Construction safety risk intelligent early warning method based on multi-source data fusion

Abstract

The application discloses a construction safety risk intelligent early warning method based on multi-source data fusion, relates to the technical field of construction safety early warning, and comprises the following steps: constructing a multi-source fusion digital twin base: a digital twin space-time field synchronized with a physical construction site is constructed, and multi-source heterogeneous construction data from a geological survey data source, an environment monitoring data source, a personnel equipment positioning data source and a multi-element physical field monitoring data source are standardized and space-time aligned, are injected into the digital twin space-time field, and form a unified dynamic fusion data base, a space-time graph is constructed to realize risk prediction: based on the dynamic fusion data base, a space-time knowledge graph is constructed with construction risk elements as nodes and space-time causal relationships between elements as edges, and a space-time graph neural network model is used to perform real-time reasoning on the space-time knowledge graph to simulate the dynamic propagation and superposition process of construction safety risks in the digital twin space-time field.

Description

Technical Field

[0001] This invention relates to the field of construction safety early warning technology, and in particular to an intelligent early warning method for construction safety risks based on multi-source data fusion. Background Technology

[0002] Tunnel blasting is a highly efficient construction technique that uses explosives to release energy instantaneously to break rocks and achieve tunnel excavation. It is characterized by high efficiency, low cost, and strong adaptability. However, it is necessary to strictly control safety risks and construction quality. Tunnel blasting involves the explosion of explosives in the blast hole, which generates high-temperature and high-pressure gas and shock waves, causing the rocks to break and be thrown out.

[0003] Traditional blasting construction relies on a single, static data source for safety early warning, neglecting multi-dimensional risk factors such as geological conditions, surrounding rock stress, groundwater, and harmful gases. It cannot perceive and correlate the specific location, activity status, and real-time environmental changes of construction personnel and equipment in real time, resulting in a disconnect between early warning and specific recipients. Therefore, an intelligent early warning method for construction safety risks based on multi-source data fusion is proposed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that rely on a single, static data source, neglect multi-dimensional risk factors such as geological conditions, surrounding rock stress, groundwater, and harmful gases, and are unable to perceive and correlate the specific location, activity status, and real-time environmental changes of construction personnel and equipment in real time, resulting in a disconnect between early warning and specific recipients. Therefore, this invention proposes an intelligent early warning method for construction safety risks based on multi-source data fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The intelligent early warning method for construction safety risks based on multi-source data fusion includes the following steps: S1. Construct a digital twin spatiotemporal field synchronized with the physical construction site, and standardize and spatiotemporally align multi-source heterogeneous construction data from geological exploration data sources, environmental monitoring data sources, personnel and equipment positioning data sources, and multi-dimensional physical field monitoring data sources, and inject them into the digital twin spatiotemporal field to form a unified dynamic fusion data base; S2. Based on the aforementioned dynamic fusion data base, a spatiotemporal knowledge graph is constructed with construction risk elements as nodes and spatiotemporal causal relationships between elements as edges. The spatiotemporal graph neural network model is used to perform real-time reasoning on the spatiotemporal knowledge graph to simulate the dynamic propagation and superposition process of construction safety risks in the digital twin spatiotemporal field, and outputs probabilistic risk prediction results for future time series. S3. Receive the probabilistic risk prediction results and integrate the adaptive early warning threshold model updated based on the federated learning mechanism to generate an intelligent early warning decision that includes risk level, spatiotemporal location and recommended control measures.

