory. This dual-output design enables the system to provide both immediate early warning judgments and forward-looking handling references for operations and maintenance personnel.

[0084] S6: Construct a dynamic correction factor based on real-time collected environmental and equipment operation data, and correct the confidence level of the fire probability value and fire development trend level. When the corrected output exceeds the preset warning threshold, output a warning signal.

[0085] In this step, real-time operating parameters such as ambient temperature and humidity, wind speed and direction, and equipment load rate of the substation are accessed. These parameters are input into a lightweight mapping network, which outputs a dynamic correction coefficient that matches the current operating conditions. This coefficient multiplicatively adjusts the fire probability output by the model and corrects the trend level. An early warning is triggered only when the corrected probability exceeds the system's preset threshold and the corrected trend level meets or exceeds the risk level requirements. The early warning information includes the coordinates of the abnormal heat source, the determined fire type, and the handling priority. It is sent to the on-site audible and visual alarms, the back-end monitoring screen, and the handheld terminals of maintenance personnel through different communication methods to ensure accurate information transmission.

[0086] Specifically, in this embodiment, the confidence correction and early warning signal output are implemented in the following ways: S6.1 Real-time acquisition of substation environmental temperature, humidity, wind speed and direction, and current equipment load rate data to construct a multi-dimensional environment-operation status vector.

[0087] This step involves real-time access to the environmental monitoring sensors and equipment monitoring system at the substation site to collect data such as ambient temperature, air humidity, wind speed and direction, and real-time load rate of primary equipment. These heterogeneous data are then integrated into a multi-dimensional state vector describing the current external conditions and internal operating conditions, providing an objective basis for subsequent confidence level adjustments.

[0088] S6.2 Input the environment-operational state vector into a pre-established correction factor mapping network and output a dynamic correction factor that matches the current operating condition. This factor is used to multiplicatively adjust the fire probability value and correct the offset of the development trend level.

[0089] This step inputs the multidimensional state vector into a pre-established correction factor mapping network. This network learns the distribution patterns of false alarms and missed alarms under different operating conditions in historical data, establishing a mapping relationship from operating condition state to correction coefficients, and outputting a dynamic correction factor that matches the current operating condition. This factor multiplicatively adjusts the fire probability output by the model, appropriately lowering the confidence level under operating conditions that are prone to triggering false alarms, such as high temperature and humidity, and appropriately increasing the confidence level under operating conditions where the fire risk is indeed high, such as high equipment load operation; at the same time, it performs offset correction on the development trend level, making the early warning judgment more consistent with the current actual operating scenario.

[0090] S6.3 When the corrected fire probability value exceeds the preset warning threshold and the corrected development trend level is not lower than the preset risk level, the warning triggering condition is determined to be met.

[0091] This step applies a joint decision condition to the two corrected outputs. The warning trigger condition is only met when the corrected fire probability value exceeds the system's preset warning threshold and the corrected development trend level reaches or exceeds the preset minimum risk level. This dual-condition design effectively prevents false triggering caused by accidental changes in a single indicator.

[0092] S6.4 The early warning signal includes the spatial coordinates of the abnormal heat source, the fire type prediction label, and the priority level information for handling, and is pushed to the local audible and visual alarm device, the centralized monitoring platform, and the mobile terminal through different communication protocols.

[0093] Once the warning is triggered, the system generates complete warning information, including the spatial coordinates of the abnormal heat source, a fire type prediction label, and a response priority level. Depending on the priority level, the information is pushed to local audible and visual alarm devices for on-site warning, to a centralized monitoring platform for dispatching and decision-making, and to the mobile terminals of maintenance personnel for rapid response, all via different communication protocols. This diversified push method ensures that the warning information is delivered to all responsible parties in a timely and accurate manner.

[0094] Example 2 This embodiment, based on Embodiment 1 above, provides a substation fire early warning system with multispectral fusion and anti-interference AI. Please refer to [link to previous document]. Figure 2 The substation fire early warning method for implementing multispectral fusion and anti-interference AI in Embodiment 1 specifically includes a data acquisition and preprocessing module 210, an anti-interference fusion image generation module 220, a hyperspectral feature reconstruction module 230, a feature vector set generation module 240, a fire prediction module 250, and a confidence correction and early warning output module 260.

