Electrical fire monitoring and detection method and system based on image processing

Abstract

The present invention discloses an electrical fire monitoring and detection method and system based on image processing, which belongs to the field of electrical fire monitoring technology. The method collects images of the electrical equipment monitoring area to obtain continuous video frames; extracts multiple candidate areas as spatial graph nodes of the graph neural network, and constructs time edges based on spatial position relationships; calculates the regional deformation rate and edge drift vector difference of the corresponding candidate areas in the continuous frames, inputs them into the region-edge coupling analysis model, and calculates the time edge confidence; during the propagation of the graph neural network, performs weighted processing or gate control on the time edge based on the confidence; finally, classification and identification are performed based on the node characteristics to determine whether there are electrical fire characteristics such as smoke or flames in the area; the method has the advantages of strong dynamic recognition ability, low false alarm rate, and adaptability to complex scenarios, and is suitable for intelligent fire monitoring in places such as distribution boxes, low-voltage switch cabinets, and cable joints.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical fire monitoring, and in particular to an electrical fire monitoring and detection method and system based on image processing. Background Art

[0002] Electrical fire monitoring and detection uses specialized equipment and systems to monitor and analyze in real time any abnormalities (such as overloads, short circuits, leakage, poor contact, and elevated temperatures) that may arise during the operation of electrical circuits and equipment. This provides early warnings before a fire occurs, preventing fires caused by electrical faults. This system is widely used in construction, power generation, and industrial settings, and is a key technical tool for fire prevention.

[0003] Existing electrical fire monitoring technologies based on graph neural networks typically construct spatiotemporal graphs by directly creating "temporal edges" between pixels or regions at the same location in images at different times to model temporal evolution relationships. However, in real smoke or flame scenes, these regions often exhibit dynamic characteristics such as drift and deformation. Simply assuming that their spatial position remains unchanged across frames can easily lead to distortion of temporal correlations. As a result, the model may mistakenly identify the natural flow of smoke as a continuous anomaly or misidentify discontinuous changes in light and shadow as flames, significantly reducing the system's recognition accuracy and reliability. Summary of the Invention

[0004] The purpose of the present invention is to provide an electrical fire monitoring and detection method and system based on image processing to solve the shortcomings of the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an electrical fire monitoring and detection method based on image processing, comprising:

[0006] Capture images of the electrical equipment monitoring area and obtain continuous video frames;

[0007] Preprocess each frame of the image, extract multiple candidate regions as spatial graph nodes of the graph neural network, and construct temporal edges based on the preliminary spatial position relationship;

[0008] For any pair of corresponding candidate regions in consecutive frames, their region deformation rate and edge drift vector difference are calculated respectively. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames, and the edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames.

[0009] Inputting the regional deformation rate and edge drift vector difference into the regional-edge coupling analysis model to calculate the confidence value of the corresponding time edge;

[0010] During the information propagation process of the graph neural network, the time edges are weighted or gated based on the confidence values ​​of the time edges;

[0011] Classification and identification are performed based on the node features processed by the graph neural network to determine whether the monitored area has smoke or flame characteristics of electrical fires.

[0012] Preferably, the step of collecting images of the electrical equipment monitoring area includes:

[0013] Install at least one image acquisition device in the electrical equipment monitoring area;

[0014] The image acquisition device is an industrial-grade high-definition camera, an infrared thermal imager or a multispectral imaging device;

[0015] The image acquisition device acquires continuous video frames at a frame rate of 10 to 30 frames per second;

[0016] The captured video frames are then transmitted to a local or remote processing platform.

[0017] Preferably, the image preprocessing step includes:

[0018] Perform grayscale conversion, noise filtering and histogram equalization on each frame of image;

[0019] Extract image edge information and use Canny or Sobel algorithm to enhance boundary features;

[0020] Extract multiple candidate regions based on superpixel segmentation algorithm or sliding window mechanism;

[0021] Each candidate region is regarded as a spatial graph node, and its position, edge, texture or thermal feature information is recorded.