[0006] The above technical solution further includes: The construction of the digital twin spatiotemporal field in S1 specifically includes: S1.1: Based on the BIM model and real-scene scanned point cloud data of the construction site, a three-dimensional computable twin model with geometric, physical, and engineering attributes is constructed. The BIM model from the design phase is imported as the basic framework, and based on the real-scene scanned point cloud data collected periodically or on demand, the geometry of the BIM model is dynamically corrected and updated through a point cloud registration algorithm to ensure that the geometric state of the three-dimensional computable twin model is synchronized with the physical construction site. For the structural components and spatial units in the three-dimensional computable twin model, according to their design specifications and material information, physical attribute parameters including density, elastic modulus, and thermal performance are added, as well as engineering attribute parameters including component number, design strength, construction batch, planned and current construction status, so that the model has the computable capability to support physical simulation and engineering logic analysis. S1.2: Deploy and access multi-source heterogeneous construction data. According to monitoring needs, deploy an Internet of Things (IoT) sensor network at the construction site, including ground-penetrating radar, stress and strain sensors, microseismic monitoring instruments, gas concentration sensors, high-precision positioning base stations, and equipment status sensors. Through multi-protocol adapters configured on edge computing gateways or central servers, collect and access multi-source heterogeneous construction data from the sensor network in real time or near real time. The data includes advanced geological forecast data characterizing geological conditions, microclimate and harmful gas monitoring data reflecting environmental conditions in the tunnel, high-precision positioning trajectory data of personnel and equipment tracking dynamic targets, microseismic event sequences and surrounding rock stress and strain monitoring data sensing rock mass response, and IoT data characterizing the status of construction machinery. S1.3: Design unified spatiotemporal coding rules and data interface specifications to clean, transform, and timestamp the multi-source heterogeneous construction data, and map it to the corresponding spatiotemporal coordinates of the three-dimensional computable twin model, forming a fused data field that reflects the dynamics of the entire construction process. Define and implement a unified set of spatiotemporal coding rules, assigning a unique spatial grid code to each smallest spatial voxel or logical component covered by the three-dimensional computable twin model, and adding a precise timestamp based on a unified time reference to all accessed monitoring data streams. Preprocess the raw data according to predefined data cleaning rules, including: Outlier removal: Statistical characteristics of each sensor data stream are calculated based on a sliding time window. Data points that exceed the range of ±3 standard deviations of the historical mean are identified as outliers caused by sensor malfunctions and removed. Short-series data missing information: For data missing segments caused by network packet loss and lasting less than a preset threshold, linear interpolation algorithms or predictions based on autoregressive models are used to fill in the missing data to maintain the continuity of the data sequence. High-frequency noise smoothing filter: High-frequency random noise caused by environmental interference in the monitoring data is processed by a convolutional smoothing filter, which effectively smooths high-frequency fluctuations while preserving the true trend of data change. The cleaned data is associated with the three-dimensional computable twin model based on its spatiotemporal encoding, specifically including: Spatiotemporal coding parsing and matching: Parse the spatiotemporal coding attached to each cleaned data, which includes at least a spatial grid and a timestamp; determine the target geometric entity or logical location corresponding to the data based on the spatial grid ID and the pre-established spatial index mapping table in the 3D computable twin model; Dynamic attribute association and update: Through application programming interface or database connection, data values ​​are dynamically associated as attributes and written into the attribute table of the target geometric entity or logical location in real time, while overwriting or versioning historical data under the same spatiotemporal coordinates; Data field generation and maintenance: Continuously execute the above parsing, matching and association operations so that all multi-source data are attached to the twin model according to their inherent spatiotemporal logic, thereby constructing and maintaining a data field that supports multi-dimensional querying and analysis by time, space and entity type.

