Edge calculation method and system applied to construction site safety management

Abstract

The invention discloses an edge calculation method and system applied to construction site safety management. The method comprises the following steps: S1, collecting and preprocessing multi-modal data; s2, extracting and sharing risk features, and generating a preliminary risk feature map; s3, generating a local abnormal projection result by using a distributed high-dimensional abnormal mode projection module, and splicing to form a high-dimensional risk feature curved surface; s4, monitoring the data flow by using a dynamic event sensing module, identifying a high-risk event and outputting a high-risk event mark; s5, calling a dynamic risk weight tree search algorithm in a path optimization module to generate a dynamic security path; s6, generating a local risk prediction result by using the space-time hybrid prediction model; s7, generating a global dynamic prediction result through a layered space-time cooperation mechanism; and S8, providing real-time risk early warning and path optimization suggestions for construction management personnel through the mobile terminal. According to the invention, edge calculation and artificial intelligence technologies are combined, so that construction site dynamic risk identification and safety management are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of construction site safety management, and in particular to an edge computing method and system for construction site safety management. Background Art

[0002] With the rapid development of the construction industry, the scale and complexity of construction sites are increasing, and the resulting safety management issues are becoming more severe. The construction site is usually a dynamic and changing environment, involving a large number of personnel, equipment, and complex operating procedures. Possible risks in the construction process include illegal operations by personnel, equipment failure, environmental changes, etc. These risks not only threaten the lives of construction workers, but may also cause project delays, economic losses, and even environmental pollution. Therefore, how to effectively manage the safety of construction sites, especially the real-time assessment and response to risks, has become a research focus in the field of construction safety management.

[0003] Existing construction site safety management technologies mainly rely on manual inspections, surveillance video analysis, and rule-based safety assessment systems. However, these traditional methods have many limitations. First, manual inspections are inefficient, with limited inspection scope, and it is difficult to detect hidden safety hazards in a timely manner. At the same time, manual inspections are highly dependent on the experience and professional skills of personnel, and are prone to missed inspections or misjudgments in complex construction site environments. Secondly, although traditional monitoring systems can collect video data, they lack intelligent data processing capabilities. Most monitoring still relies on manual viewing of video playback, and cannot identify abnormal behaviors or high-risk events in real time. In addition, rule-based safety assessment systems usually use fixed thresholds or preset rules to assess construction risks, while the construction site is a dynamically changing environment. Such static rules are difficult to adapt to the complex and changing actual conditions on site.

[0004] In recent years, with the rapid development of the Internet of Things, edge computing and artificial intelligence technologies, new technical means have been provided for construction site safety management. For example, the Internet of Things technology can be used to collect multimodal data of the construction site in real time, including video streams, environmental sensor data and personnel trajectory data; edge computing can realize local processing of data and reduce the latency of cloud computing; artificial intelligence technology can perform intelligent analysis of data through deep learning models. However, there are still many problems with the existing construction safety management system based on these new technologies. First, data processing and analysis are mostly concentrated in the cloud. Transmitting a large amount of construction site data to the cloud will increase the pressure on network bandwidth. At the same time, data delays or losses may occur during the transmission process, which will affect real-time performance. Secondly, most existing systems use analysis methods for single-modal data, which makes it difficult to integrate the correlation information between multimodal data, resulting in inaccurate risk assessment.

[0005] Therefore, how to provide an edge computing method and system for construction site safety management is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] One purpose of the present invention is to propose an edge computing method and system for construction site safety management. The present invention combines edge computing with artificial intelligence technology, and realizes dynamic risk identification and path optimization at the construction site through recursive attention analysis, high-dimensional abnormal pattern projection, and spatiotemporal hybrid prediction methods. The system makes full use of multimodal data fusion and layered spatiotemporal collaboration mechanisms to conduct real-time assessment and intelligent management of risks in complex construction environments, generate dynamic safety paths and accurate risk prediction results, and ensure the safety and efficiency of construction personnel and equipment in high-risk environments. The present invention has the significant advantages of strong real-time performance, high accuracy, excellent adaptability, and fast response speed, providing an intelligent solution for construction safety management.

[0007] According to an embodiment of the present invention, an edge computing method for construction site safety management includes the following steps:

[0008] S1. Collect multimodal data through edge nodes deployed at the construction site, and perform local preprocessing on the collected multimodal data sets to generate a preliminary cleansing data set;

[0009] S2. Running a lightweight recursive attention module on the edge node, performing recursive attention analysis on the preliminary cleaned data set to extract risk features, and sharing the risk features through a low-latency communication protocol to generate a preliminary risk feature map;

[0010] S3, using the preliminary risk feature map to perform abnormal feature extraction and high-dimensional risk projection in the distributed high-dimensional abnormal pattern projection module, generate local abnormal projection results and transmit them to the master node, and splice them to form a high-dimensional risk feature surface;

[0011] S4. Use the dynamic event perception module to monitor the construction site data flow in real time, identify high-risk events through the event trigger mechanism, and output high-risk event markers;

[0012] S5. Based on the high-risk event marker, a dynamic risk weight tree search algorithm is called in a path optimization module to generate a dynamic safety path;

[0013] S6. Based on the high-dimensional risk characteristic surface and the dynamic safety path, a space-time hybrid prediction model is run at the edge node to perform short-term risk prediction on the construction site and generate a local risk prediction result;

[0014] S7. Aggregate the local risk prediction results to the master node through a layered spatiotemporal collaboration mechanism to generate a global dynamic prediction result, and update the edge node parameters through an edge-cloud collaboration strategy;

[0015] S8. Based on the global dynamic prediction results and the updated edge node parameters, provide real-time risk warnings and path optimization suggestions to construction management personnel through mobile terminals.

