Intelligent hoisting construction risk data processing system based on BIM

Abstract

The invention discloses a BIM-based hoisting construction risk data intelligent processing system, and the system comprises an Internet of Things sensing layer which carries out the sensing through the deployment of a multi-mode sensor network; the BIM-IoT data fusion layer is used for directly embedding sensor time sequence data into a BIM component attribute set by using an IFC + + spatio-temporal extension framework, and dynamically updating the sensor time sequence data through a native data pipeline; the spatio-temporal feature extraction layer is used for fusing and enhancing BIM topological features and IoT time sequence features by adopting an ST-Encoder network architecture through double-flow feature extraction and a cross-modal attention fusion mechanism in combination with prior information; the hybrid reasoning decision-making layer is used for fusing a PINN-Transformer coupling model and an FEM-Neural joint calculation mechanism, and predicting stress through physical constraint enhancement, nonlinear local modeling and an alternate execution strategy; and the digital twinning execution layer is used for real-time rendering, historical playback, prediction simulation, component-level information superposition and virtual security fencing of the construction process.

Description

Technical Field

[0001] The present invention relates to the technical field of hoisting construction, and in particular to a BIM-based intelligent processing system for hoisting construction risk data. Background Art

[0002] In the field of pumped storage power station projects, with the increase in the capacity of single units and the expansion of project construction scale, the lifting and construction of large components of pumped storage power station units, such as generator stators, rotors, turbine runners, top covers, water inlet valves and GCBs (generator outlet switches), as key links, faces unprecedented challenges in risk management. The maturity of BIM (Building Information Modeling) technology provides a technical foundation for intelligent risk management, which has given rise to the demand for a BIM-based intelligent processing system for risk data of large component lifting construction of pumped storage power station units. BIM technology uses three-dimensional digital modeling and full life cycle data integration. On the one hand, it realizes the geometric, physical and rule information association of large components of units, plant structure space, etc. through parametric models, providing an accurate data foundation for risk analysis. On the other hand, combined with Internet of Things technology, it can access sensor data (such as strain, displacement, wind speed, lifting speed, etc.) in real time to build a two-way mapping between physical entities and digital models.

[0003] In the existing technology, pure machine learning models (such as LSTM) lack explicit constraints on the mechanical principles of large-scale hoisting of pumped-storage power station units and may output physically unreliable results. Therefore, a BIM-based intelligent processing system for hoisting construction risk data is proposed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the existing technology and propose a BIM-based intelligent processing system for lifting construction risk data.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A BIM-based intelligent processing system for risk data of hoisting construction, comprising:

[0007] IoT perception layer: By deploying a multimodal sensor network, including fiber Bragg grating strain gauges, laser scanners, micro weather stations, crane encoders, UWB base stations, and 4K panoramic cameras, and using 5G+LoRaWAN hybrid networking technology, it can perceive structural health, geometry, environmental parameters, equipment status, personnel location, and video surveillance.

[0008] BIM-IoT data fusion layer: Leveraging the IFC++ spatiotemporal extension framework, sensor time series data is directly embedded into BIM component attribute sets and dynamically updated through native data pipelines, eliminating middleware conversion overhead.

[0009] Spatiotemporal feature extraction layer: Using the ST-Encoder network architecture, dual-stream feature extraction and cross-modal attention fusion mechanism, combined with prior information, fuse and enhance BIM topological features and IoT temporal features. The dual-stream feature extraction includes a BIM branch and an IoT branch. The BIM branch uses a graph attention network to process component topological relationships, and the IoT branch uses a temporal convolutional network to capture dynamic response patterns.

[0010] Hybrid reasoning decision layer: Integrates the PINN-Transformer coupling model and the FEM-Neural joint calculation mechanism to predict stress through physical constraint enhancement, nonlinear constitutive modeling, and alternating execution strategies;

[0011] Digital twin execution layer: Through the Unreal Engine and AR-assisted decision-making system, real-time rendering, historical playback, predictive simulation, component-level information overlay and virtual safety fences of the construction process are provided.