[0007] The specific aspects of constructing risk reasoning using spatiotemporal knowledge graphs in S2 include: S2.1: Define the node categories and attributes of construction risk elements; S2.2: Define causal relationship edges and spatiotemporal impact edges between nodes. The causal relationship edges are used to describe the inherent influence logic of geological conditions on construction stability. The spatiotemporal impact edges are used to describe the spatial attenuation and temporal delay effects of dynamic physical processes such as the propagation of blasting vibration waves. The causal relationship edges are defined to express static or quasi-static logical influence relationships, including risk aggravation edges from geological defect units to construction activity units or risk receptor units. Their weights are quantified based on the empirical or theoretical influence of geological conditions on engineering stability. The spatiotemporal impact edges are defined to express the propagation relationship of dynamic physical fields, including vibration generation or load application edges from construction activity units to monitoring physical quantity units or risk receptor units. Their weights are dynamically calculated using a function model that attenuates with distance and has a time delay to describe the spatiotemporal propagation effect of the risk field. S2.3: The structure of the spatiotemporal knowledge graph is used as the input graph structure of the spatiotemporal graph neural network model, and the real-time monitoring data in the dynamic fusion data base is used as node features and edge weights. Through the forward propagation calculation of the neural network, the risk state probability distribution of each spatial unit under multiple construction steps in the future is deduced in real time. The nodes and edges defined in the spatiotemporal knowledge graph are used to construct the initial topology graph of the spatiotemporal graph neural network model. The node vectors are initialized according to their attributes, and the edge weights are initialized according to their relationship type and current weight value. The real-time monitoring data obtained in the dynamic fusion data base is used as the dynamic feature input of the corresponding monitoring physical quantity unit node, and the weights of the relevant spatiotemporal influence edges are dynamically updated according to physical laws or data associations. Spatial dependencies are captured through the multi-layer graph convolution operation of the spatiotemporal graph neural network model, and temporal dependencies are captured by combining temporal convolution or recurrent neural network units. The node state is iteratively updated throughout the graph using a message passing mechanism, and finally the risk state probability distribution of each spatial unit facing different levels of danger under a specified future time sequence is output.

[0008] The risk state probability distribution is used to drive the visualization rendering of the digital twin spatiotemporal field. On the three-dimensional computable twin model, the real-time and future risk levels of each region are overlaid and displayed in the form of dynamic risk heat maps with different colors and transparency. Specifically, this includes the following steps: Visual mapping configuration: A color mapping table and a transparency mapping table are predefined to correspond to the risk probability value range, where high risk probability values ​​are mapped to red tones and high opacity, and low risk probability values ​​are mapped to green tones and low opacity.

[0009] Heatmap generation and overlay: Based on the risk state probability distribution data, query and assign corresponding color and transparency values ​​to each spatial unit in the digital twin spatiotemporal field; call the 3D graphics rendering engine, and use the processed color and transparency data as an independent semi-transparent layer to accurately overlay and render on the corresponding geometric surface of the 3D computable twin model to form a dynamic risk heatmap.

[0010] Dynamic time-series update: When the spatiotemporal graph neural network model outputs a new probability distribution of risk states for future prediction time series, the system automatically updates the visual attributes of the heat map according to steps S4.1 and S4.2, and controls it to periodically change or rotate according to the prediction time series to intuitively present the evolution of risk in the spatiotemporal dimension. It also provides a visual query function for risk tracing and transmission paths, specifically including the following steps: Interaction Triggering and Target Location: In response to the user's selection of a specific high-risk area on the dynamic risk heat map, the system calculates the corresponding three-dimensional spatial location and the management unit to which it belongs based on the screen coordinates of the click.

[0011] Path calculation and visualization: Taking the selected unit as the endpoint, based on the node association relationships stored in the spatiotemporal knowledge graph, reverse query and calculate all possible causal relationship edges and spatiotemporal influence edges pointing to the endpoint to form one or more risk transmission chains; on the three-dimensional computable twin model, the transmission chains are animated in the form of highlighted three-dimensional arrow lines or particle flow.

[0012] Linked information display: While visualizing the path, the sidebar of the interactive interface simultaneously displays the detailed attributes of the risk source nodes involved in the path, the associated historical monitoring data curves, and related early warning records, completing a full risk tracing analysis.

