Hoisting construction safety monitoring and early warning system based on BIM

Abstract

The invention discloses a BIM (Building Information Modeling)-based hoisting construction safety monitoring and early warning system. The system comprises a terminal sensing layer which is used for collecting environmental parameters and personnel behavior data in a closed space in real time; the edge computing layer is used for carrying out cleaning, compression and encrypted transmission on original data by utilizing an explosion-proof edge computing gateway; the cloud collaboration layer is used for storing full data based on a BIM digital twinborn platform, constructing a'danger mode-construction feature-disposal measure 'three-dimensional meta-knowledge graph by adopting an MAML + + algorithm, meanwhile, coupling a physical mechanism data enhancement engine with a multi-physics field coupling model and a physical constraint generative adversarial network, generating virtual data conforming to mass conservation and energy conservation, and sending the virtual data to the cloud collaboration layer; performing mixed training with real data; according to the intelligent decision-making layer, a space-time adaptive threshold evolutionary algorithm encodes a space-time context through a graph attention network and Transform, an alarm threshold is dynamically optimized through deep reinforcement learning, meanwhile, a digital twin deduction engine calculates a shortest safety path in real time, and rescue resource allocation is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of hoisting construction safety, and in particular to a hoisting construction safety monitoring and early warning system based on BIM. Background Art

[0002] Modern pumped-storage power stations primarily consist of major components such as the generator stator, rotor, turbine runner, inlet valve, top cover, and GCB (generator discharge switch). High-voltage switchgear often uses SF6 gas as an insulating medium. Leakage of SF6 gas can not only damage the environment but, in severe cases, cause suffocation and death. A dedicated GIS room, essential for hydropower stations, primarily houses switchgear connecting the power station to the grid. Modern high-voltage switchgear all uses SF6 gas as an insulating medium.

[0003] The inherent high-risk characteristics of confined space operations and the severe challenges of safety management and control of hoisting operations. Due to the limited entrances and exits, poor natural ventilation, potential toxic and harmful gases and oxygen-deficient environments in confined spaces, hoisting operations face multiple risks, such as collision risks during the handling of heavy equipment, operational errors under limited vision, and explosion hazards caused by gas leaks. The BIM-based safety monitoring and early warning system for large-scale hoisting construction of pumped-storage power station units builds a three-dimensional visualization model, integrates IoT sensor data, and uses algorithms such as deep reinforcement learning to achieve real-time optimization of hoisting paths, dynamic monitoring of hazardous gases, and adaptive adjustment of thresholds. It effectively solves the problems of the lag of traditional manual inspections, the limitations of fixed sensor coverage, and the poor adaptability of fixed threshold mechanisms to changes in construction stages, providing accurate, real-time, and intelligent safety guarantees for hoisting operations in confined spaces.

[0004] In existing technologies, accident data in confined space construction scenarios exhibits a typical "long-tail distribution" characteristic. While data on common risks (such as oxygen deficiency) is sufficient, samples of high-risk conditions (such as hydrogen explosion and structural collapse) are scarce. The decision boundary between normal samples and high-risk condition samples is too steep, resulting in a high misjudgment rate for high-risk conditions. Therefore, a BIM-based hoisting construction safety monitoring and early warning system is proposed. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a BIM-based hoisting construction safety monitoring and early warning system.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A BIM-based hoisting construction safety monitoring and early warning system, including:

[0008] Terminal perception layer: By deploying explosion-proof multimodal sensor networks, lightweight BIM terminals, and AR glasses, environmental parameters and human behavior data within confined spaces are collected in real time.

[0009] Edge computing layer: Utilizes explosion-proof edge computing gateways to clean, compress, and encrypt raw data for transmission. BIM geometric features, construction log semantic features, and sensor timing features are extracted locally. Lightweight CFD simulations are then performed using a fast physical field solver to provide real-time early warnings and localized decision-making.

