Intelligent fire-fighting scene risk assessment system based on AI

Abstract

The invention relates to the technical field of fire safety, and discloses an AI-based intelligent fire scene risk assessment system, which comprises a data acquisition module used for acquiring multi-source heterogeneous data in a fire scene; the knowledge graph construction module is used for constructing and updating a dynamic heterogeneous knowledge graph in real time based on the multi-source heterogeneous data; the cross-modal fusion processing module is used for carrying out feature extraction and fusion on the multi-source heterogeneous data by adopting a BioNet algorithm, distributing dynamic weights to different modal data by adopting a multi-head attention mechanism, and generating a fusion feature vector; the space-time risk prediction module is used for predicting the risk evolution trend of the fire-fighting scene in a future preset time period by adopting a space-time Transform model; the dynamic allocation module is used for scheduling fire-fighting resources by adopting a DBO algorithm based on a result output by the space-time Transform model, and dynamically allocating fire extinguishing equipment, emergency channels and rescue strategies; according to the invention, the accuracy and timeliness of risk assessment are improved.

Description

Technical Field

[0001] This invention relates to the field of fire safety technology, specifically to an AI-based intelligent fire scenario risk assessment system. Background Technology

[0002] Fire safety is a core component of the social public safety system. Traditional fire risk assessment methods have long relied on static means such as manual inspections and fixed threshold alarms, which have significant technical limitations. On the one hand, manual inspections are limited by labor costs and time intervals, making it difficult to achieve real-time coverage of large and complex scenarios, and the assessment results are easily affected by subjective experience. On the other hand, single sensor threshold alarms can only capture local parameter anomalies and cannot be linked to multi-dimensional information such as equipment status and personnel distribution, resulting in a high false alarm rate and a high risk of missed alarms. With the development of IoT and AI technologies, some intelligent fire protection systems have begun to integrate multi-source data, but key technical bottlenecks remain. First, data fusion remains at a simple stitching level, failing to consider the heterogeneity of different modalities such as visual images, infrared thermal imaging, and environmental sensors, making it difficult to form a unified risk characteristic representation. Second, scenario modeling is static; most systems use fixed equipment topology relationships, failing to update dynamic factors such as personnel flow and temporary item stacking in real time, leading to a disconnect between risk assessment and actual scenarios. Third, risk prediction lacks spatiotemporal correlation analysis, only outputting the current risk level and failing to predict the risk's spread path and evolution trend over a future period, limiting the initiative of emergency response. Fourth, resource scheduling relies on preset plans, failing to dynamically adjust in conjunction with real-time risks, making it difficult to achieve optimal configuration of fire-fighting equipment and evacuation routes. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing an AI-based intelligent fire scene risk assessment system.

[0004] The first aspect of this invention provides an AI-based intelligent fire scene risk assessment system, the system comprising: The data acquisition module is used to collect multi-source heterogeneous data in fire-fighting scenarios. The multi-source heterogeneous data includes at least visual image data, infrared thermal imaging data, environmental sensor data, equipment status data, and personnel trajectory data. The knowledge graph construction module is used to build and update a dynamic heterogeneous knowledge graph containing equipment entities, environmental entities, personnel entities and fire protection code knowledge based on multi-source heterogeneous data. The dynamic heterogeneous knowledge graph is represented and updated using a graph neural network. The cross-modal fusion processing module is used to extract and fuse features from multi-source heterogeneous data using the BioNet algorithm, and to assign dynamic weights to different modal data using a multi-head attention mechanism to generate fused feature vectors. The spatiotemporal risk prediction module is used to predict the risk evolution trend of fire scenarios within a preset time period based on heterogeneous knowledge graphs and fused feature vectors, using a spatiotemporal Transformer model. The dynamic allocation module is used to perform fire resource scheduling based on the results output by the spatiotemporal Transformer model using the DBO algorithm, and to dynamically allocate fire extinguishing equipment, emergency access routes and rescue strategies.