[0013] The specific operation of the federated learning mechanism in S3 includes: S3.1: Within the early warning system of each local construction project, the localized adaptive early warning threshold model is trained and fine-tuned using locally generated multi-source heterogeneous construction data and early warning feedback results. Temporal feature segments related to risk events are extracted from the multi-source heterogeneous construction data, and these segments are combined with the actual risk occurrence, early warning response timeliness, and handling effect in historical early warning records to form labeled training sample pairs. Using these training sample pairs, the localized adaptive early warning threshold model is periodically trained through a supervised learning algorithm to optimize its model parameters. The training aims to minimize the difference between the model's early warning results and the actual risk situation, and dynamically adjust the early warning sensitivity weights for different risk types based on the early warning feedback results. S3.2: Periodically encrypt and upload the key parameters of each local model and execute a secure model aggregation algorithm to generate a global risk warning model with strong generalization ability. When the aggregation cycle is triggered, the early warning system of each local construction project uses homomorphic encryption or secure multi-party computation technology to encrypt the gradient or parameter update of the model parameters that need to be uploaded. The encrypted parameter data is uploaded to the central coordination server. Without decrypting the original data of each participant, the server performs aggregation calculation on all received encrypted parameters by executing the federated averaging algorithm or the optimization algorithm based on secure aggregation, and generates a new set of global model parameters that integrate the experience of multiple projects, thereby constructing the global risk warning model with strong generalization ability. S3.3: The aggregated global risk warning model parameters are distributed to each local project, and the warning model knowledge of each project is shared and evolved together. The central coordination server distributes the aggregated global risk warning model parameters to the warning systems of all local construction projects participating in federated learning. The warning systems of each local project receive the global parameters and integrate them with the local model as the initial parameters or prior knowledge for the next round of local model training and fine-tuning. This allows each local project to retain its adaptability to specific scenarios while continuously absorbing common risk knowledge from other projects, thereby achieving the collaborative improvement and co-evolution of the warning model capabilities of all participating projects.

[0014] The intelligent early warning decision is converted into executable control instructions, including: automatically triggering interlocking shutdown control of construction machinery in a specific area based on risk level and spatiotemporal location; sending directional sound and light alarms and evacuation guidance to smart wearable devices of personnel in the affected area; and pushing control suggestions to the construction management platform.

[0015] The self-optimizing loop of the parallel system of the federated learning mechanism collects feedback data on the actual intervention effects of historical early warning decisions and uses it as new training samples to periodically fine-tune the parameters of the spatiotemporal graph neural network model and correct the weights of the relationships between nodes in the spatiotemporal knowledge graph.

[0016] The node categories include geological defect units, construction activity units, monitoring physical quantity units, and risk receptor units. Geological defect unit nodes are defined as geological anomalies that characterize faults, densely jointed zones, weak interlayers, or groundwater-rich areas. Their attributes include defect type, spatial orientation, scale, and strength reduction factor. Construction activity unit nodes are defined as specific procedures such as drilling, charging, blasting, excavation, and support. Their attributes include activity type, planned and actual spatiotemporal range, and construction parameters. Monitoring physical quantity unit nodes are defined as monitoring points for vibration velocity, stress value, displacement, and gas concentration. Their attributes include monitoring type, real-time and historical time-series data, and early warning threshold. Risk receptor unit nodes are defined as construction personnel, key equipment, or important buildings. Their attributes include receptor type, spatial location, status, and vulnerability level.

[0017] The present invention has the following beneficial effects: In this invention, a multi-source fusion digital twin base is constructed to integrate heterogeneous data from multiple sources, such as geology, environment, equipment status, personnel location, and structural response, under a unified spatiotemporal reference. This allows for the complete and real-time characterization of complex risk scenarios. The system uses a spatiotemporal knowledge graph to express risk logic and employs a spatiotemporal graph neural network for simulation and deduction. Before blasting is carried out, the system can predict changes in physical fields such as vibration waves and stress redistribution, as well as the chain reactions they may trigger, achieving early warning. Furthermore, the intelligent early warning decision output not only includes risk level and location but also directly associates recommended control measures. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of the intelligent early warning method for construction safety risks based on multi-source data fusion proposed in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, the intelligent early warning method for construction safety risks based on multi-source data fusion proposed in this invention includes the following steps: S1. Construct a digital twin spatiotemporal field synchronized with the physical construction site, and standardize and spatiotemporally align multi-source heterogeneous construction data from geological exploration data sources, environmental monitoring data sources, personnel and equipment positioning data sources, and multi-dimensional physical field monitoring data sources, and inject them into the digital twin spatiotemporal field to form a unified dynamic fusion data base; S2. Based on the aforementioned dynamic fusion data base, a spatiotemporal knowledge graph is constructed with construction risk elements as nodes and spatiotemporal causal relationships between elements as edges. The spatiotemporal graph neural network model is used to perform real-time reasoning on the spatiotemporal knowledge graph to simulate the dynamic propagation and superposition process of construction safety risks in the digital twin spatiotemporal field, and outputs probabilistic risk prediction results for future time series. S3. Receive the probabilistic risk prediction results and integrate the adaptive early warning threshold model updated based on the federated learning mechanism to generate an intelligent early warning decision that includes risk level, spatiotemporal location and recommended control measures.

