Fire early warning method and system based on AI video image recognition

Abstract

The application provides a fire early warning method and system based on AI video image recognition, relates to the technical field of fire alarm, obtains a visible light image and a thermal infrared image, performs registration and fusion to obtain a multi-modal feature, performs time sequence modeling to construct a space-time feature map, and extracts a fire mode evolution track; a diffusion rate and a shape change distribution are calculated to determine a main diffusion direction and a shape stability index; a combustion behavior constraint rule is extracted from a physical constraint knowledge graph; the space-time feature map is optimized based on the constraint rule to obtain a corrected feature map and is classified to obtain a fire source area and a smoke diffusion area, and a heat source center coordinate and a threatened area are analyzed; a risk propagation graph is constructed, the shape stability index is used as an initial activation strength, iterative diffusion is performed to obtain a risk situation assessment result, and an alarm signal is generated in combination with the heat source center coordinate and the threatened area.

Description

Technical Field

[0001] This invention relates to the field of fire alarm technology, and in particular to a method and system for early fire alarm based on AI video image recognition. Background Technology

[0002] Current early fire alarm technologies primarily rely on single-modal visual sensors, such as visible light cameras or thermal infrared imagers, which monitor independently. Visible light cameras identify flames or smoke by analyzing color, texture, and motion features in video frames, while thermal infrared imagers determine the location of fire sources by detecting areas of abnormal temperature. They typically use fixed thresholds or simple machine learning models to classify image features and can only identify fire events that have developed into obvious open flames or high-temperature areas.

[0003] In existing technologies, single sensors have limited ability to extract features of early small fires or smoldering stages. For example, visible light cameras are prone to false alarms or missed alarms when obscured by smoke or changes in ambient light. Thermal infrared imagers have a large response delay to slow combustion at low temperatures. They lack spatiotemporal correlation modeling of the continuous process of fire development, cannot accurately capture key dynamic information such as the direction of fire spread and changes in combustion morphology, and have difficulty distinguishing between real fires and similar heat source interference, resulting in a high false alarm rate. In addition, because the trend of fire spread cannot be predicted, alarm information often lags behind the actual risk spread, limiting the timeliness of early warning. Summary of the Invention

[0004] This invention provides a fire early warning method and system based on AI video image recognition, which can at least solve some of the problems existing in the prior art.

[0005] A first aspect of this invention provides a fire early warning method based on AI video image recognition, comprising:

[0006] Visible light and thermal infrared images within the monitoring area are acquired, registered, and fused to obtain multimodal features. Temporal modeling is performed on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map. Graph structure encoding is performed on the spatiotemporal feature map to extract the fire mode evolution trajectory.

[0007] Based on the fire mode evolution trajectory, the diffusion rate distribution and morphological change distribution are calculated. The diffusion rate distribution is decomposed into a main diffusion direction. The morphological change distribution is evaluated for stability to obtain a morphological stability index. The fire development stage identifier is determined based on the main diffusion direction and the morphological stability index, and combustion behavior constraint rules are extracted from a preset physical constraint knowledge graph.

[0008] Based on the combustion behavior constraint rules, the spatiotemporal feature map is constrained and optimized to obtain a modified feature map, and pixel-level classification is performed to obtain the fire source region and the smoke diffusion region. Spatial analysis is performed on the fire source region and the smoke diffusion region to obtain the coordinates of the heat source center and the threatened region.

[0009] A risk propagation map is constructed based on the threatened area. The morphological stability index is used as the initial activation intensity of the nodes for iterative diffusion calculation until convergence. The number of activated nodes and the distribution of activation intensity after convergence are statistically analyzed to obtain the risk situation assessment result. An alarm signal is generated and output by combining the coordinates of the heat source center and the threatened area.

[0010] In one alternative implementation,

[0011] Visible light and thermal infrared images within the monitored area are acquired, registered, and fused to obtain multimodal features. Temporal modeling is performed on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map. Graph structure encoding is then performed on the spatiotemporal feature map to extract the fire mode evolution trajectory, including:

[0012] Visible light images of the monitored area are acquired by a visible light camera, and thermal infrared images of the monitored area are acquired by a thermal infrared sensor. Texture and color features of the visible light images are extracted to construct a visible light feature map, and temperature gradient features of the thermal infrared images are extracted to construct a thermal infrared feature map. Spatial transformation correction is performed on the visible light feature map and the thermal infrared feature map, and the feature response differences at corresponding positions are calculated to construct a difference mask. Multimodal features are constructed based on the difference mask, the corrected visible light feature map, and the thermal infrared feature map.

[0013] Multimodal features from multiple consecutive frames are arranged in chronological order to construct a feature sequence, and differential operations are performed on adjacent frames to obtain temporal variation features. Temporal convolution is performed on the temporal variation features to extract short-term motion patterns. Long-range dependency modeling is performed on the feature sequence to extract long-term evolution trends, and the sequence is concatenated with the short-term motion patterns to obtain the spatiotemporal feature map.

[0014] The spatiotemporal feature map is divided into spatial grids to obtain multiple spatial nodes. The feature similarity between different spatial nodes is calculated and the node connection relationship is constructed. The feature of each node is updated by multi-level propagation to obtain the node embedding representation. The key nodes corresponding to the node embedding representation are extracted and the change trajectory of the embedding representation in the time dimension is tracked to obtain the fire mode evolution trajectory.

[0015] In one alternative implementation,

[0016] Based on the fire mode evolution trajectory, the diffusion rate distribution and morphological change distribution are calculated. The diffusion rate distribution is decomposed directionally to obtain the main diffusion direction. The morphological change distribution is evaluated for stability to obtain morphological stability indices, including:

[0017] The key node positions at different times are extracted from the fire mode evolution trajectory and the displacement vector between key nodes at adjacent times is calculated. The instantaneous diffusion rate is calculated based on the displacement vector and the diffusion rate distribution is constructed according to the spatial location. The fire source area contour at different times in the fire mode evolution trajectory is extracted and the shape difference degree of the contour at adjacent times is calculated. The shape difference degree is statistically aggregated according to the spatial location to obtain the morphological change distribution.

[0018] Anisotropic filtering decomposition is performed on the diffusion rate distribution to obtain diffusion components in different directions. The direction corresponding to the energy peak in the diffusion component is extracted as the main diffusion direction and the directional concentration is calculated. Divergence calculation is performed on the diffusion rate distribution to obtain the diffusion source intensity distribution.

[0019] Wavelet multi-resolution decomposition is performed on the morphological change distribution to obtain change components at different scales, and an autocorrelation matrix is ​​constructed using time autocorrelation coefficients. Eigenvalue decomposition is performed on the autocorrelation matrix to extract the dominant stable mode and calculate the energy proportion to obtain a global fluctuation index. Persistent homology analysis is performed on the morphological change distribution to obtain the birth and death cycle of morphological features and calculate the persistence entropy. An initial stability value is calculated based on the global fluctuation index and the persistence entropy. A directional dynamic factor is constructed based on the directional concentration and the diffusion source intensity distribution to correct the initial stability value and obtain a morphological stability index.

[0020] In one alternative implementation,

[0021] Based on the main diffusion direction and the morphological stability index, fire development stage identifiers are determined, and combustion behavior constraint rules are extracted from a preset physical constraint knowledge graph, including:

[0022] The directional angle parameters are obtained by spherical coordinate projection on the main diffusion direction, and the directional similarity distribution is obtained by calculating the angular deviation from each typical diffusion mode in the preset combustion mode library. Based on the morphological stability index and the directional similarity distribution, a state feature tensor is constructed and multilinear singular value decomposition is performed to determine the core tensor. Based on the core tensor, the state evolution velocity field is calculated and numerical integration is performed along the time dimension to obtain the state evolution trajectory. The curvature and torsion of the state evolution trajectory in phase space are calculated to obtain the dynamic feature vector. The distance between the dynamic feature vector and the preset stage boundary hyperplane is measured to determine the fire development stage identifier.

[0023] Based on the fire development stage identifier, the corresponding stage node is retrieved from the preset physical constraint knowledge graph, and the first-order neighbor node is extracted to obtain a constraint entity set. The constraint entity set is then filtered and separated by type to obtain thermal radiation constraint entities and convective heat transfer constraint entities. The radiation flux attenuation relationship and the viewpoint factor calculation method in the thermal radiation constraint entities are extracted to construct thermal radiation constraint rules. The convective heat transfer coefficient function relationship and the boundary layer temperature gradient calculation method in the convective heat transfer constraint entities are extracted to construct convective heat transfer constraint rules. These rules are then logically combined with the thermal radiation constraint rules to obtain combustion behavior constraint rules.

[0024] In one alternative implementation,

[0025] Based on the combustion behavior constraint rules, the spatiotemporal feature map is constrained and optimized to obtain a modified feature map, and pixel-level classification is performed to obtain the fire source region and the smoke diffusion region. Spatial analysis of the fire source region and the smoke diffusion region is performed to obtain the coordinates of the heat source center and the threatened area, including:

[0026] The combustion behavior constraint rules are converted into spatial propagation constraint conditions, and the spatiotemporal feature map is subjected to pixel-by-pixel constraint verification to obtain the physical consistency deviation distribution. The physical consistency deviation distribution is subjected to adaptive threshold segmentation to identify abnormal feature locations and construct a constraint mask. Based on the constraint mask, the abnormal feature locations in the spatiotemporal feature map are subjected to neighborhood feature weighted smoothing and iterative correction combined with the physical consistency deviation to obtain a corrected feature map. The corrected feature map is subjected to multi-scale segmentation to extract candidate regions of different scales and combined with temperature gradient and smoke concentration gradient for dual screening to obtain the fire source region and smoke diffusion region.

