A building fire hazard automatic detection method

Abstract

The present application relates to the field of building fire protection technology, and particularly relates to a building fire hazard automatic detection method, comprising the following steps: S100, obtaining multi-source heterogeneous perception data and static building information model (BIM) data of a target building, and after spatial alignment of the multi-source heterogeneous perception data, mapping to a three-dimensional space coordinate system of the static BIM data. The present application constructs a dynamic three-dimensional space fire knowledge graph containing entity state, distance relationship and shielding relationship, and inputs it into a space-time graph convolution network to extract evolution features in combination with a physical heat conduction model and dynamic passenger flow density, effectively solving the technical defects of lack of spatial semantics and multi-modal cross verification in the prior art, realizing automatic and quantitative early warning of space coupling hazards such as combustible material approaching heat source and blocked fire passage in the fire incubation stage, and improving the foresight and linkage response efficiency of fire hazard investigation in complex building scenarios.

Description

Technical Field

[0001] This invention relates to the field of building fire protection technology, specifically to an automatic detection method for building fire hazards. Background Technology

[0002] Existing building fire monitoring and hazard investigation systems mainly rely on independently operating smoke detectors, heat detectors, and two-dimensional visible light monitoring equipment. In complex building scenarios such as large commercial complexes, fire hazards often stem from the coupled evolution of unauthorized changes in physical spatial locations and abnormal equipment conditions. For example, temporary structures may dynamically occupy fire escape routes, or flammable materials may be illegally stored near high-load electrical equipment and abnormal heat sources.

[0003] Existing monitoring methods suffer from core flaws: two-dimensional vision algorithms and single-node IoT sensors exist in data silos, lacking integration with the three-dimensional spatial structure of buildings. This prevents cross-modal extraction of the three-dimensional spatial topological relationships between flammable materials, heat sources, and static building components. Current technologies cannot quantify the distance, degree of obstruction, and heat conduction probability between entities in the physical space dimension, nor do they possess a risk assessment mechanism that combines dynamic passenger flow density with the effective width of passageways. This results in existing systems only passively triggering alarms after a fire has started, failing to proactively identify complex fire hazards involving intersecting spatial topologies during the fire's incubation stage. Furthermore, they cannot accurately assess the actual degree of obstruction caused by spatial obstacles to evacuation routes, exhibiting severe early warning lag and limitations. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: This invention provides an automatic detection method for building fire hazards, comprising the following steps: S100, acquiring multi-source heterogeneous sensing data and static building information model (BIM) data of the target building, and spatially aligning the multi-source heterogeneous sensing data and mapping it to the three-dimensional spatial coordinate system of the static BIM data; S200, perform cross-modal feature extraction on the aligned multi-source heterogeneous sensing data, and identify dynamic spatial obstacles, fire risk entities and their corresponding state attributes in the target building; S300, based on the identified dynamic spatial obstacles, fire risk entities, state attributes, and spatial elements in the static BIM data, construct and dynamically update a three-dimensional spatial fire knowledge graph, wherein the node features of the three-dimensional spatial fire knowledge graph include entity states, and the edge features represent the spatial topology and physical relationship between entities. S400, the dynamically updated three-dimensional spatial fire protection knowledge graph is input into the pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction, the comprehensive fire hazard index of each spatial node is calculated, and the hazard classification early warning information is output based on the comprehensive fire hazard index.

[0005] Furthermore, the multi-source heterogeneous sensing data includes visible light image streams, infrared thermal imaging video streams, and IoT environmental sensing data; Acquiring multi-source heterogeneous sensing data of the target building, specifically including: The visible light image stream and the infrared thermal imaging video stream are simultaneously acquired from the same field of view using global monitoring equipment and thermal imaging equipment deployed within the target building. The IoT environmental sensing data is acquired through an environmental sensing network. The IoT environmental sensing data includes at least the internal load status data of electrical equipment and the environmental status data of the smoke exhaust duct in the catering area.

[0006] Furthermore, after spatial alignment of the multi-source heterogeneous sensing data, it is mapped to the three-dimensional spatial coordinate system of the static BIM data, including the following processing steps: Based on the pre-calibrated camera intrinsic and extrinsic parameter matrices, a transformation mapping function is constructed between the two-dimensional pixel coordinate system, the physical world coordinate system, and the three-dimensional spatial coordinate system of the static BIM data; Spatial homonym feature points are extracted from the synchronously acquired visible light image stream and infrared thermal imaging video stream. The affine transformation matrix is ​​calculated using the spatial homonym feature points to achieve pixel-level spatial registration of heterogeneous visual data. The transformation mapping function is invoked to project the registered heterogeneous visual data and the IoT environmental sensing data with spatial labels onto the corresponding nodes in the three-dimensional spatial coordinate system, thereby generating a multimodal spatial data base.

[0007] Furthermore, S100 also includes an alignment compensation mechanism: Real-time monitoring of the gimbal pose parameters of the global monitoring device and the thermal imaging device; When a shift in the gimbal pose parameters is detected, the static environmental reference features under the current field of view are extracted, and the extrinsic parameter matrix and the transformation mapping function are adaptively updated in combination with the offset to maintain the real-time consistency between the dynamically updated multimodal spatial data base and the static BIM data in the three-dimensional spatial coordinate system.

[0008] Furthermore, the identification of dynamic spatial obstacles within the target building includes a dynamic spatial reconstruction mechanism for temporary structures, with the following specific steps: The instance segmentation network is used to parse the visual features mapped to the three-dimensional spatial coordinate system in real time, extract semantic labels with non-fixed structure, and generate corresponding three-dimensional bounding boxes. Perform spatial Boolean operations on the three-dimensional bounding box and the preset fire evacuation route model and fireproof roller shutter forced landing area model in the static BIM data; If the calculation results show a spatial intersection, the non-fixed structure that generates the spatial intersection is identified as the dynamic spatial obstacle, and its occupancy area and spatial volume are calculated as the corresponding state attributes.