[0005] Optionally, in a first implementation of the present invention, the knowledge graph construction module includes: The identification submodule is used to identify equipment entities, environmental entities, and personnel entities from multi-source heterogeneous data, and extract entity attributes; The first calculation submodule is used to determine the relationships between entities through semantic analysis and spatial calculation. The relationships between entities include at least the connection relationship between devices, the location relationship between personnel and areas, and the influence relationship between the environment and devices. A submodule is formed to use the identified entities as nodes and the extracted relationships as edges to form an initial heterogeneous knowledge graph containing fire protection code knowledge. The allocation submodule is used to assign initial feature vectors to the nodes and edges of the initial heterogeneous knowledge graph and initialize the graph neural network; The iterative update submodule is used to iteratively update node features and edge weights through a graph neural network, and output the updated dynamic heterogeneous knowledge graph.

[0006] Optionally, in a second implementation of the present invention, the iterative update submodule includes: For each node, the message passing layer of the graph neural network collects the feature information of its neighboring nodes, and the received feature information of all neighbors is fused with the node's own current features to form an intermediate feature vector. The intermediate feature vector is input into the feature update layer of the graph neural network, and a non-linear transformation is performed through the ReLU function to generate new feature vectors for the nodes. The new feature vector based on the node is used to recalculate the weights of the edges to be updated through the edge weight update layer of the graph neural network. The iteration ends when the preset number of iterations is reached, and the new feature vectors of all nodes and the new weight matrices of the edges are integrated to form an updated dynamic heterogeneous knowledge graph.

[0007] Optionally, in a third implementation of the present invention, the cross-modal fusion processing module includes: The extraction submodule is used to input multi-source heterogeneous data into different subnetworks of the BioNet algorithm to extract features; The capture submodule is used to input the modal feature vectors output by the BioNet algorithm into the multi-head attention mechanism. Each attention head calculates the similarity of different modal features through matrix operations in order to capture the correlation between different modal features. The splicing submodule is used to assign dynamic weights to each modality feature based on similarity. It then uses a multi-head attention mechanism to splice and linearly transform the weighted modality feature vectors to generate a fused feature vector.

[0008] Optionally, in the fourth implementation of the present invention, the sub-networks of the BioNet algorithm include at least a convolutional sub-network, a temperature feature sub-network, a temporal sub-network, an attribute sub-network, and a motion sub-network.

[0009] Optionally, in the fifth implementation of the present invention, visual image data is input into a convolutional subnetwork, and texture and contour features are extracted by simulating biological visual perception mechanisms; Infrared thermal imaging data is input into a temperature feature subnetwork to extract temperature gradients and high-temperature region features. Environmental sensor data is input into a time-series subnetwork to extract the changing trend characteristics of temperature, humidity, and gas concentration. The device status data is input into the attribute subnetwork to extract operating parameter features; Personnel trajectory data is input into the motion subnetwork to extract features such as movement direction and dwell time.

[0010] Optionally, in a sixth implementation of the present invention, the spatiotemporal risk prediction module includes: A submodule is constructed to build a spatiotemporal feature matrix based on the spatial features of a dynamic heterogeneous knowledge graph and the temporal features of a fused feature vector. The encoding submodule is used to add timestamp encoding and spatial coordinate encoding to the spatiotemporal feature matrix to obtain the encoded spatiotemporal features. The second computational submodule is used to input the encoded spatiotemporal features into the encoder layer of the spatiotemporal Transformer, and calculate the association weights of features at different time steps and different spatial locations through multiple attention heads to capture long-distance spatiotemporal dependencies. The output submodule is used to perform nonlinear transformation and aggregation on the features after attention calculation through the feedforward neural network of the encoder layer. The aggregated features are then input into the decoder layer of the Transformer. The decoder outputs the risk features at each time point based on a preset time step. The features are then converted into the risk level and risk diffusion path of each region within the corresponding time period by the classifier, so as to output the risk evolution trend.

[0011] Optionally, in a seventh implementation of the present invention, the dynamic allocation module includes: The initialization submodule is used to collect information on currently available fire-fighting resources, determine constraints and optimization objectives, and initialize the DBO algorithm population. The mutation and crossover submodule is used to calculate the fitness value of each allocation scheme according to the optimization objective, update the population through the search mechanism of the DBO algorithm, retain the scheme with high fitness as the high-quality solution, and perform mutation and crossover on the remaining schemes to generate new allocation schemes. The convergence submodule is used to iterate repeatedly until the fitness value converges, select the allocation scheme with the highest fitness as the optimal solution, and perform fire resource scheduling based on the optimal solution, dynamically allocating fire extinguishing equipment, emergency access routes and rescue strategies.

