Construction safety risk monitoring method and system

Abstract

The invention relates to the technical field of safety monitoring and intelligent early warning, discloses a construction safety risk monitoring method and system, and aims to solve the problems that a traditional construction safety monitoring means is isolated, static and lagged. The method comprises the following steps: deploying a multi-modal sensing network to collect environment, personnel, equipment and structural data; constructing a risk factor time sequence matrix through space-time alignment and feature normalization; constructing a dynamic association graph in combination with the risk conduction rule base and historical cases; predicting risk propagation potential energy of each node by using a graph neural network; triggering early warning according to the multi-level threshold value and outputting a conduction path and a blocking suggestion; the field devices are linked to perform hierarchical suppression actions. The system comprises a sensing deployment module, a data fusion module, a map construction module, a potential energy prediction module and an early warning decision and execution linkage module. Through space-time modeling and intelligent deduction, active identification and second-level intervention of a risk link are realized, the transformation of construction safety from passive response to active prevention is promoted, and the accident probability and the loss scale are reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of safety monitoring and intelligent early warning, and particularly relates to a construction safety risk monitoring method and system. BACKGROUND

[0002] With the acceleration of urbanization and the intensive landing of large infrastructure projects, construction safety has become the core issue of engineering management. Modern construction sites involve personnel, equipment, materials, environment and other multi-dimensional dynamic elements, and their interaction is complex and highly coupled. Traditional safety monitoring systems rely on sensor threshold alarms and manual inspection mechanisms, and their core logic is based on the static identification of isolated risk points, lacking the ability to model the spatial dimension diffusion and time dimension evolution of risks. This passive response mode can capture explicit abnormalities, but it cannot predict the systemic chain reaction caused by small hidden dangers, resulting in delayed safety intervention and accidents prevention and control falling into a passive situation of after-the-fact remediation.

[0003] Among them, the transmission mechanism of construction safety risk is particularly critical. Risks do not exist independently in a single work point or equipment unit, but are dynamically transmitted among different construction elements through time and space correlation paths such as physical proximity, process connection, and resource dependence. For example, local loosening of scaffolding may cause instability of adjacent support structures through mechanical transmission, and then affect the safety of high-altitude workers; electrical equipment short-circuit sparks may spread along the distribution path of flammable materials, eventually triggering a large-scale fire. Existing monitoring methods cannot quantify the risk diffusion intensity, predict the scope of the spread, or deduce the evolution time sequence because they have not built a correlation map between elements and have not introduced a propagation dynamics model.

[0004] Existing technologies generally ignore the dynamic modeling of risk transmission paths, resulting in a system that can only deal with isolated nodes that have triggered alarms, and cannot block the risk propagation chain in advance. Especially in high-density, multi-trade cross-operation scenarios, the speed of risk transmission grows exponentially, and the response delay of traditional methods can easily lead to major safety accidents. An intelligent monitoring system that can depict the spatio-temporal correlation of construction elements, simulate risk transmission paths, and output proactive intervention strategies is urgently needed to realize the paradigm shift from passive alarm to proactive foresight in safety management and control. SUMMARY

[0005] The application provides a construction safety risk monitoring method and system, which is characterized in that an intelligent monitoring system capable of performing spatio-temporal correlation modeling, dynamic transmission path deduction and forward-looking risk warning on multi-source heterogeneous risk factors in a construction site is constructed.

[0006] The application provides a construction safety risk monitoring method, which comprises the following steps: deploying a multi-modal perception network in the whole construction site, which is used for synchronously collecting physical environment parameters, personnel behavior trajectories, equipment operation states and structure deformation data; performing spatio-temporal alignment and feature normalization processing on the collected multi-source heterogeneous data, and constructing a risk factor time sequence matrix in a unified spatio-temporal coordinate system; constructing a dynamic correlation graph between risk factors based on a preset risk transmission rule library and a historical accident case library, wherein the correlation graph comprises risk source nodes, transmission path edges and path weight coefficients; performing joint embedding learning on the risk factor time sequence matrix and the dynamic correlation graph by using a graph neural network, so as to generate a potential risk propagation potential value of each risk node at the current time; comparing the propagation potential value with a preset multi-level risk threshold interval, and triggering a warning signal of the corresponding level when the propagation potential value of any risk node falls into a high-risk interval, and simultaneously outputting a visual topological structure of a risk transmission path and a suggestion of a key blocking node; pushing the warning signal and the blocking suggestion to a site control terminal, and connecting a site sound and light alarm device, a personnel positioning system and an equipment emergency stop interface, so as to execute a preset risk suppression action sequence.