[0009] Furthermore, the cross-modal feature extraction of the aligned multi-source heterogeneous sensing data is performed to identify fire risk entities, and a feature fusion verification strategy based on a two-stream convolutional neural network is adopted: The visible light branch of the dual-stream convolutional neural network is used to extract flammable objects and fire-fighting facilities in the scene, and the infrared branch is used to extract the surface temperature gradient distribution. The pixel region of the flammable object is fused with the abnormally heated region in the surface temperature gradient distribution using region of interest (ROI) pooling. When the temperature rise slope of the abnormal heating area exceeds a preset threshold, and the Euclidean distance between its spatial enclosure and the spatial enclosure of the flammable material entity is less than a safety threshold, the flammable material entity and the adjacent heat source are jointly marked as the fire risk entity of the high-risk level. The acquisition of its corresponding state attributes adopts cross-modal data coupling extraction rules, specifically as follows: Visual features of oil stain thickness on the surface of the smoke exhaust pipe in the catering defense zone were extracted based on the image gray-level co-occurrence matrix. The visual characteristics of the oil stain thickness, the abnormal temperature fluctuation characteristics inside the pipeline obtained by the infrared thermal imaging video stream, and the gas pipeline pressure characteristics in the IoT environmental sensing data are vectorized and stitched together. Based on the spliced ​​feature vectors, the fire risk entities within the catering protection zone are assigned multi-dimensional fire coupling state attributes to characterize the potential fire risk level caused by oil accumulation and high-temperature smoke exhaust.

[0010] Furthermore, based on the identified dynamic spatial obstacles, fire risk entities, state attributes, and spatial elements in the static BIM data, a three-dimensional spatial fire safety knowledge graph is constructed. This is achieved using a hierarchical graph construction strategy based on ontology and instances, specifically including: The fire compartments, evacuation routes, and fixed fire protection facilities in the static BIM data are abstracted into static body nodes of the graph; The identified dynamic spatial obstacles and fire risk entities are abstracted into dynamic instance nodes of a graph, and the corresponding state attributes are assigned as feature vectors to the corresponding nodes. The relative spatial distribution between each node is calculated based on the three-dimensional spatial coordinate system, and the edge features connecting the nodes are constructed. The edge features include distance edges, occlusion edges, and containment edges. The occlusion edges are generated based on the projection overlap between the three-dimensional bounding box of the dynamic instance node and the static body node from the evacuation viewpoint. The edge features also include dynamic physical association weights that characterize the trend of danger spread, and the calculation mechanism of the dynamic physical association weights is as follows: Extract the surface temperature gradient value associated with the fire risk entity from the state attributes; Combining the relative spatial distance corresponding to the distance edge, and the current ventilation status data of the protection zone obtained from the IoT environmental sensing data, the physical probability of the fire risk entity transmitting heat energy or harmful gas to its adjacent nodes is calculated using a preset heat conduction attenuation model. After normalizing the physical probabilities, they are assigned as dynamic physical association weights to the edges connecting the corresponding nodes to quantitatively characterize the spatial coupling hazard between entities during the fire incubation stage.

[0011] Furthermore, a subgraph update strategy based on local spatial topology events is adopted: When real-time sensing data detects that the spatial position of the dynamic space obstacle has shifted, or the state attribute of the fire risk entity has changed abruptly, and the amount of displacement or the magnitude of the change exceeds a preset stability threshold, a local space anomaly event is generated. In response to the local spatial anomaly event, only the target node where the anomaly occurred is locked, and the edge features and node features in the first-order and second-order connected domains of the target node are reconstructed and calculated, while the topology of the remaining unaffected subgraphs in the three-dimensional spatial fire protection knowledge graph remains unchanged.

[0012] Furthermore, the dynamically updated 3D spatial fire protection knowledge graph is input into a pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction. This process employs a spatiotemporal joint graph convolution strategy incorporating self-attention mechanisms and includes the following steps: By aggregating the node features and edge features of the target node and its neighboring nodes through the spatial graph convolutional layer in the spatiotemporal graph convolutional network model, spatial topological features that characterize the spread trend of fire risk in three-dimensional space are extracted. The spatial topology features within multiple consecutive time steps are extracted using the temporal convolutional layer in the spatiotemporal graph convolutional network model to capture the dynamic mutation rate of the state attributes (including temperature, oil or gas concentration). A self-attention mechanism module is introduced to assign higher attention weights to high-risk nodes in the spatial topology and temporal features, and output the physical fire incubation probability index of each target node in the future preset time window.

[0013] Furthermore, the calculation of the comprehensive fire hazard index for each spatial node adopts a multi-dimensional parameter aggregation risk quantification evaluation model, the calculation formula of which is as follows: in, The comprehensive fire hazard index; This refers to the physical fire spawning probability index; The first obtained through visual passenger flow statistics Real-time population density within each defense zone; The first in the static BIM data The initial design of the total width of the escape routes in each defense zone; The width of the current passage being illegally obstructed by the dynamic spatial obstacle, calculated from the three-dimensional spatial fire protection knowledge graph; As a gathering point for people Spatial topological distance to predicted high-risk target nodes; , , Weighting coefficients preset for the target building structure; The output of hazard classification and early warning information based on the comprehensive fire hazard index adopts a virtual-real mapping linkage response mechanism based on the BIM digital twin platform: The calculated comprehensive fire hazard index is compared with the preset dynamic grading threshold to determine the hazard level; The three-dimensional spatial coordinates of the target nodes that exceed the comprehensive fire hazard index are projected in reverse to the BIM digital twin visualization platform, and the corresponding physical defense zone and hazard spread radiation circle are highlighted and rendered in the virtual twin model. Based on the determined hazard level, a hazard rectification order is automatically generated and sent to the mobile terminal. When the hazard level is determined to be the highest, the pre-execution linkage strategy of the hardware equipment in the corresponding defense zone is triggered. The pre-execution linkage strategy includes the early activation of the smoke exhaust system preheating or non-fire protection power cut-off action in the defense zone.

