Intelligent building fire identification and simulation early warning method and system

Abstract

The invention discloses an intelligent building fire identification and simulation early warning method and system, belongs to the technical field of building informatization and disaster prevention and control, and aims to solve the technical problems of how to realize fire early identification, fire intelligent prediction and evacuation path dynamic planning, improve fire identification accuracy and response speed, and improve the safety and reliability of a building. According to the technical scheme, the method comprises the steps of BIM modeling, wherein a high-precision three-dimensional space model is established based on building information modeling, and the high-precision three-dimensional space model is deeply coupled with an FDS fire numerical simulation engine; multi-modal data acquisition: performing data acquisition by adopting a visual, environmental and spatial multi-modal sensor, and performing feature fusion on the acquired data by using a multi-channel deep convolutional network to generate unified space-time fire characterization; a video-sensing-geometric data collaborative sensing network is constructed on the basis of space-time fire characterization, so that the recognition robustness in a complex environment is improved; intelligent identification and decision making; performing early warning control linkage; and post-disaster assessment feedback.

Description

Technical Field

[0001] The present invention relates to the technical field of building informatization and disaster prevention and control, and in particular to an intelligent building fire identification and simulation early warning method and system. Background Art

[0002] Current building fire monitoring technology has the following problems:

[0003] ① Traditional sensor systems: rely on smoke / temperature threshold alarms, have high false alarm rates, and cannot provide spatial fire analysis;

[0004] ②Static BIM application: Existing BIM models are mostly used in the design phase and lack dynamic integration with real-time monitoring data;

[0005] ③ Limitations of fire simulation: Physical simulation methods such as CFD (computational fluid dynamics) are time-consuming and cannot meet the needs of real-time warning.

[0006] ④Data island problem: Sensor data, building structure data, and historical fire cases have not formed multimodal collaborative analysis.

[0007] Therefore, how to achieve early fire identification, intelligent fire prediction and dynamic planning of evacuation routes, and improve fire identification accuracy and response speed are technical problems that need to be solved urgently. Summary of the Invention

[0008] The technical task of this invention is to provide an intelligent building fire identification and simulation warning method and system to solve the problem of how to achieve early fire identification, intelligent fire prediction and dynamic planning of evacuation routes, and improve fire identification accuracy and response speed.

[0009] The technical task of the present invention is achieved in the following manner: a method for identifying and simulating fire in intelligent buildings, which is specifically as follows:

[0010] BIM modeling: Based on Building Information Modeling (BIM), a high-precision 3D spatial model is established and deeply coupled with the FDS fire numerical simulation engine. BIM carries information on building structure and functional zoning, visualizes the building model, and displays real-time updates of the simulated boundaries of the room. The fire spread simulation process is dynamically adjusted based on information and parameters such as internal building materials, opening structure, wind speed, and internal stacking items, enabling accurate responses to different types of fire (e.g., fast-burning, slow-burning, and explosive).

[0011] Multimodal data acquisition: Multimodal sensors for vision (RGB / IR video), environment (temperature, humidity, CO concentration), and space (radar / laser scanning) are used for data acquisition. A multi-channel deep convolutional network is used to fuse the collected data to generate a unified spatiotemporal fire representation. Based on this spatiotemporal fire representation, a video-sensing-geometric data collaborative perception network is constructed to improve recognition robustness in complex environments.

[0012] Intelligent Identification and Decision-Making: An interpretable reinforcement learning model is introduced to construct an intelligent strategy optimizer, using the evolution of building fires as a state space and the operations of fire extinguishing, evacuation, and lockdown as an action space. Real-time reward function feedback is provided through the BIM+FDS joint simulation environment, allowing the reinforcement learning agent to continuously learn and optimize fire response paths and control measures. The attention mechanism and causal graph modeling are then used to output the strategy basis, allowing firefighters to trace the model's decision logic, improving trust and controllability.

[0013] Early warning and control linkage: Automatically trigger alarm broadcasting, evacuation guidance, sprinkler system, and access control management equipment according to the fire level and location, achieving precise linkage;

[0014] Post-disaster assessment feedback: Summarize operational data and conduct a comprehensive assessment of response efficiency, equipment performance, and building damage to optimize subsequent response plans.

[0015] As a preferred option, BIM modeling is as follows:

[0016] BIM model standardization and lightweighting: Each BIM component (wall / door / pipe / equipment) is assigned a unique identifier, INT_ID, and basic information configuration is written to the database and IFC property set. The data is automatically generated and stored through the Element.UniqueId of the Autodesk Revit API. After storing the basic data, the data format is geometrically processed to ensure that FDS can be parsed. Specifically, the component type corresponding to FDS is classified and complex three-dimensional structures are decomposed into individual, quantifiable specific structures, allowing FDS to directly parse them.

[0017] FDS simulation execution and result analysis: First, a BIM model is created to simulate the real-world environment, striving to restore the real-world environment. The data format is then geometrically processed to ensure FDS parsing. After simplifying the data format, an automatic file generator is established to store real-time change information and write the BIM INT_ID into the FDS grid attributes. The computational grid is then automatically divided according to the geometric boundaries, dividing the complex grid into data attributes that can be associated with INT_ID, reducing the complexity of the real-world data. Finally, FDS is used to calculate changes in the transmitted data and execute fire simulations based on variable values ​​and labeled grids, such as physical field calculations: temperature / smoke / heat flux, or data storage: each grid cell records the component ID and physical quantity time history.

[0018] The BIM mapping of spatiotemporal data domains is as follows: for surface temperature or smoke concentration scenarios, point sampling is used. Specifically, a sampling point cloud is generated on the wall surface, the grid cell where the point is located is queried through KD-Tree, and IDW inverse distance weighted interpolation is performed. For heat flux or cumulative thermal damage, surface association is used. Specifically, FDS boundary slices are extracted, the intersection of the BIM surface and the slices is calculated through the Shapely library, and the average is obtained through numerical integration.

[0019] Intelligent metadata writing and visualization in BIM: Design extended attribute sets by referencing external massive time-series data through URLs to avoid BIM model bloat. Import the calculated metadata back into the BIM model to intuitively observe the information responses of temperature, smoke and fire caused by environmental factors.