[0027] The centroid of the fire source area is calculated to obtain the coordinates of the heat source center. Based on the coordinates of the heat source center and the smoke diffusion area, a distance attenuation weight field is constructed and the propagation intensity value at each spatial location is calculated. The propagation intensity value is hierarchically clustered to extract spatial regions with different intensity levels. The hazard level distribution is calculated by combining the temperature distribution in the modified feature map. Based on the hazard level distribution, connected regions exceeding the preset critical hazard level are identified to obtain the threatened area.

[0028] In one alternative implementation,

[0029] Adaptive threshold segmentation is performed on the physical consistency deviation distribution to identify anomalous feature locations and construct a constraint mask. Based on the constraint mask, neighborhood feature weighted smoothing is applied to the anomalous feature locations in the spatiotemporal feature map, and iterative correction is performed in conjunction with the physical consistency deviation to obtain a corrected feature map, including:

[0030] The cumulative distribution curve of the physical consistency deviation distribution is calculated and the slope change position is extracted as an adaptive threshold for binarization segmentation to obtain an initial anomaly mask. The initial anomaly mask is subjected to connected component analysis to obtain compactness and filter anomaly feature positions. The spatial coordinate encoding of the anomaly feature positions is used to construct a constraint mask.

[0031] Based on the constraint mask, the abnormal feature positions in the spatiotemporal feature map are located, and a multi-scale neighborhood window is constructed to extract the normal feature values ​​in the neighborhood corresponding to the abnormal feature positions. The spatial decay weight and feature similarity weight of the normal feature values ​​are calculated and fused to obtain a comprehensive weight. Based on the comprehensive weight, the normal feature values ​​in the neighborhood are weighted and fused to obtain smooth feature values. The smooth feature values ​​are weighted and combined with the physical consistency deviation values ​​of the corresponding positions to obtain a correction value and update the abnormal feature positions. Based on the updated spatiotemporal feature map, the physical consistency deviation distribution is recalculated and the constraint mask is updated. This process is repeated until the number of abnormal feature positions in the constraint mask converges to obtain the corrected feature map.

[0032] In one alternative implementation,

[0033] A risk propagation map is constructed based on the threatened area. The morphological stability index is used as the initial activation intensity of the nodes for iterative diffusion calculation until convergence. The number of activated nodes and the distribution of activation intensity after convergence are statistically analyzed to obtain the risk situation assessment result. An alarm signal is generated and output based on the coordinates of the heat source center and the threatened area, including:

[0034] Based on the threatened area, a set of risk nodes is obtained by spatial grid division, and the spatial location and danger level of each risk node are extracted. The spatial distance between different risk nodes is calculated, and the radial distance of each risk node relative to the heat source is calculated by combining the coordinates of the heat source center. Based on the radial distance and the spatial distance, the propagation weight between nodes is determined, and the node connection relationship is established to construct a risk propagation map. The morphological stability index is allocated according to the danger level of each risk node to obtain the initial activation intensity of the node.

[0035] The initial activation intensity of each node in the risk propagation diagram is calculated and updated according to the node connection relationship and propagation weight. This process is repeated until the change in activation intensity of each node converges to obtain a converged activation intensity distribution. The number of nodes exceeding a preset activation threshold in the converged activation intensity distribution is counted to obtain the number of activated nodes. The gradient of the converged activation intensity distribution is calculated to obtain the activation intensity gradient distribution. Based on the number of activated nodes and the activation intensity gradient distribution, the risk propagation speed and risk coverage area are calculated and fused to obtain the risk situation assessment result. Based on the risk situation assessment result, the alarm level is determined, and alarm information is constructed by combining the coordinates of the heat source center and the spatial range of the threatened area. The alarm information is encoded into an alarm signal and output.

[0036] A second aspect of this invention provides a fire early warning system based on AI video image recognition, comprising:

[0037] The evolution modeling unit is used to acquire visible light images and thermal infrared images within the monitoring area and perform registration and fusion to obtain multimodal features. It performs temporal modeling on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map, and performs graph structure encoding on the spatiotemporal feature map to extract the fire mode evolution trajectory.

[0038] The evolution analysis unit is used to calculate the diffusion rate distribution and morphological change distribution based on the fire mode evolution trajectory, perform directional decomposition on the diffusion rate distribution to obtain the main diffusion direction, perform stability assessment on the morphological change distribution to obtain the morphological stability index, determine the fire development stage identifier based on the main diffusion direction and the morphological stability index, and extract combustion behavior constraint rules from a preset physical constraint knowledge graph.

[0039] The classification and localization unit is used to perform constraint optimization on the spatiotemporal feature map based on the combustion behavior constraint rules to obtain a modified feature map and perform pixel-level classification to obtain the fire source area and the smoke diffusion area. It also performs spatial analysis on the fire source area and the smoke diffusion area to obtain the coordinates of the heat source center and the threatened area.

[0040] The risk assessment unit is used to construct a risk propagation map based on the threatened area, use the morphological stability index as the initial activation intensity of the node for iterative diffusion calculation until convergence, and obtain the risk situation assessment result by statistically analyzing the number of activated nodes and the distribution of activation intensity after convergence. It also generates and outputs an alarm signal by combining the coordinates of the heat source center and the threatened area.

[0041] A third aspect of the present invention provides an electronic device, comprising:

[0042] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0043] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0044] In this invention, based on multimodal feature fusion and spatiotemporal feature map construction, the complementary utilization of visible light and thermal infrared information is achieved, significantly improving the sensitivity of early fire detection in complex environments. The extraction of fire pattern evolution trajectories accurately captures the dynamic changes in flames and smoke, enabling the quantification of the spatiotemporal evolution of combustion behavior and providing a high-fidelity data foundation for subsequent trend prediction. Through dual constraints of the main diffusion direction and morphological stability index, combined with combustion behavior rules in the physical constraint knowledge graph, feature drift caused by environmental interference is effectively suppressed. After pixel-level classification of the corrected feature map, the identification accuracy of the fire source area and smoke diffusion area is significantly improved, and the positioning errors of the heat source center coordinates and threatened areas are significantly reduced, providing accurate spatial basis for subsequent emergency response. The risk propagation map, combined with the morphological stability index, performs iterative diffusion calculations to simulate the potential paths and intensity distribution of fire spread. The number and intensity distribution of activated nodes after convergence directly reflect the spatial heterogeneity of the risk situation, enabling the generation of alarm signals to have dynamic adaptive capabilities and avoiding false alarms or missed alarms caused by fixed thresholds. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the early fire alarm method based on AI video image recognition, as described in an embodiment of the present invention.

[0046] Figure 2 This is a flowchart illustrating the fire mode evolution analysis of the early fire alarm method based on AI video image recognition, as described in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0049] Figure 1 This is a flowchart illustrating the early fire alarm method based on AI video image recognition, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0050] Visible light and thermal infrared images within the monitoring area are acquired, registered, and fused to obtain multimodal features. Temporal modeling is performed on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map. Graph structure encoding is performed on the spatiotemporal feature map to extract the fire mode evolution trajectory.

[0051] Based on the fire mode evolution trajectory, the diffusion rate distribution and morphological change distribution are calculated. The diffusion rate distribution is decomposed into a main diffusion direction. The morphological change distribution is evaluated for stability to obtain a morphological stability index. The fire development stage identifier is determined based on the main diffusion direction and the morphological stability index, and combustion behavior constraint rules are extracted from a preset physical constraint knowledge graph.

[0052] Based on the combustion behavior constraint rules, the spatiotemporal feature map is constrained and optimized to obtain a modified feature map, and pixel-level classification is performed to obtain the fire source region and the smoke diffusion region. Spatial analysis is performed on the fire source region and the smoke diffusion region to obtain the coordinates of the heat source center and the threatened region.

[0053] A risk propagation map is constructed based on the threatened area. The morphological stability index is used as the initial activation intensity of the nodes for iterative diffusion calculation until convergence. The number of activated nodes and the distribution of activation intensity after convergence are statistically analyzed to obtain the risk situation assessment result. An alarm signal is generated and output by combining the coordinates of the heat source center and the threatened area.

[0054] In one alternative implementation,

[0055] Visible light and thermal infrared images within the monitored area are acquired, registered, and fused to obtain multimodal features. Temporal modeling is performed on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map. Graph structure encoding is then performed on the spatiotemporal feature map to extract the fire mode evolution trajectory, including:

[0056] Visible light images of the monitored area are acquired by a visible light camera, and thermal infrared images of the monitored area are acquired by a thermal infrared sensor. Texture and color features of the visible light images are extracted to construct a visible light feature map, and temperature gradient features of the thermal infrared images are extracted to construct a thermal infrared feature map. Spatial transformation correction is performed on the visible light feature map and the thermal infrared feature map, and the feature response differences at corresponding positions are calculated to construct a difference mask. Multimodal features are constructed based on the difference mask, the corrected visible light feature map, and the thermal infrared feature map.