[0007] Further, the multi-modal perception network comprises a distributed temperature and humidity sensor array, a dust concentration detection unit, a noise monitoring probe, a high-definition video monitoring camera, an infrared thermal imager, a laser range finder, an inclination sensor, a vibration accelerometer, a personnel radio frequency identification tag reader, a heavy machinery operation state acquisition module and a structure stress and strain monitoring piece; the distributed temperature and humidity sensor array is arranged at intervals of 50 m 2A collection point is arranged, the dust concentration detection unit is arranged at each 3m of the upwind direction and the downwind direction of the main dust operation area, the noise monitoring probe is installed on the surface of the high noise equipment shell and within the 1m radius range of the ear side of the operation personnel, the high-definition video monitoring camera covers all the main operation surfaces and the intersection of the passages, the lens focal length is adjustable in the range of 35mm to 150mm, the infrared thermal imager is aimed at the core heat dissipation components and the surface of the electrical cabinet of the large mechanical equipment, the temperature measurement range is-20℃ to 500℃, the laser range finder is installed at the end of the tower crane jib and the top of the base edge protective rail column, the measurement accuracy is ±1mm, the inclination sensor is fixed at the top of the scaffold vertical rod and the main beam node of the formwork support system, the measuring range is ±15°, the vibration accelerometer is pasted on the handle of the concrete vibrator and the base of the pile foundation construction hammering equipment, the sampling frequency is greater than 1KHz, the personnel radio frequency identification tag reader is arranged at the entrance of each operation area and the dangerous operation permission issuing point, the reading distance is adjustable in the range of 0.5m to 3m, the heavy machinery operation state collection module is connected with the engine control unit and the hydraulic system pressure sensor of the excavator, the crane and the concrete pump truck through the controller area network bus interface, the structure stress and strain monitoring sheet is pasted on the surface of the main reinforcement of the deep foundation pit support pile body and the middle part of the key compression rod of the large-span formwork support system, and the strain measurement range is ±5000με.

[0008] Further, the space-time alignment and feature normalization processing includes: taking the coordinated universal time output by the global satellite navigation system timing module as the reference time source, stamping all sensor data with a uniform timestamp, and the time synchronization error is less than 10ms; taking the two-dimensional Cartesian coordinate system established by the construction site general plan as the reference space framework, mapping the spatial positions of each sensor collection point to the unified coordinate plane through joint calibration of the laser range finder and the video monitoring camera, and the spatial positioning error is less than 5cm; the sliding window mean filtering algorithm is used for temperature and humidity data to eliminate instantaneous fluctuations, and the window length is 30s; the dust concentration data is trend smoothed by using the exponential weighted moving average method, and the weight coefficient is 0.85; the noise data is integrated and calculated according to the equivalent continuous sound level, and the integration period is 1min; the background difference method is used for video image data to extract the motion target contour, and the personnel identity and behavior intention are bound in combination with the radio frequency identification tag data; the regional maximum temperature extraction method is used for infrared thermal imaging data to obtain the device hotspot temperature value; the fast Fourier transform is used for inclination and vibration data to extract the main frequency component and amplitude feature; the temperature compensation algorithm is used for structure strain data to eliminate the influence of environmental temperature drift, and the compensation coefficient is obtained through laboratory calibration; finally, all the processed feature vectors are spliced into a unified format risk factor time sequence matrix according to the preset dimension, the number of rows corresponds to the time step, and the number of columns corresponds to the total number of risk factor types.