[0020] Preferably, multimodal data collection is specifically as follows:

[0021] Multimodal data preprocessing: Time is synchronized through the NTP protocol to achieve unified data time standards. RGB frame sequences and IR sequences are spatially aligned through SIFT feature point matching and homography transformation. External parameter calibration and coordinate system conversion are then used to unify them into a common coordinate system. The calibrated values ​​of the temperature sensor are then assigned to the point information of the laser point cloud.

[0022] Multi-channel deep convolutional network: The Conv3D_1 layer extracts spatiotemporal features from RGB frame sequences, the Conv3D_2 layer enhances temperature-sensitive areas in IR sequences, the PointNet++ layer extracts the spatial structure of laser point clouds, and the ConvLSTM layer captures the evolution trend of environmental parameters in sensor data. The feature concatenation layer concatenates spatial features, temperature-sensitive areas, spatial features, and environmental parameter trends to obtain concatenated features. This concatenation is then fused using cross-modal attention to obtain 256-dimensional fused features. The 256-dimensional fused features are then refined to obtain a 128-dimensional spatiotemporal representation, which is then used for fire classification.

[0023] Feature fusion: The spatiotemporal features of the RGB frame sequence are used as the query, and the temperature-sensitive areas of the IR sequence and the evolution trends of environmental parameters in the sensor data are used as the key and value. Weighted fusion is calculated through attention weights. Then, through laser point cloud spatial structure extraction and optimization, the spatial structure features of disordered point clouds are efficiently extracted to adapt to the complex geometric structures that may appear in fire scenes. Specifically, the laser point cloud spatial structure extraction and optimization involves the original point cloud being processed through PointNetConv1, ReLU activation, PointNetConv2, and global maximum pooling to obtain 1024-dimensional features.

[0024] Spatiotemporal fire representation generation: Fusion and decoding of multi-source heterogeneous sensor data (such as thermal imaging, visible light, gas concentration, structural vibration, etc.) into dynamic fire representations with clear physical meaning; the dynamic fire representations include spatial thermal maps, time evolution curves, risk vectors, and event markers.

[0025] Preferably, the spatiotemporal fire condition representation is generated as follows:

[0026] The input layer receives raw data streams from different sensors, including thermal imaging temperature fields, visible light smoke characteristics, gas concentrations, and vibration signals.

[0027] The aligned multimodal features are concatenated in the channel dimension (e.g., heat map channel + gas channel + vibration channel) through the feature fusion layer.

[0028] The 128-dimensional feature vector output by the network is fused through the decoding layer. Each pixel value represents the probability of fire intensity at the corresponding location (0-1), and the color is mapped to the temperature range (red>650℃, yellow 300-650℃, blue<80℃);

[0029] The physical parameter generation layer is used to reversely map the "abstract values" decoded from deep learning features to actual parameters with physical units (e.g., mapping the [0,1] value output by the network to the actual temperature value or CO concentration ppm value).

[0030] The output layer classifies and identifies key events (such as flashover and collapse), and outputs the event type and confidence level.

[0031] As a preferred option, the intelligent strategy optimizer is constructed as follows:

[0032] The 128-dimensional vector generated by multimodal feature fusion is used as the core input. The spatial heat map (deconvolution decoding) and temporal risk curve (LSTM decoding) in the vector are analyzed. The BIM component status (structural temperature / CO concentration) and personnel distribution data are integrated to output a three-dimensional fire situation matrix.

[0033] Action space design defines a set of actionable strategies, such as fire extinguishing, evacuation, and lockdown, and designs alarm notification channels;

[0034] Establish a reward function mechanism. Targeting specific fire situations, those who implement the correct strategy will be rewarded, and a multi-objective optimization function will be constructed.

[0035] The strategy evolution engine optimizes the strategy matching mechanism and improves the matching rate and speed while ensuring the correctness of strategy output.

[0036] An intelligent building fire identification and simulation early warning system, the system comprising:

[0037] The BIM modeling and simulation module is used to construct the three-dimensional structural information of the target building. Through deep coupling with the FDS fire dynamics engine, it realizes real-time physical simulation of multi-dimensional parameters such as fire spread path, temperature field, and smoke diffusion. The fire spread simulation process is dynamically adjusted based on the building's internal materials, opening structure, and wind speed parameters.

[0038] A multimodal perception module collects dynamic data including infrared / visible light video, smoke concentration, temperature and humidity, CO concentration, and structural vibration in key areas of the building as recognition input;

[0039] Intelligent identification and decision-making module, which uses deep learning and interpretable reinforcement learning networks to identify fire conditions and generate dynamic response strategies;

[0040] The linkage control module is used to automatically trigger alarm broadcasting, evacuation guidance, sprinkler systems, and access control management equipment according to the fire level and location, achieving precise linkage;

[0041] The post-disaster assessment and feedback module aggregates operational data, comprehensively assessing response efficiency, equipment performance, and building damage to optimize subsequent emergency plans. The assessment report generated by the post-disaster assessment and feedback module includes fire response timelines, evacuation efficiency and route utilization, equipment activation and effectiveness, a comparison of fire cause speculation with physical simulations, and an analysis of damage to structures and systems caused by the fire.

[0042] Preferably, the BIM modeling and simulation module includes:

[0043] The standardization and lightweight processing submodule is used to assign a unique identifier INT_ID to each BIM component (wall / door / pipe / equipment) and write basic information configuration to the database and IFC property set. The data is automatically generated and stored through the Element.UniqueId of the Autodesk Revit API. After storing the basic data, the data format is geometrically processed to ensure that FDS can be parsed. Specifically, it is classified according to the component type corresponding to FDS and complex three-dimensional structures are decomposed into individual quantifiable specific structures, allowing FDS to directly parse them.

[0044] The result analysis submodule is used to build BIM models, generate FDS input files, divide FDS meshes, and perform FDS solutions;

[0045] The mapping submodule is used for surface temperature or smoke concentration scenarios using point sampling. Specifically, it generates a sampling point cloud on the wall surface, queries the grid cell where the point is located through KD-Tree, and uses IDW inverse distance weighted interpolation. It is also used for heat flux or cumulative thermal damage using surface association. Specifically, it extracts FDS boundary slices, calculates the intersection of the BIM surface and the slices through the Shapely library, and calculates the average through numerical integration.

[0046] The metadata writing module is used to reference external massive time-series data through URLs to design extended attribute sets, avoid bloated BIM models, re-import the calculated metadata into the BIM model, and intuitively observe the information responses of temperature, smoke and fire caused by environmental factors.