[0021] The construction of the digital twin spatiotemporal field in S1 specifically includes: S1.1: Based on the BIM model and real-scene scanned point cloud data of the construction site, a three-dimensional computable twin model with geometric, physical, and engineering attributes is constructed. The BIM model from the design phase is imported as the basic framework, and based on the real-scene scanned point cloud data collected periodically or on demand, the geometry of the BIM model is dynamically corrected and updated through a point cloud registration algorithm to ensure that the geometric state of the three-dimensional computable twin model is synchronized with the physical construction site. For the structural components and spatial units in the three-dimensional computable twin model, according to their design specifications and material information, physical attribute parameters including density, elastic modulus, and thermal performance are added, as well as engineering attribute parameters including component number, design strength, construction batch, planned and current construction status, so that the model has the computable capability to support physical simulation and engineering logic analysis. S1.2: Deploy and access multi-source heterogeneous construction data. According to monitoring needs, deploy an Internet of Things (IoT) sensor network at the construction site, including ground-penetrating radar, stress and strain sensors, microseismic monitoring instruments, gas concentration sensors, high-precision positioning base stations, and equipment status sensors. Through multi-protocol adapters configured on edge computing gateways or central servers, collect and access multi-source heterogeneous construction data from the sensor network in real time or near real time. The data includes advanced geological forecast data characterizing geological conditions, microclimate and harmful gas monitoring data reflecting environmental conditions in the tunnel, high-precision positioning trajectory data of personnel and equipment tracking dynamic targets, microseismic event sequences and surrounding rock stress and strain monitoring data sensing rock mass response, and IoT data characterizing the status of construction machinery. S1.3: Design unified spatiotemporal coding rules and data interface specifications to clean, transform, and timestamp the multi-source heterogeneous construction data, and map it to the corresponding spatiotemporal coordinates of the three-dimensional computable twin model, forming a fused data field that reflects the dynamics of the entire construction process. Define and implement a unified set of spatiotemporal coding rules, assigning a unique spatial grid code to each smallest spatial voxel or logical component covered by the three-dimensional computable twin model, and adding a precise timestamp based on a unified time reference to all accessed monitoring data streams. Preprocess the raw data according to predefined data cleaning rules, including: Outlier removal: Statistical characteristics of each sensor data stream are calculated based on a sliding time window. Data points that exceed the range of ±3 standard deviations of the historical mean are identified as outliers caused by sensor malfunctions and removed. Short-series data missing information: For data missing segments caused by network packet loss and lasting less than a preset threshold, linear interpolation algorithms or predictions based on autoregressive models are used to fill in the missing data to maintain the continuity of the data sequence. High-frequency noise smoothing filter: High-frequency random noise caused by environmental interference in the monitoring data is processed by a convolutional smoothing filter, which effectively smooths high-frequency fluctuations while preserving the true trend of data change. The cleaned data is associated with the three-dimensional computable twin model based on its spatiotemporal encoding, specifically including: Spatiotemporal coding parsing and matching: Parse the spatiotemporal coding attached to each cleaned data, which includes at least a spatial grid and a timestamp; determine the target geometric entity or logical location corresponding to the data based on the spatial grid ID and the pre-established spatial index mapping table in the 3D computable twin model; Dynamic attribute association and update: Through application programming interface or database connection, data values ​​are dynamically associated as attributes and written into the attribute table of the target geometric entity or logical location in real time, while overwriting or versioning historical data under the same spatiotemporal coordinates; Data field generation and maintenance: Continuously execute the above parsing, matching and association operations so that all multi-source data are attached to the twin model according to their inherent spatiotemporal logic, thereby constructing and maintaining a data field that supports multi-dimensional querying and analysis by time, space and entity type.