[0022] Preferably, the step of calculating the regional deformation rate includes:

[0023] Extract the edge contours of the candidate regions in consecutive frames and perform equidistant sampling on the contours of the previous frame;

[0024] Find the closest matching point to each contour point in the previous frame in the next frame to form a set of perturbation vectors;

[0025] The average length of the disturbance vector is calculated as the contour deviation, and combined with the curvature weighting to form the contour deformation;

[0026] Calculate the relative rate of change of regional area;

[0027] The area change rate and the contour deformation degree are weighted and superimposed to obtain the regional deformation rate.

[0028] Preferably, the step of calculating the edge drift vector difference includes:

[0029] Perform gradient operation on the candidate area to extract the direction angle sequence of edge pixels;

[0030] Perform Fourier transform on the edge direction angle sequence to extract the main frequency feature component;

[0031] Compare the differences in the main frequency components in consecutive frames to obtain the frequency domain structure difference;

[0032] Use structural tensor analysis to obtain the edge dominant direction vector and calculate the angle difference between the two frames;

[0033] The frequency domain structure difference and the main direction angle difference are weighted and synthesized to form the edge drift vector difference.

[0034] Preferably, the step of constructing the region-edge coupling analysis model includes:

[0035] The regional deformation rate and edge drift vector difference are converted into a comprehensive feature vector, which is used as the input of the machine learning model. The machine learning model uses the confidence value label of the time edge predicted by each set of comprehensive feature vectors as the prediction target, and minimizes the sum of the prediction errors of the confidence value labels of all time edges as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The confidence value of the time edge is determined according to the model output results. The machine learning model is a polynomial regression model.

[0036] Preferably, the step of performing weighted processing or gate control on the time edge includes:

[0037] Feed the confidence value into the gating function: ; Among them, γ and δ are adjustment parameters, Represents time edge The gating factor, is the confidence value of the time edge; represents the hyperbolic tangent function;

[0038] The gating factor controls whether the temporal side information is allowed to propagate;

[0039] At the same time, calculate the temporal attention weight based on the node state features;

[0040] Multiply the gating factor by the attention weight to obtain the final time edge propagation weight;

[0041] The propagation weights are used during the propagation of graph neural networks to perform feature updates in the time dimension.

[0042] Preferably, the classification and identification step includes:

[0043] Input the final feature vector of each node into a multi-layer fully connected neural network;

[0044] Use Softmax to output category probability;

[0045] Set the category probability threshold to determine whether it is a valid fire signal;

[0046] For areas where multiple consecutive nodes are identified as fire categories, spatial clustering is performed;

[0047] If the cluster area meets the continuity or continuity conditions, it is determined to be an electrical fire event and an alarm is output.

[0048] The present invention also provides an electrical fire monitoring and detection system based on image processing, comprising:

[0049] Image acquisition module, which collects images of the electrical equipment monitoring area and obtains continuous video frames;

[0050] The graph construction module preprocesses each frame of the image, extracts multiple candidate regions as spatial graph nodes of the graph neural network, and constructs temporal edges based on the preliminary spatial position relationship;

[0051] The dynamic feature extraction module calculates the region deformation rate and edge drift vector difference for any pair of corresponding candidate regions in consecutive frames. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames, and the edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames.

[0052] A region-edge coupling analysis module, which inputs the region deformation rate and edge drift vector difference into a region-edge coupling analysis model and calculates the confidence value of the corresponding time edge;

[0053] A graph neural network propagation control module performs weighted processing or gate control on the time edges based on the confidence values ​​of the time edges during the information propagation process of the graph neural network;

[0054] The result output module classifies and identifies the node features processed by the graph neural network to determine whether the monitored area contains smoke or flame characteristics of an electrical fire.