[0047] Preferably, the result parsing submodule includes:

[0048] BIM model builder, used to build BIM models that simulate the real environment and strive to restore the real environment;

[0049] Lightweight processor for geometric processing of data formats to ensure FDS parsing;

[0050] FDS input file generator is used to simplify the data format and establish an automatic file generator to store real-time change information and write BIM INT_ID into FDS grid attributes;

[0051] FDS meshing engine is used to automatically divide the computational mesh according to the geometric boundary, dividing the complex mesh into data attributes that can be associated with INT_ID, reducing the complexity of the real data;

[0052] FDS solver, used to perform variable calculations on the transmitted data using FDS, and to perform fire simulations based on variable values ​​and labeled grids, such as physical field calculations: temperature / smoke / heat flux, or data storage: each grid cell records the component ID and physical quantity time history;

[0053] Preferably, the multimodal perception module includes:

[0054] The multimodal data preprocessing submodule is used to synchronize time through the NTP protocol to achieve unified data time standards. It spatially aligns the RGB frame sequence and the IR sequence through SIFT feature point matching and homography transformation, unifies them into a common coordinate system through external parameter calibration and coordinate system transformation, and then assigns the calibrated values ​​of the temperature sensor to the point information of the laser point cloud.

[0055] The multi-channel deep convolutional network submodule is used to extract the spatiotemporal features of the RGB frame sequence through the Conv3D_1 layer, enhance the temperature-sensitive areas of the IR sequence through the Conv3D_2 layer, extract the spatial structure of the laser point cloud through the PointNet++ layer, capture the evolution trend of environmental parameters in the sensor data through the ConvLSTM layer, and perform feature splicing on spatial features, temperature-sensitive areas, spatial features, and environmental parameter trends through the feature splicing layer to obtain spliced ​​features. The spliced ​​features are then fused with cross-modal attention to obtain 256-dimensional fused features. The 256-dimensional fused features are then refined to obtain a 128-dimensional spatiotemporal representation, and fire classification is performed based on the 128-dimensional spatiotemporal representation.

[0056] The feature fusion submodule uses the spatiotemporal features of the RGB frame sequence as the query, the temperature-sensitive areas of the IR sequence, and the evolution trends of environmental parameters in the sensor data as the key and value. It then performs weighted fusion calculations using attention weights. This is then combined with laser point cloud spatial structure extraction and optimization to efficiently extract the spatial structural features of disordered point clouds, adapting to the complex geometric structures that may occur in fire scenes. Specifically, the laser point cloud spatial structure extraction and optimization involves passing the original point cloud through PointNetConv1, ReLU activation, PointNetConv2, and global maximum pooling operations to obtain 1024-dimensional features.

[0057] The spatiotemporal fire condition representation generation submodule is used to fuse and decode multi-source heterogeneous sensor data (such as thermal imaging, visible light, gas concentration, structural vibration, etc.) into a fire condition dynamic representation with clear physical meaning; the fire condition dynamic representation includes spatial thermal maps, time evolution curves, risk vectors, and event markers.

[0058] As a preferred option, the intelligent strategy optimizer is constructed as follows:

[0059] The 128-dimensional vector generated by multimodal feature fusion is used as the core input. The spatial heat map (deconvolution decoding) and temporal risk curve (LSTM decoding) in the vector are analyzed. The BIM component status (structural temperature / CO concentration) and personnel distribution data are integrated to output a three-dimensional fire situation matrix.

[0060] Action space design defines a set of actionable strategies, such as fire extinguishing, evacuation, and lockdown, and designs alarm notification channels;

[0061] Establish a reward function mechanism. Targeting specific fire situations, those who implement the correct strategy will be rewarded, and a multi-objective optimization function will be constructed.

[0062] The strategy evolution engine optimizes the strategy matching mechanism and improves the matching rate and speed while ensuring the correctness of strategy output.

[0063] The intelligent building fire identification and simulation early warning method and system of the present invention have the following advantages:

[0064] (1) The present invention constructs a high-precision three-dimensional building model and deeply couples it with a fire dynamics simulation engine to achieve real-time physical simulation of fire spread paths, temperature fields, and smoke diffusion. It deploys multiple sensors to collect multimodal data including infrared / visible light video, smoke concentration, temperature and humidity, CO concentration, structural vibration, etc., and uses deep learning and interpretable reinforcement learning networks to identify fire conditions and generate dynamic response strategies. Then, based on the fire level and location, it automatically triggers alarm broadcasts, evacuation guidance, sprinkler systems, access control management, and other equipment to achieve precise linkage. Finally, it summarizes the operating data of each subsystem and conducts a comprehensive assessment of response efficiency, equipment performance, and building damage. This is used to optimize subsequent plans and models, thereby improving fire identification accuracy and response speed, enhancing decision-making transparency, and has broad application prospects.

[0065] (2) The present invention integrates BIM's 3D spatial data, IoT real-time sensor data, and physical drive simulation data to achieve early fire identification, intelligent fire prediction, and dynamic planning of evacuation routes, as well as adaptive dynamic optimization of evacuation routes, thereby improving the accuracy of fire identification.

[0066] (3) The present invention reverse-annotates FDS (Fire Dynamics Simulator) fire simulation results to BIM components to construct intelligent BIM metadata with physical attributes;

[0067] (4) This invention introduces SHAP (SHapley Additive Intelligent Building Fire Identification and Simulation Warning Method and System Planations) value analysis into path planning. This is a Shapley value-based machine learning model interpretation tool that decomposes prediction results, emphasizes feature contributions, and is applicable to various models and feature types to ensure that AI decisions comply with fire safety regulations.

[0068] (5) The present invention generates adversarial samples through the digital twin system and continuously optimizes the robustness of the model;

[0069] (6) The intelligent recognition and decision-making module of the present invention uses the attention mechanism and causal graph modeling to output the strategy basis, so that firefighters can trace the model decision logic and improve trust and controllability. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The present invention will be further described below with reference to the accompanying drawings.