[0057] Multimodal features from multiple consecutive frames are arranged in chronological order to construct a feature sequence, and differential operations are performed on adjacent frames to obtain temporal variation features. Temporal convolution is performed on the temporal variation features to extract short-term motion patterns. Long-range dependency modeling is performed on the feature sequence to extract long-term evolution trends, and the sequence is concatenated with the short-term motion patterns to obtain the spatiotemporal feature map.

[0058] The spatiotemporal feature map is divided into spatial grids to obtain multiple spatial nodes. The feature similarity between different spatial nodes is calculated and the node connection relationship is constructed. The feature of each node is updated by multi-level propagation to obtain the node embedding representation. The key nodes corresponding to the node embedding representation are extracted and the change trajectory of the embedding representation in the time dimension is tracked to obtain the fire mode evolution trajectory.

[0059] Visible light cameras continuously capture images of the monitored area, obtaining visual information such as flame color and smoke texture. Thermal infrared sensors acquire thermal infrared images of the same area, carrying information about the target surface temperature distribution and reflecting changes in thermal radiation invisible to the human eye. For the visible light images, texture and color features are extracted: texture features are calculated using local binary mode or multi-scale filter banks to describe the irregular texture structure of the flame edge; color features are extracted in the HSV color space, focusing on the hue distribution and saturation changes of the orange-red hues, thus constructing a visible light feature map. For the thermal infrared images, the temperature gradient features of each pixel are calculated, i.e., the rate of change of temperature values ​​is calculated in both the horizontal and vertical directions. The gradient magnitude and direction are used to describe the heat source boundary and heat diffusion trend, constructing a thermal infrared feature map.

[0060] Due to differences in installation location, focal length, and imaging plane between the visible light camera and the thermal infrared sensor, directly superimposing the two types of feature maps would introduce spatial misalignment errors. To address this, spatial transformation corrections are performed on both the visible light and thermal infrared feature maps. Specifically, a homography matrix is ​​used to map the two images to a unified coordinate system, ensuring that the same physical location corresponds to the same pixel coordinate in both feature maps. After spatial alignment, the feature response differences between the visible light and thermal infrared feature maps at corresponding locations are calculated pixel-by-pixel. Regions with difference amplitudes exceeding a preset threshold are marked as valid sensing areas, and a difference mask is constructed. The purpose of the difference mask is to highlight the complementary areas of the two sensors. For example, thermal infrared images can detect the thermal radiation of an obscured fire source, while visible light images can capture the visual texture of smoke; areas with significant differences are often where fire characteristics are most prominent. Based on the difference mask, the corrected visible light and thermal infrared feature maps are weighted and fused, with higher fusion weights assigned to regions with larger difference mask values, thus obtaining multimodal features that take into account both visual and thermal radiation information.

[0061] Multimodal features acquired from multiple consecutive frames are arranged chronologically to construct a feature sequence. A difference operation is then performed on the multimodal features of adjacent frames, i.e., the first feature sequence is calculated. Frame and the The difference in feature values ​​at corresponding positions across frames yields temporal variation features, reflecting dynamic changes in the flame region, such as expansion, contraction, or displacement, within adjacent timeframes. A one-dimensional convolution operation is applied to these temporal variation features along the time axis, with the convolution kernel covering a short time window (e.g., 3 to 5 frames) to extract short-term motion patterns, including high-frequency dynamic information such as rapid flame flickering and localized smoke surges. Long-range dependency modeling is then performed on the complete feature sequence using an attention-based temporal coding structure to calculate the association weights between any two frames in the sequence. This effectively captures the evolutionary relationships between distant frames, thereby extracting long-term evolutionary trends, such as the slow changes in flame expansion and gradual increase in smoke concentration. The short-term motion patterns and long-term evolutionary trends are concatenated along the feature dimension to form a spatiotemporal feature map that simultaneously contains both local dynamic information and global evolutionary information.

[0062] The spatiotemporal feature map is spatially gridded, uniformly dividing the entire monitoring area into several rectangular sub-regions. Each sub-region corresponds to a spatial node, and the node features are represented by the mean pooling result of the spatiotemporal feature map within that sub-region. Feature similarity between different spatial nodes is calculated, using cosine similarity to measure the directional consistency between node feature vectors: Let the node... The feature vector is ,node The feature vector is The similarity between the two nodes is... for .when When the set threshold is exceeded, at the node With nodes Establish a connection edge between them, with the edge weight set to a value of 1. This allows for the construction of node connections, forming a graph structure that describes the relationships between spatial regions.

[0063] In a graph structure, the features of each node are updated through multi-layer message propagation. In each layer of propagation, a node aggregates the feature information of its neighboring nodes and combines it with its own features through a nonlinear transformation to obtain an updated node representation. After multi-layer propagation, the representation of each node not only includes the spatiotemporal features of its own region but also incorporates contextual information from the surrounding region, ultimately yielding a node embedding representation. Node embedding representations reflect the comprehensive characteristic state of each spatial region during fire development; nodes with higher embedding values ​​typically correspond to the flame core or areas of high smoke concentration. Key nodes with significantly higher embedding representations than the background region are selected from all nodes; these nodes spatially correspond to active fire-related areas. The embedding representation change trajectory of key nodes is tracked across multiple consecutive frames, recording the embedding value and spatial location of each key node at each moment, forming a fire pattern evolution trajectory describing the joint spatial and temporal evolution of fire features. This trajectory includes both the displacement path of the flame core and the increase / decrease in fire feature intensity over time, providing a structured input basis for subsequent diffusion rate analysis and morphological change assessment.

[0064] Figure 2 This is a flowchart illustrating the fire mode evolution analysis of the early fire alarm method based on AI video image recognition, as described in an embodiment of the present invention.

[0065] In one alternative implementation,

[0066] Based on the fire mode evolution trajectory, the diffusion rate distribution and morphological change distribution are calculated. The diffusion rate distribution is decomposed directionally to obtain the main diffusion direction. The morphological change distribution is evaluated for stability to obtain morphological stability indices, including:

[0067] The key node positions at different times are extracted from the fire mode evolution trajectory and the displacement vector between key nodes at adjacent times is calculated. The instantaneous diffusion rate is calculated based on the displacement vector and the diffusion rate distribution is constructed according to the spatial location. The fire source area contour at different times in the fire mode evolution trajectory is extracted and the shape difference degree of the contour at adjacent times is calculated. The shape difference degree is statistically aggregated according to the spatial location to obtain the morphological change distribution.

[0068] Anisotropic filtering decomposition is performed on the diffusion rate distribution to obtain diffusion components in different directions. The direction corresponding to the energy peak in the diffusion component is extracted as the main diffusion direction and the directional concentration is calculated. Divergence calculation is performed on the diffusion rate distribution to obtain the diffusion source intensity distribution.

[0069] Wavelet multi-resolution decomposition is performed on the morphological change distribution to obtain change components at different scales, and an autocorrelation matrix is ​​constructed using time autocorrelation coefficients. Eigenvalue decomposition is performed on the autocorrelation matrix to extract the dominant stable mode and calculate the energy proportion to obtain a global fluctuation index. Persistent homology analysis is performed on the morphological change distribution to obtain the birth and death cycle of morphological features and calculate the persistence entropy. An initial stability value is calculated based on the global fluctuation index and the persistence entropy. A directional dynamic factor is constructed based on the directional concentration and the diffusion source intensity distribution to correct the initial stability value and obtain a morphological stability index.

[0070] Extracting key node locations at different times from the fire mode evolution trajectory requires reading the coordinates of the graph nodes corresponding to each time step in the trajectory graph and recording the position of each node in the spatial coordinate system. For adjacent times... and Calculate the displacement vector of the same key node. ,in Indicates the node at time [time]. Spatial coordinates. Based on displacement vector and time interval. The ratio of the two values ​​yields the instantaneous diffusion rate. The instantaneous diffusion rates at various spatial locations were aggregated into a two-dimensional diffusion rate distribution map. Each pixel location corresponds to a velocity scalar value. Simultaneously, fire source region contours at different times are extracted from the fire pattern evolution trajectory. By calculating the shape difference between contours at adjacent times, such as the Hausdorff distance based on the contour or the matching error of the shape context descriptor, the shape difference at each spatial location is statistically aggregated to construct a morphological change distribution map. This reflects the dramatic dynamic changes in the boundary of the fire source in different regions.

[0071] For diffusion rate distribution Anisotropic filtering decomposition is performed using a directional Gaussian filter kernel across multiple preset directional angles, covering 12 directions from 0° to 165° with a step size of 15°, to obtain diffusion component response maps for different directions. The total energy is calculated for the response map in each direction, and the direction corresponding to the energy peak is identified as the main diffusion direction. Directional Concentration The degree of concentration in the direction of fire spread is reflected by calculating the complement of the normalized distribution entropy of energy in each direction. A lower entropy indicates a higher concentration, suggesting a clear directional advantage in fire spread. (This refers to the distribution of spread rates.) Discrete divergence calculations are performed to obtain the intensity distribution of the diffusion source. The positive divergence region corresponds to the source of the fire's outward expansion, while the negative divergence region corresponds to the area where the fire converges, thus spatially locating the dynamic driving center of the fire source.