[0009] Further, the risk transmission rule base includes four types of basic rules of physical transmission rules, behavior transmission rules, equipment coupling rules and structure failure rules; the physical transmission rules define the transmission path of dust concentration rising leading to visibility decline and then causing mechanical collision accidents, the transmission delay time is 30s to two minutes after the dust concentration exceeds the standard, and the path weight coefficient is inversely proportional to the wind speed; the behavior transmission rules define the transmission path of personnel entering high work area without wearing safety helmets leading to falling objects hitting the head, the transmission delay time is within 0.5s after the personnel enter the dangerous area, and the path weight coefficient is proportional to the square of the work height; the equipment coupling rules define the transmission path of tower crane overload operation leading to steel wire rope fracture and then causing the falling of hoisted objects, the transmission delay time is after the overload lasts more than 5s, and the path weight coefficient is proportional to the product of the mass of hoisted objects and the lifting height; the structure failure rules define the transmission path of foundation pit support pile body strain exceeding the limit leading to soil instability and then causing collapse, the transmission delay time is 10min to 30min after the strain value lasts more than the threshold value, and the path weight coefficient is proportional to the product of the water content of the soil and the spacing of the support piles; the historical accident case base stores the whole process data of typical accidents of the same type of project occurring in the past five years, including the environmental parameter change curve within 12h before the accident, the personnel and equipment activity log, the abnormal fluctuation mode of structure monitoring data, and the final accident type and loss degree, which are used to dynamically calibrate and update the path weight coefficients in the risk transmission rule base.

[0010] Further, the graph neural network adopts a spatio-temporal graph convolution network architecture, the input layer receives the risk factor time series matrix and the dynamic correlation graph, the hidden layer includes three graph convolution modules, the neighborhood aggregation range of each graph convolution module is first-order adjacent nodes, second-order adjacent nodes and third-order adjacent nodes in turn, the activation function adopts a rectified linear unit, and the output layer generates the propagation potential prediction value of each risk node at the next time step; the calculation formula of the propagation potential prediction value is: the propagation potential is equal to the current node risk intensity value multiplied by the sum of the path weight coefficients and then multiplied by the time decay factor, and the time decay factor is the negative time step power of the natural constant; the multi-level risk threshold interval is divided into a low risk interval, a medium risk interval, a high risk interval and an urgent risk interval, corresponding to the propagation potential value less than 0.3, greater than or equal to 0.3 and less than 0.6, greater than or equal to 0.6 and less than 0.9, and greater than or equal to 0.9 respectively; when the propagation potential value falls into the high risk interval, a yellow warning is triggered, and a risk transmission path topological graph is pushed to the project manager mobile terminal; when the propagation potential value falls into the urgent risk interval, a red warning is triggered, and the on-site sound and light alarm is started synchronously, the evacuation instruction is sent to the smart safety helmet worn by the relevant operating personnel, the speed reduction or shutdown instruction is sent to the heavy machinery controller, and the increased pumping flow instruction is sent to the foundation pit dewatering system.

[0011] Further, the risk inhibition action sequence includes a first level inhibition action, a second level inhibition action and a third level inhibition action; the first level inhibition action is executed when the yellow pre-warning is triggered, and the content is to broadcast voice prompts to the relevant area, to highlight the risk transmission path on the monitoring large screen, and to push the inspection task to the handheld terminal of the safety officer; the second level inhibition action is executed within 5s after the red pre-warning is triggered, and the content is to cut off the unnecessary power supply of the relevant area, to start the emergency lighting system, to open all safety passage access control, and to send emergency braking instructions to the tower crane and the elevator; the third level inhibition action is executed within 15s after the red pre-warning is triggered, and the content is to start the automatic spraying dust reduction system around the foundation pit, to activate the prestressed tension device of the structural support system, to send a linkage request to the fire control center, and to send the estimated number and the injury type information of the wounded to the hospital emergency center.

[0012] According to another aspect of the present application, a construction safety risk monitoring system is provided, which comprises: a multi-modal sensing network deployment unit for arranging various types of environmental, personnel, equipment and structural monitoring sensors at a preset density and position in a construction site; a space-time data fusion processing unit for performing time synchronization, space mapping, noise filtering and feature extraction on multi-source heterogeneous sensing data to generate a standardized risk factor time sequence matrix; a risk transmission graph construction unit for loading a preset risk transmission rule library and a historical accident case library, dynamically generating and updating the correlation graph between risk factors; a risk propagation potential prediction unit for calling a graph neural network model to perform propagation path deduction and potential value calculation on the current risk situation; a multi-level pre-warning decision unit for comparing the propagation potential value with a preset threshold interval, determining the pre-warning level and generating the corresponding pre-warning signal and blocking node suggestion; a risk inhibition execution unit for receiving the pre-warning signal, and executing risk intervention measures according to a preset action sequence to link various types of security, mechanical and electrical and emergency equipment in the field; a man-machine interactive display unit for synchronously displaying the risk heat map, the transmission path animation, the pre-warning information list and the equipment linkage state on the central control room large screen and the mobile terminal.