[0071] Attachment Figure 1 Provide a flowchart for BIM model standardization and lightweight processing;

[0072] Attachment Figure 2 Schematic diagram of multimodal data preprocessing;

[0073] Attachment Figure 3 Schematic diagram of the multi-channel deep convolutional network structure;

[0074] Attachment Figure 4 Optimize schematic diagram for point cloud feature extraction;

[0075] Attachment Figure 5 Generate schematic diagrams for spatiotemporal fire regime representation;

[0076] Attachment Figure 6 This is a structural diagram of intelligent building fire identification and simulation warning. DETAILED DESCRIPTION

[0077] The intelligent building fire identification and simulation warning method and system of the present invention are described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] Example 1:

[0079] This embodiment provides a method for fire identification and simulation warning in intelligent buildings, which is specifically as follows:

[0080] S1. BIM Modeling: Based on Building Information Modeling (BIM), a high-precision 3D spatial model is established and deeply coupled with the FDS fire numerical simulation engine. BIM carries information on building structure, functional zoning, building model visualization, and real-time updates of the simulated boundaries of the room. The fire spread simulation process is dynamically adjusted based on information and parameters such as internal building materials, opening structure, wind speed, and internal stacked items, enabling accurate responses to different types of fire conditions (such as fast burning, slow burning, and explosion).

[0081] S2. Multimodal Data Acquisition: Multimodal sensors for visual (RGB / IR video), environmental (temperature, humidity, CO concentration), and spatial (radar / laser scanning) data acquisition are used. A multi-channel deep convolutional network is used to fuse the collected data to generate a unified spatiotemporal fire representation. Based on this spatiotemporal fire representation, a video-sensing-geometric data collaborative perception network is constructed to improve recognition robustness in complex environments.

[0082] S3. Intelligent Identification and Decision-making: Introducing an explainable reinforcement learning model, we construct an intelligent strategy optimizer based on the evolution of building fires as a state space and the actions of fire extinguishing, evacuation, and containment as an action space. Through the BIM+FDS joint simulation environment, we provide real-time feedback on the reward function, enabling the reinforcement learning agent to continuously learn and optimize fire response paths and control measures. We then employ an attention mechanism and causal graph modeling to output the strategy basis, enabling firefighters to trace the model's decision logic, improving trust and controllability.

[0083] S4, early warning and control linkage: According to the fire level and location, the alarm broadcast, evacuation guidance, sprinkler system, and access control management equipment are automatically triggered to achieve precise linkage;

[0084] S5. Post-disaster assessment feedback: Summarize operational data and conduct a comprehensive assessment of response efficiency, equipment performance, and building damage to optimize subsequent plans.

[0085] The BIM modeling in step S1 of this embodiment is specifically as follows:

[0086] S101. BIM model standardization and lightweight processing: Assign a unique identifier INT_ID to each BIM component (wall / door / pipe / equipment), and write basic information configuration into the database and IFC property set. The data is automatically generated and stored through the Element.UniqueId of the Autodesk Revit API. After storing the basic data, the data format is geometrically processed to ensure that FDS can be parsed. Specifically, classify according to the component type corresponding to FDS, and decompose the complex three-dimensional structure into individual quantifiable specific structures so that FDS can directly parse it. For different structural types, create a mapping table between BIM materials and FDS thermal physical properties, for example:

[0087] #Material mapping table example (CSV format);

[0088] BIM_Material_ID,FDS_Material_Name,Density,Conductivity,Specific_Heat;

[0089] "MAT_WALL_01","CONCRETE",2300,1.4,880.

[0090] S102, as attached Figure 1 As shown in the figure, the FDS simulation execution and result analysis are as follows:

[0091] S10201. Establish a BIM model that simulates the real environment and strive to restore the real environment;

[0092] S10202, geometrically process the data format to ensure FDS parsing;

[0093] S10203. After simplifying the data format, establish an automatic file generator to store real-time change information and write the BIM INT_ID into the FDS grid attribute;

[0094] S10204. Automatically divide the computational grid based on the geometric boundary, dividing the complex grid into data attributes that can be associated with INT_ID, thereby reducing the complexity of the real data;

[0095] S10205. Utilize FDS to perform change calculations on the transmitted data and execute fire simulation based on variable values ​​and labeled grids, such as performing physical field calculations: temperature / smoke / heat flux, or performing data storage: recording the component ID and physical quantity time history for each grid cell;

[0096] S103. The BIM construction mapping of the spatiotemporal data domain is as follows: for surface temperature or smoke concentration scenarios, point sampling is used. Specifically, a sampling point cloud is generated on the wall surface, the grid cell where the point is located is queried through KD-Tree, and IDW inverse distance weighted interpolation is performed. For heat flux or cumulative thermal damage, surface association is used. Specifically, FDS boundary slices are extracted, the intersection of the BIM surface and the slice is calculated through the Shapely library, and the average is obtained through numerical integration, as shown in Table 1:

[0097] Table 1 Comparison between point sampling strategy and surface sampling strategy

[0098]

[0099]

[0100] S104. Intelligent metadata writing into BIM and visualization: Design extended attribute sets through URL reference of external massive time series data to avoid BIM model bloat, re-import the calculated metadata into the BIM model, and intuitively observe the information response of temperature, smoke and fire caused by environmental factors.

[0101] The multimodal data collection in step S2 of this embodiment is specifically as follows:

[0102] S201. Multimodal data preprocessing: as shown in the attached Figure 2 As shown in the figure, the time is synchronized through the NTP protocol to achieve the unification of data time standards; the RGB frame sequence and the IR sequence are spatially aligned through SIFT feature point matching and homography transformation, and then unified into a common coordinate system through external parameter calibration and coordinate system transformation. The value calibrated by the temperature sensor is then assigned to the point information of the laser point cloud;

[0103] S202, Multi-channel Deep Convolutional Network: as attached Figure 3 As shown in the figure, the Conv3D_1 layer is used to extract the spatiotemporal features of the RGB frame sequence, the Conv3D_2 layer is used to enhance the temperature sensitive area of ​​the IR sequence, the PointNet++ layer is used to extract the spatial structure of the laser point cloud, the ConvLSTM layer is used to capture the evolution trend of the environmental parameters in the sensor data, and the feature splicing layer is used to perform feature splicing on the spatial features, temperature sensitive areas, spatial features and environmental parameter trends to obtain the spliced ​​features. The spliced ​​features are then cross-modal fused to obtain 256-dimensional fused features. The 256-dimensional fused features are refined to obtain 128-dimensional spatiotemporal representations. Fire classification is performed based on the 128-dimensional spatiotemporal representations, as shown in Table 2.