[0012] Optionally, in the eighth implementation of the present invention, the fire-fighting resource information includes the quantity and location of fire-fighting equipment, the unobstructed status of emergency passages, and the configuration of rescue personnel; the constraints include the maximum movement range of equipment and the carrying capacity of passages; and the optimization objectives include the fastest response time, maximizing resource utilization, and maximizing rescue coverage.

[0013] Optionally, in a ninth implementation of the present invention, the method includes the following steps: Collect multi-source heterogeneous data in fire-fighting scenarios, including at least visual image data, infrared thermal imaging data, environmental sensor data, equipment status data, and personnel trajectory data. A dynamic heterogeneous knowledge graph containing equipment entities, environmental entities, personnel entities, and fire protection code knowledge is constructed and updated in real time based on multi-source heterogeneous data. The dynamic heterogeneous knowledge graph is represented and updated using a graph neural network. The BioNet algorithm is used to extract and fuse features from multi-source heterogeneous data, and a multi-head attention mechanism is used to assign dynamic weights to data of different modalities to generate fused feature vectors. Based on heterogeneous knowledge graphs and fused feature vectors, a spatiotemporal Transformer model is used to predict the risk evolution trend of fire scenarios within a preset time period in the future. The DBO algorithm is used to schedule fire-fighting resources based on the output of the spatiotemporal Transformer model, dynamically allocating fire-fighting equipment, emergency access routes, and rescue strategies.

[0014] The technical solution provided by this invention involves collecting multi-source heterogeneous data from fire scenarios, constructing and updating a dynamic heterogeneous knowledge graph containing equipment entities, environmental entities, personnel entities, and fire safety regulations based on this data, employing the BioNet algorithm for feature extraction and fusion of the multi-source heterogeneous data, and using a multi-head attention mechanism to assign dynamic weights to different modalities to generate fused feature vectors. Based on the heterogeneous knowledge graph and the fused feature vectors, a spatiotemporal Transformer model is used to predict the risk evolution trend of the fire scenario within a preset time period. Based on the output of the spatiotemporal Transformer model, the DBO algorithm is used for fire resource scheduling, dynamically allocating fire extinguishing equipment, emergency exits, and rescue strategies. This invention constructs a dynamic... The heterogeneous knowledge graph integrates knowledge of various entities such as equipment, personnel, and environment, as well as fire protection regulations. It can reflect scene changes in real time, overcoming the shortcomings of static scene modeling in existing technologies. Compared with traditional real-time alarm modes, the spatiotemporal Transformer model prediction provides managers with more decision-making time, allowing for targeted adjustments to prevention and control strategies, avoiding passive responses, and significantly improving the effectiveness of emergency response. Compared with traditional fixed-scheme scheduling, the DBO algorithm for fire resource scheduling significantly shortens response time, improves resource utilization, and ensures precise intervention at critical nodes in risk evolution. It significantly improves the accuracy and timeliness of risk assessment, reduces false alarm rates, enhances adaptability to complex dynamic scenarios, and provides intelligent decision support for fire safety management. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 A schematic diagram of the structure of an AI-based intelligent fire scene risk assessment system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the knowledge graph construction module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the cross-modal fusion processing module provided in an embodiment of the present invention. Detailed Implementation

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

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of the structure of an AI-based intelligent fire scene risk assessment system provided in this embodiment of the invention. The system includes: The data acquisition module is used to collect multi-source heterogeneous data in fire-fighting scenarios. The multi-source heterogeneous data includes at least visual image data, infrared thermal imaging data, environmental sensor data, equipment status data, and personnel trajectory data. The knowledge graph construction module is used to build and update a dynamic heterogeneous knowledge graph containing equipment entities, environmental entities, personnel entities and fire protection code knowledge based on multi-source heterogeneous data. The dynamic heterogeneous knowledge graph is represented and updated using a graph neural network. The cross-modal fusion processing module is used to extract and fuse features from multi-source heterogeneous data using the BioNet algorithm, and to assign dynamic weights to different modal data using a multi-head attention mechanism to generate fused feature vectors. The spatiotemporal risk prediction module is used to predict the risk evolution trend of fire scenarios within a preset time period based on heterogeneous knowledge graphs and fused feature vectors, using a spatiotemporal Transformer model. The dynamic allocation module is used to perform fire resource scheduling based on the results output by the spatiotemporal Transformer model using the DBO algorithm, and to dynamically allocate fire extinguishing equipment, emergency access routes and rescue strategies.