[0022] The specific aspects of constructing risk reasoning using spatiotemporal knowledge graphs in S2 include: S2.1: Define the node categories and attributes of construction risk elements; S2.2: Define causal relationship edges and spatiotemporal impact edges between nodes. The causal relationship edges are used to describe the inherent influence logic of geological conditions on construction stability. The spatiotemporal impact edges are used to describe the spatial attenuation and temporal delay effects of dynamic physical processes such as the propagation of blasting vibration waves. The causal relationship edges are defined to express static or quasi-static logical influence relationships, including risk aggravation edges from geological defect units to construction activity units or risk receptor units. Their weights are quantified based on the empirical or theoretical influence of geological conditions on engineering stability. The spatiotemporal impact edges are defined to express the propagation relationship of dynamic physical fields, including vibration generation or load application edges from construction activity units to monitoring physical quantity units or risk receptor units. Their weights are dynamically calculated using a function model that attenuates with distance and has a time delay to describe the spatiotemporal propagation effect of the risk field. S2.3: The structure of the spatiotemporal knowledge graph is used as the input graph structure of the spatiotemporal graph neural network model, and the real-time monitoring data in the dynamic fusion data base is used as node features and edge weights. Through the forward propagation calculation of the neural network, the risk state probability distribution of each spatial unit under multiple construction steps in the future is deduced in real time. The nodes and edges defined in the spatiotemporal knowledge graph are used to construct the initial topology graph of the spatiotemporal graph neural network model. The node vectors are initialized according to their attributes, and the edge weights are initialized according to their relationship type and current weight value. The real-time monitoring data obtained in the dynamic fusion data base is used as the dynamic feature input of the corresponding monitoring physical quantity unit node, and the weights of the relevant spatiotemporal influence edges are dynamically updated according to physical laws or data associations. Spatial dependencies are captured through the multi-layer graph convolution operation of the spatiotemporal graph neural network model, and temporal dependencies are captured by combining temporal convolution or recurrent neural network units. The node state is iteratively updated throughout the graph using a message passing mechanism, and finally the risk state probability distribution of each spatial unit facing different levels of danger under a specified future time sequence is output.

[0023] The risk state probability distribution is used to drive the visualization rendering of the digital twin spatiotemporal field. On the three-dimensional computable twin model, the real-time and future risk levels of each region are overlaid and displayed in the form of dynamic risk heat maps with different colors and transparency. Specifically, this includes the following steps: Visual mapping configuration: A color mapping table and a transparency mapping table are predefined to correspond to the risk probability value range, where high risk probability values ​​are mapped to red tones and high opacity, and low risk probability values ​​are mapped to green tones and low opacity.

[0024] Heatmap generation and overlay: Based on the risk state probability distribution data, query and assign corresponding color and transparency values ​​to each spatial unit in the digital twin spatiotemporal field; call the 3D graphics rendering engine, and use the processed color and transparency data as an independent semi-transparent layer to accurately overlay and render on the corresponding geometric surface of the 3D computable twin model to form a dynamic risk heatmap.

[0025] Dynamic time-series update: When the spatiotemporal graph neural network model outputs a new probability distribution of risk states for future prediction time series, the system automatically updates the visual attributes of the heat map according to steps S4.1 and S4.2, and controls it to periodically change or rotate according to the prediction time series to intuitively present the evolution of risk in the spatiotemporal dimension. It also provides a visual query function for risk tracing and transmission paths, specifically including the following steps: Interaction Triggering and Target Location: In response to the user's selection of a specific high-risk area on the dynamic risk heat map, the system calculates the corresponding three-dimensional spatial location and the management unit to which it belongs based on the screen coordinates of the click.

[0026] Path calculation and visualization: Taking the selected unit as the endpoint, based on the node association relationships stored in the spatiotemporal knowledge graph, reverse query and calculate all possible causal relationship edges and spatiotemporal influence edges pointing to the endpoint to form one or more risk transmission chains; on the three-dimensional computable twin model, the transmission chains are animated in the form of highlighted three-dimensional arrow lines or particle flow.