[0014] Beneficial effects This invention constructs a dynamic 3D spatial fire safety knowledge graph that includes entity states, distance relationships, and occlusion relationships by aligning multi-source heterogeneous sensing data such as visible light and infrared thermal imaging with static BIM models at the pixel level and mapping them in 3D space. This knowledge graph is then input into a spatiotemporal graph convolutional network and combined with a physical heat conduction model and dynamic passenger flow density for evolutionary feature extraction. This effectively solves the technical defects of existing technologies that lack spatial semantics and multimodal cross-validation. It enables automated and quantitative early warning of spatially coupled hazards such as flammable materials approaching heat sources and blocked fire lanes during the fire incubation stage, improving the foresight and linkage response efficiency of fire hazard investigation in complex building scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a schematic diagram illustrating the principle of multimodal sensing data spatial alignment and mapping in this invention. Figure 3 This is the flow graph for cross-modal feature extraction and entity logic binding of the present invention; Figure 4 This is a diagram of the spatiotemporal graph network prediction and virtual-real linkage architecture of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] The present invention will now be described in further detail with reference to the accompanying drawings: Example: like Figures 1-4As shown, an automatic detection method for building fire hazards includes the following steps: S100: Acquire multi-source heterogeneous sensing data and static building information model (BIM) data of the target building, and after spatial alignment of the multi-source heterogeneous sensing data, map it into the three-dimensional spatial coordinate system of the static BIM data. S200 performs cross-modal feature extraction on aligned multi-source heterogeneous sensing data to identify dynamic spatial obstacles, fire risk entities and their corresponding state attributes within the target building. S300 constructs and dynamically updates a three-dimensional spatial fire protection knowledge graph based on the identified dynamic spatial obstacles, fire risk entities, status attributes, and spatial elements in static BIM data. The node features of the three-dimensional spatial fire protection knowledge graph include entity status, and the edge features represent the spatial topology and physical relationship between entities. S400 inputs the dynamically updated 3D spatial fire safety knowledge graph into a pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction, calculates the comprehensive fire hazard index for each spatial node, and outputs hazard classification early warning information based on the comprehensive fire hazard index.

[0019] Furthermore, the specific implementation process of step S100 is as follows: Step S100: Obtain multi-source heterogeneous sensing data and static building information model (BIM) data of the target building, and after spatial alignment of the multi-source heterogeneous sensing data, map it to the three-dimensional spatial coordinate system of the static BIM data.

[0020] The target building is equipped with global monitoring equipment, thermal imaging equipment, and an environmental sensor network. The global monitoring equipment continuously outputs visible light image streams, while the thermal imaging equipment continuously outputs infrared thermal imaging video streams. The environmental sensor network includes current and temperature sensors deployed inside electrical equipment, as well as pressure and temperature sensors deployed in the exhaust ducts of the catering area. The visible light image streams, infrared thermal imaging video streams, and IoT environmental sensor data are time-stamped at the acquisition end via a network time protocol, ensuring that the multi-source heterogeneous sensing data are sampled at the same time dimension. The static BIM data adopts the Industrial Basic Class (IFC) standard data format and has a built-in absolute three-dimensional spatial coordinate system.

[0021] The process involves spatially aligning multi-source heterogeneous sensing data and mapping it to the three-dimensional spatial coordinate system of static BIM data, specifically including calibration transformation, pixel registration, and node projection operations.

[0022] In the calibration transformation operation, calibration parameters of the global monitoring equipment and thermal imaging equipment are extracted, and a transformation mapping function is constructed between the two-dimensional pixel coordinate system, the physical world coordinate system, and the three-dimensional spatial coordinate system of the static BIM data. The calibration parameters include the camera intrinsic parameter matrix and the extrinsic parameter matrix. The intrinsic parameter matrix represents the camera's focal length, principal point coordinates, and distortion coefficients; the extrinsic parameter matrix represents the rotation matrix and translation vector of the optical center of the global monitoring equipment and the thermal imaging equipment relative to the origin of the physical world coordinate system.

[0023] In the pixel registration operation, single-frame extraction is performed on the synchronously acquired visible light image stream and infrared thermal imaging video stream, and spatial homonymous feature points are matched in the corresponding extracted frames. A scale-invariant feature transformation algorithm or an accelerated robust feature algorithm is used to extract feature points with significant local structural attributes from the two types of heterogeneous visual data. Specifically, the extraction targets are static physical corner points in the environment, such as wall corners, fire door frames, and the outer contours of fire hydrants. Using the matched spatial homonymous feature points, the affine transformation matrix between the visible light image coordinate system and the infrared image coordinate system is calculated to complete the pixel-level spatial registration of the heterogeneous visual data. In the registered multimodal image matrix, each pixel unit's data channel simultaneously contains red, green, and blue (RGB) color values ​​and a surface temperature scalar value.

[0024] In the node projection operation, the constructed transformation mapping function is called. Combined with the depth information calculated by the binocular vision depth estimation algorithm, the 3D spatial position of each pixel unit in the registered multimodal image matrix is ​​transformed into the 3D spatial coordinate system of the static BIM data through coordinate system multiplication, generating a visual spatial point cloud. During the sensor hardware deployment phase, IoT environmental sensing data has already been entered with absolute 3D coordinate labels matching the static BIM data model. The system directly mounts the IoT environmental sensing data to the corresponding node positions in the 3D spatial coordinate system based on these coordinate labels. The visual spatial point cloud and the mounted sensor node data are then merged into a multimodal spatial data base.