[0104] Table 2 Multi-channel deep convolutional network content table

[0105]

[0106] S203, Feature Fusion: As shown in the attached Figure 4 As shown in the figure, the spatiotemporal features of the RGB frame sequence are used as the query (Query), the temperature-sensitive areas of the IR sequence and the evolution trend of environmental parameters in the sensor data are used as the key (Key) and value (Value), and weighted fusion is calculated through attention weights. Then, the spatial structure features of the disordered point cloud are efficiently extracted through laser point cloud spatial structure extraction optimization to adapt to the complex geometric structures that may appear in fire scenes. Specifically, the laser point cloud spatial structure extraction optimization is as follows: the original point cloud is subjected to PointNetConv1, ReLU activation, PointNetConv2 and global maximum pooling operations to obtain 1024-dimensional features;

[0107] S204. Generation of spatiotemporal fire condition representation: Fusion and decoding of multi-source heterogeneous sensor data (such as thermal imaging, visible light, gas concentration, structural vibration, etc.) into a dynamic fire condition representation with clear physical meaning; the dynamic fire condition representation includes spatial thermal maps, time evolution curves, risk vectors, and event markers.

[0108] As attached Figure 5 As shown, the spatiotemporal fire condition representation generation in step S204 of this embodiment is specifically as follows:

[0109] S20401. Receive raw data streams from different sensors through the input layer; the raw data streams include thermal imaging temperature fields, visible light smoke characteristics, gas concentrations, and vibration signals;

[0110] S20402. Using a feature fusion layer, the aligned multimodal features are concatenated in the channel dimension (e.g., heat map channel + gas channel + vibration channel).

[0111] S20403, the 128-dimensional feature vector output by the decoding layer fusion network, each pixel value represents the probability of fire intensity at the corresponding location (0-1), and the color is mapped to the temperature range (red>650℃, yellow 300-650℃, blue<80℃);

[0112] S20404. Reversely map the "abstract values" decoded from the deep learning features to actual parameters with physical units through the physical parameter generation layer (e.g., mapping the [0,1] value output by the network to the actual temperature value or CO concentration ppm value).

[0113] S20405. Classify and identify key events (such as flashover and collapse) through the output layer, and output the event type and confidence level.

[0114] The intelligent strategy optimizer in step S3 of this embodiment is constructed as follows:

[0115] S301: Use the 128-dimensional vector generated by multimodal feature fusion as the core input, analyze the spatial heat map (deconvolution decoding) and temporal risk curve (LSTM decoding) in the vector, integrate BIM component status (structural temperature / CO concentration) and personnel distribution data, and output a three-dimensional fire situation matrix;

[0116] S302. Action space design defines a set of operational strategies, such as fire extinguishing, evacuation, and lockdown strategies, and designs alarm notification channels.

[0117] S303. Establish a reward function mechanism. Targeting a specific fire situation, those who implement the correct strategy will be rewarded, and a multi-objective optimization function is constructed.

[0118] S304. The strategy evolution engine optimizes the strategy matching mechanism to improve the matching rate and matching speed on the basis of ensuring the correctness of strategy output.

[0119] In step S3 of this embodiment, the reinforcement learning agent acts as a "tactical advisor": it proposes basic strategies based on the state space (e.g., "turn on the sprinklers in area A"). The intelligent strategy optimizer acts as the "command center": it evaluates strategy effectiveness (calculates expected casualties / structural damage) → optimizes action parameters (adjusts water pressure / angle) → iterates and upgrades strategies.

[0120] Based on traditional fire identification and simulation, this embodiment introduces a deep coupling mechanism of BIM and physical simulation, an explainable AI decision-making engine, and an interactive module that supports real-scene fusion training. It effectively improves the system's recognition accuracy, response efficiency, and decision-making transparency for complex building fires, and has significant engineering application prospects and promotion value.

[0121] Example 2:

[0122] As attached Figure 6 As shown, this embodiment provides an intelligent building fire identification and simulation warning system, which includes:

[0123] The BIM modeling and simulation module is used to construct the three-dimensional structural information of the target building. Through deep coupling with the FDS fire dynamics engine, it realizes real-time physical simulation of multi-dimensional parameters such as fire spread path, temperature field, and smoke diffusion. The fire spread simulation process is dynamically adjusted based on the building's internal materials, opening structure, and wind speed parameters.

[0124] A multimodal perception module collects dynamic data including infrared / visible light video, smoke concentration, temperature and humidity, CO concentration, and structural vibration in key areas of the building as recognition input;

[0125] Intelligent identification and decision-making module, which uses deep learning and interpretable reinforcement learning networks to identify fire conditions and generate dynamic response strategies;

[0126] The linkage control module is used to automatically trigger alarm broadcasting, evacuation guidance, sprinkler systems, and access control management equipment according to the fire level and location, achieving precise linkage;

[0127] The post-disaster assessment and feedback module aggregates operational data, comprehensively assessing response efficiency, equipment performance, and building damage to optimize subsequent emergency plans. The assessment report generated by the post-disaster assessment and feedback module includes fire response timelines, evacuation efficiency and route utilization, equipment activation and effectiveness, a comparison of fire cause speculation with physical simulations, and an analysis of damage to structures and systems caused by the fire.

[0128] The BIM modeling and simulation module in this embodiment includes:

[0129] The standardization and lightweight processing submodule is used to assign a unique identifier INT_ID to each BIM component (wall / door / pipe / equipment) and write basic information configuration to the database and IFC property set. The data is automatically generated and stored through the Element.UniqueId of the Autodesk Revit API. After storing the basic data, the data format is geometrically processed to ensure that FDS can be parsed. Specifically, it is classified according to the component type corresponding to FDS and complex three-dimensional structures are decomposed into individual quantifiable specific structures, allowing FDS to directly parse them.

[0130] The result analysis submodule is used to build BIM models, generate FDS input files, divide FDS meshes, and perform FDS solutions;

[0131] The mapping submodule is used for surface temperature or smoke concentration scenarios using point sampling. Specifically, it generates a sampling point cloud on the wall surface, queries the grid cell where the point is located through KD-Tree, and uses IDW inverse distance weighted interpolation. It is also used for heat flux or cumulative thermal damage using surface association. Specifically, it extracts FDS boundary slices, calculates the intersection of the BIM surface and the slices through the Shapely library, and calculates the average through numerical integration.