[0072] Distribution of morphological changes Wavelet multi-resolution decomposition is performed, and two-dimensional discrete wavelet transform is used to extract low-frequency approximate components and high-frequency detail components at multiple decomposition levels to obtain morphological change components at different scales. ,in This represents the decomposition hierarchical index. For morphological changes in a time series, the temporal autocorrelation coefficients of morphological changes at each spatial location are calculated, and an autocorrelation matrix is ​​constructed. Matrix elements Indicates time lag as The autocorrelation value at time step [time]. For the autocorrelation matrix... Perform eigenvalue decomposition to extract the dominant stable mode corresponding to the largest eigenvalue, and calculate the energy proportion of this mode. , which is the ratio of the dominant eigenvalue to the sum of all eigenvalues. A larger value indicates the presence of a dominant, stable cyclical pattern in the morphological changes, from which a global volatility index is derived. , The smaller the value, the more regular the morphological changes and the higher the stability; The larger the value, the more chaotic and disordered the morphological changes are, and the more unstable and expanding the fire is.

[0073] The distribution of morphological changes was analyzed through continuous homology analysis. Treating it as a filtering function, simplex complexes are constructed under different thresholds to track the generation time of topological features. With the moment of extinction , Indexing topological features to form a persistent set of pairs Persistence The survival duration reflects the topological characteristics; features with longer durations correspond to truly stable structures in the morphology, while features with shorter durations correspond to noise or transient perturbations. Based on this, the persistence entropy is calculated. It is defined as calculating the information entropy of a normalized persistence probability distribution: ,in It is the sum of all persistences. The larger the value, the more uniform the distribution of the birth and death of topological features, the lack of a dominant structure in morphological changes, and the lower the stability. The smaller the value, the more likely the morphology has a few significant and persistent topological structures, indicating higher stability.

[0074] Based on global volatility indicators and persistent entropy The initial stability value is calculated through weighted fusion. ,in The normalized persistence entropy, and For the weighting coefficients, satisfying The value can be calibrated based on historical fire data. The range of values ​​is A higher value indicates a more stable morphological evolution and that the fire is in a relatively stable combustion stage.

[0075] Directional dynamics factor By directional concentration and diffusion source intensity distribution Jointly constructed. Specifically, extracting the intensity distribution of the diffusion source. The mean value of the region of neutral divergence is used as the mean value of the diffusion source intensity. After normalization, it is compared with the directional concentration. Perform product fusion to obtain ,in The Sigmoid normalization function is used to map the diffusion source intensity to... Interval. The larger the value, the more it indicates that the fire has a clear direction of spread and a strong driving force from the source, and the fire is in a rapid directional spread stage. At this time, the initial stability value should be adjusted downward. The smaller the value, the more dispersed the fire spread and the weaker the driving force at the source, indicating that the fire is in a relatively stable or decaying stage. Therefore, the initial stability value should be adjusted upwards. Final morphological stability index. The calculation method is as follows ,in The strength coefficient, used to control the influence of the directional dynamics factor on the initial stability value, is also calibrated using historical data. As a core input for judging the subsequent fire development stage and the initial activation intensity of nodes in the risk propagation map, its accurate calculation is of decisive significance for accurately assessing the fire situation.

[0076] In one alternative implementation,

[0077] Based on the main diffusion direction and the morphological stability index, fire development stage identifiers are determined, and combustion behavior constraint rules are extracted from a preset physical constraint knowledge graph, including:

[0078] The directional angle parameters are obtained by spherical coordinate projection on the main diffusion direction, and the directional similarity distribution is obtained by calculating the angular deviation from each typical diffusion mode in the preset combustion mode library. Based on the morphological stability index and the directional similarity distribution, a state feature tensor is constructed and multilinear singular value decomposition is performed to determine the core tensor. Based on the core tensor, the state evolution velocity field is calculated and numerical integration is performed along the time dimension to obtain the state evolution trajectory. The curvature and torsion of the state evolution trajectory in phase space are calculated to obtain the dynamic feature vector. The distance between the dynamic feature vector and the preset stage boundary hyperplane is measured to determine the fire development stage identifier.

[0079] Based on the fire development stage identifier, the corresponding stage node is retrieved from the preset physical constraint knowledge graph, and the first-order neighbor node is extracted to obtain a constraint entity set. The constraint entity set is then filtered and separated by type to obtain thermal radiation constraint entities and convective heat transfer constraint entities. The radiation flux attenuation relationship and the viewpoint factor calculation method in the thermal radiation constraint entities are extracted to construct thermal radiation constraint rules. The convective heat transfer coefficient function relationship and the boundary layer temperature gradient calculation method in the convective heat transfer constraint entities are extracted to construct convective heat transfer constraint rules. These rules are then logically combined with the thermal radiation constraint rules to obtain combustion behavior constraint rules.

[0080] By projecting spherical coordinates onto the main diffusion direction, the direction vector in three-dimensional space is mapped into two components: polar angle and azimuth angle, denoted as [reference needed]. and Forming direction and angle parameter pairs The preset combustion mode library stores reference direction vectors for several typical diffusion modes, with each typical mode corresponding to a standard direction angle parameter pair. The currently calculated direction angle parameters are compared one by one with the standard angle parameters of each typical diffusion mode, and the angular deviation between the two is calculated using the spherical angle distance formula. ,in This is an index of typical diffusion patterns. After normalization and inversion, the current main diffusion direction and the first... Directional similarity between typical patterns ,all Together they form a directional similarity distribution vector .

[0081] morphological stability index Similarity distribution vector with direction The common organization is the state feature tensor The moduli of the tensor correspond to the time dimension, spatial dimension, and feature type dimension, respectively. For the state feature tensor... Perform multilinear singular value decomposition, i.e., Tucker decomposition, to decompose it into the core tensor. The product of the core tensor and each modulo factor matrix. The coupling relationships between state features across dimensions are captured, and the absolute value of each element reflects the contribution of the corresponding multilinear singular values. During decomposition, components corresponding to the main singular values ​​are retained based on a truncation threshold, while higher-order components with smaller contributions are discarded. This achieves a low-rank approximation of the state features, reducing the complexity of subsequent calculations.

[0082] Based on core tensor The difference between core tensor slices at adjacent time points is calculated along the time dimension to obtain the state evolution velocity field. The physical meaning is the distribution of the rate of change in the state characteristic space per unit time. For a velocity field... Numerical integration is performed along the time dimension, and the integration step size is adaptively controlled using the fourth-order Runge-Kutta method. Starting from the initial moment, the process is gradually advanced to obtain the state evolution trajectory. State evolution trajectory It describes the path that a fire state takes in a feature space over time, and the geometric shape of the path directly reflects the dynamic characteristics of the fire development process.

[0083] In phase space, the trajectory of state evolution Perform differential geometric analysis to calculate the curvature of the trajectory curve point by point. With torsion Curvature The degree of local curvature of the trajectory is measured, characterizing the drastic change in fire condition; torsion. Measuring the degree to which the trajectory deviates from its plane characterizes the three-dimensional dynamic complexity of the fire's development. The mean curvature is then used. Curvature variance Mean torsion and torsion variance Concatenate into dynamic feature vectors This vector comprehensively characterizes the dynamic characteristics of the fire evolution trajectory from a geometric perspective.

[0084] The pre-defined stage-boundary hyperplanes divide the feature space into regions corresponding to different fire development stages. Each hyperplane is determined by fitting data from a large number of historical fire cases during the training phase, corresponding to the boundaries between the initial combustion stage, growth stage, complete combustion stage, and decay stage, respectively. Dynamic feature vectors are then calculated. Signed distance to the hyperplanes of each stage boundary ,in For hyperplane indices, the symbol indicates which side of the hyperplane the vector is located on. Based on the sign and absolute value of each distance value, the region where the current fire state is located is determined, and the corresponding fire development stage is marked. The assigned value is output, and this identifier is used to retrieve the corresponding node from the physical constraint knowledge graph later.

[0085] The physical constraint knowledge graph is stored in the form of a graph database. Nodes in the graph represent physical entities, constraint rules, or calculation methods involved in the combustion process, and edges represent the relationships between nodes. This is based on fire development stage identification. Locate the corresponding stage node in the graph, and traverse along the associated edges of the graph to extract the first-order neighbor nodes of that node, thus obtaining the set of constrained entities. .gather The entities in the data encompass all physical constraints relevant to this stage of development, including various types such as thermal radiation, convective heat transfer, and mass transfer.

[0086] For the set of constrained entities Filtering and separation are performed based on entity type labels, with entities labeled "thermal radiation" grouped into the thermal radiation-constrained entity subset. Entities with the type label "convective heat transfer" are grouped into the convection heat transfer constraint entity subset. Thermal radiation-constrained entity subset The system stores the radiative flux attenuation relationship and the method for calculating the viewing angle factor. The radiative flux attenuation relationship describes the decrease in thermal radiation intensity as the propagation distance increases, and is usually given as a power function of the radiation distance. The viewing angle factor calculation method describes the geometric efficiency of energy transfer between radiating surfaces with different geometries, and is calculated numerically based on the shape parameters of the fire source region. Combining these two types of information, a thermal radiation constraint rule is constructed. This is used for subsequent constraint optimization of thermal radiation-related features in the spatiotemporal feature map.