[0012] The above technical solution further includes:

[0013] Furthermore, the fiber Bragg grating strain gauge performs structural health monitoring, the laser scanner performs geometric shape perception, the crane encoder performs equipment status monitoring, the UWB base station performs personnel positioning, and the camera performs video monitoring. The structural health, equipment status and personnel positioning data are key data transmitted via 5G uplink, and the remaining data are regular data transmitted via the LoRaWAN wide area network.

[0014] Furthermore, the BIM-IoT data fusion layer receives data transmitted by the IoT perception layer and performs secondary verification on the received data. For critical data, if it does not arrive within the preset time, the retransmission mechanism is triggered and an alarm is generated. For conventional data, a delay threshold is set, the actual delay value is allowed to be less than the delay threshold, and the missing values ​​are supplemented through the data interpolation algorithm.

[0015] Furthermore, the BIM-IoT data fusion layer uses the IFC++ spatiotemporal extension framework to directly embed sensor time series data into the BIM component attribute set and dynamically update it through the native data pipeline.

[0016] IFC++ spatiotemporal extension framework design: Define the IfcStructuralLoadTimeSeries extension entity, build spatiotemporal encoding rules, and define the spatial binding relationship between sensors and BIM components;

[0017] Data embedding and binding: Collect time series data through sensors and dynamically bind the collected time series data to the corresponding BIM component attribute set using the AIXM standard;

[0018] Native data pipeline implementation: The raw data is pre-processed using a data cleansing algorithm deployed on edge computing nodes. The cleansed data is then converted into a format recognizable by the BIM model via the OPC UA protocol. The IFC update engine monitors data changes in real time and dynamically updates the BIM model's attribute set.

[0019] Dynamic update mechanism: The data update frequency is dynamically adjusted according to the different construction stages, and a timestamp alignment algorithm is used to synchronize the BIM model with the site status in real time. At the same time, the NTP protocol is used to synchronize the device clock to further control the synchronization error.

[0020] Furthermore, the spatiotemporal feature extraction layer utilizes a graph attention network to process the specific steps of the component topological relationship;

[0021] Graph construction: Abstract the components in the BIM model into graph nodes, abstract the connection relationships between components into edges, initialize the node feature vectors containing geometric properties, material properties and initial state, and establish a mathematical expression of the component topological relationship;

[0022] Attention coefficient calculation: Use a shared weight matrix to linearly transform node features, concatenate projected features, and calculate attention weights through a feedforward neural network. Softmax function is then used to normalize the weights and dynamically quantify the degree of mutual influence between components.

[0023] Feature aggregation: The projected features of neighboring nodes are weighted summed according to the normalized attention weights, and new features are generated through a nonlinear activation function, so that each node can integrate the information of its neighboring nodes;

[0024] Multi-head attention: Compute multiple sets of independent attention weights in parallel and concatenate or average the results.

[0025] Furthermore, the spatiotemporal feature extraction layer uses a temporal convolutional network to capture the specific steps of the dynamic response pattern:

[0026] Causal convolution: Use a one-dimensional convolution kernel to process the input sequence in time step order, and perform time recursion through the formula, which is expressed as where K∈R k×1 is a one-dimensional convolution kernel with a size of k, X t is the sensor input value at time t, is the output feature after convolution;

[0027] Dilated convolution: Introducing the dilation factor d to control the sampling interval, expanding the receptive field to a pyramid scale through sparse sampling to capture long-term dependencies;

[0028] Residual connection: skip-add the convolution output to the original input to alleviate the vanishing gradient and enhance nonlinear expression capabilities;

[0029] Dynamic response pattern capture: Stacking multiple layers of TCN for multi-scale feature extraction, combined with the attention mechanism to automatically learn mutation detection, periodicity identification and trend prediction, and by integrating with BIM topological features, real-time risk monitoring, predictive maintenance and decision support enhancement are carried out.

[0030] Furthermore, the spatiotemporal feature extraction layer fuses and enhances the BIM topological features and IoT temporal features, including the following steps:

[0031] Feature mapping: Project the topological features output by the BIM branch and the temporal features output by the IoT branch to the same dimension using the learnable weight matrices W_Q, W_K, and W_V∈R^(F′×d). Q_BIM = W_Q·H_GAT, K_IoT = W_K·H_TCN, and V_IoT = W_V·H_TCN, where Q_BIM is the query vector of the BIM feature, K_IoT and V_IoT are the key and value vectors of the IoT feature, respectively, and d is the dimension after projection.