[0132] The metadata writing module is used to reference external massive time-series data through URLs to design extended attribute sets, avoid bloated BIM models, re-import the calculated metadata into the BIM model, and intuitively observe the information responses of temperature, smoke and fire caused by environmental factors.

[0133] As attached Figure 1 As shown, the result parsing submodule in this embodiment includes:

[0134] BIM model builder, used to build BIM models that simulate the real environment and strive to restore the real environment;

[0135] Lightweight processor for geometric processing of data formats to ensure FDS parsing;

[0136] FDS input file generator is used to simplify the data format and establish an automatic file generator to store real-time change information and write BIM INT_ID into FDS grid attributes;

[0137] FDS meshing engine is used to automatically divide the computational mesh according to the geometric boundary, dividing the complex mesh into data attributes that can be associated with INT_ID, reducing the complexity of the real data;

[0138] FDS solver, used to perform variable calculations on the transmitted data using FDS, and to perform fire simulations based on variable values ​​and labeled grids, such as physical field calculations: temperature / smoke / heat flux, or data storage: each grid cell records the component ID and physical quantity time history;

[0139] The multimodal perception module in this embodiment includes:

[0140] The multimodal data preprocessing submodule is used to synchronize time through the NTP protocol to achieve the unification of data time standards; the RGB frame sequence and the IR sequence are spatially aligned based on SIFT feature point matching and homography transformation, and unified into a common coordinate system through external parameter calibration and coordinate system transformation, and then the value calibrated by the temperature sensor is assigned to the point information of the laser point cloud, as shown in the attached figure. Figure 2 As shown;

[0141] The multi-channel deep convolutional network submodule is used to extract the spatiotemporal features of the RGB frame sequence through the Conv3D_1 layer, enhance the temperature-sensitive areas of the IR sequence through the Conv3D_2 layer, extract the spatial structure of the laser point cloud through the PointNet++ layer, capture the evolution trend of environmental parameters in the sensor data through the ConvLSTM layer, and perform feature splicing on the spatial features, temperature-sensitive areas, spatial features, and environmental parameter trends through the feature splicing layer to obtain the spliced ​​features. The spliced ​​features are then cross-modally fused to obtain 256-dimensional fused features. The 256-dimensional fused features are refined to obtain a 128-dimensional spatiotemporal representation. Fire classification is performed based on the 128-dimensional spatiotemporal representation, as shown in the attached figure. Figure 3 As shown; Among them, the content of the multi-channel deep convolutional network is shown in Table 2;

[0142] Table 2 Multi-channel deep convolutional network content table

[0143]

[0144]

[0145] The feature fusion submodule is used to use the spatiotemporal features of the RGB frame sequence as the query (Query), the temperature-sensitive areas of the IR sequence and the evolution trend of the environmental parameters in the sensor data as the key (Key) and value (Value), calculate the weighted fusion through attention weight, and then extract the spatial structure features of the disordered point cloud through laser point cloud spatial structure extraction optimization to efficiently extract the spatial structure features of the disordered point cloud and adapt to the complex geometric structures that may appear in the fire scene; Among them, the laser point cloud spatial structure extraction optimization is specifically as follows: the original point cloud is subjected to PointNetConv1, ReLU activation, PointNetConv2 and global maximum pooling operations to obtain 1024-dimensional features, as shown in the attached figure. Figure 4 As shown;

[0146] The spatiotemporal fire condition representation generation submodule is used to fuse and decode multi-source heterogeneous sensor data (such as thermal imaging, visible light, gas concentration, structural vibration, etc.) into a fire condition dynamic representation with clear physical meaning; the fire condition dynamic representation includes spatial thermal map, time evolution curve, risk vector, event marker, as shown in the attached figure. Figure 4 shown.

[0147] The intelligent strategy optimizer in this embodiment is constructed as follows:

[0148] ① The 128-dimensional vector generated by multimodal feature fusion is used as the core input. The spatial heat map (deconvolution decoding) and temporal risk curve (LSTM decoding) in the vector are analyzed. The BIM component status (structural temperature / CO concentration) and personnel distribution data are integrated to output a three-dimensional fire situation matrix.

[0149] ② Action space design defines a set of actionable strategies, such as fire extinguishing, evacuation, and lockdown, and designs alarm notification channels;

[0150] ③ Establish a reward function mechanism. For specific fire situations, those who release the correct strategy will be rewarded, and build a multi-objective optimization function;

[0151] ④ The strategy evolution engine optimizes the strategy matching mechanism to improve the matching rate and speed while ensuring the correctness of strategy output.

[0152] This embodiment supports multi-dimensional fire emergency plan configuration based on area and fire type (fast burning / slow burning, smoldering / explosion). Each type of plan includes the following:

[0153] Initial response level setting (local alarm, whole-building evacuation, etc.); evacuation route recommendation and crowd guidance strategy (based on BIM spatial analysis and real-time crowd flow simulation); fire zoning containment plan and smoke exhaust and ventilation equipment linkage logic; fire interface scheduling strategy, including fire elevator priority and fire source location heat map generation.

[0154] The plan can be optimized and iterated based on historical drill data and FDS simulation results, and can be dynamically updated through the system management platform.

[0155] The fire response process in this embodiment is as follows:

[0156] (1) The multimodal perception module collects data in real time and feeds it into the recognition engine;

[0157] (2) When the intelligent identification and decision-making module determines that the fire probability exceeds the threshold, the system is triggered to enter the early warning state;

[0158] (3) Accurately locate the fire source and calculate the spread trend through the BIM model;

[0159] (4) Generate the optimal escape route and linkage plan, issue an alarm, and activate equipment;

[0160] (5) Upload the fire status, identification results, and response measures to the smart city fire command platform to achieve human-machine collaborative response.

[0161] After the fire is over, this embodiment performs data backtracking and analysis on the entire incident process and generates a post-disaster assessment report, including:

[0162] ① Fire response timeline (early warning → response → control → release);

[0163] ② Personnel evacuation efficiency and route utilization rate;

[0164] ③ The startup response and effectiveness of each device;

[0165] ④Comparison between fire cause speculation and physical simulation;

[0166] ⑤ Analysis of damage caused by fire to structures and systems (combined with BIM structural annotation);

[0167] ⑥The evaluation report will be automatically archived, and the next round of training strategies will be optimized through the model feedback mechanism to continuously improve the system robustness and emergency response capabilities.