[0087] Convection heat transfer constrained entity subset The system stores the functional relationship of the convective heat transfer coefficient and the method for calculating the boundary layer temperature gradient. The functional relationship of the convective heat transfer coefficient provides an empirical formula for how the heat transfer coefficient varies with flow velocity, temperature difference, and geometric dimensions. The relevant parameters are estimated based on the thermal infrared image characteristics of the current fire scene and then substituted into the calculation. The boundary layer temperature gradient calculation method describes the spatial distribution characteristics of the temperature field in the near-wall region around the fire source, and is solved by fitting the temperature distribution data from the thermal infrared image. Combining these two types of information, a convective heat transfer constraint rule is constructed. .

[0088] thermal radiation confinement rules With convective heat transfer constraint rules Performing logical conjunction requires that the modified spatiotemporal feature map simultaneously satisfy the physical constraints imposed by both types of rules, resulting in complete combustion behavior constraint rules. The logical conjunction is implemented by adding the penalty terms corresponding to the two types of constraint rules to form a joint constraint loss function. When the spatiotemporal feature map is corrected in the subsequent constraint optimization process, the output of the feature map is required to minimize the joint constraint loss function, thereby ensuring that the corrected feature map physically conforms to both the laws of thermal radiation propagation and the laws of convective heat transfer, thus improving the physical interpretability and accuracy of fire identification results.

[0089] In one alternative implementation,

[0090] Based on the combustion behavior constraint rules, the spatiotemporal feature map is constrained and optimized to obtain a modified feature map, and pixel-level classification is performed to obtain the fire source region and the smoke diffusion region. Spatial analysis of the fire source region and the smoke diffusion region is performed to obtain the coordinates of the heat source center and the threatened area, including:

[0091] The combustion behavior constraint rules are converted into spatial propagation constraint conditions, and the spatiotemporal feature map is subjected to pixel-by-pixel constraint verification to obtain the physical consistency deviation distribution. The physical consistency deviation distribution is subjected to adaptive threshold segmentation to identify abnormal feature locations and construct a constraint mask. Based on the constraint mask, the abnormal feature locations in the spatiotemporal feature map are subjected to neighborhood feature weighted smoothing and iterative correction combined with the physical consistency deviation to obtain a corrected feature map. The corrected feature map is subjected to multi-scale segmentation to extract candidate regions of different scales and combined with temperature gradient and smoke concentration gradient for dual screening to obtain the fire source region and smoke diffusion region.

[0092] The centroid of the fire source area is calculated to obtain the coordinates of the heat source center. Based on the coordinates of the heat source center and the smoke diffusion area, a distance attenuation weight field is constructed and the propagation intensity value at each spatial location is calculated. The propagation intensity value is hierarchically clustered to extract spatial regions with different intensity levels. The hazard level distribution is calculated by combining the temperature distribution in the modified feature map. Based on the hazard level distribution, connected regions exceeding the preset critical hazard level are identified to obtain the threatened area.

[0093] Combustion behavior constraints are transformed into spatial propagation constraints, mapping rules describing thermal radiation propagation and convective heat transfer in the physical knowledge graph into pixel-level spatial constraint expressions. Specifically, thermal radiation constraints are transformed into constraints that the temperature gradient between adjacent pixels does not exceed the physical upper limit, and convective heat transfer constraints are transformed into spatial constraints that smoke concentration monotonically decreases in the wind direction. For each pixel location in the spatiotemporal feature map, its feature value is checked pixel by pixel to see if it satisfies the above spatial propagation constraints. If there is a deviation between the feature value of a pixel and the constraint conditions, the deviation is recorded, ultimately obtaining a physical consistency deviation distribution covering the entire image plane. Locations with large deviation values ​​indicate that the features at that pixel are inconsistent with physical laws, possibly due to noise pollution, sensor malfunction, or feature extraction errors.

[0094] An adaptive thresholding method is used to segment the distribution of physical consistency deviations. Instead of a globally fixed threshold, the threshold is dynamically determined based on the statistical characteristics of the deviation values ​​within each local region. Specifically, the image is divided into several non-overlapping local windows. Within each window, the mean and standard deviation of the deviation values ​​are calculated. The mean plus a certain multiple of the standard deviation is used as the local threshold for that window. Pixels with deviation values ​​exceeding the local threshold are marked as anomalous feature locations. The marking results of all anomalous feature locations are aggregated into a binary image, which is then filled with isolated small holes using morphological closing operations to obtain a constraint mask. In the constraint mask, regions with a value of 1 correspond to anomalous locations that need correction, while regions with a value of 0 correspond to normal locations with good physical consistency. This mask will guide the scope of feature smoothing in subsequent correction steps.

[0095] Based on constraint masks, neighborhood feature weighted smoothing is applied to anomalous feature locations in the spatiotemporal feature map. For each anomalous pixel with a mask value of 1, a neighborhood window of fixed radius is selected centered on it. Within this window, all normal neighboring pixels with a mask value of 0 are counted. The reciprocal of the spatial distance between each normal neighboring pixel and the central anomalous pixel is used as the weight to calculate the weighted average of the feature values ​​of the normal neighboring pixels. This average value replaces the original feature value of the anomalous pixel, thereby achieving the propagation of information from physically consistent regions to anomalous regions. After weighted smoothing, the corrected feature values ​​are combined with the corresponding physical consistency deviation values ​​for iterative correction. If the corrected feature values ​​still deviate from the constraint conditions, the neighborhood weighted smoothing process is repeated for the remaining deviation locations until the deviation values ​​at all locations are below the convergence threshold or the maximum number of iterations is reached. Finally, the corrected feature map is output. The iterative correction mechanism ensures that physical consistency deviations are fully eliminated after multiple rounds of smoothing, while limiting the maximum number of iterations avoids feature blurring caused by over-smoothing.

[0096] Multi-scale segmentation is performed on the corrected feature map. An image pyramid approach is used to perform region segmentation at multiple resolution levels. The coarse-scale level captures candidate regions with large-scale connectivity, while the fine-scale level captures local, fine boundaries. The segmentation results at each scale are upsampled and aligned to the original resolution before being fused to obtain a candidate region set. The candidate region set undergoes dual screening: temperature gradient screening requires that the temperature gradient distribution within the candidate region matches the physical characteristics of the fire source region, i.e., the temperature at the center of the region is significantly higher than the edge, and the gradient direction decreases from the inside out; smoke concentration gradient screening requires that the smoke concentration gradient distribution within the candidate region matches the physical characteristics of the smoke diffusion region, i.e., the concentration gradually decreases along the diffusion direction. Candidate regions that simultaneously meet the temperature gradient screening condition are classified as fire source regions, and candidate regions that simultaneously meet the smoke concentration gradient screening condition are classified as smoke diffusion regions. A certain degree of spatial overlap is allowed between the two types of regions to cover the transition zone at the boundary between flame and smoke.

[0097] Centroids of the heat source region are calculated, and the temperature feature values ​​of the corresponding pixels in the corrected feature map are used as weights to calculate the weighted centroid coordinates, thus obtaining the coordinates of the heat source center. ,in and These are the x and y coordinates of the heat source center in the image plane, respectively. Compared to the geometric centroid, the weighted centroid better reflects the true location of heat concentration and is more robust when the heat source shape is irregular or there are multiple local high-temperature zones.

[0098] Based on the coordinates of the heat source center Construct a distance attenuation weight field with the smoke diffusion area for each spatial location in the image. Calculate the coordinates to the center of the heat source Euclidean distance With an exponential decay function Define the distance decay weight, where The attenuation coefficient controls the rate at which the weight decreases with increasing distance. Based on this, the weight field is modulated by considering the smoke concentration distribution within the smoke diffusion area: locations with higher smoke concentrations are multiplied by an enhancement factor greater than 1, while locations with zero smoke concentration are not modulated. This ensures that the weight field simultaneously reflects both the distance from the heat source and the smoke concentration. The propagation intensity value at each spatial location... Determined by the product of distance attenuation weight and smoke concentration modulation factor, a higher propagation intensity value indicates a greater likelihood that the location is affected by fire propagation.

[0099] For propagation intensity value Hierarchical clustering is performed, using a natural breakpoint method based on intensity value distribution to divide the entire image plane into several intensity level intervals, each interval corresponding to an intensity level label, thus obtaining a discretized intensity level map. Combining the temperature distribution in the modified feature map, a temperature-weighted hazard score is calculated for the spatial location within each intensity level interval. The intensity level and temperature-weighted hazard score are then fused into a comprehensive hazard level, resulting in a hazard level distribution covering the entire map. In the hazard level distribution, a higher level value indicates that the location simultaneously has higher propagation intensity and higher local temperature, belonging to a high-risk area.

[0100] Based on the identification of connected regions exceeding a preset critical hazard level according to the hazard level distribution, the hazard level distribution map is first binarized. Pixels with a hazard level exceeding the preset critical hazard level threshold are marked as 1, and all other pixels are marked as 0. Then, connected component analysis is performed on the binarized result to extract all connected components. Connected components with excessively small areas are considered noise interference and are removed. Connected components with areas exceeding a minimum area threshold are retained, and the spatial range covered by these connected components is determined as the threatened area. The range of the threatened area and the hazard level distribution together provide the spatial basis for the subsequent construction of the risk propagation map, ensuring that alarm signals can accurately locate high-risk spatial locations.