[0032] Attention calculation: Calculate the correlation between BIM features and IoT features, fuse the features through the attention mechanism, calculate the correlation using dot product attention, and normalize it through the Softmax function;

[0033] Prior information injection: The prior information is embedded into a learnable vector and combined with the fusion feature through a gating mechanism. The prior information is represented as a learnable vector E∈R^(C×d), where C is the number of categories of the prior information. The gating mechanism G is used to control the degree of fusion between the prior information and the fusion feature.

[0034] Furthermore, the hybrid reasoning decision layer performs the following specific steps to predict stress:

[0035] Physical constraint enhancement: Through physical constraint enhancement technology, the equilibrium equations in continuum mechanics are Loss functions incorporated into neural networks Where α and β are weight coefficients, L_MSE is the mean square error between the predicted stress and the true stress, forcing the network prediction results to satisfy the laws of physics;

[0036] Nonlinear constitutive modeling: Constructing neural network driven material cards, through neural network Approximating the stress-strain relationship of a material Complete simulation of nonlinear material behavior;

[0037] FEM-Neural Co-calculation Mechanism: An alternating execution strategy is adopted to alternately execute the finite element method and neural network prediction through the FEM-Neural Co-calculation mechanism. In each iteration, the FEM solver updates the stress field based on the current boundary conditions and material parameters, while the neural network predicts the material parameters or boundary conditions until the convergence criteria are met.

[0038] The present invention has the following beneficial effects:

[0039] In this paper, the PINN-Transformer coupling model and FEM-Neural combined calculation mechanism are used to reduce stress prediction errors to meet engineering precision requirements. By integrating the equilibrium equation into the loss function, the prediction results are ensured to conform to the principles of continuum mechanics, avoiding the physically unreasonable results that may occur in purely data-driven models. Through the IFC++ spatiotemporal extension framework, sensor time series data is directly embedded in the BIM component attribute set, avoiding the data delays and errors caused by traditional middleware conversion and ensuring real-time synchronization of the BIM model with the site status. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a system block diagram of a BIM-based intelligent processing system for lifting construction risk data proposed in the present invention. DETAILED DESCRIPTION

[0041] 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.

[0042] See also Figure 1 As shown, the present invention is a BIM-based intelligent processing system for hoisting construction risk data, comprising:

[0043] IoT perception layer: By deploying a multimodal sensor network, including fiber Bragg grating strain gauges, laser scanners, micro weather stations, crane encoders, UWB base stations, and 4K panoramic cameras, and using 5G+LoRaWAN hybrid networking technology, critical data (security-related) is transmitted via 5G uplink, while conventional data is covered by LoRaWAN over a wide area. This layer senses structural health, geometry, environmental parameters, equipment status, personnel location, and video surveillance, ensuring that critical data is transmitted with low latency and high reliability.

[0044] BIM-IoT data fusion layer: Leveraging the IFC++ Spatiotemporal Extension Framework (ST-IFC), sensor time series data is directly embedded in BIM component attribute sets and dynamically updated through native data pipelines, eliminating middleware conversion overhead and ensuring real-time synchronization between the BIM model and field status.

[0045] Spatiotemporal feature extraction layer: Using the ST-Encoder network architecture, dual-stream feature extraction and cross-modal attention fusion mechanisms are used to fuse and enhance BIM topological features and IoT temporal features, combined with prior information such as bridge type codes and construction stage labels. The dual-stream feature extraction includes a BIM branch and an IoT branch. The BIM branch uses a graph attention network (GAT) to process component topological relationships, while the IoT branch uses a temporal convolutional network (TCN) to capture dynamic response patterns.

[0046] Hybrid reasoning decision layer: Integrates the PINN-Transformer coupling model and the FEM-Neural joint calculation mechanism to predict stress through physical constraint enhancement, nonlinear constitutive modeling, and alternating execution strategies;

[0047] Digital twin execution layer: Through Unreal Engine (integrated with Unity Reflect + Omniverse platform) and AR decision-making support system (HoloLens 2), real-time rendering, historical playback, predictive simulation, component-level information overlay and virtual safety fences of the construction process are achieved.