[0025] To address the issue of spatial mapping failure caused by changes in the viewing angle of the PTZ (pan-tilt-zoom) camera in large commercial complexes, the embodiment of step S100 includes an alignment compensation mechanism. Through the application programming interfaces (APIs) at the bottom layers of the global monitoring equipment and the thermal imaging equipment, the current horizontal yaw angle, pitch angle, and zoom parameters of the PTZ camera are read in real time. When the program detects a difference between the current value of the horizontal yaw angle, pitch angle, or zoom parameter and its initial calibration value, alignment compensation calculation is triggered.

[0026] The system extracts static environmental reference features from the current field of view. These features are specifically defined as the edges of load-bearing columns or fixed stair railings of the 3D structure recorded in the static BIM data. The extracted static environmental reference features are then compared with their current pixel coordinates in a 2D pixel coordinate system using perspective N-point (PnP) pose estimation. This process is used to inversely calculate the real-time rotation matrix and translation vector after the device's pan-tilt unit (PTZ) changes. The solved real-time rotation matrix and translation vector are then overwritten into the extrinsic parameter matrix, completing the adaptive update of the calibration parameters. The system synchronously reconstructs the transformation mapping function based on the updated extrinsic parameter matrix to maintain numerical consistency between the dynamically updated multimodal spatial data base and the static BIM data in the 3D spatial coordinate system.

[0027] Furthermore, the specific implementation process of step S200 is as follows: Step S200: Perform cross-modal feature extraction on the aligned multi-source heterogeneous sensing data to identify dynamic spatial obstacles, fire risk entities and their corresponding state attributes within the target building.

[0028] The system identifies dynamic spatial obstacles within target buildings, including a dynamic spatial reconstruction mechanism for temporary structures. A multimodal spatial data base mapped to a 3D coordinate system is input into an instance segmentation network model. The backbone of the instance segmentation network extracts downsampled features from the visual spatial point cloud in the multimodal spatial data base using residual convolutional layers, outputting multi-scale feature maps. A region proposal network generates candidate bounding boxes on the multi-scale feature maps, and a mask prediction branch outputs pixel-level class masks in parallel. The system extracts pixel clusters with non-fixed-structure semantic labels from the class masks; these non-fixed-structure semantic labels specifically include temporary booths, stacked goods, and amusement facilities.

[0029] Based on the extracted pixel clusters and their corresponding depth information, the minimum bounding cube algorithm is used to calculate the boundary extrema of the pixel clusters in the three-dimensional spatial coordinate system, generating the corresponding three-dimensional bounding boxes. The system reads static BIM data and extracts the three-dimensional coordinate sets of the preset fire evacuation route model and the three-dimensional coordinate sets of the fireproof roller shutter forced landing area model from the static BIM data.

[0030] The system invokes the computational geometry component to perform a spatial Boolean intersection operation on the generated 3D bounding box, the 3D coordinate set of the fire evacuation route model, and the 3D coordinate set of the fireproof roller shutter forced landing area model. If the spatial Boolean intersection operation outputs a non-empty subset of 3D spatial coordinates, then the non-fixed structure that produces the non-empty subset of 3D spatial coordinates is determined to be a dynamic spatial obstacle.

[0031] The system counts the total number of voxels within a non-empty 3D spatial coordinate subset and calculates the corresponding spatial volume using the voxel integral formula. Simultaneously, it orthogonally projects the non-empty 3D spatial coordinate subset onto a 2D horizontal plane along the direction of gravity and calculates the area of ​​the orthogonally projected polygon as the occlusion area. The system combines the occlusion area and spatial volume values ​​as the state attribute corresponding to the dynamic spatial obstacle and writes them into the memory dictionary.

[0032] Cross-modal feature extraction and fire risk entity identification are performed on aligned multi-source heterogeneous sensing data using a feature fusion mechanism based on a two-stream convolutional neural network. The two-stream convolutional neural network includes parallel visible light data processing branches and infrared data processing branches.

[0033] The red, green, and blue color value channels in the registered multimodal image matrix are input into the visible light data processing branch. Texture gradient features and shape contour features are extracted through multi-layer convolutional kernels, and a visible light feature map is output.

[0034] Classification regression was performed based on visible light feature maps to identify flammable material entities and fire protection facility entities. Flammable material entities included cardboard boxes and stacks of wooden boards; fire protection facility entities included fire extinguishers and indoor fire hydrants.

[0035] The surface temperature scalar value channel from the registered multimodal image matrix is ​​simultaneously input to the infrared data processing branch to extract the temperature gradient matrix within consecutive frames and output an infrared feature map. The system extracts the set of consecutive pixels in the infrared feature map whose temperature scalar value is higher than the ambient reference temperature threshold, generates a surface temperature gradient distribution matrix, and marks the corresponding abnormal heating areas in a three-dimensional spatial coordinate system.

[0036] The system extracts the pixel coordinate range of the flammable object in the visible light feature map and the pixel coordinate range of the abnormally heated area in the infrared feature map. It then performs region of interest (ROI) pooling fusion on the pixel coordinate ranges of the flammable object and the abnormally heated area, mapping the heterogeneous feature maps into feature vectors of uniform, fixed size. In the feature vector extraction layer, the system calculates the temperature peak value of the abnormally heated area at multiple consecutive sampling times within a preset time window, and calculates the temporal slope of the temperature peak value by taking the derivative.