[0055] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0056] 1. This invention introduces two custom parameters, the regional deformation rate and the edge drift vector difference, to reflect the dynamic state of candidate regions. It also constructs a region-edge coupling analysis model to calculate temporal edge confidence, thereby achieving dynamic weighted control of the propagation paths of temporal information in graph neural networks. This method effectively avoids erroneous temporal edge connections caused by non-rigid changes such as smoke drift and flame deformation, significantly improving the system's accuracy and robustness in identifying fire signs in complex scenarios.

[0057] 2. Compared to traditional image fire recognition methods that rely on fixed temporal connections or static feature judgment, this invention possesses stronger spatiotemporal modeling and dynamic perception capabilities. By integrating a gating mechanism with an attention propagation strategy, it not only improves the ability to capture fine-grained anomalies such as faint smoke and early flames, but also reduces the false alarm rate caused by factors such as background disturbances and lighting changes. This makes it suitable for fire monitoring in unattended electrical equipment environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0059] Figure 1 Flow chart of the method of the present invention.

[0060] Figure 2 It is a flow chart of the system modules of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0062] Example 1, please refer to Figure 1 As shown, the electrical fire monitoring and detection method based on image processing described in this embodiment includes:

[0063] Capture images of the electrical equipment monitoring area and obtain continuous video frames;

[0064] Preprocess each frame of the image, extract multiple candidate regions as spatial graph nodes of the graph neural network, and construct temporal edges based on the preliminary spatial position relationship;

[0065] For any pair of corresponding candidate regions in consecutive frames, their region deformation rate and edge drift vector difference are calculated respectively. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames, and the edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames.

[0066] Inputting the regional deformation rate and edge drift vector difference into the regional-edge coupling analysis model to calculate the confidence value of the corresponding time edge;

[0067] During the information propagation process of the graph neural network, the time edges are weighted or gated based on the confidence values ​​of the time edges;

[0068] Classification and identification are performed based on the node features processed by the graph neural network to determine whether the monitored area has smoke or flame characteristics of electrical fires.

[0069] In this embodiment of the present invention, images of electrical equipment within a target monitoring area are first captured for subsequent analysis and identification of electrical fire hazards. This monitoring area can include distribution boxes, low-voltage switchgear, cable connectors, electrical control cabinets, and other locations prone to electrical failures. To achieve real-time, continuous image data acquisition, at least one image acquisition device is deployed within the monitoring area. This device can be an industrial-grade high-definition camera, infrared thermal imager, or multispectral imaging device, and can be flexibly configured based on the lighting conditions and monitoring requirements of the monitoring environment.

[0070] The acquisition device continuously captures the monitored area at a set frame rate (e.g., 10-30 frames per second), generating an image sequence or video stream. The collected data is transmitted via the network to a local or remote computing platform, serving as input for fire detection algorithms. In some practical scenarios, optional low-light imaging technology, autofocus capabilities, or bandwidth-compression encoding technologies (such as H.264 / H.265) can be added to improve image quality and acquisition efficiency at night or in high-noise environments.

[0071] Each frame of the continuous video frame will serve as the input basis for subsequent image preprocessing, candidate region extraction, graph construction and graph neural network analysis, ensuring the system has continuity, timeliness and spatial coverage, and meeting the real-time image recognition requirements of electrical fire hazards.

[0072] After completing the image acquisition, the present invention performs preprocessing operations on each frame of the acquired image to enhance the image quality, extract key areas and construct the spatial graph structure required by the graph neural network.

[0073] First, perform conventional preprocessing on the original image, including image denoising (such as median filtering or Gaussian filtering), grayscale conversion, histogram equalization, and edge enhancement, to improve the accuracy and stability of subsequent region extraction. For thermal imaging or low-light scenarios, contrast enhancement or multi-channel fusion strategies can be used.

[0074] Then, perform candidate region extraction in each frame of the image, which can be based on one of the following methods:

[0075] Divide the image into multiple regions with similar texture or color features based on superpixel segmentation (such as the SLIC algorithm);

[0076] Or based on sliding window scanning and edge detection, identify target areas that may have abnormal features (high temperature, smoke, bright spots, etc.);

[0077] Alternatively, a binarization method based on thermal gradient or brightness threshold can be used to extract fire-related areas.