[0019] In this embodiment, please refer to Figure 2 The knowledge graph construction module includes: The identification submodule is used to identify equipment entities, environmental entities, and personnel entities from multi-source heterogeneous data, and extract entity attributes; The first calculation submodule is used to determine the relationships between entities through semantic analysis and spatial calculation. The relationships between entities include at least the connection relationship between devices, the location relationship between personnel and areas, and the influence relationship between the environment and devices. A submodule is formed to use the identified entities as nodes and the extracted relationships as edges to form an initial heterogeneous knowledge graph containing fire protection code knowledge. The allocation submodule is used to assign initial feature vectors to the nodes and edges of the initial heterogeneous knowledge graph and initialize the graph neural network; The iterative update submodule is used to iteratively update node features and edge weights through a graph neural network, and output the updated dynamic heterogeneous knowledge graph.

[0020] In this embodiment, the iterative update submodule includes: For each node, the message passing layer of the graph neural network collects the feature information of its neighboring nodes, and the received feature information of all neighbors is fused with the node's own current features to form an intermediate feature vector. The intermediate feature vector is input into the feature update layer of the graph neural network, and a non-linear transformation is performed through the ReLU function to generate new feature vectors for the nodes. The new feature vector based on the node is used to recalculate the weights of the edges to be updated through the edge weight update layer of the graph neural network. The iteration ends when the preset number of iterations is reached, and the new feature vectors of all nodes and the new weight matrices of the edges are integrated to form an updated dynamic heterogeneous knowledge graph.

[0021] In this embodiment, please refer to Figure 3 The cross-modal fusion processing module includes: The extraction submodule is used to input multi-source heterogeneous data into different subnetworks of the BioNet algorithm to extract features; The capture submodule is used to input the modal feature vectors output by the BioNet algorithm into the multi-head attention mechanism. Each attention head calculates the similarity of different modal features through matrix operations in order to capture the correlation between different modal features. The splicing submodule is used to assign dynamic weights to each modality feature based on similarity. It then uses a multi-head attention mechanism to splice and linearly transform the weighted modality feature vectors to generate a fused feature vector.

[0022] In this embodiment, the subnetworks of the BioNet algorithm include at least a convolutional subnetwork, a temperature feature subnetwork, a temporal subnetwork, an attribute subnetwork, and a motion subnetwork.

[0023] In this embodiment, after visual image data is input into the convolutional subnetwork, the subnetwork simulates the hierarchical processing mechanism from retinal photoreceptor cells to the visual cortex in the biological visual system. First, the bottom convolutional kernels capture basic visual signals such as edges and brightness changes in the image, just as human vision first perceives the outline of an object. Then, the middle convolutional layers combine these basic signals to gradually extract the texture patterns of the object's surface, such as the material texture of a wall or the surface texture of an object. Finally, the high-level convolutional layers integrate the texture and edge information to form complete object outline features, such as the shape of a table and chair or the outline of a door or window, thus achieving accurate capture of the morphological features of various objects in the scene. After the infrared thermal imaging data is input into the temperature feature subnetwork, the subnetwork first divides the thermal imaging image into regions to simulate the biological sensitivity to temperature differences, focusing on regions with abrupt temperature changes. Then, it calculates the temperature difference between adjacent regions through a sliding window to extract temperature gradient features such as the rate and direction of temperature change from high-temperature to low-temperature regions. At the same time, based on a preset temperature threshold, regions with temperatures higher than the environmental baseline are selected. Combining the area, shape, and temperature distribution uniformity of the regions, the core features of high-temperature regions are extracted, such as the location, size, and peak temperature of high-temperature points, thereby accurately depicting the abnormal temperature distribution in the scene. After the environmental sensor data is input into the time-series subnetwork, the subnetwork adopts a continuous sensing mode similar to that of organisms to environmental changes. First, the continuously collected temperature, humidity, and gas concentration data are sorted according to time series to form a complete dynamic change curve. Then, the mean, variance, and rate of change in different time periods are calculated through a sliding time window to capture short-term fluctuation characteristics, such as a sudden rise or fall in humidity within a 5-minute period. Then, through long-term trend analysis, the periodic change patterns of the data are extracted, such as the daily rise and fall trend of temperature and long-term drift characteristics, such as the slow accumulation of gas concentration over several hours, to comprehensively reflect the dynamic evolution process of environmental parameters. After the equipment status data is input into the attribute subnetwork, the subnetwork first classifies and sorts out various parameters of the equipment, such as the pressure value of the fire extinguisher, the sensitivity of the alarm, and the operating power of the water pump, based on the operating characteristics of the fire-fighting equipment, and establishes attribute dimensions corresponding to the equipment functions; then, through feature mapping, the various parameters are converted into standardized feature values ​​to eliminate the dimensional differences between different parameters; then, through attribute association analysis, the cooperative change features between parameters, such as the matching relationship between water pump power and pipeline pressure, and abnormal deviation features, such as the extent to which the operating current of a certain piece of equipment exceeds the normal range, are extracted to accurately reflect the operating status and health of the equipment. After the personnel trajectory data is input into the motion subnetwork, the subnetwork simulates the perception logic of organisms regarding their own and others' movement states. First, based on the timestamp and coordinate information of the positioning points, it calculates the movement speed and direction vectors of personnel at different times, capturing instantaneous movement characteristics such as sudden turns, acceleration, or deceleration. Then, through fitting analysis of continuous trajectory points, it extracts the overall trend of the movement path, such as straight-line movement or detours. At the same time, it statistically analyzes the characteristics such as the duration of personnel's stay in each area and the frequency of entry and exit, and combines these with the functional attributes of the areas, such as dangerous areas and safety passages, to form feature vectors that reflect personnel behavior patterns, comprehensively depicting the movement state and activity patterns of personnel in the scene.