[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying and simulating fire in intelligent buildings, characterized in that: The method is as follows: BIM modeling: Based on building information modeling, a high-precision 3D spatial model is established and deeply coupled with the FDS fire numerical simulation engine. BIM carries information on building structure, functional zoning, building model visualization, and real-time updates of simulated boundaries within the building. The fire spread simulation process is dynamically adjusted based on information and parameters such as internal building materials, opening structure, wind speed, and internal stacking items, enabling accurate responses to different types of fires. Multimodal data acquisition: Data is collected using multimodal sensors that combine vision, environment, and space. A multi-channel deep convolutional network is used to fuse the collected data to generate a unified spatiotemporal fire representation. A video-sensor-geometric data collaborative perception network is constructed based on the spatiotemporal fire representation to improve recognition robustness in complex environments. Intelligent Identification and Decision-Making: Introducing an explainable reinforcement learning model, we construct an intelligent strategy optimizer based on the evolution of building fires as a state space and the actions of fire extinguishing, evacuation, and containment as an action space. Through the BIM+FDS joint simulation environment, we provide real-time feedback on the reward function, enabling the reinforcement learning agent to continuously learn and optimize fire response paths and control measures. We then employ an attention mechanism and causal graph modeling to output the strategy basis. Early warning and control linkage: Automatically trigger alarm broadcasting, evacuation guidance, sprinkler system, and access control management equipment according to the fire level and location, achieving precise linkage; Post-disaster assessment feedback: Summarize operational data and conduct a comprehensive assessment of response efficiency, equipment performance, and building damage.

2. The intelligent building fire identification and simulation warning method according to claim 1 is characterized in that: BIM modeling is as follows: BIM model standardization and lightweighting: Each BIM component is assigned a unique identifier, INT_ID, and basic information configuration is written to the database and IFC property set. The data is automatically generated and stored through the Element.UniqueId of the Autodesk Revit API. After storing the basic data, the data format is geometrically processed to ensure that FDS can be parsed. Specifically, the component type corresponding to FDS is classified and complex three-dimensional structures are decomposed into individual, quantifiable specific structures, allowing FDS to directly parse them. FDS simulation execution and result analysis: First, a BIM model simulating the real-world environment is established, striving to restore the real-world environment. The data format is then geometrically processed to ensure FDS parsing. After simplifying the data format, an automatic file generator is established to store real-time change information and write the BIM INT_ID into the FDS grid attributes. The computational grid is then automatically divided according to the geometric boundaries, dividing the complex grid into data attributes that can be associated with INT_ID, reducing the complexity of the real-world data. Finally, FDS is used to calculate changes in the transmitted data, and fire simulations are performed based on variable values ​​and labeled grids, or data is stored: each grid cell records the component ID and physical quantity time history. The BIM mapping of spatiotemporal data domains is as follows: for surface temperature or smoke concentration scenarios, point sampling is used. Specifically, a sampling point cloud is generated on the wall surface, the grid cell where the point is located is queried through KD-Tree, and IDW inverse distance weighted interpolation is performed. For heat flux or cumulative thermal damage, surface association is used. Specifically, FDS boundary slices are extracted, the intersection of the BIM surface and the slices is calculated through the Shapely library, and the average is obtained through numerical integration. Intelligent metadata writing and visualization in BIM: Design extended attribute sets by referencing external massive time-series data through URLs, re-import the calculated metadata into the BIM model, and visually observe the information responses to temperature, smoke, and fire caused by environmental factors.

3. The intelligent building fire identification and simulation warning method according to claim 1 is characterized in that: The details of multimodal data collection are as follows: Multimodal data preprocessing: Time is synchronized through the NTP protocol to achieve unified data time standards. RGB frame sequences and IR sequences are spatially aligned through SIFT feature point matching and homography transformation. External parameter calibration and coordinate system conversion are then used to unify them into a common coordinate system. The calibrated values ​​of the temperature sensor are then assigned to the point information of the laser point cloud. Multi-channel deep convolutional network: The Conv3D_1 layer extracts spatiotemporal features from RGB frame sequences, the Conv3D_2 layer enhances temperature-sensitive areas in IR sequences, the PointNet++ layer extracts the spatial structure of laser point clouds, and the ConvLSTM layer captures the evolution trend of environmental parameters in sensor data. The feature concatenation layer concatenates spatial features, temperature-sensitive areas, spatial features, and environmental parameter trends to obtain concatenated features. This concatenation is then fused using cross-modal attention to obtain 256-dimensional fused features. The 256-dimensional fused features are then refined to obtain a 128-dimensional spatiotemporal representation, which is then used for fire classification. Feature fusion: The spatiotemporal features of the RGB frame sequence are used as queries, and the temperature-sensitive areas of the IR sequence and the evolution trends of environmental parameters in the sensor data are used as keys and values. Weighted fusion is calculated through attention weights. Then, through laser point cloud spatial structure extraction and optimization, the spatial structure features of disordered point clouds are efficiently extracted to adapt to the complex geometric structures that may appear in fire scenes. Specifically, the laser point cloud spatial structure extraction and optimization involves the raw point cloud being processed through PointNetConv1, ReLU activation, PointNetConv2, and global maximum pooling to obtain 1024-dimensional features. Spatiotemporal fire representation generation: Fusion and decoding of multi-source heterogeneous sensor data into dynamic fire representations with clear physical meaning; the dynamic fire representations include spatial heat maps, time evolution curves, risk vectors, and event markers.

4. The intelligent building fire identification and simulation warning method according to claim 3 is characterized in that: The spatiotemporal fire situation representation is generated as follows: The input layer receives raw data streams from different sensors, including thermal imaging temperature fields, visible light smoke characteristics, gas concentrations, and vibration signals. The aligned multimodal features are spliced ​​in the channel dimension through the feature fusion layer; The 128-dimensional feature vector output by the network is fused through the decoding layer. Each pixel value represents the probability of fire intensity at the corresponding location, and the color is mapped to the temperature range; The physical parameter generation layer reversely maps the "abstract values" decoded from deep learning features to actual parameters with physical units; The output layer classifies and identifies key events, and outputs event types and confidence levels.