[0010] Cloud collaboration layer: Based on the BIM digital twin platform to store all data, the MAML++ algorithm is used to construct a three-dimensional meta-knowledge graph of "hazard mode-construction characteristics-treatment measures". At the same time, the physical mechanism data enhancement engine couples the multi-physics field coupling model and the physical constraint generative adversarial network to generate virtual data that conforms to the conservation of mass and energy, and mixes it with real data for training.

[0011] Intelligent decision-making layer: The spatiotemporal adaptive threshold evolutionary algorithm encodes the spatiotemporal context through the graph attention network and Transformer, and then uses deep reinforcement learning to dynamically optimize the alarm threshold. At the same time, the digital twin deduction engine calculates the shortest safe path in real time and optimizes the allocation of rescue resources.

[0012] The above technical solution further includes:

[0013] Furthermore, the explosion-proof multimodal sensor network includes a laser gas sensor array, a UWB-LoRa fusion positioning tag, and a fiber Bragg grating strain sensor. The laser gas sensor array collects O2, CO, H2 and S combustible gas data, the UWB-LoRa fusion positioning tag collects spatial position data of personnel / equipment, and the fiber Bragg grating strain sensor monitors the deformation of the pipeline corridor support structure.

[0014] Furthermore, by using PointNet++ combined with dynamic graph convolution, structural features are extracted for BIM geometric feature extraction. The BERT-specification knowledge graph is used to semantically understand the safety technical briefing records and thus realize the extraction of semantic features of the construction log. The TCN+ self-attention module is used to capture the timing pattern before the sudden change in gas concentration to extract the sensor timing features.

[0015] Furthermore, the specific steps of extracting BIM geometric features, construction log semantic features, and sensor timing features locally, and combining them with a physical field fast solver to perform lightweight CFD simulations for real-time warnings and localized decision-making are as follows:

[0016] Fast physical field solver: It adapts to the fluid flow problem in a confined space by simplifying the Navier-Stokes equation and discretizing the equation using the finite volume method. The simplified Navier-Stokes equation is expressed as Where u is the fluid velocity vector, p is the pressure, ρ is the fluid density, v is the kinematic viscosity, and f is the external force term;

[0017] Lightweight CFD simulation: combining model reduction and parallel computing technology to accelerate the solution;

[0018] Feature fusion: Fusing BIM geometric features, construction log semantic features, and sensor temporal features;

[0019] Early warning judgment: Based on the fused features and CFD simulation results, early warning judgment is made using the preset early warning threshold or machine learning model;

[0020] Localized decision-making: Based on the early warning results and the actual situation on site, a lightweight AI inference engine is used to make localized decisions.

[0021] Furthermore, the cloud collaboration layer stores all data based on the BIM digital twin platform and uses the MAML++ algorithm to construct a three-dimensional knowledge graph of "hazard mode-construction characteristics-treatment measures", including the following steps:

[0022] BIM digital twin platform data storage: The BIM digital twin platform integrates and stores all the data of the hoisting construction;

[0023] Extracting "hazard pattern-construction characteristics-treatment measures" triples: Extracting "hazard pattern-construction characteristics-treatment measures" triples from the full data. The hazard pattern is determined through cluster analysis or expert experience, the construction characteristics are extracted from BIM models, sensor data, and construction logs, and the treatment measures are determined based on safety regulations or historical cases.

[0024] Constructing a meta-knowledge graph: The inner loop updates the model parameters of a small number of samples on the tasks extracted by the outer loop, which is expressed as Among them, φ is the initial parameter of the meta-model, α is the learning rate of the inner loop, L τi (f φ ) is the loss function on task τi, which measures the performance of the model on the task, f φ The feature extraction network with parameter φ is used to extract features from BIM models and sensor data, and the cross-task metamodel parameter update in the outer loop is expressed as Where β is the learning rate of the outer loop, p(τ) is the task distribution, and tasks are randomly selected from historical projects. is the parameter φ′ after the inner loop update i Under this condition, the loss function on task τi enables the meta-model to quickly adapt to new tasks;

[0025] Dynamic update: Information is transmitted between nodes through the message passing mechanism of the graph neural network to update the embedding vector of the node. Incremental learning is used to update the embedding vector and edge weight of related nodes when new task data arrives to adapt to new construction scenarios and danger patterns.