[0037] The system extracts the 3D coordinates of the center point of the bounding box of the flammable material entity and the 3D coordinates of the center point of the bounding box of the abnormally heated area. Based on the 3D coordinate system distance formula, it calculates the 3D Euclidean distance between the two center points. When the temporal slope of the temperature peak exceeds a preset slope threshold, and the calculated 3D Euclidean distance is less than a preset safety threshold, the system logically binds the flammable material entity that meets the above conditions to the adjacent abnormally heated area, and marks the logically bound set of physical objects as a high-risk fire hazard entity.

[0038] For the catering defense zone in a large commercial complex, a cross-modal data coupling extraction rule is used to acquire status attributes. The system extracts multi-source heterogeneous sensing data deployed on the surface of the exhaust ducts in the catering defense zone and performs grayscale processing on its visible light image layer. The system traverses the pixel matrix in the grayscale image, calculates the joint probability distribution of pixel pairs appearing at specific pixel steps and angles, and constructs an image grayscale co-occurrence matrix. Contrast parameters, correlation parameters, energy parameters, and homogeneity parameters are extracted from the image grayscale co-occurrence matrix. The extracted texture parameters are input into a pre-fitted multiple linear regression function, and the output is a visual feature value of the oil stain thickness on the surface of the exhaust duct.

[0039] The system synchronously extracts the set of extreme temperature values ​​for the corresponding internal region of the exhaust duct in the infrared thermal imaging video stream under a time series frame, calculates the root mean square error of the extreme temperature value set, and generates characteristic values ​​of abnormal temperature fluctuations. Simultaneously, the system reads the pressure characteristic values ​​of the gas pipeline network collected by pressure sensors from the IoT environmental sensing data via a serial communication interface.

[0040] The system normalizes the visual characteristics of oil stain thickness, abnormal temperature fluctuations, and gas pipeline pressure by applying maximum and minimum values. The normalized floating-point numbers from these three dimensions are then concatenated into a vector according to a pre-defined data structure, generating a one-dimensional concatenated feature vector. This one-dimensional feature vector is then appended as a multi-dimensional fire coupling state attribute to the identified fire risk entity data nodes within the catering service area.

[0041] Furthermore, the specific implementation process of step S300 is as follows: Step S300: Based on the identified dynamic spatial obstacles, fire risk entities, status attributes, and spatial elements in static BIM data, construct and dynamically update a three-dimensional spatial fire protection knowledge graph. The node features of the three-dimensional spatial fire protection knowledge graph include entity status, and the edge features represent the spatial topology and physical relationship between entities.

[0042] A 3D spatial fire protection knowledge graph is constructed based on the identified perceptual data and static BIM data, employing a hierarchical graph construction mechanism based on ontology and instances. The system reads the static BIM data, extracting fire compartment grids, evacuation routes, and coordinates of fixed fire protection facilities with spatial physical attributes. These extracted fire compartment grids, evacuation routes, and coordinates of fixed fire protection facilities are then instantiated as static ontology nodes in the graph database. Simultaneously, dynamic spatial obstacles and fire risk entities identified in the previous steps are extracted and instantiated as dynamic instance nodes in the graph database. The system reads the state attribute data associated with the dynamic instance nodes, converts the state attribute data into multi-dimensional feature vectors, and assigns these multi-dimensional feature vectors as node features to the corresponding dynamic instance nodes and static ontology nodes.

[0043] The relative spatial 3D coordinate difference between all nodes in the 3D spatial fire protection knowledge graph is calculated based on the 3D spatial coordinate system, and the edge features connecting different nodes are constructed.

[0044] Edge features specifically include distance edges, occlusion edges, and containment edges. When generating distance edges, the Euclidean distance is calculated from the 3D coordinate center points of connected nodes, and the calculated physical distance scalar value is assigned to the distance edge. When generating containment edges, it is determined whether the 3D coordinates of a dynamic instance node are completely within the 3D coordinate bounding box corresponding to a fire compartment mesh. If the determination result is true, a directed containment edge is generated between the dynamic instance node and the corresponding static body node.

[0045] The generation of occlusion edges employs an algorithm based on the projection intersection-union ratio (IUGR). The system extracts the human eye's line-of-sight vector defined in the static BIM data for evacuation routes and constructs a virtual 2D projection plane perpendicular to this vector. The system then performs a mathematical matrix-based dimensionality reduction orthogonal projection of the 3D bounding box coordinates of the dynamic instance nodes and the 3D contour coordinates of the evacuation routes along the human eye's line-of-sight vector, mapping them onto the virtual 2D projection plane to form a first polygon and a second polygon. The system calculates the intersection area parameter of the first and second polygons and divides this parameter by the total area parameter of the second polygon to obtain the projection overlap value. This projection overlap value is then assigned as a weight parameter to the occlusion edge to quantitatively characterize the degree to which dynamic spatial obstacles physically obstruct the evacuation line of sight.

[0046] The edge features in the 3D spatial fire protection knowledge graph synchronously include dynamic physical association weights that characterize the trend of hazard spread. The system traverses dynamic instance nodes with fire risk entity labels in the graph database and reads the surface temperature gradient values ​​recorded in the status attributes. At the same time, it reads the relative spatial physical distances recorded in the distance edges and reads the ventilation status parameters collected by wind speed and direction sensors in the current protection zone from the IoT environmental sensing data through a serial interface.

[0047] The system uses surface temperature gradient values, relative spatial physical distances, and ventilation status parameters as input variables, substituting them into a pre-defined heat conduction attenuation model. The heat conduction attenuation model operates on a partial differential thermodynamic diffusion equation at its core. After numerical integration, it outputs physical probability scalars representing the heat energy propagation and harmful gas release from a fire-risk entity to its topologically adjacent nodes. The system uses a minimum-maximum normalization function to map these physical probability scalars to continuous floating-point numbers between zero and one. These continuous floating-point numbers are then used as dynamic physical association weights, overwritten into the graph edge feature data structure connecting the fire-risk entity and its topologically adjacent nodes.