[0013] Further, the multi-modal sensing network deployment unit comprises a sensor selection sub-module, an installation positioning sub-module and a communication networking sub-module; the sensor selection sub-module automatically matches the required sensor types and quantities according to the engineering type and risk level, and outputs the equipment list and the technical parameter table; the installation positioning sub-module generates an optimal sensor distribution scheme based on the building information model, including three-dimensional coordinates, installation height, orientation angle and fixing method; the communication networking sub-module uses an industrial wireless mesh network protocol to construct a self-healing data transmission network, and the maximum hop number between network nodes is not more than five hops, and the end-to-end transmission delay is less than 50ms.

[0014] Further, the spatio-temporal data fusion processing unit comprises a time synchronization submodule, a spatial registration submodule, a feature extraction submodule, and a matrix generation submodule; the time synchronization submodule acquires a time signal from a reference clock source through a network time protocol, adds a time stamp to each sensor data packet, and performs interpolation alignment; the spatial registration submodule converts sensor physical coordinates to a unified construction coordinate system using a laser scanning point cloud and video image feature point matching algorithm; the feature extraction submodule calls a corresponding signal processing algorithm library according to the sensor type, and outputs a standardized feature vector; the matrix generation submodule aligns all feature vectors according to time steps, sorts them according to risk factor categories, and splices them into a two-dimensional matrix structure.

[0015] Further, the risk propagation graph construction unit comprises a rule loading submodule, a case matching submodule, a graph initialization submodule, and a weight updating submodule; the rule loading submodule reads the latest version of the risk propagation rule library from local storage or a cloud server, parses the source node, target node, propagation condition, and initial weight in the rule; the case matching submodule performs similarity calculation on the current engineering features and the metadata in the historical accident case library, and selects the three cases with the highest similarity as weight calibration samples; the graph initialization submodule generates an initial directed weighted graph according to the loaded rules, with nodes as risk factors and edges as propagation paths; the weight updating submodule performs Bayesian posterior probability correction on the initial weight according to the actual accident evolution path in the calibration sample, and outputs the dynamically adjusted risk propagation graph.

[0016] Further, the risk propagation potential prediction unit comprises a model loading submodule, a graph embedding submodule, a convolution calculation submodule, and a potential output submodule; the model loading submodule calls a pre-trained spatio-temporal graph convolution network parameter file from a model repository to complete model instantiation; the graph embedding submodule converts the risk factor time series matrix and the dynamic correlation graph into an adjacency matrix and a node feature matrix that can be processed by a graph neural network; the convolution calculation submodule performs three-layer graph convolution operations, and performs batch normalization and residual connection after each layer operation; the potential output submodule performs full connection transformation on the final node embedding vector, and outputs the propagation potential prediction value of each risk node at the next time step.

[0017] Further, the multi-level early warning decision unit comprises a threshold comparison submodule, an early warning generation submodule, a path analysis submodule and an instruction packaging submodule; the threshold comparison submodule compares the propagation potential prediction value with the four-level threshold interval to determine the highest matching level; the early warning generation submodule generates an early warning message comprising the early warning level, the trigger time, the risk node number and the potential value according to the matching level; the path analysis submodule traces all the transmission paths from the source risk node to the current node, calculates the path cumulative weight and sorts it, and extracts the three paths with the highest weight as the key transmission links; the instruction packaging submodule packages the early warning message and the key transmission link information into a structured instruction data packet, and adds the execution priority and the timeliness label.

[0018] Further, the risk suppression execution unit comprises an instruction analysis submodule, a device addressing submodule, an action sequence scheduling submodule and an execution state feedback submodule; the instruction analysis submodule disassembles the structured instruction data packet to extract the early warning level and the key transmission link information; the device addressing submodule maps the risk node type in the key transmission link to the corresponding on-site execution device address list; the action sequence scheduling submodule matches the preset action sequence according to the early warning level to generate a device control instruction queue, which is distributed to each execution device in priority order; the execution state feedback submodule receives the execution confirmation signals returned by each device, generates an execution state report after summarizing, and returns it to the man-machine interaction display unit.