[0027] Linked information display: While visualizing the path, the sidebar of the interactive interface simultaneously displays the detailed attributes of the risk source nodes involved in the path, the associated historical monitoring data curves, and related early warning records, completing a full risk tracing analysis.

[0028] The specific operation of the federated learning mechanism in S3 includes: S3.1: Within the early warning system of each local construction project, the localized adaptive early warning threshold model is trained and fine-tuned using locally generated multi-source heterogeneous construction data and early warning feedback results. Temporal feature segments related to risk events are extracted from the multi-source heterogeneous construction data, and these segments are combined with the actual risk occurrence, early warning response timeliness, and handling effect in historical early warning records to form labeled training sample pairs. Using these training sample pairs, the localized adaptive early warning threshold model is periodically trained through a supervised learning algorithm to optimize its model parameters. The training aims to minimize the difference between the model's early warning results and the actual risk situation, and dynamically adjust the early warning sensitivity weights for different risk types based on the early warning feedback results. S3.2: Periodically encrypt and upload the key parameters of each local model and execute a secure model aggregation algorithm to generate a global risk warning model with strong generalization ability. When the aggregation cycle is triggered, the early warning system of each local construction project uses homomorphic encryption or secure multi-party computation technology to encrypt the gradient or parameter update of the model parameters that need to be uploaded. The encrypted parameter data is uploaded to the central coordination server. Without decrypting the original data of each participant, the server performs aggregation calculation on all received encrypted parameters by executing the federated averaging algorithm or the optimization algorithm based on secure aggregation, and generates a new set of global model parameters that integrate the experience of multiple projects, thereby constructing the global risk warning model with strong generalization ability. S3.3: The aggregated global risk warning model parameters are distributed to each local project, and the warning model knowledge of each project is shared and evolved together. The central coordination server distributes the aggregated global risk warning model parameters to the warning systems of all local construction projects participating in federated learning. The warning systems of each local project receive the global parameters and integrate them with the local model as the initial parameters or prior knowledge for the next round of local model training and fine-tuning. This allows each local project to retain its adaptability to specific scenarios while continuously absorbing common risk knowledge from other projects, thereby achieving the collaborative improvement and co-evolution of the warning model capabilities of all participating projects.

[0029] The intelligent early warning decision is converted into executable control instructions, including: automatically triggering interlocking shutdown control of construction machinery in a specific area based on risk level and spatiotemporal location; sending directional sound and light alarms and evacuation guidance to smart wearable devices of personnel in the affected area; and pushing control suggestions to the construction management platform.

[0030] The self-optimizing loop of the parallel system of the federated learning mechanism collects feedback data on the actual intervention effects of historical early warning decisions and uses it as new training samples to periodically fine-tune the parameters of the spatiotemporal graph neural network model and correct the weights of the relationships between nodes in the spatiotemporal knowledge graph.

[0031] The node categories include geological defect units, construction activity units, monitoring physical quantity units, and risk receptor units. Geological defect unit nodes are defined as geological anomalies that characterize faults, densely jointed zones, weak interlayers, or groundwater-rich areas. Their attributes include defect type, spatial orientation, scale, and strength reduction factor. Construction activity unit nodes are defined as specific procedures such as drilling, charging, blasting, excavation, and support. Their attributes include activity type, planned and actual spatiotemporal range, and construction parameters. Monitoring physical quantity unit nodes are defined as monitoring points for vibration velocity, stress value, displacement, and gas concentration. Their attributes include monitoring type, real-time and historical time-series data, and early warning threshold. Risk receptor unit nodes are defined as construction personnel, key equipment, or important buildings. Their attributes include receptor type, spatial location, status, and vulnerability level.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A construction safety risk intelligent early warning method based on multi-source data fusion, characterized in that, Includes the following steps: S1. Construct a multi-source fusion digital twin foundation: Construct a digital twin spatiotemporal field synchronized with the physical construction site, and standardize and spatiotemporally align multi-source heterogeneous construction data from geological exploration data sources, environmental monitoring data sources, personnel and equipment positioning data sources, and multi-dimensional physical field monitoring data sources, and inject them into the digital twin spatiotemporal field to form a unified dynamic fusion data foundation; S2. Constructing a spatiotemporal graph to achieve risk prediction: Based on the dynamic fusion data base, a spatiotemporal knowledge graph is constructed with construction risk elements as nodes and spatiotemporal causal relationships between elements as edges. The spatiotemporal knowledge graph is then used to perform real-time reasoning using a spatiotemporal graph neural network model to simulate the dynamic propagation and superposition process of construction safety risks in the digital twin spatiotemporal field, and outputs probabilistic risk prediction results for future time series. S3. Generate early warning decisions based on federated learning: Receive the probabilistic risk prediction results and integrate the adaptive early warning threshold model updated based on the federated learning mechanism to generate intelligent early warning decisions that include risk level, spatiotemporal location, and recommended control measures.

2. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 1, characterized in that, The construction of the digital twin spatiotemporal field in S1 specifically includes: S1.1: Based on the BIM model of the construction site and the real-scene scanned point cloud data, construct a three-dimensional computable twin model with geometric, physical and engineering attributes; S1.2: Deploy and access multi-source heterogeneous construction data; S1.3: Design unified spatiotemporal coding rules and data interface specifications to clean, transform, and timestamp the multi-source heterogeneous construction data, and map it to the corresponding spatiotemporal coordinates of the three-dimensional computable twin model to form a fused data field that can reflect the dynamics of the entire construction process.

3. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 2, characterized in that, The specific aspects of constructing risk reasoning using spatiotemporal knowledge graphs in S2 include: S2.1: Define the node categories and attributes of construction risk elements, wherein the node categories include geological defect units, construction activity units, monitoring physical quantity units, and risk receptor units; S2.2: Define the causal relationship edge and the spatiotemporal influence edge between nodes, wherein the causal relationship edge is used to describe the inherent influence logic of geological conditions on construction stability, and the spatiotemporal influence edge is used to describe executable control commands such as the propagation of blasting vibration waves, and the spatial attenuation and temporal delay effects of dynamic physical processes. S2.3: The structure of the spatiotemporal knowledge graph is used as the input graph structure of the spatiotemporal graph neural network model, and the real-time monitoring data in the dynamic fusion data base is used as node features and edge weights. Through the forward propagation calculation of the neural network, the risk state probability distribution of each spatial unit under multiple future construction steps is deduced in real time.

4. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 3, characterized in that, The risk state probability distribution is used to drive the visualization rendering of the digital twin spatiotemporal field. The three-dimensional computable twin model displays the real-time and future risk levels of each region in the form of dynamic risk heat maps with different colors and transparency, and provides a visualization query function for risk tracing and transmission paths.

5. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 1, characterized in that, The specific operation of the federated learning mechanism in S3 includes: S3.1: Within the early warning system of each local construction project, the localized adaptive early warning threshold model is trained and fine-tuned using locally generated multi-source heterogeneous construction data and early warning feedback results. S3.2: Periodically encrypt and upload the key parameters of each local model and execute a secure model aggregation algorithm to generate a global risk warning model with strong generalization ability; S3.3: Distribute the aggregated global risk warning model parameters to each local project, and share and evolve the warning model knowledge among the projects.

6. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 5, characterized in that, The intelligent early warning decision is converted into executable control instructions, including: automatically triggering interlocking shutdown control of construction machinery in a specific area based on risk level and spatiotemporal location; sending directional sound and light alarms and evacuation guidance to smart wearable devices of personnel in the affected area; and pushing control suggestions to the construction management platform.

7. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 1, characterized in that, The self-optimizing loop of the parallel system of the federated learning mechanism collects feedback data on the actual intervention effects of historical early warning decisions and uses it as new training samples to periodically fine-tune the parameters of the spatiotemporal graph neural network model and correct the weights of the relationships between nodes in the spatiotemporal knowledge graph.

8. The intelligent early warning method for construction safety risks based on multi-source data fusion according to claim 3, characterized in that, The node categories include geological defect units, construction activity units, monitoring physical quantity units, and risk receptor units.

Images