[0048] The three-dimensional spatial fire protection knowledge graph is dynamically updated using a subgraph reconstruction mechanism triggered by local spatial topology events. The system clock continuously detects the sensing data stream according to a preset millisecond-level sampling period. The system calculates the spatial displacement vector value of the three-dimensional coordinate center point of the dynamic spatial obstacle in the current sampling period compared to the previous period, and calculates the numerical jump difference of the multi-dimensional feature vector in the fire risk entity state attribute.

[0049] The spatial displacement vector value and the numerical jump difference are compared with the preset spatial stability threshold and attribute stability threshold, respectively. When the spatial displacement vector value exceeds the spatial stability threshold, or the numerical jump difference exceeds the attribute stability threshold, the system triggers an interrupt handler and generates a local spatial anomaly event with the target node identification number.

[0050] The system memory responds to local spatial anomaly events, locking the anomaly-affected target node in the graph database based on the target node identification number. Using the target node as the root node, the system executes a breadth-first search algorithm, traversing and extracting the first-level connected node set with a path depth of one and the second-level connected node set with a path depth of two, forming a first-order connected domain and a second-order connected domain subgraph structure composed of the target node, the first-level connected node set, and the second-level connected node set. The system calls processor resources to re-extract and overwrite the edge and node features only within the first-order and second-order connected domain subgraph structures. Simultaneously with the local subgraph reconstruction calculation, the system database controller performs read-write lock freezing operations on the topology of all other unaffected subgraphs outside the first-order and second-order connected domains in the 3D spatial fire protection knowledge graph to avoid global matrix multiplication iterations of the entire graph data.

[0051] Furthermore, the specific implementation process of step S400 is as follows: Step S400: Input the dynamically updated 3D spatial fire protection knowledge graph into the pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction, calculate the comprehensive fire hazard index of each spatial node, and output hazard classification early warning information based on the comprehensive fire hazard index.

[0052] The dynamically updated 3D spatial fire safety knowledge graph is input into a pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction, employing a spatiotemporal joint graph convolutional computation model incorporating a self-attention mechanism. The system transforms the 3D spatial fire safety knowledge graph into a sequence of adjacency matrices and node feature matrices, which are then input into the pre-trained spatiotemporal graph convolutional network model. The spatiotemporal graph convolutional network model calls its internally configured spatial graph convolutional layer, utilizing the graph Laplacian operator to perform weighted summation and aggregation operations on the features of the target node and its topologically adjacent nodes. During the aggregation operation, the spatial graph convolutional layer simultaneously extracts distance edge parameters, occlusion edge parameters, and dynamic physical association weights representing the spread trend of danger from the edge features, generating a spatial fusion feature vector matrix containing 3D spatial topological information. This completes the extraction of spatial topological features representing the spread trend of fire risk in 3D space.

[0053] The spatiotemporal graph convolutional network model invokes its internally configured temporal convolutional layer. The temporal convolutional layer receives a sequence of spatially fused feature vector matrices output at multiple consecutive time steps and performs a one-dimensional convolution operation along the time dimension. The one-dimensional convolutional kernel of the temporal convolutional layer performs derivative operations on the gradients of the surface temperature scalar value, abnormal temperature fluctuation feature value, and oil contamination thickness visual feature value contained in the state attributes along the time axis, completing temporal feature extraction and outputting a spatiotemporal joint feature matrix containing the dynamic mutation rate of the state attributes.

[0054] The system inputs the spatiotemporal joint feature matrix into the self-attention mechanism module. The self-attention mechanism module performs a linear transformation on the spatiotemporal joint feature matrix, generating a query matrix, a key matrix, and a value matrix. The system performs a dot product operation between the transpose of the query matrix and the key matrix, and then normalizes the result using the Softmax function to obtain the attention weight matrix. The system multiplies the attention weight matrix with the value matrix, adding bias weight parameters to high-risk nodes in the spatiotemporal joint feature matrix. After processing by a fully connected layer and a sigmoid activation function, the system outputs the physical fire incubation probability index for each target node within a preset time window. The preset time window is specifically set to a continuous period of five to thirty minutes in the future.

[0055] To address the high-density customer flow characteristics of large commercial complexes, a comprehensive fire hazard index for each spatial node is calculated using a multi-dimensional parameter aggregation risk quantification assessment algorithm. The system processor calls a pre-set calculation program from memory and executes the following risk quantification assessment model formula calculation: In formula calculations, Represents the comprehensive fire hazard index; This represents the physical fire incubation probability index output by the self-attention mechanism module. The system captures pedestrian silhouettes using visual passenger flow statistics cameras deployed in various defense zones, and statistically generates the [number]th [parameter]. The real-time population density value within each defense zone is denoted as... The system reads the static BIM data file and extracts the first... The initial design total width of the escape routes in each defense zone is denoted as... The system reads the state attributes of dynamic spatial obstacles from the 3D fire safety knowledge graph, extracts the obstruction width of the current passageway due to illegal occupation, and records it as... The system calculates crowd gathering points based on a three-dimensional spatial coordinate system. The three-dimensional spatial topological distance between the target node and the predicted high-risk target node is denoted as . The base of the natural constant is denoted as , , , This represents the weighting constant coefficients that are pre-calibrated for the target building structure and written into the read-only memory.

[0056] Based on the comprehensive fire hazard index, the system outputs hazard classification and early warning information, and adopts a virtual-real mapping linkage control mechanism based on the BIM digital twin platform. The system sets a primary dynamic classification threshold and a secondary dynamic classification threshold, and performs floating-point comparison calculations between the calculated comprehensive fire hazard index and the primary and secondary dynamic classification thresholds. When the comprehensive fire hazard index exceeds the secondary dynamic classification threshold, the system determines the hazard level as a general hazard; when the comprehensive fire hazard index exceeds the primary dynamic classification threshold, the system determines the hazard level as the highest hazard.