[0048] In one embodiment, the fiber Bragg grating strain gauge (resolution 1 με) performs structural health monitoring, the laser scanner (point cloud density 5000 points / ㎡) performs geometric perception, the micro weather station (synchronously acquires wind speed / temperature and humidity / rainfall) performs environmental parameter collection, the crane encoder (angle resolution 0.01°) performs equipment status monitoring, the UWB base station (positioning accuracy 10 cm) performs personnel positioning, and the 4K panoramic camera (supporting AI behavior recognition) performs video monitoring. The structural health, equipment status and personnel positioning data are key data transmitted via 5G uplink, and the remaining data are regular data transmitted via the LoRaWAN wide area network.

[0049] In one embodiment, the BIM-IoT data fusion layer receives data transmitted by the IoT perception layer and performs a secondary check on the received data. For critical data, if it does not arrive within a preset time, a retransmission mechanism is triggered and an alarm is generated. For regular data, a delay threshold is set, the actual delay value is allowed to be less than the delay threshold, and missing values ​​are supplemented through a data interpolation algorithm.

[0050] In one embodiment, the BIM-IoT data fusion layer uses the IFC++ Spatiotemporal Extension Framework (ST-IFC) to directly embed sensor time series data into BIM component attribute sets and dynamically update them through native data pipelines.

[0051] IFC++ spatiotemporal extension framework design: This defines the IfcStructuralLoadTimeSeries extension entity. This extension entity inherits the structured data storage capabilities of traditional IFC attributes and directly encapsulates sensor time series data by embedding fields such as timestamps, sensor values, units, and unique identifiers. It also establishes spatiotemporal encoding rules and defines the spatial binding relationship between sensors and BIM components, ensuring the precise positioning of data in geometric space.

[0052] Data embedding and binding: Time series data is collected through sensors such as fiber Bragg grating strain gauges, and the collected time series data is dynamically bound to the attribute set of the corresponding BIM component (such as IfcBeam) using the AIXM standard, realizing a direct association between data and model components;

[0053] Native data pipeline implementation: Data cleaning algorithms (such as wavelet denoising and outlier removal) deployed on edge computing nodes pre-process raw data. The cleaned data is then converted into a format recognizable by the BIM model through the OPC UA protocol. The IFC update engine monitors data changes in real time and dynamically updates the BIM model's attribute set, ensuring seamless data transmission and real-time model updates.

[0054] Dynamic update mechanism: The data update frequency is dynamically adjusted according to the different construction stages, and a timestamp alignment algorithm is used to synchronize the BIM model with the site status in real time. At the same time, the NTP protocol is used to synchronize device clocks to further control synchronization errors. Compared with traditional methods, the ST-IFC native binding method eliminates the overhead of middleware conversion by directly embedding data into the IFC attribute set, significantly reducing the delay in data format conversion, coordinate system conversion, and attribute association.

[0055] In one embodiment, the spatiotemporal feature extraction layer uses a graph attention network (GAT) to process the specific steps of the component topological relationship;

[0056] Graph construction: Abstract the components in the BIM model into graph nodes, abstract the connection relationships between components into edges, initialize the node feature vectors containing geometric properties, material properties and initial state, and establish a mathematical expression of the component topological relationship;

[0057] Attention coefficient calculation: Node features are linearly transformed using a shared weight matrix, projected features are concatenated, and attention weights are calculated using a feedforward neural network. The Softmax function is then used to normalize the weights, dynamically quantifying the degree of mutual influence between components and giving higher attention to key connections.

[0058] Feature aggregation: The projected features of neighboring nodes are weighted and summed according to the normalized attention weights, and new features are generated through a nonlinear activation function. This allows each node to integrate the information of its neighboring nodes, enhancing the expressiveness of the features.

[0059] Multi-head attention: Multiple sets of independent attention weights are calculated in parallel, and the results are concatenated or averaged, thereby stabilizing the training process, capturing richer topological patterns, and improving the robustness of the model. In application scenarios, GAT provides real-time and accurate basis for risk assessment by analyzing stress propagation paths, locating weak links in topological structures, and dynamically adapting to changes in construction progress.