[0016] Optionally, the S2 specifically includes:

[0017] S21, input the preliminary cleaned data set into the lightweight recursive attention module in the edge node, the lightweight recursive attention module includes an input layer, an attention calculation layer, a multi-head attention aggregation layer and an output layer; the input layer is used to receive the time series features, spatial features and modal features of the preliminary cleaned data set, and perform preliminary encoding; the attention calculation layer calculates the correlation between data through a recursive mechanism; the multi-head attention aggregation layer combines the calculation results of multiple attention heads to generate a unified feature representation; the output layer is used to generate the final risk feature;

[0018] S22. In the input layer, the preliminary cleaned data set is separated and encoded according to the data type, and the time series feature X is extracted. t , spatial feature X s and modal features X m , and concatenate to form the input feature matrix X (0) =[X t ;X s ;X m ];

[0019] S23. In the attention calculation layer, recursive attention calculation is performed on the input feature matrix:

[0020]

[0021] X (l+1) =A (l) ·V (l) +X (l) ;

[0022] Among them, A (l) represents the attention weight matrix of the lth layer, Q (l) represents the query matrix of the lth layer, K (l) represents the bond matrix of the lth layer, V (l) represents the value matrix of the lth layer, d k represents the dimension of the key vector, softmax represents the normalization function, X (l) represents the input feature matrix of the lth layer, X (l+1) Represents the input feature matrix of the l+1th layer;

[0023] S24, in the multi-head attention aggregation layer, the weight matrix and feature matrix of multiple attention heads are combined to perform attention aggregation to obtain the aggregated feature matrix X multi ;

[0024] S25, the feature matrix X after aggregation at the output layer multi Perform normalization to eliminate differences between modalities and generate the final risk feature X f ;

[0025] S26. Sharing the risk feature X via a low-latency communication protocol f To the master node, and combined with the risk features shared by other edge nodes to generate a preliminary risk feature graph F risk .

[0026] Optionally, the S3 specifically includes:

[0027] S31, inputting the preliminary risk feature map into a distributed high-dimensional abnormal pattern projection module, wherein the distributed high-dimensional abnormal pattern projection module includes a data separation unit, a feature extraction unit, a dimensionality reduction projection unit, an abnormal pattern recognition unit and a data aggregation unit;

[0028] S32, in the data separation unit, separating the preliminary risk feature map into video modal features, environmental sensor modal features and personnel trajectory modal features, and allocating the separated modal feature data to corresponding feature extraction units;

[0029] S33, in the feature extraction unit, using the data structure and spatiotemporal correlation information of the modal features, extracting dynamic behavior features from the video modal features, extracting abnormal fluctuation features from the environmental sensor modality, extracting path deviation and clustering features from the personnel trajectory modality, and generating a modal feature matrix;

[0030] S34, in the dimensionality reduction projection unit, the modal feature matrix is ​​input into the manifold learning algorithm processing pipeline, and the high-dimensional modal feature data is projected into the low-dimensional space through a nonlinear mapping method to generate a local abnormal projection matrix. The dimensionality reduction projection unit dynamically adjusts the calculation frequency according to the resource allocation and data characteristics of the current edge node;

[0031] S35. In the abnormal pattern recognition unit, the local abnormal projection matrix is ​​used to perform pattern clustering and screening on the feature matrix in combination with the time-sensitive weight to generate a local abnormal projection result;

[0032] S36. In the data aggregation unit, the local anomaly projection results are transmitted to the master node through a low-latency communication protocol. The master node aggregates the anomaly projection results from multiple edge nodes and integrates the multi-node data through a weighted splicing algorithm to finally generate a high-dimensional risk feature surface.

[0033] Optionally, the S4 specifically includes:

[0034] S41, deploying a dynamic event perception module in the edge node, wherein the dynamic event perception module includes a data acquisition unit, a feature extraction unit, an event triggering unit, an event classification unit and an event marking unit;

[0035] S42, the data acquisition unit monitors the multimodal data stream in real time, including the video stream, the environmental sensor data and the personnel trajectory data, and forms a time series data block after assigning a timestamp;

[0036] S43, in a feature extraction unit, processing the time series data block, extracting dynamic area features from the video stream, extracting gas concentration, temperature and humidity fluctuation features from the environmental sensor, extracting path deviation features from the trajectory data, and obtaining a multimodal feature set;

[0037] S44. In the event triggering unit, the anomaly score is calculated using the extracted multimodal feature set:

[0038]

[0039] Among them, S abn represents the anomaly score, x m,i,t represents the value of the i-th feature of the m-th mode at the t-th time step, μ m,i represents the historical mean of the ith feature of the mth mode, σ m,i represents the standard deviation of the ith feature of the mth mode, ∈ represents the smoothing factor, N m represents the total number of modal features, T represents the time window length, and β represents the time attenuation coefficient;

[0040] S45, when the abnormal score S abn When the set threshold is exceeded, the event classification unit is triggered, which classifies the events and associates the time, location and abnormal feature information for each event;

[0041] S46. In the event marking unit, a high-risk event mark is generated according to the classification result, wherein the high-risk event mark includes event type, triggering time, location, and feature summary.

[0042] Optionally, the S5 specifically includes:

[0043] S51, deploying a dynamic risk weight tree search algorithm in a path optimization module, wherein the path optimization module includes a path tree construction unit, a risk weight calculation unit, and a strategy update unit;

[0044] S52: The path tree construction unit divides the construction site into grid areas, each grid node n represents a location point, and the path tree starts from the initial location node s0 and extends to the target node sg , assign state s to each node of the path tree n , path cost C(n), comprehensive risk weight W(n) and exploration priority P(n);

[0045] S53. In the risk weight calculation unit, the comprehensive risk weight is calculated for each network node n in the path tree:

[0046]

[0047] Where W(n) represents the comprehensive risk weight of network node n, t0 and t n represents the time range from the initial time to node n, K represents the total number of risk features, and γ k represents the weight coefficient of the kth risk feature, f k (n,t) represents the kth risk characteristic value of network node n at time t, σ k represents the historical standard deviation of the k-th risk characteristic, ∈ represents the smoothing factor, δ k represents the nonlinear amplification coefficient of the k-th risk feature, d k (n) represents the distance from network node n to the kth risk source, α k represents the time attenuation coefficient;

[0048] S54. In the path tree, the root node is used as the initial position node, and node expansion is performed based on the dynamic risk weight tree search algorithm, and the optimal child node is selected in each expansion:

[0049]

[0050] Where P(n′) represents the comprehensive score of child node n′, W(n′) represents the risk weight of child node n′, C(n′) represents the cumulative path cost of child node n′, δ represents the path cost weight coefficient, λ represents the exploration balance coefficient, N(n) represents the number of visits to the parent node, N(n′) represents the number of visits to the child node, η represents the heuristic coefficient, and H(n′) represents the heuristic estimate from the node to the target node;

[0051] S55, dynamically adjust the node selection logic in the policy update unit, obtain high-risk event markers in real time, and optimize the path search efficiency and security to generate a dynamic security path P safe .