[0101] In one alternative implementation,

[0102] Adaptive threshold segmentation is performed on the physical consistency deviation distribution to identify anomalous feature locations and construct a constraint mask. Based on the constraint mask, neighborhood feature weighted smoothing is applied to the anomalous feature locations in the spatiotemporal feature map, and iterative correction is performed in conjunction with the physical consistency deviation to obtain a corrected feature map, including:

[0103] The cumulative distribution curve of the physical consistency deviation distribution is calculated and the slope change position is extracted as an adaptive threshold for binarization segmentation to obtain an initial anomaly mask. The initial anomaly mask is subjected to connected component analysis to obtain compactness and filter anomaly feature positions. The spatial coordinate encoding of the anomaly feature positions is used to construct a constraint mask.

[0104] Based on the constraint mask, the abnormal feature positions in the spatiotemporal feature map are located, and a multi-scale neighborhood window is constructed to extract the normal feature values ​​in the neighborhood corresponding to the abnormal feature positions. The spatial decay weight and feature similarity weight of the normal feature values ​​are calculated and fused to obtain a comprehensive weight. Based on the comprehensive weight, the normal feature values ​​in the neighborhood are weighted and fused to obtain smooth feature values. The smooth feature values ​​are weighted and combined with the physical consistency deviation values ​​of the corresponding positions to obtain a correction value and update the abnormal feature positions. Based on the updated spatiotemporal feature map, the physical consistency deviation distribution is recalculated and the constraint mask is updated. This process is repeated until the number of abnormal feature positions in the constraint mask converges to obtain the corrected feature map.

[0105] After obtaining the physical consistency deviation distribution, adaptive threshold segmentation is needed to identify the feature locations of deviation anomalies. Specifically, the deviation values ​​of all pixel locations in the physical consistency deviation distribution are sorted to construct a cumulative distribution curve, with the deviation value on the horizontal axis and the cumulative probability on the vertical axis. The slope of the cumulative distribution curve is calculated point by point, reflecting the density change of the deviation values ​​within that interval. In normal feature regions, the deviation values ​​are small and concentrated, resulting in a larger slope of the cumulative distribution curve; while in abnormal feature regions, the deviation values ​​are large and sparsely distributed, leading to a significant abrupt decrease in the slope. By detecting the abrupt change locations in the slope sequence, the deviation value corresponding to that location is used as the adaptive threshold. The physical consistency deviation distribution is binarized and segmented; deviation values ​​exceeding... The location was marked as abnormal, below The position is marked as normal, thus obtaining the initial abnormal mask.

[0106] After obtaining the initial anomaly mask, connected component analysis is performed to group spatially connected anomaly pixels into the same connected component, and the compactness of each connected component is calculated. Compactness is defined as the ratio of the area of ​​a connected region to the area of ​​its smallest bounding rectangle, and its value ranges from [value missing]. A value closer to 1 indicates a more regular and compact shape of the connected region. Since real fire anomalies typically manifest as areas with spatial clustering and regular shapes, isolated noise points or extremely discrete connected regions often lack practical physical meaning. Therefore, a lower limit threshold for compactness is set. For compactness lower than Connected components are filtered out, retaining the anomalous feature locations corresponding to the connected components that meet the conditions. For the filtered anomalous feature locations, their spatial coordinates are extracted and encoded to construct a constraint mask. The constraint mask records the spatial index information of each anomalous feature location, which is used for precise localization in subsequent neighborhood feature smoothing operations.

[0107] After locating anomalous features in the spatiotemporal feature map using constraint masks, a multi-scale neighborhood window is constructed around each anomalous location. The multi-scale neighborhood window contains multiple rectangular windows of different sizes, for example... , , This process is repeated to cover neighborhood information across different spatial ranges. Within each neighborhood window at each scale, the feature values ​​corresponding to the constrained mask-marked normal feature locations are extracted, forming a set of normal feature values ​​for smoothing repair. For each normal feature location within the neighborhood, the Euclidean distance between it and the current abnormal location is calculated. Based on this, the spatial decay weight is calculated. Normal features that are closer in distance contribute more to smoothing and repair, while those that are farther apart contribute less. The spatial decay weight is calculated using a Gaussian kernel function, i.e. ,in This is the spatial attenuation bandwidth parameter, which controls the rate at which the weights decay with distance.

[0108] In addition to spatial decay weights, feature similarity weights also need to be calculated. This measure assesses the similarity between normal feature values ​​within a neighborhood and feature values ​​surrounding anomalies in the feature space. Feature similarity weights are obtained by calculating the cosine similarity between the normal feature value vector and the mean vector of the anomaly's neighborhood; higher similarity results in higher weights, thus prioritizing the use of normal features similar to the current anomaly location in the feature space for smoothing. Spatial decay weights are then applied. With feature similarity weight Element-by-element multiplication and fusion are performed to obtain the comprehensive weight. The comprehensive weights are then normalized so that the sum of the comprehensive weights of all normal feature positions participating in the weighting within the neighborhood is 1. Based on the normalized comprehensive weights, a weighted linear fusion is performed on all normal feature values ​​within the neighborhood to obtain the smoothed feature value of the current abnormal position. After calculating smoothed feature values ​​at different scales under a multi-scale neighborhood window, the final multi-scale smoothed feature value is obtained by weighted averaging of each scale. The weight of each scale can be adaptively determined according to the proportion of effective normal feature points in the neighborhood window.

[0109] Obtain smooth eigenvalues Then, the physical consistency deviation value between it and the corresponding location is calculated. By performing a weighted combination, the correction value for this anomaly location is obtained. The calculation method is as follows ,in The fusion weight coefficients for smoothing feature values ​​have a range of values. This is used to balance the contribution ratio between the neighborhood smoothing result and the physical consistency deviation information. When the physical consistency deviation value is large, it indicates that the location feature deviates significantly from the physical constraints, and the value can be appropriately reduced. The value of increases the guiding role of physical consistency deviation in the correction; when the physical consistency deviation value is small, it should be appropriately increased. It relies more on neighborhood smoothing results. The correction value... Write back the corresponding abnormal feature locations in the spatiotemporal feature map to complete one round of feature updates.

[0110] After completing one round of correction and update of anomalous feature locations, the physical consistency deviation distribution is recalculated based on the updated spatiotemporal feature map, and the constraint mask is regenerated according to the aforementioned adaptive threshold segmentation and connected component filtering process. If the number of anomalous feature locations in the newly generated constraint mask has changed significantly compared to the previous iteration, the next round of neighborhood feature weighted smoothing and iterative correction is performed; if the number of anomalous feature locations tends to stabilize, that is, the change in the number of anomalous feature locations in the constraint mask between two adjacent iterations is less than the preset convergence threshold... If the iteration converges, the iteration process is stopped. The converged spatiotemporal feature map is the corrected feature map, in which the original abnormal feature locations have been reasonably replaced through multiple rounds of neighborhood weighted smoothing and physical consistency-guided correction. The overall feature distribution is more consistent with the combustion behavior law described by the physical constraint knowledge graph, providing more reliable feature input for subsequent pixel-level classification and identification of fire source areas and smoke diffusion areas. By combining adaptive threshold segmentation with iterative correction, it is possible to effectively suppress abnormal feature interference introduced by sensor noise, illumination changes, or multimodal registration errors, while preserving the spatial structural integrity of real fire features, thus improving the accuracy and robustness of the final fire identification and alarm.

[0111] In one alternative implementation,

[0112] A risk propagation map is constructed based on the threatened area. The morphological stability index is used as the initial activation intensity of the nodes for iterative diffusion calculation until convergence. The number of activated nodes and the distribution of activation intensity after convergence are statistically analyzed to obtain the risk situation assessment result. An alarm signal is generated and output based on the coordinates of the heat source center and the threatened area, including:

[0113] Based on the threatened area, a set of risk nodes is obtained by spatial grid division, and the spatial location and danger level of each risk node are extracted. The spatial distance between different risk nodes is calculated, and the radial distance of each risk node relative to the heat source is calculated by combining the coordinates of the heat source center. Based on the radial distance and the spatial distance, the propagation weight between nodes is determined, and the node connection relationship is established to construct a risk propagation map. The morphological stability index is allocated according to the danger level of each risk node to obtain the initial activation intensity of the node.

[0114] The initial activation intensity of each node in the risk propagation diagram is calculated and updated according to the node connection relationship and propagation weight. This process is repeated until the change in activation intensity of each node converges to obtain a converged activation intensity distribution. The number of nodes exceeding a preset activation threshold in the converged activation intensity distribution is counted to obtain the number of activated nodes. The gradient of the converged activation intensity distribution is calculated to obtain the activation intensity gradient distribution. Based on the number of activated nodes and the activation intensity gradient distribution, the risk propagation speed and risk coverage area are calculated and fused to obtain the risk situation assessment result. Based on the risk situation assessment result, the alarm level is determined, and alarm information is constructed by combining the coordinates of the heat source center and the spatial range of the threatened area. The alarm information is encoded into an alarm signal and output.