[0060] In one embodiment, the spatiotemporal feature extraction layer uses a temporal convolutional network (TCN) to capture dynamic response patterns:

[0061] Causal convolution: Use a one-dimensional convolution kernel to process the input sequence in time step order, ensuring that the current prediction depends only on historical data to avoid future information leakage, and perform time recursion through the formula, which is expressed as where K∈R k×1 is a one-dimensional convolution kernel with a size of k, X t is the sensor input value at time t (such as crane load, structural strain), is the output feature after convolution;

[0062] Dilated convolution: Introducing the dilation factor d to control the sampling interval (for example, when d = 2, the receptive field covers time t, t-2, and t-4). Through sparse sampling, the receptive field is expanded to a pyramid scale (for example, when k = 3 and d = 4, the receptive field reaches 7 time steps) to capture long-term dependencies.

[0063] Residual connection: skip-add the convolution output to the original input to alleviate the vanishing gradient and enhance nonlinear expression capabilities;

[0064] Dynamic response pattern capture: Stacking multiple layers of TCN for multi-scale feature extraction (such as capturing minute-level load mutations at the bottom layer and identifying hour-level temperature effects at the top layer), combined with the attention mechanism to automatically learn mutation detection (such as a 5% sudden increase in load triggering an early warning), periodic identification (such as daily thermal expansion and contraction patterns) and trend prediction (such as strain accumulation effects). By fusing with BIM topological features (such as coupling time series patterns with the stiffness of steel box girder connections), real-time risk monitoring (such as overload suspension), predictive maintenance (such as wire rope fatigue prediction) and decision support enhancement (such as AR display of stress evolution in the next 15 minutes) are carried out.

[0065] In one embodiment, the spatiotemporal feature extraction layer fuses and enhances BIM topological features and IoT temporal features, including the following steps:

[0066] Feature mapping: The topological features output by the BIM branch and the temporal features output by the IoT branch are projected onto the same dimension using a learnable weight matrix W_Q, W_K, and W_V∈R^(F′×d). This is expressed as Q_BIM = W_Q·H_GAT, K_IoT = W_K·H_TCN, and V_IoT = W_V·H_TCN, where Q_BIM is the query vector of the BIM feature, K_IoT and V_IoT are the key and value vectors of the IoT feature, respectively, and d is the dimension after projection.

[0067] Attention calculation: Calculate the correlation between BIM features and IoT features, fuse the features through the attention mechanism, calculate the correlation using dot product attention, and normalize it through the Softmax function;

[0068] Prior information injection: Prior information such as the type code of large components and the marking of the construction stage of the pumped storage power station unit is embedded into a learnable vector and combined with the fusion feature through a gating mechanism to enhance the expressiveness and interpretability of the model. The prior information such as the type code of large components and the marking of the construction stage of the pumped storage power station unit is represented as a learnable vector E∈R^(C×d), where C is the number of categories of the prior information. The gating mechanism G is used to control the degree of fusion of the prior information and the fusion feature.

[0069] In one embodiment, the hybrid reasoning decision layer performs the following specific steps to predict stress:

[0070] Physical constraint enhancement: Through physical constraint enhancement technology, the equilibrium equations in continuum mechanics are Loss functions incorporated into neural networks Where α and β are weight coefficients, L_MSE is the mean square error between the predicted stress and the true stress, which forces the network prediction results to satisfy the physical laws, thus ensuring the physical consistency of the stress prediction;

[0071] Nonlinear constitutive modeling: Constructing a neural network driven material card (NeuralMaterialCard) through a neural network Approximating the stress-strain relationship of a material Complete simulation of nonlinear material behavior;

[0072] FEM-Neural Hybrid Calculation Mechanism: This system employs an alternating execution strategy, alternately executing the finite element method (FEM) and neural network prediction. In each iteration, the FEM solver updates the stress field based on the current boundary conditions and material parameters, while the neural network predicts the material parameters or boundary conditions until the convergence criteria are met. This combines the numerical stability of the FEM with the nonlinear fitting capabilities of the neural network, significantly improving computational efficiency and prediction accuracy.