[0057] The system extracts the 3D spatial coordinates of target nodes that exceed the comprehensive fire hazard index. Through reverse perspective projection matrix calculation, the 3D spatial coordinates are projected backwards onto the display viewport of the BIM digital twin visualization platform. Within the virtual twin model of the BIM digital twin visualization platform, the system generates a polygonal rendering layer covering the corresponding physical defense zone and the hazard spread radiation circle, with the 3D spatial coordinates of the target node as the geometric center.

[0058] Simultaneously, based on the determined hazard level, the system invokes the network communication module to generate a hazard rectification dispatch data packet containing hazard coordinate parameters and hazard type. This data packet is then sent wirelessly to the designated safety personnel's mobile terminal hardware address. When the hazard level is determined to be the highest hazard, the system's main control chip invokes the fieldbus protocol to directly issue a set of level control commands to the programmable logic controller (PLC) within the target building, triggering a pre-execution linkage strategy for the hardware devices in the corresponding protection zone. This pre-execution linkage strategy includes the main control chip sending a high-level closed signal to the start relay of the smoke extraction system in the corresponding protection zone to preheat the system, or sending a trip pulse signal to the control circuit breaker of the non-fire protection power supply in the corresponding protection zone to physically disconnect the power supply.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic detection method for building fire hazards, characterized in that, Includes the following steps: S100: Acquire multi-source heterogeneous sensing data and static building information model (BIM) data of the target building, and after spatial alignment of the multi-source heterogeneous sensing data, map it into the three-dimensional spatial coordinate system of the static BIM data. S200, perform cross-modal feature extraction on the aligned multi-source heterogeneous sensing data, and identify dynamic spatial obstacles, fire risk entities and their corresponding state attributes in the target building; S300, based on the identified dynamic spatial obstacles, fire risk entities, state attributes, and spatial elements in the static BIM data, construct and dynamically update a three-dimensional spatial fire knowledge graph, wherein the node features of the three-dimensional spatial fire knowledge graph include entity states, and the edge features represent the spatial topology and physical relationship between entities. S400, the dynamically updated three-dimensional spatial fire protection knowledge graph is input into the pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction, the comprehensive fire hazard index of each spatial node is calculated, and the hazard classification early warning information is output based on the comprehensive fire hazard index.

2. The automatic detection method for building fire hazards according to claim 1, characterized in that, The multi-source heterogeneous sensing data includes visible light image streams, infrared thermal imaging video streams, and IoT environmental sensing data. Acquiring multi-source heterogeneous sensing data of the target building, specifically including: The visible light image stream and the infrared thermal imaging video stream are simultaneously acquired from the same field of view using global monitoring equipment and thermal imaging equipment deployed within the target building. The IoT environmental sensing data is acquired through an environmental sensing network. The IoT environmental sensing data includes at least the internal load status data of electrical equipment and the environmental status data of the smoke exhaust duct in the catering area.

3. The automatic detection method for building fire hazards according to claim 2, characterized in that, After spatial alignment of the multi-source heterogeneous sensing data, it is mapped to the three-dimensional spatial coordinate system of the static BIM data, including the following processing steps: Based on the pre-calibrated camera intrinsic and extrinsic parameter matrices, a transformation mapping function is constructed between the two-dimensional pixel coordinate system, the physical world coordinate system, and the three-dimensional spatial coordinate system of the static BIM data; Spatial homonym feature points are extracted from the synchronously acquired visible light image stream and infrared thermal imaging video stream. The affine transformation matrix is ​​calculated using the spatial homonym feature points to achieve pixel-level spatial registration of heterogeneous visual data. The transformation mapping function is invoked to project the registered heterogeneous visual data and the IoT environmental sensing data with spatial labels onto the corresponding nodes in the three-dimensional spatial coordinate system, thereby generating a multimodal spatial data base.

4. The automatic detection method for building fire hazards according to claim 3, characterized in that, S100 also includes an alignment compensation mechanism: Real-time monitoring of the gimbal pose parameters of the global monitoring device and the thermal imaging device; When a shift in the gimbal pose parameters is detected, the static environmental reference features under the current field of view are extracted, and the extrinsic parameter matrix and the transformation mapping function are adaptively updated in combination with the offset to maintain the real-time consistency between the dynamically updated multimodal spatial data base and the static BIM data in the three-dimensional spatial coordinate system.

5. The automatic detection method for building fire hazards according to claim 4, characterized in that, The identification of dynamic spatial obstacles within the target building includes a dynamic spatial reconstruction mechanism for temporary structures, with the following specific steps: The instance segmentation network is used to parse the visual features mapped to the three-dimensional spatial coordinate system in real time, extract semantic labels with non-fixed structure, and generate corresponding three-dimensional bounding boxes. Perform spatial Boolean operations on the three-dimensional bounding box and the preset fire evacuation route model and fireproof roller shutter forced landing area model in the static BIM data; If the calculation results show a spatial intersection, the non-fixed structure that generates the spatial intersection is identified as the dynamic spatial obstacle, and its occupancy area and spatial volume are calculated as the corresponding state attributes.