[0052] Optionally, the S6 specifically includes:

[0053] S61, running the space-time hybrid prediction model in the edge node, taking the high-dimensional risk feature surface and the dynamic safety path as input, passing them to the time feature extraction unit and the spatial relationship modeling unit respectively, and initializing the space-time hybrid prediction model parameters;

[0054] S62, performing time series segmentation on the high-dimensional risk feature surface, dividing it into multiple subsequences according to time steps; extracting the time dependency relationship of each subsequence using a time feature extraction unit, and generating a time feature matrix containing time dynamic change features;

[0055] S63, constructing a spatial feature map of the construction site based on the dynamic safety path, defining the gridded area of ​​the construction site as a node, generating an adjacency graph based on the spatial connection relationship between the nodes, calculating the feature relationship of the nodes using a spatial relationship modeling unit, and generating a spatial feature matrix;

[0056] S64, inputting the time feature matrix and the space feature matrix into the time-space feature fusion module, using the feature splicing method to combine the time dynamic change and spatial risk distribution information to generate the time-space feature matrix;

[0057] S65, generating a local risk prediction result using the spatiotemporal feature matrix, wherein the local risk prediction result includes the potential risk value, occurrence time and impact range of each node;

[0058] S66. Sort the local risk prediction results according to the time series, and mark the risk level and confidence level for each prediction result.

[0059] Optionally, the S7 specifically includes:

[0060] S71. Collect local risk prediction results of edge nodes through a layered spatiotemporal collaboration mechanism, annotate the local risk prediction result of each edge node with the corresponding time step and spatial position index, and form an edge node prediction data set;

[0061] S72, integrating the local risk prediction results of all edge nodes in the master node, classifying the local risk prediction results of each time step according to the spatial grid area, and screening according to the spatiotemporal consistency of the prediction data set to eliminate abnormal prediction values;

[0062] S73, based on the integrated prediction data set, generate a global time feature matrix T through a time aggregation algorithm global ,The time aggregation algorithm smoothes the risk prediction values ​​at different time steps by ,accumulating and weighted sliding window methods to eliminate the interference of short-term fluctuations on the ,global prediction;

[0063] S74, running a spatial feature aggregation algorithm in the master node to generate a global spatial feature matrix according to the topological relationship of the spatial grid area, and analyzing the global spatial feature matrix to generate a global dynamic prediction result;

[0064] S75. Synchronize the spatiotemporal feature weights, risk level parameters and time step information in the global dynamic prediction results to the edge nodes.

[0065] An edge computing system for construction site safety management according to an embodiment of the present invention includes the following modules:

[0066] The data acquisition module is used to collect multimodal data from the construction site and perform localized preprocessing to generate a preliminary cleansed data set;

[0067] A lightweight recursive attention module, which is used to perform recursive attention analysis on the preliminary cleaned dataset to generate a preliminary risk feature map;

[0068] Distributed high-dimensional anomaly pattern projection module, used for anomaly feature extraction and high-dimensional risk projection, generating local anomaly projection results and forming a high-dimensional risk feature surface;

[0069] Dynamic event perception module, which is used to monitor the data flow of the construction site, identify high-risk events and generate high-risk event markers;

[0070] A path optimization module for generating dynamic safe paths based on high-risk event markers;

[0071] The spatiotemporal hybrid prediction module is used to generate local risk prediction results based on high-dimensional risk feature surfaces and dynamic safety paths;

[0072] A hierarchical spatiotemporal collaboration module is used to integrate local risk prediction results, generate global dynamic prediction results and update edge node parameters;

[0073] The risk warning and path optimization module is used to provide real-time risk warnings and path optimization suggestions to construction managers.

[0074] The beneficial effects of the present invention are:

[0075] First, the present invention realizes the localized collection and processing of multimodal data by deploying edge nodes at the construction site. Compared with traditional cloud computing methods, the edge computing architecture can avoid the delay and bandwidth pressure caused by data transmission, significantly improving the real-time performance of risk identification. In addition, combined with the lightweight recursive attention mechanism, the present invention has a strong weak signal capture capability in a high-noise construction environment, and can accurately extract the correlation information between temporal features, spatial features, and modal features, thereby greatly improving the system's analysis accuracy of risk features.

[0076] Secondly, the application of the distributed high-dimensional abnormal pattern projection module and the dynamic event perception module enables the present invention to efficiently detect abnormal features of the construction site and quickly identify high-risk events. The distributed high-dimensional abnormal pattern projection module generates a dynamic risk feature surface through dimensionality reduction projection and abnormal pattern analysis of multimodal data, laying a solid foundation for global risk assessment. The dynamic event perception module monitors the data stream in real time and introduces an event trigger mechanism. When an abnormal event is detected, it quickly starts a high-computational intensity task and generates a high-risk event marker, providing a key basis for subsequent path optimization and short-term risk prediction.