[0073] Real-time sensor data (such as strain, displacement, temperature, etc.) is obtained from the BIM-IoT data fusion layer, and the geometric information, material properties and boundary conditions of the components are obtained from the BIM model. Then, the ST-Encoder network is used to extract spatiotemporal features, providing a comprehensive data basis for stress prediction. Based on the extracted features, the stress field is predicted through the PINN-Transformer coupling model and the FEM-Neural joint calculation mechanism, and the safety status of the component is evaluated according to the prediction results. If the stress exceeds the safety threshold, an alarm is triggered and optimization suggestions are generated. The prediction results and actual construction data are fed back to the BIM-IoT data fusion layer, the BIM model and sensor network are updated, and the parameters of the PINN-Transformer coupling model and the FEM-Neural joint calculation mechanism are optimized, achieving continuous optimization of the model and continuous improvement of prediction accuracy.

[0074] 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 intelligent processing system for hoisting construction risk data, characterized by: include: IoT perception layer: By deploying a multimodal sensor network, including fiber Bragg grating strain gauges, laser scanners, crane encoders, UWB base stations, and cameras, and using 5G+LoRaWAN hybrid networking technology, it can perceive structural health, geometry, environmental parameters, equipment status, personnel location, and video surveillance. BIM-IoT data fusion layer: Leveraging the IFC++ spatiotemporal extension framework, sensor time series data is directly embedded into BIM component attribute sets and dynamically updated through native data pipelines, eliminating middleware conversion overhead. Spatiotemporal feature extraction layer: Using the ST-Encoder network architecture, dual-stream feature extraction and cross-modal attention fusion mechanism, combined with prior information, fuse and enhance BIM topological features and IoT temporal features. The dual-stream feature extraction includes a BIM branch and an IoT branch. The BIM branch uses a graph attention network to process component topological relationships, and the IoT branch uses a temporal convolutional network to capture dynamic response patterns. Hybrid reasoning decision layer: Integrates the PINN-Transformer coupling model and the FEM-Neural joint calculation mechanism to predict stress through physical constraint enhancement, nonlinear constitutive modeling, and alternating execution strategies; Digital twin execution layer: Through the Unreal Engine and AR-assisted decision-making system, real-time rendering, historical playback, predictive simulation, component-level information overlay and virtual safety fences of the construction process are provided.

2. The BIM-based intelligent processing system for lifting construction risk data according to claim 1 is characterized in that: The fiber Bragg grating strain gauge is used for structural health monitoring, the laser scanner is used for geometric shape perception, the crane encoder is used for equipment status monitoring, the UWB base station is used for personnel positioning, and the camera is used for video monitoring. The structural health, equipment status and personnel positioning data are key data transmitted via 5G uplink, and the remaining data are regular data transmitted via the LoRaWAN wide area network.

3. The BIM-based intelligent processing system for lifting construction risk data according to claim 2 is characterized in that: The BIM-IoT data fusion layer receives data transmitted by the IoT perception layer and performs secondary verification on the received data. For critical data, if it does not arrive within the preset time, the retransmission mechanism is triggered and an alarm is generated. For regular data, a delay threshold is set, the actual delay value is allowed to be less than the delay threshold, and the missing values ​​are supplemented through the data interpolation algorithm.

4. The BIM-based intelligent processing system for lifting construction risk data according to claim 1 is characterized in that: The BIM-IoT data fusion layer uses the IFC++ spatiotemporal extension framework to directly embed sensor time series data into the BIM component attribute set and dynamically update it through the native data pipeline. IFC++ spatiotemporal extension framework design: Define the IfcStructuralLoadTimeSeries extension entity, build spatiotemporal encoding rules, and define the spatial binding relationship between sensors and BIM components; Data embedding and binding: Collect time series data through sensors and dynamically bind the collected time series data to the corresponding BIM component attribute set using the AIXM standard; Native data pipeline implementation: The raw data is pre-processed using a data cleansing algorithm deployed on edge computing nodes. The cleansed data is then converted into a format recognizable by the BIM model via the OPC UA protocol. The IFC update engine monitors data changes in real time and dynamically updates the BIM model's attribute set. Dynamic update mechanism: The data update frequency is dynamically adjusted according to the different construction stages, and a timestamp alignment algorithm is used to synchronize the BIM model with the site status in real time. At the same time, the NTP protocol is used to synchronize the device clock to further control the synchronization error.