[0026] Furthermore, the physical mechanism data enhancement engine couples the multi-physics field coupling model and the physical constraint generative adversarial network to generate virtual data that conforms to mass conservation and energy conservation, and mixes it with real data for training, including the following steps:

[0027] Multi-physics coupling model construction: By establishing the hydrogen diffusion equation and coupling the turbulence model, the computational fluid dynamics solver is used to dynamically simulate the hydrogen leakage and diffusion process in the confined space. The hydrogen diffusion equation is expressed as Among them, C is the hydrogen concentration, u is the flow field velocity vector, calculated by the turbulence model, D is the diffusion coefficient, S is the source term, and at the same time, the structural mechanics equation is introduced and combined with finite element analysis to calculate the deformation of the lifting equipment under wind load. The structural mechanics equation is expressed as Where EI is the bending stiffness, y is the deflection, μ is the mass linear density, and q(x) is the distributed load, including wind load and the gravity of the hoisted object;

[0028] Data fusion and feature extraction: Mapping the CFD and FEM calculation results to the BIM model to generate a spatiotemporal feature matrix containing multi-physics field information;

[0029] Physically Constrained Generative Adversarial Network (PC-GAN) training: The generator uses random noise and physical parameters as input, combined with a deep neural network, to generate data that conforms to mass and energy conservation. The discriminator then evaluates the authenticity of the data and uses a loss function that includes a physical consistency term to screen the generated data and eliminate data that violates conservation laws.

[0030] Hybrid training: Combine physical simulation data, adversarially generated data, and real data to build a dataset containing various working conditions.

[0031] Furthermore, the spatiotemporal adaptive threshold evolutionary algorithm encodes spatiotemporal context through a graph attention network and a Transformer, including the following steps:

[0032] BIM model graph structuring: Build a dynamic graph structure based on the BIM model, abstract entities into nodes, abstract spatial relationships between entities into edges, initialize node feature vectors and edge weights, and convert construction scenarios into structured graph data;

[0033] Spatial embedding vector output: A graph attention network is used to dynamically learn the importance between entities by calculating attention coefficients, aggregating neighbor node information and updating node features to generate a spatial embedding vector.

[0034] Sensor sequence input: Historical sensor sequences retain temporal information by adding learnable positional encodings;

[0035] Temporal embedding vector output: The Transformer's self-attention mechanism is used to calculate query, key, and value vectors, capturing long-term dependencies in sensor data and generating a temporal embedding vector.

[0036] Cross-attention: Using the spatial embedding vector as the query and the temporal embedding vector as the key, the attention score is calculated and the features are fused to generate a spatiotemporal context embedding vector.

[0037] 2. Furthermore, the specific steps of using deep reinforcement learning to dynamically optimize the alarm threshold are:

[0038] Step 1: The policy network receives the current state vector as input, and the policy network calculates the output action probability distribution based on the current state vector through the neural network;

[0039] Step 2: Select specific actions based on the distribution using random sampling or greedy strategy and execute them, and adjust the threshold;

[0040] Step 3: After executing the action, observe the new state and reward value;

[0041] Step 4: Calculate the advantage function to evaluate the quality of the action;

[0042] Step 5: Update the policy network parameters φ based on the new state, reward, and advantage function to maximize the policy loss.