[0077] In addition, the present invention combines the dynamic risk weight tree search algorithm to significantly optimize the path planning efficiency of construction personnel and equipment. The path optimization module can generate dynamic safe paths in real time, especially when working in high-risk areas, to quickly avoid potential dangers and provide strong protection for construction safety. At the same time, the spatiotemporal hybrid prediction model accurately predicts the short-term risks of the construction site and generates local risk prediction results through the deep integration of temporal and spatial characteristics. Combined with the layered spatiotemporal collaboration mechanism and the edge-cloud collaboration strategy, the present invention effectively integrates the local prediction results of each edge node to form a global dynamic prediction result, and dynamically updates the edge node parameters, further enhancing the system's adaptability and response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0079] Figure 1 A flowchart of an edge computing method for construction site safety management proposed by the present invention;

[0080] Figure 2 This is a schematic diagram of the structure of an edge computing system for construction site safety management proposed by the present invention. DETAILED DESCRIPTION

[0081] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0082] refer to Figure 1 , an edge computing method for construction site safety management, comprising the following steps:

[0083] S1. Collect multimodal data through edge nodes deployed at the construction site, and perform local preprocessing on the collected multimodal data sets to generate a preliminary cleansing data set;

[0084] S2. Running a lightweight recursive attention module on the edge node, performing recursive attention analysis on the preliminary cleaned data set to extract risk features, and sharing the risk features through a low-latency communication protocol to generate a preliminary risk feature map;

[0085] S3, using the preliminary risk feature map to perform abnormal feature extraction and high-dimensional risk projection in the distributed high-dimensional abnormal pattern projection module, generate local abnormal projection results and transmit them to the master node, and splice them to form a high-dimensional risk feature surface;

[0086] S4. Use the dynamic event perception module to monitor the construction site data flow in real time, identify high-risk events through the event trigger mechanism, and output high-risk event markers;

[0087] S5. Based on the high-risk event marker, a dynamic risk weight tree search algorithm is called in a path optimization module to generate a dynamic safety path;

[0088] S6. Based on the high-dimensional risk characteristic surface and the dynamic safety path, a space-time hybrid prediction model is run at the edge node to perform short-term risk prediction on the construction site and generate a local risk prediction result;

[0089] S7. Aggregate the local risk prediction results to the master node through a layered spatiotemporal collaboration mechanism to generate a global dynamic prediction result, and update the edge node parameters through an edge-cloud collaboration strategy;

[0090] S8. Based on the global dynamic prediction results and the updated edge node parameters, provide real-time risk warnings and path optimization suggestions to construction management personnel through mobile terminals.

[0091] In this implementation, S2 specifically includes:

[0092] S21, input the preliminary cleaned data set into the lightweight recursive attention module in the edge node, the lightweight recursive attention module includes an input layer, an attention calculation layer, a multi-head attention aggregation layer and an output layer; the input layer is used to receive the time series features, spatial features and modal features of the preliminary cleaned data set, and perform preliminary encoding; the attention calculation layer calculates the correlation between data through a recursive mechanism; the multi-head attention aggregation layer combines the calculation results of multiple attention heads to generate a unified feature representation; the output layer is used to generate the final risk feature;

[0093] S22. In the input layer, the preliminary cleaned data set is separated and encoded according to the data type, and the time series feature X is extracted. t , spatial feature X s and modal features X m , and concatenate to form the input feature matrix X (0) =[X t;X s ;X m ];

[0094] S23. In the attention calculation layer, recursive attention calculation is performed on the input feature matrix:

[0095]

[0096] X (l+1) =A (l) ·V (l) +X (l) ;

[0097] Among them, A (l) represents the attention weight matrix of the lth layer, Q (l) represents the query matrix of the lth layer, K (l) represents the bond matrix of the lth layer, V (l) represents the value matrix of the lth layer, d k represents the dimension of the key vector, softmax represents the normalization function, X (l) represents the input feature matrix of the lth layer, X (l+1) Represents the input feature matrix of the l+1th layer;

[0098] S24, in the multi-head attention aggregation layer, the weight matrix and feature matrix of multiple attention heads are combined to perform attention aggregation to obtain the aggregated feature matrix X multi ;

[0099] S25, the feature matrix X after aggregation at the output layer multi Perform normalization to eliminate differences between modalities and generate the final risk feature X f ;

[0100] S26. Sharing the risk feature X via a low-latency communication protocol f To the master node, and combined with the risk features shared by other edge nodes to generate a preliminary risk feature graph F risk .

[0101] In this implementation, S3 specifically includes:

[0102] S31, inputting the preliminary risk feature map into a distributed high-dimensional abnormal pattern projection module, wherein the distributed high-dimensional abnormal pattern projection module includes a data separation unit, a feature extraction unit, a dimensionality reduction projection unit, an abnormal pattern recognition unit and a data aggregation unit;

[0103] S32, in the data separation unit, separating the preliminary risk feature map into video modal features, environmental sensor modal features and personnel trajectory modal features, and allocating the separated modal feature data to corresponding feature extraction units;

[0104] S33, in the feature extraction unit, using the data structure and spatiotemporal correlation information of the modal features, extracting dynamic behavior features from the video modal features, extracting abnormal fluctuation features from the environmental sensor modality, extracting path deviation and clustering features from the personnel trajectory modality, and generating a modal feature matrix;

[0105] S34, in the dimensionality reduction projection unit, the modal feature matrix is ​​input into the manifold learning algorithm processing pipeline, and the high-dimensional modal feature data is projected into the low-dimensional space through a nonlinear mapping method to generate a local abnormal projection matrix. The dimensionality reduction projection unit dynamically adjusts the calculation frequency according to the resource allocation and data characteristics of the current edge node;

[0106] S35. In the abnormal pattern recognition unit, the local abnormal projection matrix is ​​used to perform pattern clustering and screening on the feature matrix in combination with the time-sensitive weight to generate a local abnormal projection result;

[0107] S36. In the data aggregation unit, the local anomaly projection results are transmitted to the master node through a low-latency communication protocol. The master node aggregates the anomaly projection results from multiple edge nodes and integrates the multi-node data through a weighted splicing algorithm to finally generate a high-dimensional risk feature surface.