[0024] In this embodiment, the spatiotemporal risk prediction module includes: A submodule is constructed to build a spatiotemporal feature matrix based on the spatial features of a dynamic heterogeneous knowledge graph and the temporal features of a fused feature vector. The encoding submodule is used to add timestamp encoding and spatial coordinate encoding to the spatiotemporal feature matrix to obtain the encoded spatiotemporal features. The second computational submodule is used to input the encoded spatiotemporal features into the encoder layer of the spatiotemporal Transformer, and calculate the association weights of features at different time steps and different spatial locations through multiple attention heads to capture long-distance spatiotemporal dependencies. The output submodule is used to perform nonlinear transformation and aggregation on the features after attention calculation through the feedforward neural network of the encoder layer. The aggregated features are then input into the decoder layer of the Transformer. The decoder outputs the risk features at each time point based on a preset time step. The features are then converted into the risk level and risk diffusion path of each region within the corresponding time period by the classifier, so as to output the risk evolution trend.

[0025] In this embodiment, spatial features are extracted from the dynamic heterogeneous knowledge graph, including the location coordinates of entities such as equipment, personnel, and regions, the topological relationships between entities such as the relationship between equipment and regions, the distance relationship between personnel and equipment, and the spatial attributes of entities such as region area and equipment distribution density. These features are then converted into structured spatial feature vectors. Simultaneously, temporal features are extracted from the fused feature vectors, covering the changes in various modal data over time, such as temperature rise and fall curves and the temporal trajectory of personnel movement, forming temporal feature vectors. Subsequently, the submodule uses a feature alignment mechanism to associate and match spatial feature vectors and temporal feature vectors at the same time and spatial location, ultimately combining them into a feature matrix containing information in both spatiotemporal dimensions, providing a foundation for subsequent encoding and analysis. For timestamp encoding, the submodule uses sine and cosine functions to generate periodic encoding vectors of different frequencies, converting the specific time of data collection, such as hours, minutes, and seconds, into corresponding encoding values. This enables the model to identify the chronological order and periodic patterns of time series, such as the differences in risk characteristics across different time periods. For spatial coordinate encoding, the submodule generates location-discriminative encoding vectors based on the two-dimensional or three-dimensional spatial coordinates of entities. The differences in encoding values ​​reflect the spatial distance and orientation between entities; for example, entities within the same region have higher encoding similarity than entities in different regions. Finally, the submodule superimposes and fuses the timestamp encoding and spatial coordinate encoding with the corresponding features in the spatiotemporal feature matrix to obtain encoded spatiotemporal features that contain both the original feature information and clear spatiotemporal location markers. After the encoded spatiotemporal features are input into the encoder layer, multiple attention heads in the encoder layer perform parallel computation simultaneously. Each attention head autonomously focuses on different time steps and spatial locations in the feature matrix, generating corresponding association weights by calculating the similarity between features, such as the feature correlation degree of the same region at different time steps and the feature influence degree of different regions at the same time step. Specifically, for the time dimension, the attention head focuses on analyzing the continuity and correlation of features at different times, such as the correlation between the high temperature feature of the previous time step and the smoke feature of the current time step. For the spatial dimension, the attention head focuses on the mutual influence of features at different spatial locations, such as the potential impact of the fire feature of region A on the adjacent region B. Through this multi-dimensional association weight calculation, the submodule can effectively capture feature dependencies spanning long time intervals and long spatial distances, providing deep correlation evidence for the prediction of risk evolution trends. The feedforward neural network in the encoder layer performs a non-linear transformation on the features after attention calculation, enhancing the non-linear expressive power of the features through activation functions such as ReLU. Simultaneously, it aggregates the transformed features to extract the most critical core features for risk prediction, such as features of high-risk areas and abrupt changes at key time points. Subsequently, the aggregated features are input into the decoder layer of the Transformer. The decoder generates sequences based on preset time steps, such as the next 5 minutes, 15 minutes, and 30 minutes, combining existing feature information to output a risk feature vector corresponding to each time point. Finally, the classifier converts these risk feature vectors into specific risk levels, such as low, medium, and high risk. By analyzing the spatial distribution changes in the feature vectors, it predicts the spread path of risk in different regions, such as the direction and range of spread from the initial high-temperature point to the surrounding areas, ultimately forming a complete risk evolution trend output.

[0026] In this embodiment, the dynamic allocation module includes: The initialization submodule is used to collect information on currently available fire-fighting resources, determine constraints and optimization objectives, and initialize the DBO algorithm population. The mutation and crossover submodule is used to calculate the fitness value of each allocation scheme according to the optimization objective, update the population through the search mechanism of the DBO algorithm, retain the scheme with high fitness as the high-quality solution, and perform mutation and crossover on the remaining schemes to generate new allocation schemes. The convergence submodule is used to iterate repeatedly until the fitness value converges, select the allocation scheme with the highest fitness as the optimal solution, and perform fire resource scheduling based on the optimal solution, dynamically allocating fire extinguishing equipment, emergency access routes and rescue strategies.

[0027] In this embodiment, fire resource information includes the quantity and location of fire extinguishing equipment, the unobstructed status of emergency access routes, and the configuration of rescue personnel. Constraints include the maximum movement range of equipment and the carrying capacity of access routes. Optimization objectives include the fastest response time, maximizing resource utilization, and maximizing rescue coverage.