6. The automatic detection method for building fire hazards according to claim 5, characterized in that, The aligned multi-source heterogeneous sensing data is subjected to cross-modal feature extraction to identify fire risk entities, and a feature fusion verification strategy based on a dual-stream convolutional neural network is adopted. The visible light branch of the dual-stream convolutional neural network is used to extract flammable objects and fire-fighting facilities in the scene, and the infrared branch is used to extract the surface temperature gradient distribution. The pixel region of the flammable object is fused with the abnormally heated region in the surface temperature gradient distribution using region of interest (ROI) pooling. When the temperature rise slope of the abnormal heating area exceeds a preset threshold, and the Euclidean distance between its spatial enclosure and the spatial enclosure of the flammable material entity is less than a safety threshold, the flammable material entity and the adjacent heat source are jointly marked as the fire risk entity of the high-risk level. The acquisition of its corresponding state attributes adopts cross-modal data coupling extraction rules, specifically as follows: Visual features of oil stain thickness on the surface of the smoke exhaust pipe in the catering defense zone were extracted based on the image gray-level co-occurrence matrix. The visual characteristics of the oil stain thickness, the abnormal temperature fluctuation characteristics inside the pipeline obtained by the infrared thermal imaging video stream, and the gas pipeline pressure characteristics in the IoT environmental sensing data are vectorized and stitched together. Based on the spliced ​​feature vectors, the fire risk entities within the catering protection zone are assigned multi-dimensional fire coupling state attributes to characterize the potential fire risk level caused by oil accumulation and high-temperature smoke exhaust.

7. The automatic detection method for building fire hazards according to claim 6, characterized in that, The process involves constructing a three-dimensional spatial fire safety knowledge graph based on the identified dynamic spatial obstacles, fire risk entities, state attributes, and spatial elements in the static BIM data. This is achieved using a hierarchical graph construction strategy based on ontology and instances, specifically including: The fire compartments, evacuation routes, and fixed fire protection facilities in the static BIM data are abstracted into static body nodes of the graph; The identified dynamic spatial obstacles and fire risk entities are abstracted into dynamic instance nodes of a graph, and the corresponding state attributes are assigned as feature vectors to the corresponding nodes. The relative spatial distribution between each node is calculated based on the three-dimensional spatial coordinate system, and the edge features connecting the nodes are constructed. The edge features include distance edges, occlusion edges, and containment edges. The occlusion edges are generated based on the projection overlap between the three-dimensional bounding box of the dynamic instance node and the static body node from the evacuation viewpoint. The edge features also include dynamic physical association weights that characterize the trend of danger spread, and the calculation mechanism of the dynamic physical association weights is as follows: Extract the surface temperature gradient value associated with the fire risk entity from the state attributes; Combining the relative spatial distance corresponding to the distance edge, and the current ventilation status data of the protection zone obtained from the IoT environmental sensing data, the physical probability of the fire risk entity transmitting heat energy or harmful gas to its adjacent nodes is calculated using a preset heat conduction attenuation model. After normalizing the physical probabilities, they are assigned as dynamic physical association weights to the edges connecting the corresponding nodes to quantitatively characterize the spatial coupling hazard between entities during the fire incubation stage.

8. The automatic detection method for building fire hazards according to claim 7, characterized in that, The dynamically updated 3D spatial fire safety knowledge graph employs a subgraph update strategy triggered by local spatial topology events. When real-time sensing data detects that the spatial position of the dynamic space obstacle has shifted, or the state attribute of the fire risk entity has changed abruptly, and the amount of displacement or the magnitude of the change exceeds a preset stability threshold, a local space anomaly event is generated. In response to the local spatial anomaly event, only the target node where the anomaly occurred is locked, and the edge features and node features in the first-order and second-order connected domains of the target node are reconstructed and calculated, while the topology of the remaining unaffected subgraphs in the three-dimensional spatial fire protection knowledge graph remains unchanged.

9. The automatic detection method for building fire hazards according to claim 8, characterized in that, The dynamically updated 3D spatial fire protection knowledge graph is input into a pre-trained spatiotemporal graph convolutional network model for evolutionary feature extraction. This process employs a spatiotemporal joint graph convolution strategy incorporating self-attention mechanisms and includes the following steps: By aggregating the node features and edge features of the target node and its neighboring nodes through the spatial graph convolutional layer in the spatiotemporal graph convolutional network model, spatial topological features that characterize the spread trend of fire risk in three-dimensional space are extracted. The spatial topological features are extracted from the temporal convolutional layer in the spatiotemporal graph convolutional network model to capture the dynamic mutation rate of the state attributes. A self-attention mechanism module is introduced to assign higher attention weights to high-risk nodes in the spatial topology and temporal features, and output the physical fire incubation probability index of each target node in the future preset time window.

10. The automatic detection method for building fire hazards according to claim 9, characterized in that, The calculation of the comprehensive fire hazard index for each spatial node adopts a risk quantification evaluation model with multi-dimensional parameter aggregation, and the calculation formula is as follows: in, The comprehensive fire hazard index; This refers to the physical fire spawning probability index; The first obtained through visual passenger flow statistics Real-time population density within each defense zone; The first in the static BIM data The initial design of the total width of the escape routes in each defense zone; The width of the current passage being illegally obstructed by the dynamic spatial obstacle, calculated from the three-dimensional spatial fire protection knowledge graph; As a gathering point for people Spatial topological distance to predicted high-risk target nodes; , , Weighting coefficients preset for the target building structure; The output of hazard classification and early warning information based on the comprehensive fire hazard index adopts a virtual-real mapping linkage response mechanism based on the BIM digital twin platform: The calculated comprehensive fire hazard index is compared with the preset dynamic grading threshold to determine the hazard level; The three-dimensional spatial coordinates of the target nodes that exceed the comprehensive fire hazard index are projected in reverse to the BIM digital twin visualization platform, and the corresponding physical defense zone and hazard spread radiation circle are highlighted and rendered in the virtual twin model. Based on the determined hazard level, a hazard rectification order is automatically generated and sent to the mobile terminal. When the hazard level is determined to be the highest, the pre-execution linkage strategy of the hardware equipment in the corresponding defense zone is triggered. The pre-execution linkage strategy includes the early activation of the smoke exhaust system preheating or non-fire protection power cut-off action in the defense zone.

Images