[0078] Each extracted candidate region will serve as a spatial graph node of the graph neural network, and its node features include color statistics, texture features, thermal features, region edge shape information, etc.

[0079] In order to track the evolution of candidate regions in the time dimension, the present invention further establishes preliminary time edge connections in adjacent frames based on the consistency rule of spatial position. Specifically, for a certain region node in frame t If the spatial position of the image in frame t+1 overlaps or the center point displacement is within the preset threshold, it is considered to have temporal continuity and the temporal edge is established. .

[0080] The spatial graph nodes and temporal edges together form a preliminary spatiotemporal graph structure, which is used for information dissemination and feature learning in subsequent graph neural networks. This structure can express the dynamic changes in a region across time and space, providing a data foundation for determining the evolution of flames or smoke in electrical fires.

[0081] For any pair of corresponding candidate regions in consecutive frames, their region deformation rate and edge drift vector difference are calculated respectively. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames. The edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames. Specifically, it includes:

[0082] The region deformation rate is used to measure the area change of the candidate region and the overall morphological change of its contour between two consecutive frames. It specifically includes the following steps:

[0083] First, edge contours are extracted for the corresponding candidate areas in the current frame and the next frame respectively, and conventional edge detection methods such as the Canny algorithm can be used.

[0084] The pixel areas of the two regions are calculated, and the ratio of their difference to the maximum value is taken to obtain the area change rate, which is used to reflect whether the region has significant deformation behaviors such as scaling and diffusion.

[0085] The contour in the previous frame is sampled into a series of feature points at equal intervals, and the closest corresponding points are found in the contour of the next frame. This forms a set of contour perturbation vectors, which are used to measure the spatial offset between the two contours.

[0086] The changes of contour points on the edge curvature are considered and the curvature is used as a weighting factor of the perturbation vector to enhance the perception of structural changes such as corners.

[0087] The regional deformation rate is calculated by weightedly superimposing the area change rate and the average value of the contour disturbance. This metric exhibits a high response when the regional contour undergoes deformations such as displacement, expansion, drift, or distortion, and can accurately reflect the non-rigid expansion behavior of smoke or flames.

[0088] The edge drift vector difference is used to measure the overall motion direction and local structure changes of the edge of the candidate region in consecutive frames. It is mainly achieved through frequency domain analysis and main direction offset calculation. Specifically, it includes the following steps:

[0089] Perform gradient calculation on the internal edge of the region to obtain the direction information of each edge pixel, including the horizontal and vertical gradient components, and thus calculate the gradient direction vector sequence of each region , and maps it into a sequence of angles. Indicates the grayscale change of the pixel in the horizontal direction (x-axis), that is, the horizontal gradient; Indicates the grayscale change of the pixel in the vertical direction (y-axis), that is, the longitudinal gradient; Represents the gradient direction angle of the k-th edge pixel point, which is used to indicate the edge direction of the point;

[0090] Perform Fourier transform on the direction angle sequence to obtain the main frequency characteristic component of the edge 、 , reflecting the overall change trend of edge texture or structure.

[0091] Perform difference analysis on the edge spectrum of the same area in two consecutive frames to obtain the degree of change of the edge structure in the frequency domain , the expression is: ; Where M is the number of selected main frequencies.

[0092] Use local window method or structure tensor analysis to extract the dominant direction vector of the region edge in each frame 、 , and calculate the angle difference , the expression is: .

[0093] The edge drift vector difference is calculated by weighted summing the spectrum difference and the change in the main direction angle. This indicator can effectively identify dynamic processes such as flame edge jitter and smoke migration, and distinguish between natural disturbances and persistent anomalies.