[0028] In this embodiment, the DBO algorithm population is initialized, and each possible fire resource allocation scheme, such as allocating a certain type of fire extinguisher to area A or prioritizing the use of an emergency passage for personnel evacuation in area B, is converted into an algorithm-recognizable encoding form. Multiple different allocation schemes are randomly generated as the initial population to provide a starting point for subsequent optimization iterations. Based on the optimization objective determined by the initialization submodule, the fitness value is calculated for each allocation scheme in the population. For example, if a scheme enables fire-fighting equipment to reach all high-risk areas in the shortest time and maximizes resource utilization, its fitness value is high; conversely, if a scheme has problems such as resource waste or response delay, its fitness value is low. Subsequently, the submodule retains the schemes with the highest fitness ranking in the population as high-quality solutions, as these schemes already possess better resource allocation logic. For the remaining schemes with lower fitness, the submodule updates them according to the search rules of the DBO algorithm: during mutation, a portion of the code in the scheme is randomly selected for adjustment, such as changing the allocation location of a fire-fighting device from area A to the closer area C, or adjusting the opening priority of an emergency passage; during crossover, two different schemes are selected as parents, and their partial code fragments are exchanged, such as combining the part about fire truck dispatch in scheme 1 with the part about emergency passage planning in scheme 2. In this way, a new allocation scheme is generated, which retains the effective information of the parent schemes while introducing new possibilities, ensuring that the population always maintains diversity and avoids getting trapped in local optima. The new solutions generated after mutation and crossover are incorporated into the population, and the process of fitness calculation, retention of high-quality solutions, mutation, and crossover is repeated to continuously optimize the solution iteratively. In each iteration, the submodule compares the fitness value of the current best solution in the population with the best fitness value of the previous iteration, and determines whether the difference between the two is less than a preset threshold. If the difference continues to shrink and stabilizes within the threshold range, it indicates that the fitness value has converged, and the solutions in the population are close to the optimal solution. If it has not converged, the iteration continues. When the fitness value converges, the submodule selects the allocation scheme with the highest fitness in the population as the optimal solution. This scheme can best achieve the preset optimization goal under the premise of satisfying all constraints. Finally, specific dispatch instructions are generated based on the optimal solution, including specifying which fire-fighting equipment should be prioritized for dispatch to which high-risk areas, which emergency passages should be opened or closed to ensure evacuation efficiency, and what rescue strategies should be adopted for different areas, such as prioritizing the evacuation of personnel or controlling the spread of fire, to achieve dynamic and precise allocation of fire-fighting resources.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent fire scene risk assessment system, characterized in that, The system includes: The data acquisition module is used to collect multi-source heterogeneous data in fire-fighting scenarios. The multi-source heterogeneous data includes at least visual image data, infrared thermal imaging data, environmental sensor data, equipment status data, and personnel trajectory data. The knowledge graph construction module is used to build and update a dynamic heterogeneous knowledge graph containing equipment entities, environmental entities, personnel entities and fire