[0108] In this implementation manner, the S4 specifically includes:

[0109] S41, deploying a dynamic event perception module in the edge node, wherein the dynamic event perception module includes a data acquisition unit, a feature extraction unit, an event triggering unit, an event classification unit and an event marking unit;

[0110] S42, the data acquisition unit monitors the multimodal data stream in real time, including the video stream, the environmental sensor data and the personnel trajectory data, and forms a time series data block after assigning a timestamp;

[0111] S43, in a feature extraction unit, processing the time series data block, extracting dynamic area features from the video stream, extracting gas concentration, temperature and humidity fluctuation features from the environmental sensor, extracting path deviation features from the trajectory data, and obtaining a multimodal feature set;

[0112] S44. In the event triggering unit, the anomaly score is calculated using the extracted multimodal feature set:

[0113]

[0114] Among them, S abn represents the anomaly score, x m,i,t represents the value of the i-th feature of the m-th mode at the t-th time step, μ m,i represents the historical mean of the ith feature of the mth mode, σ m,irepresents the standard deviation of the ith feature of the mth mode, ∈ represents the smoothing factor, N m represents the total number of modal features, T represents the time window length, and β represents the time attenuation coefficient;

[0115] S45, when the abnormal score S abn When the set threshold is exceeded, the event classification unit is triggered, which classifies the events and associates the time, location and abnormal feature information for each event;

[0116] S46. In the event marking unit, a high-risk event mark is generated according to the classification result, wherein the high-risk event mark includes event type, triggering time, location, and feature summary.

[0117] In this implementation manner, S5 specifically includes:

[0118] S51, deploying a dynamic risk weight tree search algorithm in a path optimization module, wherein the path optimization module includes a path tree construction unit, a risk weight calculation unit, and a strategy update unit;

[0119] S52: The path tree construction unit divides the construction site into grid areas, each grid node n represents a location point, and the path tree starts from the initial location node s0 and extends to the target node s g , assign state s to each node of the path tree n , path cost C(n), comprehensive risk weight W(n) and exploration priority P(n);

[0120] S53. In the risk weight calculation unit, the comprehensive risk weight is calculated for each network node n in the path tree:

[0121]

[0122] Where W(n) represents the comprehensive risk weight of network node n, t0 and t n represents the time range from the initial time to the network node n, K represents the total number of risk features, γ k represents the weight coefficient of the kth risk feature, f k (n,t) represents the kth risk characteristic value of network node n at time t, σ k represents the historical standard deviation of the k-th risk characteristic, ∈ represents the smoothing factor, δ k represents the nonlinear amplification coefficient of the k-th risk feature, d k (n represents the distance from network node n to the kth risk source, α k represents the time attenuation coefficient;

[0123] S54. In the path tree, the root node is used as the initial position node, and node expansion is performed based on the dynamic risk weight tree search algorithm, and the optimal child node is selected in each expansion:

[0124]

[0125] Among them, P(n ′ ) represents child node n ′ The comprehensive score, W(n ′ ) represents child node n ′ The risk weight of ′ ) represents child node n ′ The cumulative path cost, δ represents the path cost weight coefficient, λ represents the exploration balance coefficient, N(n) represents the number of visits to the parent node, N(n ′ ) represents the number of visits to the child node, η represents the heuristic coefficient, H(n ′ ) represents the heuristic estimate of the distance from the node to the target node;

[0126] S55, dynamically adjust the node selection logic in the policy update unit, obtain high-risk event markers in real time, and optimize the path search efficiency and security to generate a dynamic security path P safe .

[0127] In this implementation manner, S6 specifically includes:

[0128] S61, running the space-time hybrid prediction model in the edge node, taking the high-dimensional risk feature surface and the dynamic safety path as input, passing them to the time feature extraction unit and the spatial relationship modeling unit respectively, and initializing the space-time hybrid prediction model parameters;

[0129] S62, performing time series segmentation on the high-dimensional risk feature surface, dividing it into multiple subsequences according to time steps; extracting the time dependency relationship of each subsequence using a time feature extraction unit, and generating a time feature matrix containing time dynamic change features;

[0130] S63, constructing a spatial feature map of the construction site based on the dynamic safety path, defining the gridded area of ​​the construction site as a node, generating an adjacency graph based on the spatial connection relationship between the nodes, calculating the feature relationship of the nodes using a spatial relationship modeling unit, and generating a spatial feature matrix;

[0131] S64, inputting the time feature matrix and the space feature matrix into the time-space feature fusion module, using the feature splicing method to combine the time dynamic change and spatial risk distribution information to generate the time-space feature matrix;

[0132] S65, generating a local risk prediction result using the spatiotemporal feature matrix, wherein the local risk prediction result includes the potential risk value, occurrence time and impact range of each node;

[0133] S66. Sort the local risk prediction results according to the time series, and mark the risk level and confidence level for each prediction result.

[0134] In this implementation manner, the S7 specifically includes:

[0135] S71. Collect local risk prediction results of edge nodes through a layered spatiotemporal collaboration mechanism, annotate the local risk prediction result of each edge node with the corresponding time step and spatial position index, and form an edge node prediction data set;

[0136] S72, integrating the local risk prediction results of all edge nodes in the master node, classifying the local risk prediction results of each time step according to the spatial grid area, and screening according to the spatiotemporal consistency of the prediction data set to eliminate abnormal prediction values;

[0137] S73, based on the integrated prediction data set, generate a global time feature matrix T through a time aggregation algorithm global ,The time aggregation algorithm smoothes the risk prediction values ​​at different time steps by ,accumulating and weighted sliding window methods to eliminate the interference of short-term fluctuations on the ,global prediction;

[0138] S74, running a spatial feature aggregation algorithm in the master node to generate a global spatial feature matrix according to the topological relationship of the spatial grid area, and analyzing the global spatial feature matrix to generate a global dynamic prediction result;

[0139] S75. Synchronize the spatiotemporal feature weights, risk level parameters and time step information in the global dynamic prediction results to the edge nodes.