[0094] The regional deformation rate and edge drift vector difference are input into the regional-edge coupling analysis model to calculate the confidence value of the corresponding time edge, specifically including:

[0095] The regional deformation rate and edge drift vector difference are converted into a comprehensive feature vector, which is used as the input of the machine learning model. The machine learning model uses the confidence value label of the time edge predicted by each set of comprehensive feature vectors as the prediction target, and minimizes the sum of the prediction errors of the confidence value labels of all time edges as the training target. The machine learning model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The confidence value of the time edge is determined according to the model output results. The machine learning model is a polynomial regression model.

[0096] The confidence value represents the "trusted transfer degree" of the current time edge: the confidence value of the time edge is compared with the preset threshold. If the confidence value of the time edge is greater than or equal to the preset threshold, it means that the regional state remains stable between consecutive frames and is more suitable for establishing temporal dependency relationships. If the confidence value of the time edge is less than the preset threshold, it indicates that there may be misconnection caused by deformation drift, and information propagation should be suppressed.

[0097] In graph neural networks, nodes represent regions, and edges represent spatial or temporal dependencies. For information propagation in the temporal direction, a gated weighted structure is used to update features. The control method is as follows: Where, Representation node Feature vector at time t; Representation node Feature vector at time t; 、 represents the time edge weight matrix and the spatial edge weight matrix; Represents the set of adjacent nodes of node i in the spatial graph; Represents time edge The dynamic confidence value of ; σ represents the activation function, such as ReLU or Tanh. Represents the features of node i after being processed by the gating mechanism at time t+1.

[0098] Spatial information propagation is based on the graph state at time t, and is aggregated based on the time features of the nodes after gating and adjustment, and updated to generate status.

[0099] This invention further supports incorporating the confidence of temporal edges into the feedback mechanism for graph structure learning. During training, if a temporal edge repeatedly triggers low-confidence propagation but has a high misjudgment rate in the corresponding region, the system can record this structure as a "high-risk misconnection pattern" and, through an adaptive edge update strategy, adjust the subsequent temporal edge establishment method (for example, enabling graph matching instead of position matching).

[0100] During the information propagation process of the graph neural network, the time edges are weighted or gated based on their confidence values ​​to adjust the propagation strength of the temporal information. Specifically:

[0101] Based on the confidence value, a gate function is introduced to control the propagation of time edges, which is defined as follows: ; Among them, γ and δ are learnable adjustment parameters, which ensure that when the confidence is low, the gating factor approaches zero and automatically suppresses propagation. Represents time edge The gating factor ranges from (-1, 1) and is actually used to adjust the strength of the side information propagation at that time. is the confidence value of the time edge. Represents the hyperbolic tangent function, which compresses the output to (-1, 1) and is used to generate a continuous and differentiable gating response curve.

[0102] Using the historical state of the node itself, the temporal attention weight is constructed in the following way, and the expression is: ;in: , ; 、 is a learnable linear transformation matrix; Represents the relative weight of the current time edge in all adjacent time paths. T represents the matrix transpose.

[0103] The gating factor is integrated with the attention mechanism to obtain the final time edge propagation weight ; This weighted value is used to control the proportion of information transmitted in the time direction. High-confidence time edges will be strengthened and propagated, while low-confidence time edges will be suppressed or even blocked.

[0104] Based on the above propagation weights, the feature update in the time direction is completed: ; Where: σ is the activation function; is the final feature of node i at time t+1.

[0105] After completing the spatial feature aggregation and temporal information propagation of the graph neural network, the present invention uses the high-dimensional feature vector generated by each node in the network to perform classification and recognition operations to determine whether there are electrical fire features in the current monitoring image.

[0106] Specifically, the feature vector corresponding to each graph node (i.e., candidate region) output by the graph neural network is connected to one or more fully connected layers and combined with a nonlinear activation function (such as ReLU or Softmax) for multi-category classification. The classification categories include at least "normal region," "smoke region," "flame region," and optionally "abnormally high temperature region" or "unknown region."

[0107] The classification model can be trained and inferred in the following ways:

[0108] Using a pre-collected and annotated fire image dataset, we train the graph neural network end-to-end, enabling the model to distinguish different fire types. We use cross-entropy loss, focal loss, or class-balanced loss during training to enhance the ability to identify edge and small fires.