[0019] Further, the man-machine interaction display unit comprises a risk heat map rendering submodule, a transmission path animation submodule, an early warning information list submodule and a device state dashboard submodule; the risk heat map rendering submodule maps the propagation potential value of each risk node to a color gradient and displays it on the construction plan; the transmission path animation submodule dynamically displays the key transmission link in the form of flowing arrows, and the arrow thickness is proportional to the path weight; the early warning information list submodule arranges all early warning events in reverse chronological order, and supports filtering by level, area and type; the device state dashboard submodule displays the online state, instruction reception state and action execution progress of each linked device in real time.

[0020] Compared with the prior art, the advantages and positive effects of the present application are: 1. By constructing a multi-modal perception network covering the environment, personnel, equipment and structure, full-factor risk factor acquisition of the construction site is realized. 2. By using spatio-temporal data fusion processing technology, heterogeneous data is unified to a standard spatio-temporal framework to provide high-quality input for risk modeling. 3. By introducing dynamic risk transmission atlas and graph neural network, the quantitative deduction of risk space-time transmission path and the prediction of transmission potential are realized for the first time, breaking through the limitation of traditional methods that can only identify isolated risk points; 4. By setting multi-level early warning threshold and hierarchical inhibition action sequence, the safety control measures are accurately matched with the risk severity, avoiding excessive intervention or insufficient response; 5. Through automatic linkage execution of the whole system, the time delay from risk identification to intervention measure landing is greatly shortened, and the accident prevention window period is improved from minutes to seconds; 6. Ultimately, the construction safety risk is fundamentally changed from post-tracing to pre-judgment, from point alarm to link blocking, and from manual decision to intelligent linkage, reducing the probability of occurrence of major safety accidents and the potential loss scale. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the overall technical scheme architecture schematic diagram of a construction safety risk monitoring method and system proposed by the present application; Figure 2 is the core principle framework schematic diagram of risk transmission path deduction and transmission potential prediction based on graph neural network in the present application; Figure 3 is the logic flow framework diagram of spatio-temporal alignment and feature normalization processing of multi-modal perception data in the present application; Figure 4 is the logic flow framework diagram of dynamic risk transmission atlas construction and weight calibration mechanism in the present application; Figure 5 is the logic flow framework diagram of multi-level early warning decision and risk inhibition action linkage execution in the present application. DETAILED DESCRIPTION

[0022] Please refer to Figures 1 to 5 The present application provides a construction safety risk monitoring method and system, which is characterized by constructing an intelligent monitoring system capable of modeling the spatio-temporal correlation of multi-source heterogeneous risk factors in the construction site, deducing the dynamic transmission path and providing early warning of potential risks. The system breaks through the limitation of traditional monitoring methods that can only identify isolated risk points and provide post-alarm, establishes a spatio-temporal coupling relationship atlas among risk factors, combines real-time environmental perception data and historical accident evolution patterns, realizes active identification of potential risk transmission links, quantitative evaluation of transmission intensity and dynamic triggering of multi-level early warning thresholds, thereby providing risk intervention decision support before accidents occur, and achieving a fundamental change from passive response to active prevention in construction safety control.

[0023] In the embodiment, the construction safety risk monitoring method comprises: a multi-modal perception network deployed in the whole construction site, which is used to synchronously collect physical environment parameters, personnel behavior trajectories, equipment operation states and structure deformation data; The multi-source heterogeneous data collected is subjected to spatio-temporal alignment and feature normalization processing to construct a risk factor time sequence matrix in a unified spatio-temporal coordinate system; Based on a preset risk transmission rule library and a historical accident case library, a dynamic correlation graph between risk factors is constructed, the correlation graph comprising risk source nodes, transmission path edges and path weight coefficients; a graph neural network is used to jointly embed and learn the risk factor time sequence matrix and the dynamic correlation graph to generate a potential risk transmission potential value of each risk node at the current time; According to the transmission potential value and a preset multi-level risk threshold interval, when the transmission potential value of any risk node falls into a high risk interval, a corresponding level of early warning signal is triggered, and a visual topological structure of the risk transmission path and a key blocking node suggestion are synchronously output; the early warning signal and the blocking suggestion are pushed to a site control terminal, and a site sound-light alarm device, a personnel positioning system and an equipment emergency stop interface are linked to execute a preset risk suppression action sequence.