[0043] The present invention has the following beneficial effects:

[0044] In the present invention, the full amount of data is stored on the BIM digital twin platform, and the MAML++ algorithm is used to construct a three-dimensional knowledge graph of "hazard mode-construction characteristics-treatment measures" to support causal reasoning under small samples. At the same time, the physical mechanism data enhancement engine couples the multi-physics field coupling model and the physical constraint generative adversarial network (PC-GAN) to generate virtual data that conforms to the conservation of mass and energy, and mixed training with real data to improve the model's recognition accuracy of rare working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a system block diagram of a BIM-based hoisting construction safety monitoring and early warning system proposed in this invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See also Figure 1 As shown, the present invention is a BIM-based hoisting construction safety monitoring and early warning system, comprising:

[0048] Terminal perception layer: By deploying explosion-proof multimodal sensor networks, lightweight BIM terminals, and AR glasses, environmental parameters (such as O2 / CO / H2S / combustible gas concentrations, temperature and humidity, and liquid levels) and personnel behavior data (such as precise positioning and trajectory tracking) within confined spaces are collected in real time. BIM lightweight technology compresses complex models, allowing on-site personnel to view models, query attribute information, and receive hazard annotations at any time through explosion-proof terminals or AR glasses. AR glasses also overlay the BIM model with the on-site field of view, providing intuitive guidance on escape routes and hazard source locations.

[0049] Edge computing layer: Utilizes explosion-proof edge computing gateways to clean, compress, and encrypt raw data for transmission. BIM geometric features, construction log semantic features, and sensor timing features are extracted locally. Lightweight CFD simulations are then performed using a fast physical field solver to provide real-time early warnings and localized decision-making.

[0050] Cloud collaboration layer: Based on the BIM digital twin platform to store full data, the MAML++ algorithm is used to construct a three-dimensional meta-knowledge graph of "hazard mode-construction characteristics-treatment measures". At the same time, the physical mechanism data enhancement engine couples the multi-physics field coupling model (hydrogen diffusion equation, gas phase change equation, structural mechanics equation) and the physical constraint generative adversarial network (PC-GAN) to generate virtual data that conforms to the conservation of mass and energy, and mixes it with real data for training.

[0051] Intelligent decision-making layer: The spatiotemporal adaptive threshold evolutionary algorithm (ST-ETA) encodes the spatiotemporal context through the graph attention network (GAT) and Transformer, and then uses deep reinforcement learning (PPO algorithm) to dynamically optimize the alarm threshold. At the same time, the digital twin deduction engine calculates the shortest safe path in real time, optimizes the allocation of rescue resources, and interacts with the on-site AR glasses through the emergency command system.

[0052] In one embodiment, the explosion-proof multimodal sensor network includes a laser gas sensor array, a UWB-LoRa fusion positioning tag, and a fiber Bragg grating strain sensor. The laser gas sensor array collects O2, CO, H2, and S combustible gas data, the UWB-LoRa fusion positioning tag collects spatial position data of personnel / equipment, and the fiber Bragg grating strain sensor monitors the deformation of the pipeline corridor support structure.

[0053] In one embodiment, PointNet++ is combined with dynamic graph convolution to extract complex structural features such as pipeline corridor nodes and tank surfaces for BIM geometric feature extraction. The BERT-specification knowledge graph is used to semantically understand the safety technical briefing records and thus realize the extraction of semantic features of construction logs. The TCN+ self-attention module is used to capture the timing pattern before the sudden change in gas concentration to extract the sensor timing features.

[0054] In one embodiment, the specific steps of extracting BIM geometric features, construction log semantic features, and sensor timing features locally, and combining them with a physical field fast solver to perform lightweight CFD simulations for real-time warning and localized decision-making are as follows:

[0055] Fast physical field solver: It adapts to the fluid flow problem in a confined space by simplifying the Navier-Stokes equation and discretizing the equation using the finite volume method. The simplified Navier-Stokes equation is expressed as Where u is the fluid velocity vector, p is the pressure, ρ is the fluid density, v is the kinematic viscosity, and f is the external force term;

[0056] Lightweight CFD simulation: Combines model order reduction (POD) with parallel computing technology to accelerate solution, thus achieving rapid simulation of physical fields such as fluid flow and heat transfer;

[0057] Feature fusion: Fusing BIM geometric features, construction log semantic features, and sensor temporal features;

[0058] Early warning judgment: Based on the fused features and CFD simulation results, early warning judgment is made using the preset early warning threshold or machine learning model.