5. The intelligent building fire identification and simulation warning method according to claim 1 is characterized in that: The construction of the intelligent strategy optimizer is as follows: The 128-dimensional vector generated by multimodal feature fusion is used as the core input. The spatial heat map and time-series risk curve in the vector are analyzed, and the BIM component status and personnel distribution data are integrated to output a three-dimensional fire situation matrix. Action space design defines the set of actionable strategies; Establish a reward function mechanism. Targeting specific fire situations, those who implement the correct strategy will be rewarded, and a multi-objective optimization function will be constructed. The strategy evolution engine optimizes the strategy matching mechanism and improves the matching rate and speed while ensuring the correctness of strategy output.

6. An intelligent building fire identification and simulation warning system, characterized in that: The system includes: The BIM modeling and simulation module is used to construct the three-dimensional structural information of the target building. Through deep coupling with the FDS fire dynamics engine, it realizes real-time physical simulation of multi-dimensional parameters such as fire spread path, temperature field, and smoke diffusion. The fire spread simulation process is dynamically adjusted based on the building's internal materials, opening structure, and wind speed parameters. A multimodal perception module collects dynamic data including infrared / visible light video, smoke concentration, temperature and humidity, CO concentration, and structural vibration in key areas of the building as recognition input; Intelligent identification and decision-making module, which uses deep learning and interpretable reinforcement learning networks to identify fire conditions and generate dynamic response strategies; The linkage control module is used to automatically trigger alarm broadcasting, evacuation guidance, sprinkler systems, and access control management equipment according to the fire level and location, achieving precise linkage; The post-disaster assessment feedback module is used to summarize operational data, conduct a comprehensive assessment of response efficiency, equipment performance, and building damage, and optimize subsequent emergency plans.

7. The intelligent building fire identification and simulation warning system according to claim 6, characterized in that: BIM modeling and simulation modules include: The standardization and lightweight processing submodule is used to assign a unique identifier INT_ID to each BIM component and write basic information configuration to the database and IFC property set. The data is automatically generated and stored through the Element.UniqueId of the Autodesk Revit API. After storing the basic data, the data format is geometrically processed to ensure that FDS can be parsed. Specifically, it is classified according to the component type corresponding to FDS and complex three-dimensional structures are decomposed into individual quantifiable specific structures, allowing FDS to directly parse them. The result analysis submodule is used to build BIM models, generate FDS input files, divide FDS meshes, and perform FDS solutions; The mapping submodule is used for surface temperature or smoke concentration scenarios using point sampling. Specifically, it generates a sampling point cloud on the wall surface, queries the grid cell where the point is located through KD-Tree, and uses IDW inverse distance weighted interpolation. It is also used for heat flux or cumulative thermal damage using surface association. Specifically, it extracts FDS boundary slices, calculates the intersection of the BIM surface and the slices through the Shapely library, and calculates the average through numerical integration. The metadata writing module is used to reference external massive time-series data through URLs to design extended attribute sets, re-import the calculated metadata into the BIM model, and intuitively observe the information responses of temperature, smoke and fire caused by environmental factors.

8. The intelligent building fire identification and simulation warning system according to claim 7 is characterized in that: The result parsing submodules include: BIM model builder, used to build BIM models that simulate the real environment and strive to restore the real environment; Lightweight processor for geometric processing of data formats to ensure FDS parsing; FDS input file generator is used to simplify the data format and establish an automatic file generator to store real-time change information and write BIM INT_ID into FDS grid attributes; FDS meshing engine is used to automatically divide the computational mesh according to the geometric boundary, dividing the complex mesh into data attributes that can be associated with INT_ID, reducing the complexity of the real data; FDS solver, used to calculate the changes of the transmitted data using FDS, perform fire simulation based on variable values ​​and labeled grids, or store data: each grid cell records the component ID and physical quantity time history.

9. The intelligent building fire identification and simulation warning system according to claim 6, characterized in that: The multimodal perception module includes: The multimodal data preprocessing submodule is used to synchronize time through the NTP protocol to achieve unified data time standards. It spatially aligns the RGB frame sequence and the IR sequence through SIFT feature point matching and homography transformation, unifies them into a common coordinate system through external parameter calibration and coordinate system transformation, and then assigns the calibrated values ​​of the temperature sensor to the point information of the laser point cloud. The multi-channel deep convolutional network submodule is used to extract the spatiotemporal features of the RGB frame sequence through the Conv3D_1 layer, enhance the temperature-sensitive areas of the IR sequence through the Conv3D_2 layer, extract the spatial structure of the laser point cloud through the PointNet++ layer, capture the evolution trend of environmental parameters in the sensor data through the ConvLSTM layer, and perform feature splicing on spatial features, temperature-sensitive areas, spatial features, and environmental parameter trends through the feature splicing layer to obtain spliced ​​features. The spliced ​​features are then fused with cross-modal attention to obtain 256-dimensional fused features. The 256-dimensional fused features are then refined to obtain a 128-dimensional spatiotemporal representation, and fire classification is performed based on the 128-dimensional spatiotemporal representation. The feature fusion submodule uses the spatiotemporal features of the RGB frame sequence as a query, the temperature-sensitive areas of the IR sequence, and the evolution trends of environmental parameters in the sensor data as keys and values. This is weighted fusion calculated through attention weights, and then optimized through laser point cloud spatial structure extraction to efficiently extract the spatial structural features of disordered point clouds, adapting to the complex geometric structures that may appear in fire scenes. Specifically, the laser point cloud spatial structure extraction optimization involves the raw point cloud being processed through PointNetConv1, ReLU activation, PointNetConv2, and global maximum pooling to obtain 1024-dimensional features. The spatiotemporal fire situation representation generation submodule is used to fuse and decode multi-source heterogeneous sensor data into a fire situation dynamic representation with clear physical meaning; the fire situation dynamic representation includes spatial thermal maps, time evolution curves, risk vectors, and event markers.

10. The intelligent building fire identification and simulation warning system according to claim 6, characterized in that: The construction of the intelligent strategy optimizer is as follows: The 128-dimensional vector generated by multimodal feature fusion is used as the core input. The spatial heat map and time-series risk curve in the vector are analyzed, and the BIM component status and personnel distribution data are integrated to output a three-dimensional fire situation matrix. Action space design defines the set of actionable strategies; Establish a reward function mechanism. Targeting specific fire situations, those who implement the correct strategy will be rewarded, and a multi-objective optimization function will be constructed. The strategy evolution engine optimizes the strategy matching mechanism and improves the matching rate and speed while ensuring the correctness of strategy output.

Images