[0115] A risk propagation map is constructed based on the threatened area. The threatened area is spatially gridded, dividing it into several spatial units using a uniform or adaptive grid. Each spatial unit corresponds to a risk node, resulting in a set of risk nodes. The spatial coordinates and corresponding hazard level of each risk node are extracted. The hazard level is comprehensively assessed based on the temperature distribution, smoke concentration, and relative position to the fire source in the node's area, typically categorized into several discrete levels (e.g., low, medium, high, and extremely high) or expressed as a continuous numerical value. After extracting the spatial locations of each risk node, the Euclidean spatial distance between any two risk nodes is calculated. Simultaneously, the radial distance of each risk node relative to the heat source center is calculated using the heat source center coordinates. The radial distance reflects the proximity of the node to the fire spread path and is an important basis for assessing the priority of fire propagation.

[0116] The determination of propagation weights between nodes considers both spatial distance and radial distance. Node pairs with closer spatial distances have higher propagation weights, indicating that fire heat or smoke is more easily transferred between adjacent nodes. Radial distance is used to correct the propagation direction; node pairs radiating outwards from the heat source have higher propagation weights to reflect the physical law that fire preferentially spreads along the main diffusion direction. Let nodes... With nodes The spatial distance between them is ,node The radial distance is ,node The radial distance is Then the weights are propagated between nodes. It can be represented as ,in This is the distance attenuation coefficient. To prevent extremely small positive numbers from being divided by zero, valid connections are filtered based on a propagation weight threshold. Establish a node when the preset connection threshold is exceeded. With nodes The directed edges between these edges ultimately construct a complete risk propagation graph.

[0117] morphological stability index The initial activation intensity of each node is determined by allocating resources according to its hazard level. Specifically, nodes with higher hazard levels are assigned a larger proportion of the initial activation intensity to reflect the stronger risk activation characteristics of high-risk areas in the early stages of a fire. Let nodes be... The normalized value of the hazard level is (The range of values ​​is) ), then the node initial activation strength satisfy This ensures that the sum of the initial activation intensities of all nodes equals the morphological stability index. This achieves the initial allocation in the sense of energy conservation.

[0118] After initial activation strength allocation, iterative diffusion calculations are performed on the risk propagation graph. In each iteration, the nodes... The activation strength of a node is updated by weighted accumulation based on the activation strength of its neighboring nodes and their corresponding propagation weights. The specific update rule is as follows: Node In the Activation intensity after the next iteration It equals the sum of its previous activation strength and the activation amounts from all neighboring nodes, i.e. ,in Represents a node The set of incoming neighbor nodes. To prevent the activation intensity from growing indefinitely, the activation intensity of each node is normalized after each update to keep it within a reasonable range. This diffusion process is repeated until the change in activation intensity of all nodes between two adjacent iterations is less than the convergence threshold. ,Right now When the diffusion process is considered to converge, the convergent activation intensity distribution is obtained. .

[0119] After obtaining the convergent activation intensity distribution, the activation intensity exceeding the preset activation threshold is statistically analyzed. The number of nodes is denoted as the number of active nodes. The number of activated nodes directly reflects the spatial extent of fire risk; a higher number indicates a wider range of threatened nodes. Simultaneously, spatial gradient calculations are performed on the convergent activation intensity distribution to obtain the activation intensity gradient distribution. Regions with large gradients indicate drastic changes in activation intensity along that direction, corresponding to the rapid spread front of fire risk. Based on the activation intensity gradient distribution, regions with gradient magnitudes exceeding a gradient threshold are identified. The region is analyzed to determine its spatial extent in order to estimate the rate of risk spread; simultaneously, the number of activated nodes is considered. Area of ​​a single grid cell The product calculation of risk coverage area The speed of risk spread and the area covered by the risk are linearly weighted and fused according to preset fusion weights to obtain a comprehensive risk situation assessment result. This result quantifies the severity and development trend of the current fire situation in scalar or vector form.

[0120] The alarm level is determined based on the risk situation assessment results. The assessment results are compared with preset multi-level alarm thresholds to determine whether the first-level (highest level) to multi-level alarm conditions are triggered, and the corresponding alarm level identifier is output. The alarm information is constructed by comprehensively utilizing the coordinates of the heat source center. (Including the x-axis) and the vertical axis The data includes the location of the heat source, the coordinates of the boundary of the threatened area, the alarm level, and the number of activated nodes. Risk coverage area Information such as timestamps is encapsulated into structured alarm information according to a predetermined format. Finally, the structured alarm information is encoded into alarm signals according to communication protocols (such as JSON serialization or binary encoding) and output to the fire control center, mobile terminals, or linked equipment through wired or wireless communication interfaces to achieve a closed-loop response for early fire warning.

[0121] A second aspect of this invention provides a fire early warning system based on AI video image recognition, comprising:

[0122] The evolution modeling unit is used to acquire visible light images and thermal infrared images within the monitoring area and perform registration and fusion to obtain multimodal features. It performs temporal modeling on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map, and performs graph structure encoding on the spatiotemporal feature map to extract the fire mode evolution trajectory.

[0123] The evolution analysis unit is used to calculate the diffusion rate distribution and morphological change distribution based on the fire mode evolution trajectory, perform directional decomposition on the diffusion rate distribution to obtain the main diffusion direction, perform stability assessment on the morphological change distribution to obtain the morphological stability index, determine the fire development stage identifier based on the main diffusion direction and the morphological stability index, and extract combustion behavior constraint rules from a preset physical constraint knowledge graph.

[0124] The classification and localization unit is used to perform constraint optimization on the spatiotemporal feature map based on the combustion behavior constraint rules to obtain a modified feature map and perform pixel-level classification to obtain the fire source area and the smoke diffusion area. It also performs spatial analysis on the fire source area and the smoke diffusion area to obtain the coordinates of the heat source center and the threatened area.

[0125] The risk assessment unit is used to construct a risk propagation map based on the threatened area, use the morphological stability index as the initial activation intensity of the node for iterative diffusion calculation until convergence, and obtain the risk situation assessment result by statistically analyzing the number of activated nodes and the distribution of activation intensity after convergence. It also generates and outputs an alarm signal by combining the coordinates of the heat source center and the threatened area.

[0126] A third aspect of the present invention provides an electronic device, comprising:

[0127] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0128] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0129] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fire early warning method based on AI video image recognition, characterized in that, include: Visible light and thermal infrared images within the monitoring area are acquired, registered, and fused to obtain multimodal features. Temporal modeling is performed on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map. Graph structure encoding is performed on the spatiotemporal feature map to extract the fire mode evolution trajectory. Based on the fire mode evolution trajectory, the diffusion rate distribution and morphological change distribution are calculated. The diffusion rate distribution is decomposed into a main diffusion direction. The morphological change distribution is evaluated for stability to obtain a morphological stability index. The fire development stage identifier is determined based on the main diffusion direction and the morphological stability index, and combustion behavior constraint rules are extracted from a preset physical constraint knowledge graph. Based on the combustion behavior constraint rules, the spatiotemporal feature map is constrained and optimized to obtain a modified feature map, and pixel-level classification is performed to obtain the fire source region and the smoke diffusion region. Spatial analysis is performed on the fire source region and the smoke diffusion region to obtain the coordinates of the heat source center and the threatened region. A risk propagation map is constructed based on the threatened area. The morphological stability index is used as the initial activation intensity of the node for iterative diffusion calculation until convergence. The number of activated nodes and the distribution of activation intensity after convergence are statistically analyzed to obtain the risk situation assessment result. An alarm signal is generated and output by combining the coordinates of the heat source center and the threatened area. Based on the main diffusion direction and the morphological stability index, fire development stage identifiers are determined, and combustion behavior constraint rules are extracted from a preset physical constraint knowledge graph, including: The directional angle parameters are obtained by spherical coordinate projection on the main diffusion direction, and the directional similarity distribution is obtained by calculating the angular deviation from each typical diffusion mode in the preset combustion mode library. Based on the morphological stability index and the directional similarity distribution, a state feature tensor is constructed and multilinear singular value decomposition is performed to determine the core tensor. Based on the core tensor, the state evolution velocity field is calculated and numerical integration is performed along the time dimension to obtain the state evolution trajectory. The curvature and torsion of the state evolution trajectory in phase space are calculated to obtain the dynamic feature vector. The distance between the dynamic feature vector and the preset stage boundary hyperplane is measured to determine the fire development stage identifier. Based on the fire development stage identifier, the corresponding stage node is retrieved from the preset physical constraint knowledge graph, and the first-order neighbor node is extracted to obtain a constraint entity set. The constraint entity set is then filtered and separated by type to obtain thermal radiation constraint entities and convective heat transfer constraint entities. The radiation flux attenuation relationship and the viewpoint factor calculation method in the thermal radiation constraint entities are extracted to construct thermal radiation constraint rules. The convective heat transfer coefficient function relationship and the boundary layer temperature gradient calculation method in the convective heat transfer constraint entities are extracted to construct convective heat transfer constraint rules. These rules are then logically combined with the thermal radiation constraint rules to obtain combustion behavior constraint rules.