[0024] The multi-modal perception network comprises a distributed temperature and humidity sensor array, a dust concentration detection unit, a noise monitoring probe, a high-definition video monitoring camera, an infrared thermal imager, a laser range finder, an inclination sensor, a vibration accelerometer, a personnel radio frequency identification tag reader, a heavy machinery operation state acquisition module and a structure stress and strain monitoring sheet.

[0025] The distributed temperature and humidity sensor array is arranged at intervals of 50 m 2A collection point is arranged, the dust concentration detection unit is arranged at each 3m of the upwind direction and the downwind direction of the main dust operation area, the noise monitoring probe is installed on the surface of the high noise equipment shell and within the 1m radius range of the ear side of the operation personnel, the high-definition video monitoring camera covers all the main operation surfaces and channel intersections, the lens focal length is adjustable in the range of 35mm to 150mm, the infrared thermal imager is aimed at the core heat dissipation components and the surface of the electrical cabinet of the large mechanical equipment, the temperature measurement range is-20℃ to 500℃, the laser range finder is installed at the end of the tower crane jib and the top of the base pit edge protection rail column, the measurement accuracy is ±1mm, the inclination sensor is fixed at the top of the scaffold stand and the main beam node of the formwork support system, the measurement range is ±15°, the vibration accelerometer is pasted on the handle of the concrete vibrator and the base of the pile foundation construction hammering equipment, the sampling frequency is greater than 1KHz, the personnel radio frequency identification tag reader is arranged at the entrance of each operation area and the dangerous operation permission issuing point, the reading distance is adjustable in the range of 0.5m to 3m, the heavy machinery operating state collection module is connected to the engine control unit and the hydraulic system pressure sensor of the excavator, the crane and the concrete pump truck through the controller area network bus interface, the structural stress and strain monitoring sheet is pasted on the surface of the main reinforcement of the deep foundation pit support pile and the middle part of the key compression rod of the large-span formwork support system, and the strain measurement range is ±5000με.

[0026] In the data collection stage, all sensors continuously collect raw data with millisecond-level time granularity. The temperature and humidity sensor outputs relative humidity percentage and Celsius temperature value, the dust concentration detection unit outputs milligrams per cubic meter concentration value, the noise monitoring probe outputs decibel value, the high-definition video monitoring camera outputs 25fps compressed video stream, the infrared thermal imager outputs per pixel temperature matrix, the laser range finder outputs millimeter-level distance value, the inclination sensor outputs angle value, the vibration accelerometer outputs three-axis acceleration components, the personnel radio frequency identification tag reader outputs tag unique identifier and reading timestamp, the heavy machinery operating state collection module outputs engine speed, hydraulic pressure, oil temperature, load percentage and other parameters, and the structural stress and strain monitoring sheet outputs micro-strain value. All raw data is transmitted to the edge computing node through the industrial wireless mesh network protocol, the maximum hop number between network nodes is not more than five, and the end-to-end transmission delay is less than 50ms.

[0027] In the space-time alignment and feature normalization processing stage, the coordinated universal time output by the global satellite navigation system timing module is taken as the reference time source, and all sensor data is stamped with a unified timestamp, and the time synchronization error is less than 10ms.

[0028] A two-dimensional Cartesian coordinate system established by the total plan of the construction site is taken as the reference space framework. The spatial positions of the collection points of various sensors are mapped to a unified coordinate plane through joint calibration of a laser range finder and a video monitoring camera, and the spatial positioning error is less than 5 cm. The sliding window mean filtering algorithm is used to eliminate instantaneous fluctuations of the temperature and humidity data, and the window length is 30 s, and an average value is output every 30 s.

[0029] The exponential weighted moving average method is used to perform trend smoothing on the dust concentration data, the weight coefficient is 0.85, and the smoothed concentration value is output.

[0030] The equivalent continuous sound level is calculated by integrating the noise data, the integration period is 1 min, and the equivalent sound level value is output every minute.