[0095] The data acquisition and preprocessing module 210 is used to acquire multispectral image data of the substation, registered infrared thermal imaging reference images and low spatial resolution hyperspectral data, construct multispectral data cubes and reference image data, and perform radiometric and geometric preprocessing.

[0096] The anti-interference fusion image generation module 220 is used to extract high-pass components from the preprocessed multispectral image and the reference image, respectively, construct a local dual-scale coefficient matrix to adaptively fit the high-pass domain relationship, and combine spectral correction and total variation regularization to iteratively solve and output the anti-interference fusion image.

[0097] The hyperspectral feature reconstruction module 230 is used to input the anti-interference fused image and the low spatial resolution hyperspectral data into a functional data analysis framework, estimate the degradation parameters and apply the nuclear norm constraint, and reconstruct the hyperspectral feature cube.

[0098] The feature vector set generation module 240 is used to construct a dual-graph structure based on the hyperspectral feature cube, introduce horseshoe prior to dynamically adjust the smoothing intensity, and jointly determine the abnormal heat source region with low-rank constraints. Multiple fire-sensitive features are extracted in the abnormal heat source region to generate a fire-sensitive feature vector set.

[0099] The fire prediction module 250 is used to input the fire sensitive feature vector set into the AI ​​model embedded in the interference suppression module, and output the fire probability value and the fire development trend level. The confidence correction and early warning output module 260 is used to construct a dynamic correction factor based on real-time collected environmental and equipment operation data, and to perform confidence correction on the fire probability value and fire development trend level. When the corrected output exceeds the preset early warning threshold, an early warning signal is output.

[0100] Example 3 Based on Embodiment 1 described above, this embodiment also provides an electronic device, please refer to the appendix. Figure 3 , Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this disclosure. The electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0101] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0102] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0103] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0104] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0105] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0106] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0107] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0108] Example 4 Based on Embodiment 1 above, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above method.

[0109] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), or any suitable combination thereof.

[0110] In this embodiment, the client and server can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol), and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0111] The aforementioned computer-readable medium may be included in the aforementioned device or may exist independently without being assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire substation multispectral image data, registered infrared thermal imaging reference images, and low spatial resolution hyperspectral data; construct a multispectral data cube and reference image data, and perform radiometric and geometric preprocessing; extract high-pass components from the preprocessed multispectral images and reference images respectively; construct a local dual-scale coefficient matrix to adaptively fit the high-pass domain relationship; combine spectral correction and total variational regularization; iteratively solve and output an anti-interference fused image; and input the anti-interference fused image and the low spatial resolution hyperspectral data into a functional data analysis box. The system reconstructs a hyperspectral feature cube by estimating degradation parameters and applying nuclear norm constraints. Based on this cube, a dual-graph structure is constructed, and a horseshoe prior is introduced to dynamically adjust the smoothing intensity. Combined with low-rank constraints, abnormal heat source regions are identified. Multiple fire-sensitive features are extracted from these regions to generate a fire-sensitive feature vector set. This vector set is then input into an AI model embedded in an interference suppression module, outputting a fire probability value and a fire development trend level. A dynamic correction factor is constructed based on real-time collected environmental and equipment operation data to adjust the confidence level of the fire probability value and fire development trend level. When the corrected output exceeds a preset warning threshold, a warning signal is output.

[0112] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0113] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0114] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), etc.

[0115] Obviously, those skilled in the art will understand that the various steps of the present invention described above can be performed in a manner different from that described above, and the simulation methods and experimental equipment include, but are not limited to, the above description. The steps of the present invention described above can be performed in a different order in certain circumstances, and the steps shown or described above can be performed separately. Therefore, the present invention is not limited to any particular combination of hardware and software.

[0116] The above description, in conjunction with specific embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered within the scope of protection of the present invention.