[0059] Localized decision-making: Based on the warning results and the actual on-site conditions, a lightweight AI inference engine is used to make localized decisions to determine the safety of the current construction status. Once a potential risk is detected, the system will immediately trigger an early warning and, based on the actual on-site conditions, such as construction progress and personnel location, quickly generate localized decisions through the lightweight AI inference engine, such as adjusting the construction plan, suspending dangerous operations, or initiating emergency plans.

[0060] In one embodiment, the cloud collaboration layer stores full data based on the BIM digital twin platform and uses the MAML++ algorithm to construct a three-dimensional knowledge graph of "hazard mode-construction characteristics-treatment measures", including the following steps:

[0061] BIM digital twin platform data storage: The BIM digital twin platform integrates and stores all the data of the hoisting construction. This data is precisely associated with the BIM model through unique identifiers or spatial coordinates, ensuring the accurate positioning of the data in three-dimensional space. Database management systems and cloud storage services are used to achieve efficient storage of structured and unstructured data.

[0062] Extracting "hazard pattern - construction feature - treatment measure" triples: Extracting "hazard pattern - construction feature - treatment measure" triples from the full data. The hazard pattern is determined through cluster analysis or expert experience, such as boom overload and cargo swing. The construction features are extracted from BIM models, sensor data, and construction logs, such as wind speed, lifting weight, and operation time. The treatment measures are determined based on safety regulations or historical cases, such as suspending operations or adjusting the lifting plan.

[0063] Constructing a meta-knowledge graph: The inner loop updates the model parameters of a small number of samples on the tasks extracted by the outer loop, which is expressed as Among them, φ is the initial parameter of the meta-model, α is the learning rate of the inner loop, L τi (f φ ) is the loss function on task τi, which measures the performance of the model on the task, f φ The feature extraction network (PointNet++) with parameter φ is used to extract features from the BIM model and sensor data, as well as the cross-task metamodel parameter update in the outer loop, which is expressed as Where β is the learning rate of the outer loop, p(τ) is the task distribution, and tasks are randomly selected from historical projects. is the parameter φ′ after the inner loop update i Under this condition, the loss function on task τi enables the meta-model to quickly adapt to new tasks;

[0064] Dynamic update: Information is transmitted between nodes through the message passing mechanism of the graph neural network (GNN), the node embedding vector is updated, and incremental learning is used to update the embedding vectors and edge weights of related nodes when new task data arrives, adapting to new construction scenarios and danger patterns.

[0065] In one embodiment, the physical mechanism data enhancement engine couples a multi-physics field coupling model (hydrogen diffusion equation, gas phase change equation, structural mechanics equation) and a physical constraint generative adversarial network (PC-GAN) to generate virtual data that conforms to mass conservation and energy conservation, and mixes it with real data for training, including the following steps:

[0066] Multi-physics coupling model construction: By establishing the hydrogen diffusion equation and coupling the turbulence model, the computational fluid dynamics (CFD) solver is used to dynamically simulate the hydrogen leakage and diffusion process in the confined space. The hydrogen diffusion equation is expressed as Where C is the hydrogen concentration (mol / m 3 ), u is the velocity vector of the flow field (m / s), calculated by the turbulence model (such as the k-ε model), and D is the diffusion coefficient (m 2 / s), S is the source term (leakage rate, mol / (m 3 ·s), the hydrogen diffusion equation not only describes the change of concentration with time and space, but also accurately captures the velocity distribution of the flow field through the turbulence model, providing key data for the prediction of dangerous areas. At the same time, the structural mechanics equation is introduced and combined with the finite element analysis (FEM) to calculate the deformation of the lifting equipment under wind load. The structural mechanics equation is expressed as Where EI is the bending stiffness (N·m 2 ), y is the deflection (m), μ is the mass linear density (kg / m), q(x) is the distributed load (N / m), including wind load and the gravity of the hoisted object. The structural mechanics equation provides a scientific basis for structural safety assessment by describing the relationship between boom deflection and distributed load;