[0109] A dynamic threshold is set based on the category probability output by the model. Only when the confidence of a certain category exceeds the preset threshold (such as flame > 0.8 or smoke > 0.7) will the system consider that there are actual signs of fire in the area, thus avoiding false alarms caused by slight noise.

[0110] For low-intensity fire signals distributed across multiple nodes in the graph, the present invention further provides a regional clustering mechanism (such as based on adjacency merging or spatial connectivity analysis) to improve the consistency and stability of overall fire identification.

[0111] Once the model determines that one or more regional nodes are classified as "smoke" or "flame" and meet the continuity or continuity judgment rules, the system will output an alarm signal, triggering subsequent warning actions, recording images, or linkage control instructions.

[0112] Through the above-mentioned method, the present invention effectively realizes the dynamic intelligent recognition of smoke and flames in electrical fire images, has high early warning capability and false alarm suppression effect, and is suitable for the intelligent fire monitoring needs of unmanned places, power distribution systems and industrial equipment.

[0113] Example 2, please refer to Figure 2 As shown, the electrical fire monitoring and detection system based on image processing described in this embodiment includes:

[0114] Image acquisition module, which collects images of the electrical equipment monitoring area and obtains continuous video frames;

[0115] The graph construction module preprocesses each frame of the image, extracts multiple candidate regions as spatial graph nodes of the graph neural network, and constructs temporal edges based on the preliminary spatial position relationship;

[0116] The dynamic feature extraction module calculates the region deformation rate and edge drift vector difference for any pair of corresponding candidate regions in consecutive frames. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames, and the edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames.

[0117] A region-edge coupling analysis module, which inputs the region deformation rate and edge drift vector difference into a region-edge coupling analysis model and calculates the confidence value of the corresponding time edge;

[0118] A graph neural network propagation control module performs weighted processing or gate control on the time edges based on the confidence values ​​of the time edges during the information propagation process of the graph neural network;

[0119] The result output module classifies and identifies the node features processed by the graph neural network to determine whether the monitored area contains smoke or flame characteristics of an electrical fire.

[0120] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. An electrical fire monitoring and detection method based on image processing, characterized in that: include: Capture images of the electrical equipment monitoring area and obtain continuous video frames; Preprocess each frame of the image, extract multiple candidate regions as spatial graph nodes of the graph neural network, and construct temporal edges based on the preliminary spatial position relationship; For any pair of corresponding candidate regions in consecutive frames, their region deformation rate and edge drift vector difference are calculated respectively. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames, and the edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames. Inputting the regional deformation rate and edge drift vector difference into a regional-edge coupling analysis model to calculate the confidence value of the corresponding time edge; The steps of constructing the region-edge coupling analysis model include: converting the regional deformation rate and the edge drift vector difference into a comprehensive feature vector, using the comprehensive feature vector as the input of a machine learning model, the machine learning model using the confidence value label of the time edge predicted by each group of comprehensive feature vectors as a prediction target, minimizing the sum of the prediction errors of the confidence value labels of all time edges as a training target, training the machine learning model until the sum of the prediction errors reaches convergence, stopping the model training, and determining the confidence value of the time edge based on the model output result, wherein the machine learning model is a polynomial regression model; During the information propagation process of the graph neural network, the time edges are weighted or gated based on the confidence values ​​of the time edges; Classification and identification are performed based on the node features processed by the graph neural network to determine whether the monitored area has smoke or flame characteristics of electrical fires.

2. The electrical fire monitoring and detection method based on image processing according to claim 1, characterized in that: The step of collecting images of the electrical equipment monitoring area includes: Install at least one image acquisition device in the electrical equipment monitoring area; The image acquisition device is an industrial-grade high-definition camera, an infrared thermal imager or a multispectral imaging device; The image acquisition device acquires continuous video frames at a frame rate of 10 to 30 frames per second; The captured video frames are then transmitted to a local or remote processing platform.