2. The method according to claim 1, characterized in that, Visible light and thermal infrared images within the monitored area are acquired, registered, and fused to obtain multimodal features. Temporal modeling is performed on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map. Graph structure encoding is then performed on the spatiotemporal feature map to extract the fire mode evolution trajectory, including: Visible light images of the monitored area are acquired by a visible light camera, and thermal infrared images of the monitored area are acquired by a thermal infrared sensor. Texture and color features of the visible light images are extracted to construct a visible light feature map, and temperature gradient features of the thermal infrared images are extracted to construct a thermal infrared feature map. Spatial transformation correction is performed on the visible light feature map and the thermal infrared feature map, and the feature response differences at corresponding positions are calculated to construct a difference mask. Multimodal features are constructed based on the difference mask, the corrected visible light feature map, and the thermal infrared feature map. Multimodal features from multiple consecutive frames are arranged in chronological order to construct a feature sequence, and differential operations are performed on adjacent frames to obtain temporal variation features. Temporal convolution is performed on the temporal variation features to extract short-term motion patterns. Long-range dependency modeling is performed on the feature sequence to extract long-term evolution trends, and the sequence is concatenated with the short-term motion patterns to obtain the spatiotemporal feature map. The spatiotemporal feature map is divided into spatial grids to obtain multiple spatial nodes. The feature similarity between different spatial nodes is calculated and the node connection relationship is constructed. The feature of each node is updated by multi-level propagation to obtain the node embedding representation. The key nodes corresponding to the node embedding representation are extracted and the change trajectory of the embedding representation in the time dimension is tracked to obtain the fire mode evolution trajectory.

3. The method according to claim 1, characterized in that, Based on the fire mode evolution trajectory, the diffusion rate distribution and morphological change distribution are calculated. The diffusion rate distribution is decomposed directionally to obtain the main diffusion direction. The morphological change distribution is evaluated for stability to obtain morphological stability indices, including: The key node positions at different times are extracted from the fire mode evolution trajectory and the displacement vector between key nodes at adjacent times is calculated. The instantaneous diffusion rate is calculated based on the displacement vector and the diffusion rate distribution is constructed according to the spatial location. The fire source area contour at different times in the fire mode evolution trajectory is extracted and the shape difference degree of the contour at adjacent times is calculated. The shape difference degree is statistically aggregated according to the spatial location to obtain the morphological change distribution. Anisotropic filtering decomposition is performed on the diffusion rate distribution to obtain diffusion components in different directions. The direction corresponding to the energy peak in the diffusion component is extracted as the main diffusion direction and the directional concentration is calculated. Divergence calculation is performed on the diffusion rate distribution to obtain the diffusion source intensity distribution. Wavelet multi-resolution decomposition is performed on the morphological change distribution to obtain change components at different scales, and an autocorrelation matrix is ​​constructed using time autocorrelation coefficients. Eigenvalue decomposition is performed on the autocorrelation matrix to extract the dominant stable mode and calculate the energy proportion to obtain a global fluctuation index. Persistent homology analysis is performed on the morphological change distribution to obtain the birth and death cycle of morphological features and calculate the persistence entropy. An initial stability value is calculated based on the global fluctuation index and the persistence entropy. A directional dynamic factor is constructed based on the directional concentration and the diffusion source intensity distribution to correct the initial stability value and obtain a morphological stability index.

4. The method according to claim 1, characterized in that, Based on the combustion behavior constraint rules, the spatiotemporal feature map is constrained and optimized to obtain a modified feature map, and pixel-level classification is performed to obtain the fire source region and the smoke diffusion region. Spatial analysis of the fire source region and the smoke diffusion region is performed to obtain the coordinates of the heat source center and the threatened area, including: The combustion behavior constraint rules are converted into spatial propagation constraint conditions, and the spatiotemporal feature map is subjected to pixel-by-pixel constraint verification to obtain the physical consistency deviation distribution. The physical consistency deviation distribution is subjected to adaptive threshold segmentation to identify abnormal feature locations and construct a constraint mask. Based on the constraint mask, the abnormal feature locations in the spatiotemporal feature map are subjected to neighborhood feature weighted smoothing and iterative correction combined with the physical consistency deviation to obtain a corrected feature map. The corrected feature map is subjected to multi-scale segmentation to extract candidate regions of different scales and combined with temperature gradient and smoke concentration gradient for dual screening to obtain the fire source region and smoke diffusion region. The centroid of the fire source area is calculated to obtain the coordinates of the heat source center. Based on the coordinates of the heat source center and the smoke diffusion area, a distance attenuation weight field is constructed and the propagation intensity value at each spatial location is calculated. The propagation intensity value is hierarchically clustered to extract spatial regions with different intensity levels. The hazard level distribution is calculated by combining the temperature distribution in the modified feature map. Based on the hazard level distribution, connected regions exceeding the preset critical hazard level are identified to obtain the threatened area.

5. The method according to claim 4, characterized in that, Adaptive threshold segmentation is performed on the physical consistency deviation distribution to identify anomalous feature locations and construct a constraint mask. Based on the constraint mask, neighborhood feature weighted smoothing is applied to the anomalous feature locations in the spatiotemporal feature map, and iterative correction is performed in conjunction with the physical consistency deviation to obtain a corrected feature map, including: The cumulative distribution curve of the physical consistency deviation distribution is calculated and the slope change position is extracted as an adaptive threshold for binarization segmentation to obtain an initial anomaly mask. The initial anomaly mask is subjected to connected component analysis to obtain compactness and filter anomaly feature positions. The spatial coordinate encoding of the anomaly feature positions is used to construct a constraint mask. Based on the constraint mask, the abnormal feature positions in the spatiotemporal feature map are located, and a multi-scale neighborhood window is constructed to extract the normal feature values ​​in the neighborhood corresponding to the abnormal feature positions. The spatial decay weight and feature similarity weight of the normal feature values ​​are calculated and fused to obtain a comprehensive weight. Based on the comprehensive weight, the normal feature values ​​in the neighborhood are weighted and fused to obtain smooth feature values. The smooth feature values ​​are weighted and combined with the physical consistency deviation values ​​of the corresponding positions to obtain a correction value and update the abnormal feature positions. Based on the updated spatiotemporal feature map, the physical consistency deviation distribution is recalculated and the constraint mask is updated. This process is repeated until the number of abnormal feature positions in the constraint mask converges to obtain the corrected feature map.

6. The method according to claim 1, characterized in that, A risk propagation map is constructed based on the threatened area. The morphological stability index is used as the initial activation intensity of the nodes for iterative diffusion calculation until convergence. The number of activated nodes and the distribution of activation intensity after convergence are statistically analyzed to obtain the risk situation assessment result. An alarm signal is generated and output based on the coordinates of the heat source center and the threatened area, including: Based on the threatened area, a set of risk nodes is obtained by spatial grid division, and the spatial location and danger level of each risk node are extracted. The spatial distance between different risk nodes is calculated, and the radial distance of each risk node relative to the heat source is calculated by combining the coordinates of the heat source center. Based on the radial distance and the spatial distance, the propagation weight between nodes is determined, and the node connection relationship is established to construct a risk propagation map. The morphological stability index is allocated according to the danger level of each risk node to obtain the initial activation intensity of the node. The initial activation intensity of each node in the risk propagation diagram is calculated and updated according to the node connection relationship and propagation weight. This process is repeated until the change in activation intensity of each node converges to obtain a converged activation intensity distribution. The number of nodes exceeding a preset activation threshold in the converged activation intensity distribution is counted to obtain the number of activated nodes. The gradient of the converged activation intensity distribution is calculated to obtain the activation intensity gradient distribution. Based on the number of activated nodes and the activation intensity gradient distribution, the risk propagation speed and risk coverage area are calculated and fused to obtain the risk situation assessment result. Based on the risk situation assessment result, the alarm level is determined, and alarm information is constructed by combining the coordinates of the heat source center and the spatial range of the threatened area. The alarm information is encoded into an alarm signal and output.

7. A fire early warning system based on AI video image recognition, used to implement the method of any one of claims 1-6, characterized in that, include: The evolution modeling unit is used to acquire visible light images and thermal infrared images within the monitoring area and perform registration and fusion to obtain multimodal features. It performs temporal modeling on the multimodal features corresponding to multiple consecutive frames to construct a spatiotemporal feature map, and performs graph structure encoding on the spatiotemporal feature map to extract the fire mode evolution trajectory. The evolution analysis unit is used to calculate the diffusion rate distribution and morphological change distribution based on the fire mode evolution trajectory, perform directional decomposition on the diffusion rate distribution to obtain the main diffusion direction, perform stability assessment on the morphological change distribution to obtain the morphological stability index, determine the fire development stage identifier based on the main diffusion direction and the morphological stability index, and extract combustion behavior constraint rules from a preset physical constraint knowledge graph. The classification and localization unit is used to perform constraint optimization on the spatiotemporal feature map based on the combustion behavior constraint rules to obtain a modified feature map and perform pixel-level classification to obtain the fire source area and the smoke diffusion area. It also performs spatial analysis on the fire source area and the smoke diffusion area to obtain the coordinates of the heat source center and the threatened area. The risk assessment unit is used to construct a risk propagation map based on the threatened area, use the morphological stability index as the initial activation intensity of the node for iterative diffusion calculation until convergence, and obtain the risk situation assessment result by statistically analyzing the number of activated nodes and the distribution of activation intensity after convergence. It also generates and outputs an alarm signal by combining the coordinates of the heat source center and the threatened area.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images