[0067] Data fusion and feature extraction: Mapping the CFD and FEM calculation results to the BIM model generates a spatiotemporal feature matrix containing multi-physics field information such as concentration field and deformation field;

[0068] Physically Constrained Generative Adversarial Network (PC-GAN) training: The generator uses random noise and physical parameters as input, combined with a deep neural network, to generate data that conforms to mass and energy conservation. The discriminator evaluates the authenticity of the data and uses a loss function that includes a physical consistency term to screen the generated data, eliminating data that violates conservation laws and ensuring the quality of the generated data.

[0069] Hybrid training: Combining physical simulation data, adversarially generated data, and real data to build a dataset containing various working conditions, providing strong support for the model's generalization ability.

[0070] 3. In one embodiment, the spatiotemporal adaptive threshold evolutionary algorithm (ST-ETA) encodes spatiotemporal context via a graph attention network (GAT) and a Transformer, comprising the following steps:

[0071] BIM model graph structuring: A dynamic graph structure is constructed based on the BIM model. Entities such as cranes and components are abstracted as nodes, and spatial relationships between entities (such as distance and collision volume) are abstracted as edges. Node feature vectors (including geometric attributes such as position and size, and state attributes such as inclination and tension) and edge weights (using learnable parameters to quantify spatial relationships) are initialized to convert the construction scene into structured graph data.

[0072] Spatial embedding vector output: A Graph Attention Network (GAT) is used to dynamically learn the importance of entities by calculating attention coefficients (for example, the spatial relationship between a crane and surrounding buildings may be more weighted than that with distant components). This approach aggregates neighboring node information and updates node features to generate a spatial embedding vector, effectively capturing the spatial heterogeneity of lifting operations (for example, the spatial correlation between the crane's swing area and the component storage area).

[0073] Sensor sequence input: Historical sensor sequences (such as tilt, wind speed, and tension data) retain temporal information by adding learnable positional encodings;

[0074] Temporal embedding vector output: The Transformer's self-attention mechanism is used to calculate query, key, and value vectors, capturing long-term dependencies in sensor data (such as the relationship between the precursor to a sudden change in wind speed and the subsequent change in tilt angle) and generating a temporal embedding vector.

[0075] Cross-attention: Using the spatial embedding vector as the query and the temporal embedding vector as the key, the attention score is calculated and the features are fused to generate a spatiotemporal context embedding vector, achieving a deep fusion of spatial heterogeneity and temporal trends (for example, the crane node has a higher attention weight for the time step of wind speed mutation, indicating that the system focuses on the spatiotemporal correlation at that moment).

[0076] 4. In one embodiment, the specific steps of dynamically optimizing the alarm threshold using deep reinforcement learning (PPO algorithm) are:

[0077] Step 1: The policy network receives the current state vector as input. The current state vector incorporates multi-dimensional information such as the crane tilt angle threshold, false alarm rate, missed alarm rate, and construction stage code. The policy network calculates the output action probability distribution based on the current state vector through a neural network.

[0078] Step 2: Select specific actions based on the distribution using random sampling or greedy strategy and execute them, and adjust the threshold.

[0079] Step 3: After executing the action, observe the new state (false alarm rate, false negative rate) and reward value;

[0080] Step 4: Calculate the advantage function to evaluate the quality of the action;

[0081] Step 5: Update the policy network parameters φ based on the new state, reward, and advantage function to maximize the policy loss.