[0140] refer to Figure 2 , an edge computing system for construction site safety management, including the following modules:

[0141] The data acquisition module is used to collect multimodal data from the construction site and perform localized preprocessing to generate a preliminary cleansed data set;

[0142] A lightweight recursive attention module, which is used to perform recursive attention analysis on the preliminary cleaned dataset to generate a preliminary risk feature map;

[0143] Distributed high-dimensional anomaly pattern projection module, used for anomaly feature extraction and high-dimensional risk projection, generating local anomaly projection results and forming a high-dimensional risk feature surface;

[0144] Dynamic event perception module, which is used to monitor the data flow of the construction site, identify high-risk events and generate high-risk event markers;

[0145] A path optimization module for generating dynamic safe paths based on high-risk event markers;

[0146] The spatiotemporal hybrid prediction module is used to generate local risk prediction results based on high-dimensional risk feature surfaces and dynamic safety paths;

[0147] A hierarchical spatiotemporal collaboration module is used to integrate local risk prediction results, generate global dynamic prediction results and update edge node parameters;

[0148] The risk warning and path optimization module is used to provide real-time risk warnings and path optimization suggestions to construction managers.

[0149] Embodiment 1:

[0150] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a subway construction site in the city center. The site covers an area of ​​about 5,000 square meters, with a complex operation area, involving multiple high-risk operation points, including deep foundation pits, scaffolding installation areas, and large equipment hoisting areas. During the construction period, the area was densely populated, the operating equipment was diverse, and the environmental conditions were complex and changeable. Traditional safety management methods could no longer meet the real-time and efficient safety management needs, and there were problems such as delayed identification of high-risk events, unreasonable path planning, and frequent personnel accidents.

[0151] In this embodiment, the construction site safety management edge computing system of the present invention is deployed to verify its performance. The system deploys 5 edge nodes in different areas of the construction site, covering key risk points and surrounding areas. Each edge node is equipped with video acquisition equipment, environmental sensors and wireless communication modules, which can collect multimodal data of the construction site in real time, including video streams, gas concentrations, temperature and humidity, and personnel positioning data. At the same time, the master node is deployed in the cloud, which is responsible for integrating local data and generating global prediction results.

[0152] The data acquisition module collects video data, environmental data, and personnel trajectory data of the construction area in real time, and performs local preprocessing through edge nodes. The collected data is input into the lightweight recursive attention module to extract the time series characteristics and spatial distribution characteristics of each modal data, thereby generating a preliminary risk feature map. The distributed high-dimensional abnormal pattern projection module performs dimensionality reduction and abnormal pattern recognition on the risk feature map, and detects abnormal fluctuations, such as abnormal vibration of equipment around deep foundation pits, excessive density of personnel gathering in scaffolding work areas, and rapid increase in concentration near gas leaks.

[0153] In one test, the dynamic event perception module detected irregular jitters during the operation of a certain hoisting equipment, and the gas sensor detected a trend of increasing concentrations of volatile organic compounds in the hoisting area. The system calculated the anomaly score and generated a high-risk event marker, treating abnormal operation of the hoisting equipment, excessive volatile organic compounds, and construction workers approaching the hoisting area as high-risk events. Subsequently, the path optimization module generated a dynamic safe path based on the event markers, suggesting that construction workers bypass the hoisting area and temporarily block the high-risk area to minimize the probability of accidents.

[0154] At the same time, the spatiotemporal hybrid prediction module predicted the short-term risks in the deep foundation pit area based on multimodal data collected in real time. The results showed that the frequent start-up of construction equipment increased the settlement rate of the surrounding soil and there was a risk of foundation pit collapse. The system transmits the prediction results to the master node through a layered spatiotemporal collaboration mechanism. The master node integrates the data and issues a warning and adjusts the risk level, while sending real-time notifications to on-site construction managers.

[0155] Table 1 Comparison of test results between traditional system and the system of the present invention

[0156]

[0157]

[0158] Table 1 above compares the performance of the traditional construction site safety management method and the system of the present invention in detail on multiple key performance indicators, clearly demonstrating the superiority of the system of the present invention. From the perspective of risk identification efficiency, the traditional method takes an average of 180 seconds to detect high-risk events, while the system of the present invention completes risk identification in only 10 seconds through edge computing and intelligent analysis, which improves efficiency by 1700%. This significant gap is mainly due to the rapid response capabilities of the lightweight recursive attention module and dynamic event perception module in the system.

[0159] In terms of risk identification accuracy, the accuracy of the traditional method is 85%, while the accuracy of the system of the present invention reaches 98%, an increase of 15.3%. This improvement is due to the use of multimodal data fusion and high-dimensional abnormal pattern projection technology in the present invention, which effectively improves the system's ability to capture and analyze multidimensional risk characteristics in complex construction environments.

[0160] In terms of path optimization time, the traditional method takes an average of 120 seconds to calculate a reasonable path, while the system of the present invention only takes 6 seconds, which is 1900% more efficient. This optimization is due to the application of the dynamic risk weight tree search algorithm, which combines real-time event marking and dynamic weight evaluation to quickly generate a dynamic safe path.

[0161] In terms of the processing capacity of high-risk events, the traditional method can only handle 15 high-risk events per week, while the system of the present invention can handle 28 times per week, with a processing capacity increase of 86.7%. This advantage ensures that the system can quickly respond to various risk events at the construction site, thereby significantly reducing the occurrence of accidents.

[0162] In terms of accident avoidance rate, the system of the present invention is also significantly better than the traditional method. The accident avoidance rate of the traditional method is 65%, while the system of the present invention reaches 90%, an increase of 25%. This result reflects the excellent capabilities of the system of the present invention in real-time prediction, dynamic adjustment and global collaboration.

[0163] In addition, the data processing delay of the system of the present invention is only 0.5 seconds, which is 900% lower than the 5 seconds of the traditional system. This fast response capability is mainly due to the localized calculation of the edge node, which avoids the time consumption of data transmission. At the same time, the system response time is shortened from 7 seconds of the traditional method to 0.8 seconds, further demonstrating the significant advantage of the system of the present invention in terms of efficiency.

[0164] In terms of multimodal data processing capabilities, traditional methods usually only support single-mode data processing, while the system of the present invention can simultaneously process three modal data: video, environmental sensor, and personnel trajectory, increasing data processing capabilities by 200%. This capability ensures that the system can comprehensively analyze multiple risk factors in complex construction sites, providing more reliable support for safety management.