3. The electrical fire monitoring and detection method based on image processing according to claim 2, characterized in that: The preprocessing of each frame of image comprises: Perform grayscale conversion, noise filtering and histogram equalization on each frame of image; Extract image edge information and use Canny or Sobel algorithm to enhance boundary features; Extract multiple candidate regions based on superpixel segmentation algorithm or sliding window mechanism; Each candidate region is regarded as a spatial graph node, and its position, edge, texture or thermal feature information is recorded.

4. The electrical fire monitoring and detection method based on image processing according to claim 1, characterized in that: The step of calculating the regional deformation rate includes: Extract the edge contours of the candidate regions in consecutive frames and perform equidistant sampling on the contours of the previous frame; Find the closest matching point to each contour point in the previous frame in the next frame to form a set of perturbation vectors; The average length of the disturbance vector is calculated as the contour deviation, and combined with the curvature weighting to form the contour deformation; Calculate the relative rate of change of regional area; The area change rate and the contour deformation degree are weighted and superimposed to obtain the regional deformation rate.

5. The electrical fire monitoring and detection method based on image processing according to claim 4, characterized in that: The step of calculating the edge drift vector difference includes: Perform gradient operation on the candidate area to extract the direction angle sequence of edge pixels; Perform Fourier transform on the edge direction angle sequence to extract the main frequency feature component; Compare the differences in the main frequency components in consecutive frames to obtain the frequency domain structure difference; Use structural tensor analysis to obtain the edge dominant direction vector and calculate the angle difference between the two frames; The frequency domain structure difference and the main direction angle difference are weighted and synthesized to form the edge drift vector difference.

6. The electrical fire monitoring and detection method based on image processing according to claim 1, characterized in that: The step of performing weighted processing or gate control on the time edge includes: Feed the confidence value into the gating function: ; Among them, γ and δ are adjustment parameters, Represents time edge The gating factor, is the confidence value of the time edge; represents the hyperbolic tangent function; The gating factor controls whether the temporal side information is allowed to propagate; At the same time, calculate the temporal attention weight based on the node state features; Multiply the gating factor by the attention weight to obtain the final time edge propagation weight; The propagation weights are used during the propagation of graph neural networks to perform feature updates in the time dimension.

7. The electrical fire monitoring and detection method based on image processing according to claim 1, characterized in that: The classification and identification based on the node features processed by the graph neural network includes: Input the final feature vector of each node into a multi-layer fully connected neural network; Use Softmax to output category probability; Set the category probability threshold to determine whether it is a valid fire signal; For areas where multiple consecutive nodes are identified as fire categories, spatial clustering is performed; If the cluster area meets the continuity or continuity conditions, it is determined to be an electrical fire event and an alarm is output.

8. An electrical fire monitoring and detection system based on image processing, for implementing the electrical fire monitoring and detection method based on image processing according to any one of claims 1 to 7, characterized in that: include: Image acquisition module, which collects images of the electrical equipment monitoring area and obtains continuous video frames; The graph construction module preprocesses each frame of the image, extracts multiple candidate regions as spatial graph nodes of the graph neural network, and constructs temporal edges based on the preliminary spatial position relationship; The dynamic feature extraction module calculates the region deformation rate and edge drift vector difference for any pair of corresponding candidate regions in consecutive frames. The region deformation rate is used to measure the area change and contour difference of the candidate region in adjacent frames, and the edge drift vector difference is used to measure the difference in the motion direction of the candidate region edge in consecutive frames. A region-edge coupling analysis module, which inputs the region deformation rate and edge drift vector difference into a region-edge coupling analysis model and calculates the confidence value of the corresponding time edge; A graph neural network propagation control module performs weighted processing or gate control on the time edges based on the confidence values ​​of the time edges during the information propagation process of the graph neural network; The result output module classifies and identifies the node features processed by the graph neural network to determine whether the monitored area contains smoke or flame characteristics of an electrical fire.

Images