[0082] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A BIM-based hoisting construction safety monitoring and early warning system, characterized by: include: Terminal perception layer: By deploying explosion-proof multimodal sensor networks, lightweight BIM terminals, and AR glasses, environmental parameters and human behavior data within confined spaces are collected in real time. Edge computing layer: Utilizes explosion-proof edge computing gateways to clean, compress, and encrypt raw data for transmission. BIM geometric features, construction log semantic features, and sensor timing features are extracted locally. Lightweight CFD simulations are then performed using a fast physical field solver to provide real-time early warnings and localized decision-making. Cloud collaboration layer: Based on the BIM digital twin platform, the platform stores all data and uses the MAML++ algorithm to construct a three-dimensional knowledge graph of "hazard mode-construction characteristics-treatment measures." Simultaneously, the physical mechanism data enhancement engine couples the multi-physics coupling model with the physical constraint generative adversarial network to generate virtual data that adheres to mass conservation and energy conservation, which is then mixed with real data for training. Intelligent decision-making layer: The spatiotemporal adaptive threshold evolutionary algorithm encodes the spatiotemporal context through the graph attention network and Transformer, and then uses deep reinforcement learning to dynamically optimize the alarm threshold. At the same time, the digital twin deduction engine calculates the shortest safe path in real time and optimizes the allocation of rescue resources.

2. A BIM-based hoisting construction safety monitoring and early warning system according to claim 1, characterized in that: The explosion-proof multimodal sensor network includes a laser gas sensor array, a UWB-LoRa fusion positioning tag, and a fiber Bragg grating strain sensor. The laser gas sensor array collects O2, CO, H2, and S combustible gas data; the UWB-LoRa fusion positioning tag collects spatial position data of personnel / equipment; and the fiber Bragg grating strain sensor monitors the deformation of the pipeline corridor support structure.

3. The BIM-based hoisting construction safety monitoring and early warning system according to claim 1 is characterized in that: By using PointNet++ combined with dynamic graph convolution, structural features are extracted for BIM geometric feature extraction. The BERT-specification knowledge graph is used to semantically understand the safety technical briefing records and thus extract the semantic features of the construction log. The TCN+ self-attention module is used to capture the timing pattern before the sudden change in gas concentration to extract the sensor timing features.

4. The BIM-based hoisting construction safety monitoring and early warning system according to claim 1 is characterized in that: The specific steps for extracting BIM geometric features, construction log semantic features, and sensor timing features locally, and combining them with a fast physical field solver to perform lightweight CFD simulations for real-time warnings and localized decision-making are as follows: Fast physical field solver: It adapts to the fluid flow problem in a confined space by simplifying the Navier-Stokes equation and discretizing the equation using the finite volume method. The simplified Navier-Stokes equation is expressed as Where u is the fluid velocity vector, p is the pressure, ρ is the fluid density, v is the kinematic viscosity, and f is the external force term; Lightweight CFD simulation: combining model reduction and parallel computing technology to accelerate the solution; Feature fusion: Fusing BIM geometric features, construction log semantic features, and sensor temporal features; Early warning judgment: Based on the fused features and CFD simulation results, early warning judgment is made using the preset early warning threshold or machine learning model; Localized decision-making: Based on the early warning results and the actual situation on site, a lightweight AI inference engine is used to make localized decisions.

5. The BIM-based hoisting construction safety monitoring and early warning system according to claim 1 is characterized in that: The cloud collaboration layer stores all data based on the BIM digital twin platform and uses the MAML++ algorithm to construct a three-dimensional knowledge graph of "hazard mode-construction characteristics-treatment measures", including the following steps: BIM digital twin platform data storage: The BIM digital twin platform integrates and stores all the data of the hoisting construction; Extracting "hazard pattern-construction characteristics-treatment measures" triplets: Extracting "hazard pattern-construction characteristics-treatment measures" triplets from the full data. The hazard pattern is determined through cluster analysis or expert experience, the construction characteristics are extracted from BIM models, sensor data, and construction logs, and the treatment measures are determined based on safety regulations or historical cases. Constructing a meta-knowledge graph: The inner loop updates the model parameters of a small number of samples on the tasks extracted by the outer loop, which is expressed as Among them, φ is the initial parameter of the meta-model, α is the learning rate of the inner loop, L τi (f φ ) is the loss function on task τi, which measures the performance of the model on the task, f φ The feature extraction network with parameter φ is used to extract features from BIM models and sensor data, and the cross-task metamodel parameter update in the outer loop is expressed as Where β is the learning rate of the outer loop, p(τ) is the task distribution, and tasks are randomly selected from historical projects. is the parameter φ′ after the inner loop update i Under this condition, the loss function on task τi enables the meta-model to quickly adapt to new tasks; Dynamic update: Information is transmitted between nodes through the message passing mechanism of the graph neural network to update the embedding vector of the node. Incremental learning is used to update the embedding vector and edge weight of related nodes when new task data arrives to adapt to new construction scenarios and danger patterns.