protection code knowledge based on multi-source heterogeneous data. The dynamic heterogeneous knowledge graph is represented and updated using a graph neural network. The cross-modal fusion processing module is used to extract and fuse features from multi-source heterogeneous data using the BioNet algorithm, and to assign dynamic weights to different modal data using a multi-head attention mechanism to generate fused feature vectors. The spatiotemporal risk prediction module is used to predict the risk evolution trend of fire scenarios within a preset time period based on heterogeneous knowledge graphs and fused feature vectors, using a spatiotemporal Transformer model. The dynamic allocation module is used to perform fire resource scheduling based on the results output by the spatiotemporal Transformer model using the DBO algorithm, and to dynamically allocate fire extinguishing equipment, emergency access routes and rescue strategies.

2. The AI-based intelligent fire scene risk assessment system as described in claim 1, characterized in that, The knowledge graph construction module includes: The identification submodule is used to identify equipment entities, environmental entities, and personnel entities from multi-source heterogeneous data, and extract entity attributes; The first calculation submodule is used to determine the relationships between entities through semantic analysis and spatial calculation. The relationships between entities include at least the connection relationship between devices, the location relationship between personnel and areas, and the influence relationship between the environment and devices. A submodule is formed to use the identified entities as nodes and the extracted relationships as edges to form an initial heterogeneous knowledge graph containing fire protection code knowledge. The allocation submodule is used to assign initial feature vectors to the nodes and edges of the initial heterogeneous knowledge graph and initialize the graph neural network; The iterative update submodule is used to iteratively update node features and edge weights through a graph neural network, and output the updated dynamic heterogeneous knowledge graph.

3. The AI-based intelligent fire scene risk assessment system as described in claim 2, characterized in that, The iterative update submodule includes: For each node, the message passing layer of the graph neural network collects the feature information of its neighboring nodes, and the received feature information of all neighbors is fused with the node's own current features to form an intermediate feature vector. The intermediate feature vector is input into the feature update layer of the graph neural network, and a non-linear transformation is performed through the ReLU function to generate new feature vectors for the nodes. The new feature vector based on the node is used to recalculate the weights of the edges to be updated through the edge weight update layer of the graph neural network. The iteration ends when the preset number of iterations is reached, and the new feature vectors of all nodes and the new weight matrices of the edges are integrated to form an updated dynamic heterogeneous knowledge graph.

4. The AI-based intelligent fire scene risk assessment system as described in claim 1, characterized in that, The cross-modal fusion processing module includes: The extraction submodule is used to input multi-source heterogeneous data into different subnetworks of the BioNet algorithm to extract features; The capture submodule is used to input the modal feature vectors output by the BioNet algorithm into the multi-head attention mechanism. Each attention head calculates the similarity of different modal features through matrix operations in order to capture the correlation between different modal features. The splicing submodule is used to assign dynamic weights to each modality feature based on similarity. It then uses a multi-head attention mechanism to splice and linearly transform the weighted modality feature vectors to generate a fused feature vector.