[0165] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An edge computing method for construction site safety management, characterized in that: The steps include: S1. Collect multimodal data through edge nodes deployed at the construction site, and perform local preprocessing on the collected multimodal data sets to generate a preliminary cleansing data set; S2. Running a lightweight recursive attention module on the edge node, performing recursive attention analysis on the preliminary cleaned data set to extract risk features, and sharing the risk features through a low-latency communication protocol to generate a preliminary risk feature map; S3, using the preliminary risk feature map to perform abnormal feature extraction and high-dimensional risk projection in the distributed high-dimensional abnormal pattern projection module, generate local abnormal projection results and transmit them to the master node, and splice them to form a high-dimensional risk feature surface; S4. Use the dynamic event perception module to monitor the construction site data flow in real time, identify high-risk events through the event trigger mechanism, and output high-risk event markers; S5. Based on the high-risk event marker, a dynamic risk weight tree search algorithm is called in a path optimization module to generate a dynamic safety path; S6. Based on the high-dimensional risk characteristic surface and the dynamic safety path, a space-time hybrid prediction model is run at the edge node to perform short-term risk prediction on the construction site and generate a local risk prediction result; S7. Aggregate the local risk prediction results to the master node through a layered spatiotemporal collaboration mechanism to generate a global dynamic prediction result, and update the edge node parameters through an edge-cloud collaboration strategy; S8. Based on the global dynamic prediction results and the updated edge node parameters, provide real-time risk warnings and path optimization suggestions to construction management personnel through mobile terminals.

2. The edge computing method for construction site safety management according to claim 1 is characterized in that: The S2 specifically includes: S21, input the preliminary cleaned data set into the lightweight recursive attention module in the edge node, the lightweight recursive attention module includes an input layer, an attention calculation layer, a multi-head attention aggregation layer and an output layer; the input layer is used to receive the time series features, spatial features and modal features of the preliminary cleaned data set, and perform preliminary encoding; the attention calculation layer calculates the correlation between data through a recursive mechanism; the multi-head attention aggregation layer combines the calculation results of multiple attention heads to generate a unified feature representation; the output layer is used to generate the final risk feature; S22. In the input layer, the preliminary cleaned data set is separated and encoded according to the data type, and the time series feature X is extracted. t , spatial feature X s and modal features X m , and concatenate to form the input feature matrix X (0) =[X t ;X s ;X m ]; S23. In the attention calculation layer, recursive attention calculation is performed on the input feature matrix: X (l+1) =A (l) ·V (l) +X (l) ; Among them, A (l) represents the attention weight matrix of the lth layer, Q (l) represents the query matrix of the lth layer, K (l) represents the bond matrix of the lth layer, V (l) represents the value matrix of the lth layer, d k represents the dimension of the key vector, softmax represents the normalization function, X (l) represents the input feature matrix of the lth layer, X (l+1) Represents the input feature matrix of the l+1th layer; S24, in the multi-head attention aggregation layer, the weight matrix and feature matrix of multiple attention heads are combined to perform attention aggregation to obtain the aggregated feature matrix X multi ; S25, the feature matrix X after aggregation at the output layer multi Perform normalization to eliminate differences between modalities and generate the final risk feature X f ; S26. Sharing the risk feature X via a low-latency communication protocol f To the master node, and combined with the risk features shared by other edge nodes to generate a preliminary risk feature graph F risk .

3. The edge computing method for construction site safety management according to claim 1, characterized in that: The S3 specifically includes: S31, inputting the preliminary risk feature map into a distributed high-dimensional abnormal pattern projection module, wherein the distributed high-dimensional abnormal pattern projection module includes a data separation unit, a feature extraction unit, a dimensionality reduction projection unit, an abnormal pattern recognition unit and a data aggregation unit; S32, in the data separation unit, separating the preliminary risk feature map into video modal features, environmental sensor modal features and personnel trajectory modal features, and allocating the separated modal feature data to corresponding feature extraction units; S33, in the feature extraction unit, using the data structure and spatiotemporal correlation information of the modal features, extracting dynamic behavior features from the video modal features, extracting abnormal fluctuation features from the environmental sensor modality, extracting path deviation and clustering features from the personnel trajectory modality, and generating a modal feature matrix; S34, in the dimensionality reduction projection unit, the modal feature matrix is ​​input into the manifold learning algorithm processing pipeline, and the high-dimensional modal feature data is projected into the low-dimensional space through a nonlinear mapping method to generate a local abnormal projection matrix. The dimensionality reduction projection unit dynamically adjusts the calculation frequency according to the resource allocation and data characteristics of the current edge node; S35. In the abnormal pattern recognition unit, the local abnormal projection matrix is ​​used to perform pattern clustering and screening on the feature matrix in combination with the time-sensitive weight to generate a local abnormal projection result; S36. In the data aggregation unit, the local anomaly projection results are transmitted to the master node through a low-latency communication protocol. The master node aggregates the anomaly projection results from multiple edge nodes and integrates the multi-node data through a weighted splicing algorithm to finally generate a high-dimensional risk feature surface.

4. The edge computing method for construction site safety management according to claim 1, characterized in that: The S4 specifically includes: S41, deploying a dynamic event perception module in the edge node, wherein the dynamic event perception module includes a data acquisition unit, a feature extraction unit, an event triggering unit, an event classification unit and an event marking unit; S42, the data acquisition unit monitors the multimodal data stream in real time, including the video stream, the environmental sensor data and the personnel trajectory data, and forms a time series data block after assigning a timestamp; S43, in a feature extraction unit, processing the time series data block, extracting dynamic area features from the video stream, extracting gas concentration, temperature and humidity fluctuation features from the environmental sensor, extracting path deviation features from the trajectory data, and obtaining a multimodal feature set; S44. In the event triggering unit, the anomaly score is calculated using the extracted multimodal feature set: Among them, S abn represents the anomaly score, x m,i,t represents the value of the i-th feature of the m-th mode at the t-th time step, μ m,i represents the historical mean of the ith feature of the mth mode, σ m,i represents the standard deviation of the ith feature of the mth mode, ∈ represents the smoothing factor, N m represents the total number of modal features, T represents the time window length, and β represents the time attenuation coefficient; S45, when the abnormal score S abn When the set threshold is exceeded, the event classification unit is triggered, which classifies the events and associates the time, location and abnormal feature information for each event; S46. In the event marking unit, a high-risk event mark is generated according to the classification result, wherein the high-risk event mark includes event type, triggering time, location, and feature summary.