6. A BIM-based hoisting construction safety monitoring and early warning system according to claim 5, characterized in that: The physical mechanism data enhancement engine couples the multi-physics field coupling model and the physical constraint generative adversarial network to generate virtual data that conforms to mass conservation and energy conservation, and mixes the virtual data with the real data for training, including the following steps: Multi-physics coupling model construction: By establishing the hydrogen diffusion equation and coupling the turbulence model, the computational fluid dynamics solver is used to dynamically simulate the hydrogen leakage and diffusion process in the confined space. The hydrogen diffusion equation is expressed as Among them, C is the hydrogen concentration, u is the flow field velocity vector, calculated by the turbulence model, D is the diffusion coefficient, S is the source term, and at the same time, the structural mechanics equation is introduced and combined with finite element analysis to calculate the deformation of the lifting equipment under wind load. The structural mechanics equation is expressed as Where EI is the bending stiffness, y is the deflection, μ is the mass linear density, and q(x) is the distributed load, including wind load and the gravity of the hoisted object; Data fusion and feature extraction: Mapping the CFD and FEM calculation results to the BIM model to generate a spatiotemporal feature matrix containing multi-physics field information; Physically Constrained Generative Adversarial Network (PC-GAN) training: The generator uses random noise and physical parameters as input, combined with a deep neural network, to generate data that conforms to mass and energy conservation. The discriminator then evaluates the authenticity of the data and uses a loss function that includes a physical consistency term to screen the generated data and eliminate data that violates conservation laws. Hybrid training: Combine physical simulation data, adversarially generated data, and real data to build a dataset containing various working conditions.

7. The BIM-based hoisting construction safety monitoring and early warning system according to claim 1 is characterized in that: The spatiotemporal adaptive threshold evolutionary algorithm encodes spatiotemporal context through a graph attention network and a Transformer, including the following steps: BIM model graph structuring: Build a dynamic graph structure based on the BIM model, abstract entities into nodes, abstract spatial relationships between entities into edges, initialize node feature vectors and edge weights, and convert construction scenarios into structured graph data; Spatial embedding vector output: A graph attention network is used to dynamically learn the importance between entities by calculating attention coefficients, aggregating neighbor node information and updating node features to generate a spatial embedding vector. Sensor sequence input: Historical sensor sequences retain temporal information by adding learnable positional encodings; Temporal embedding vector output: The Transformer's self-attention mechanism is used to calculate query, key, and value vectors, capturing long-term dependencies in sensor data and generating a temporal embedding vector. Cross-attention: Using the spatial embedding vector as the query and the temporal embedding vector as the key, the attention score is calculated and the features are fused to generate a spatiotemporal context embedding vector.

8. The BIM-based hoisting construction safety monitoring and early warning system according to claim 1 is characterized in that: The specific steps of using deep reinforcement learning to dynamically optimize the alarm threshold are: Step 1: The policy network receives the current state vector as input, and the policy network calculates the output action probability distribution based on the current state vector through the neural network; Step 2: Select specific actions based on the distribution using random sampling or greedy strategy and execute them, and adjust the threshold; Step 3: After executing the action, observe the new state and reward value; Step 4: Calculate the advantage function to evaluate the quality of the action; Step 5: Update the policy network parameters φ based on the new state, reward, and advantage function to maximize the policy loss.

Images