5. The AI-based intelligent fire scene risk assessment system as described in claim 4, characterized in that, The subnetworks of the BioNet algorithm include at least a convolutional subnetwork, a temperature feature subnetwork, a temporal subnetwork, an attribute subnetwork, and a motion subnetwork.

6. The AI-based intelligent fire scene risk assessment system as described in claim 5, characterized in that, Visual image data is input into a convolutional subnetwork, which extracts texture and contour features by simulating biological visual perception mechanisms. Infrared thermal imaging data is input into a temperature feature subnetwork to extract temperature gradients and high-temperature region features. Environmental sensor data is input into a time-series subnetwork to extract the changing trend characteristics of temperature, humidity, and gas concentration. The device status data is input into the attribute subnetwork to extract operating parameter features; Personnel trajectory data is input into the motion subnetwork to extract features such as movement direction and dwell time.

7. The AI-based intelligent fire scene risk assessment system as described in claim 1, characterized in that, The spatiotemporal risk prediction module includes: A submodule is constructed to build a spatiotemporal feature matrix based on the spatial features of a dynamic heterogeneous knowledge graph and the temporal features of a fused feature vector. The encoding submodule is used to add timestamp encoding and spatial coordinate encoding to the spatiotemporal feature matrix to obtain the encoded spatiotemporal features. The second computational submodule is used to input the encoded spatiotemporal features into the encoder layer of the spatiotemporal Transformer, and calculate the association weights of features at different time steps and different spatial locations through multiple attention heads to capture long-distance spatiotemporal dependencies. The output submodule is used to perform nonlinear transformation and aggregation on the features after attention calculation through the feedforward neural network of the encoder layer. The aggregated features are then input into the decoder layer of the Transformer. The decoder outputs the risk features at each time point based on a preset time step. The features are then converted into the risk level and risk diffusion path of each region within the corresponding time period by the classifier, so as to output the risk evolution trend.

8. The AI-based intelligent fire scene risk assessment system as described in claim 1, characterized in that, The dynamic allocation module includes: The initialization submodule is used to collect information on currently available fire-fighting resources, determine constraints and optimization objectives, and initialize the DBO algorithm population. The mutation and crossover submodule is used to calculate the fitness value of each allocation scheme according to the optimization objective, update the population through the search mechanism of the DBO algorithm, retain the scheme with high fitness as the high-quality solution, and perform mutation and crossover on the remaining schemes to generate new allocation schemes. The convergence submodule is used to iterate repeatedly until the fitness value converges, select the allocation scheme with the highest fitness as the optimal solution, and perform fire resource scheduling based on the optimal solution, dynamically allocating fire extinguishing equipment, emergency access routes and rescue strategies.

9. The AI-based intelligent fire scene risk assessment system as described in claim 8, characterized in that, The fire-fighting resource information includes the quantity and location of fire-fighting equipment, the unobstructed status of emergency access routes, and the configuration of rescue personnel. The constraints include the maximum movement range of equipment and the carrying capacity of access routes. The optimization objectives include the fastest response time, maximizing resource utilization, and maximizing rescue coverage.

10. A method for implementing an AI-based intelligent fire scene risk assessment system as described in claim 1, characterized in that, The method includes the following steps: Collect multi-source heterogeneous data in fire-fighting scenarios, including at least visual image data, infrared thermal imaging data, environmental sensor data, equipment status data, and personnel trajectory data. A dynamic heterogeneous knowledge graph containing equipment entities, environmental entities, personnel entities, and fire protection code knowledge is constructed and updated in real time based on multi-source heterogeneous data. The dynamic heterogeneous knowledge graph is represented and updated using a graph neural network. The BioNet algorithm is used to extract and fuse features from multi-source heterogeneous data, and a multi-head attention mechanism is used to assign dynamic weights to data of different modalities to generate fused feature vectors. Based on heterogeneous knowledge graphs and fused feature vectors, a spatiotemporal Transformer model is used to predict the risk evolution trend of fire scenarios within a preset time period in the future. The DBO algorithm is used to schedule fire-fighting resources based on the output of the spatiotemporal Transformer model, dynamically allocating fire-fighting equipment, emergency access routes, and rescue strategies.

Images