5. The BIM-based intelligent processing system for lifting construction risk data according to claim 1 is characterized in that: The spatiotemporal feature extraction layer uses a graph attention network to process the component topological relationship; Graph construction: Abstract the components in the BIM model into graph nodes, abstract the connection relationships between components into edges, initialize the node feature vectors containing geometric properties, material properties and initial state, and establish a mathematical expression of the component topological relationship; Attention coefficient calculation: Use a shared weight matrix to linearly transform node features, concatenate projected features, and calculate attention weights through a feedforward neural network. Softmax function is then used to normalize the weights and dynamically quantify the degree of mutual influence between components. Feature aggregation: The projected features of neighboring nodes are weighted summed according to the normalized attention weights, and new features are generated through a nonlinear activation function, so that each node can integrate the information of its neighboring nodes; Multi-head attention: Compute multiple sets of independent attention weights in parallel and concatenate or average the results.

6. The BIM-based intelligent processing system for lifting construction risk data according to claim 1 is characterized in that: The specific steps of the spatiotemporal feature extraction layer using a temporal convolutional network to capture dynamic response patterns are as follows: Causal convolution: Use a one-dimensional convolution kernel to process the input sequence in time step order, and perform time recursion through the formula, which is expressed as where K∈R k×1 is a one-dimensional convolution kernel with a size of k, X t is the sensor input value at time t, is the output feature after convolution; Dilated convolution: Introducing the dilation factor d to control the sampling interval, expanding the receptive field to a pyramid scale through sparse sampling to capture long-term dependencies; Residual connection: skip-add the convolution output to the original input to alleviate the vanishing gradient and enhance nonlinear expression capabilities; Dynamic response pattern capture: Stacking multiple layers of TCN for multi-scale feature extraction, combined with the attention mechanism to automatically learn mutation detection, periodicity identification and trend prediction, and by integrating with BIM topological features, real-time risk monitoring, predictive maintenance and decision support enhancement are carried out.

7. The BIM-based intelligent processing system for lifting construction risk data according to claim 1 is characterized in that: The spatiotemporal feature extraction layer fuses and enhances BIM topological features and IoT temporal features, including the following steps: Feature mapping: The topological features output by the BIM branch and the temporal features output by the IoT branch are projected onto the same dimension using a learnable weight matrix W_Q, W_K, and W_V∈R^(F′×d). This is expressed as Q_BIM = W_Q·H_GAT, K_IoT = W_K·H_TCN, and V_IoT = W_V·H_TCN, where Q_BIM is the query vector of the BIM feature, K_IoT and V_IoT are the key and value vectors of the IoT feature, respectively, and d is the dimension after projection. Attention calculation: Calculate the correlation between BIM features and IoT features, fuse the features through the attention mechanism, calculate the correlation using dot product attention, and normalize it through the Softmax function; Prior information injection: The prior information is embedded into a learnable vector and combined with the fusion feature through a gating mechanism. The prior information is represented as a learnable vector E∈R^(C×d), where C is the number of categories of the prior information. The gating mechanism G is used to control the degree of fusion between the prior information and the fusion feature.

8. The BIM-based intelligent processing system for lifting construction risk data according to claim 1 is characterized in that: The specific steps of the hybrid reasoning decision layer to predict stress are as follows: Physical constraint enhancement: Through physical constraint enhancement technology, the equilibrium equations in continuum mechanics are The loss function L = αL_MSE+ integrated into the neural network Where α and β are weight coefficients, L_MSE is the mean square error between the predicted stress and the true stress, forcing the network prediction results to satisfy the laws of physics; Nonlinear constitutive modeling: Constructing neural network driven material cards, through neural network Approximating the stress-strain relationship of a material Complete simulation of nonlinear material behavior; FEM-Neural Co-calculation Mechanism: An alternating execution strategy is adopted to alternately execute the finite element method and neural network prediction through the FEM-Neural Co-calculation mechanism. In each iteration, the FEM solver updates the stress field based on the current boundary conditions and material parameters, while the neural network predicts the material parameters or boundary conditions until the convergence criteria are met.

Images