5. The edge computing method for construction site safety management according to claim 1, characterized in that: The S5 specifically includes: S51, deploying a dynamic risk weight tree search algorithm in a path optimization module, wherein the path optimization module includes a path tree construction unit, a risk weight calculation unit, and a strategy update unit; S52: The path tree construction unit divides the construction site into grid areas, each grid node n represents a location point, and the path tree starts from the initial location node s0 and extends to the target node s g , assign state s to each node of the path tree n , path cost C(n), comprehensive risk weight W(n) and exploration priority P(n); S53, in the risk weight calculation unit, the comprehensive risk weight is calculated for each network node n in the path tree: Where W(n) represents the comprehensive risk weight of network node n, t0 and t n represents the time range from the initial time to node n, K represents the total number of risk features, and γ k represents the weight coefficient of the kth risk feature, f k (n,t) represents the kth risk characteristic value of network node n at time t, σ k represents the historical standard deviation of the k-th risk characteristic, ∈ represents the smoothing factor, δ k represents the nonlinear amplification coefficient of the k-th risk feature, d k (n) represents the distance from network node n to the kth risk source, α k represents the time attenuation coefficient; S54. In the path tree, the root node is used as the initial position node, and node expansion is performed based on the dynamic risk weight tree search algorithm, and the optimal child node is selected in each expansion: Where P(n′) represents the comprehensive score of child node n′, W(n′) represents the risk weight of child node n′, C(n′) represents the cumulative path cost of child node n′, δ represents the path cost weight coefficient, λ represents the exploration balance coefficient, N(n) represents the number of visits to the parent node, N(n′) represents the number of visits to the child node, η represents the heuristic coefficient, and H(n′) represents the heuristic estimate from the node to the target node; S55, dynamically adjust the node selection logic in the policy update unit, obtain high-risk event markers in real time, and optimize the path search efficiency and security to generate a dynamic security path P safe .

6. The edge computing method for construction site safety management according to claim 1, characterized in that: The S6 specifically includes: S61, running the space-time hybrid prediction model in the edge node, taking the high-dimensional risk feature surface and the dynamic safety path as input, passing them to the time feature extraction unit and the spatial relationship modeling unit respectively, and initializing the space-time hybrid prediction model parameters; S62, performing time series segmentation on the high-dimensional risk feature surface, dividing it into multiple subsequences according to time steps; extracting the time dependency relationship of each subsequence using a time feature extraction unit, and generating a time feature matrix containing time dynamic change features; S63, constructing a spatial feature map of the construction site based on the dynamic safety path, defining the gridded area of ​​the construction site as a node, generating an adjacency graph based on the spatial connection relationship between the nodes, calculating the feature relationship of the nodes using a spatial relationship modeling unit, and generating a spatial feature matrix; S64, inputting the time feature matrix and the space feature matrix into the time-space feature fusion module, using the feature splicing method to combine the time dynamic change and spatial risk distribution information to generate the time-space feature matrix; S65, generating a local risk prediction result using the spatiotemporal feature matrix, wherein the local risk prediction result includes the potential risk value, occurrence time and impact range of each node; S66. Sort the local risk prediction results according to the time series, and mark the risk level and confidence level for each prediction result.

7. The edge computing method for construction site safety management according to claim 1, characterized in that: The S7 specifically includes: S71. Collect local risk prediction results of edge nodes through a layered spatiotemporal collaboration mechanism, annotate the local risk prediction result of each edge node with the corresponding time step and spatial position index, and form an edge node prediction data set; S72, integrating the local risk prediction results of all edge nodes in the master node, classifying the local risk prediction results of each time step according to the spatial grid area, and screening according to the spatiotemporal consistency of the prediction data set to eliminate abnormal prediction values; S73, based on the integrated prediction data set, generate a global time feature matrix T through a time aggregation algorithm global ,The time aggregation algorithm smoothes the risk prediction values ​​at different time steps by ,accumulating and weighted sliding window methods to eliminate the interference of short-term fluctuations on the ,global prediction; S74, running a spatial feature aggregation algorithm in the master node to generate a global spatial feature matrix according to the topological relationship of the spatial grid area, and analyzing the global spatial feature matrix to generate a global dynamic prediction result; S75. Synchronize the spatiotemporal feature weights, risk level parameters and time step information in the global dynamic prediction results to the edge nodes.

8. An edge computing system for construction site safety management, executing the edge computing method for construction site safety management according to any one of claims 1 to 7, characterized in that: Includes the following modules: The data acquisition module is used to collect multimodal data from the construction site and perform localized preprocessing to generate a preliminary cleansed data set; A lightweight recursive attention module, which is used to perform recursive attention analysis on the preliminary cleaned dataset to generate a preliminary risk feature map; Distributed high-dimensional anomaly pattern projection module, used for anomaly feature extraction and high-dimensional risk projection, generating local anomaly projection results and forming a high-dimensional risk feature surface; Dynamic event perception module, which is used to monitor the data flow of the construction site, identify high-risk events and generate high-risk event markers; A path optimization module for generating dynamic safe paths based on high-risk event markers; The spatiotemporal hybrid prediction module is used to generate local risk prediction results based on high-dimensional risk feature surfaces and dynamic safety paths; A hierarchical spatiotemporal collaboration module is used to integrate local risk prediction results, generate global dynamic prediction results and update edge node parameters; The risk warning and path optimization module is used to provide real-time risk warnings and path optimization suggestions to construction managers.

Images