Claims

1. A substation fire early warning method based on multispectral fusion and anti-interference AI, characterized in that, include: Acquire multispectral image data of substations, registered infrared thermal imaging reference images and low spatial resolution hyperspectral data, construct multispectral data cubes and reference image data, and perform radiometric and geometric preprocessing; High-pass components are extracted from the preprocessed multispectral image and the reference image respectively. A local dual-scale coefficient matrix is ​​constructed to adaptively fit the high-pass domain relationship. Combining spectral correction and total variation regularization, the anti-interference fusion image is output through iterative solution. The anti-interference fused image and the low spatial resolution hyperspectral data are input into a functional data analysis framework to estimate degradation parameters and apply nuclear norm constraints to reconstruct the hyperspectral feature cube. A dual-graph structure is constructed based on the hyperspectral feature cube, and a horseshoe prior is introduced to dynamically adjust the smoothing intensity. The abnormal heat source region is determined by combining low-rank constraints. Multiple fire-sensitive features are extracted within the abnormal heat source region to generate a fire-sensitive feature vector set. The fire-sensitive feature vector set is input into the AI ​​model embedded in the interference suppression module, and the fire probability value and fire development trend level are output. A dynamic correction factor is constructed based on real-time collected environmental and equipment operation data to correct the confidence level of the fire probability value and the fire development trend level. When the corrected output exceeds the preset warning threshold, a warning signal is output.

2. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 1, characterized in that, The iterative solution outputs an anti-interference fused image, including: The preprocessed multispectral image is upsampled to the resolution of the reference image, and the high-frequency difference components of the multispectral image and the reference image are extracted respectively. Using the neighborhood window of each pixel as a unit, the ratio of the local mean of the multispectral image to the local standard deviation of the reference image is calculated, and the product of the two is used as the local dual-scale coefficient. The two are combined pixel by pixel to form a local dual-scale coefficient matrix. A high-pass injection fidelity term is constructed using the local dual-scale coefficient matrix as weights to constrain spatial detail injection, a spectral correction term for radiation intensity ratio consistency is constructed to maintain spectral fidelity, and a total variational regularization term is introduced to maintain edge structure. The high-pass injection fidelity term, the spectral correction term, and the total variation regularization term are weighted and combined into a convex energy functional, which is then iteratively solved using the alternating direction multiplier method to output the anti-interference fused image.

3. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 2, characterized in that, The size of the neighborhood window is dynamically adjusted based on the local texture complexity, including: Calculate the gradient magnitude variance within the neighborhood of each pixel as a texture metric; Small windows are used in textured areas to maintain detail fitting accuracy, large windows are used in flat areas to suppress noise interference, and linear interpolation is used in transition areas to achieve a smooth transition in window size.

4. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 1, characterized in that, The hyperspectral feature cube is reconstructed in the following manner: The anti-interference fused image and the low spatial resolution hyperspectral data are input together into the functional data analysis framework; The spectral response function and spatial point spread function are modeled as a combination of smooth basis functions, and the degradation parameters are estimated through alternating iterations. The multispectral vectors of each pixel in the anti-interference fused image are reconstructed into continuous spectral curves, and a mapping relationship to the hyperspectral image is established based on the estimated degradation parameters. The low-rank constraint of the nuclear norm is introduced to compress the estimation space, and the hyperspectral feature cube is output through iterative optimization.

5. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 4, characterized in that, The estimation of the degradation parameters employs a superpixel-guided partitioning strategy, including: Superpixel segmentation is performed on the anti-interference fused image to divide it into multiple spectrally homogeneous sub-regions; Each sub-region independently estimates the spectral response function and spatial point spread function, while adjacent sub-regions are co-optimized through smoothness constraints to avoid abrupt parameter changes; The nuclear norm constraint is applied to the coefficient matrix of each sub-region, and the sub-regions are merged after being shrunk by independent singular value thresholding.

6. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 1, characterized in that, The fire-sensitive feature vector set is generated in the following way: Based on the hyperspectral feature cube, a spatial adjacency graph is constructed with each pixel as a vertex and spatial proximity as an edge, and a spectral feature graph is constructed with the spectral angular distance between the spectral vectors of each pixel as the edge weight, thus forming the dual-graph structure. Introducing a horseshoe prior distribution into the Bayesian hierarchical model, an adaptive shrinkage prior is applied to the difference vector of each edge in the graph, so that the smoothing intensity is dynamically adjusted according to the local signal differences; By combining low-rank tensor constraints, the hyperspectral feature cube is decomposed into low-rank background components and sparse anomaly components. The normal thermal distribution field is fitted by the low-rank background components, and the heat source regions that deviate from the background are located by the sparse anomaly components, which are then used as the anomalous heat source regions. The spectral gradient features, spatial texture features, and temporal thermal accumulation rate features of each pixel are extracted within the abnormal heat source region. The three types of features are then concatenated and combined to generate the fire-sensitive feature vector set.

7. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 6, characterized in that, The construction of the dual-graph structure employs a superpixel-guided reduction strategy, including: Superpixel segmentation is performed on the spatial dimension of the hyperspectral feature cube, with superpixel blocks replacing individual pixels as graph vertices; The edges of the spatial adjacency graph connect adjacent superpixel blocks, and the edge weights are determined by the spatial distance between the superpixel blocks; the edges of the spectral feature graph connect pairs of superpixel blocks whose spectral angular distance is less than a preset threshold, and the edge weights are determined by the reciprocal of the spectral angular distance. The introduction of the horseshoe prior and the application of the low-rank tensor constraint are both performed at the superpixel level, and the identified abnormal heat source region is output with the superpixel block as the smallest unit.

8. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 1, characterized in that, The fire probability value and fire development trend level are output in the following ways: The fire-sensitive feature vector set is input into a pre-trained deep learning model. The deep learning model uses a multi-scale convolutional neural network to extract the spatial local patterns of the feature vectors and uses a Transformer temporal attention mechanism to capture the long-range dependencies of the feature vectors within a continuous time window. The interference suppression module is embedded in the middle layer of the deep learning model. The interference suppression module generates an attention mask for the interference frequency band in a learnable manner and dynamically reduces the weight of the response channels in the feature map that are related to background thermal fluctuations, ambient light drift and sensor noise. After multi-scale feature fusion and fully connected layer mapping, the deep learning model simultaneously outputs the fire probability value at the current moment and the fire development trend level at multiple future time steps.

9. The substation fire early warning method based on multispectral fusion and anti-interference AI according to claim 1, characterized in that, The confidence correction and early warning signal output are achieved through the following methods: Real-time data collection of substation ambient temperature and humidity, wind speed and direction, and current equipment load rate is used to construct a multi-dimensional environment-operation status vector. The environment-operational state vector is input into a pre-established correction factor mapping network, and the output is the dynamic correction factor that matches the current operating condition. The dynamic correction factor is used to multiplicatively adjust the fire probability value and to offset the fire development trend level. When the corrected fire probability value exceeds the preset warning threshold and the corrected fire development trend level is not lower than the preset risk level, the warning triggering condition is determined to be met. The warning signal includes the spatial coordinates of the abnormal heat source, a fire type prediction label, and a priority level information for handling. It is then pushed to local audible and visual alarm devices, a centralized monitoring platform, and mobile terminals via different communication protocols.

10. A substation fire early warning system based on multispectral fusion and anti-interference AI, used to implement the substation fire early warning method based on multispectral fusion and anti-interference AI as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and preprocessing module is used to acquire multispectral image data of substations, registered infrared thermal imaging reference images and low spatial resolution hyperspectral data, construct multispectral data cubes and reference image data, and perform radiometric and geometric preprocessing. The anti-interference fusion image generation module is used to extract high-pass components from the preprocessed multispectral image and the reference image, construct a local dual-scale coefficient matrix to adaptively fit the high-pass domain relationship, and combine spectral correction and total variation regularization to iteratively solve and output the anti-interference fusion image. The hyperspectral feature reconstruction module is used to input the anti-interference fused image and the low spatial resolution hyperspectral data into a functional data analysis framework, estimate the degradation parameters and apply the nuclear norm constraint, and reconstruct the hyperspectral feature cube. The feature vector set generation module is used to construct a dual-graph structure based on the hyperspectral feature cube, and introduce horseshoe prior to dynamically adjust the smoothing intensity. It also uses low-rank constraints to determine abnormal heat source regions, extracts multiple fire-sensitive features within the abnormal heat source regions, and generates a fire-sensitive feature vector set. The fire prediction module is used to input the fire-sensitive feature vector set into the AI ​​model embedded in the interference suppression module, and output the fire probability value and the fire development trend level. The confidence correction and early warning output module is used to construct a dynamic correction factor based on real-time collected environmental and equipment operation data, and to correct the confidence of the fire probability value and the fire development trend level. When the corrected output exceeds the preset early warning threshold, an early warning signal is output.

Images