[0031] The background difference method is used to extract the motion target contour of the video image data, and the personnel identity and behavior intention are bound in combination with the radio frequency identification tag data, and the personnel trajectory coordinate sequence and behavior classification label are output.

[0032] The regional maximum temperature extraction method is used to obtain the device hotspot temperature value of the infrared thermal imaging data, and the highest temperature point coordinate and temperature value are output.

[0033] The main frequency component and amplitude feature are extracted by using the fast Fourier transform on the inclination and vibration data, and the dominant frequency and peak acceleration are output.

[0034] The temperature compensation algorithm is used to eliminate the influence of environmental temperature drift on the structural strain data, and the compensation coefficient is obtained through laboratory calibration, and the compensated strain value is output.

[0035] Finally, all the processed feature vectors are spliced into a unified format risk factor time series matrix according to the preset dimension, the number of rows of the matrix corresponds to the time step, the number of columns corresponds to the total number of risk factor categories, and each row represents the standardized characteristic value of all risk factors in a time step.

[0036] The risk transmission rule library includes four basic rules of physical transmission rule, behavior transmission rule, device coupling rule and structure failure rule.

[0037] The physical transmission rule defines the transmission path of the dust concentration rise leading to the visibility decline and then causing mechanical collision accidents, the transmission delay time is 30 s to 2 min after the dust concentration exceeds the standard, and the path weight coefficient is inversely proportional to the wind speed.

[0038] The behavior transmission rule defines the transmission path of the personnel not wearing safety helmets entering the high-altitude work area leading to the head being hit by falling objects, the transmission delay time is within 0.5 s after the personnel enter the dangerous area, and the path weight coefficient is proportional to the square of the working height.

[0039] The device coupling rule defines the conduction path of the tower crane overload operation leading to the wire rope fracture and then causing the falling of the hoisted object, the conduction delay time is more than 5s after the overload, and the path weight coefficient is proportional to the product of the hoisted object mass and the lifting height.

[0040] The structure failure rule defines the conduction path of the foundation pit support pile body strain exceeding the threshold value leading to soil instability and then causing collapse, the conduction delay time is 10min to 30min after the strain value continuously exceeds the threshold value, and the path weight coefficient is proportional to the product of the soil water content and the support pile spacing.

[0041] The historical accident case library stores the whole process data of typical accidents of the same type of project occurring in the past five years, including the environmental parameter change curve within 12h before the accident, the personnel equipment activity log, the abnormal fluctuation mode of the structure monitoring data, and the final accident type and loss degree, which is used to dynamically calibrate and update the path weight coefficient in the risk conduction rule library.

[0042] In the risk conduction map construction stage, first, the preset risk conduction rule library is loaded, and the source node, target node, conduction condition and initial weight in each rule are parsed.

[0043] Subsequently, the similarity of the current engineering characteristics such as engineering type, geological condition, climate division, construction stage and the metadata in the historical accident case library is calculated, and the three cases with the highest similarity are selected as the weight calibration samples.

[0044] According to the loaded rules, an initial directed weighted graph is generated, the nodes are risk factors, and the edges are conduction paths. According to the actual accident evolution path in the calibration sample, the initial weight is modified by the Bayesian posterior probability, and the dynamically adjusted risk conduction map is output. For example, if the historical case shows that under the same geological conditions, the foundation pit collapse accident is directly caused by the support pile strain exceeding the threshold value, and the conduction delay time is generally shorter than the preset value, then the weight coefficient of the corresponding path will be adjusted upwards, and the conduction delay time will be shortened.

[0045] In the risk propagation potential prediction stage, a spatio-temporal graph convolution network architecture is adopted. The input layer receives the risk factor time series matrix and the dynamic association map, the hidden layer includes three graph convolution modules, the neighborhood aggregation range of each graph convolution module is first-order adjacent nodes, second-order adjacent nodes and third-order adjacent nodes in turn, the activation function adopts the rectified linear unit, and the output layer generates the propagation potential prediction value of each risk node at the next time step. The calculation formula of the propagation potential prediction value is: The propagation potential is equal to the sum of the product of the current node risk intensity value and the path weight coefficient and the time decay factor, and the time decay factor is the negative time step power of the natural constant.

[0046] The multi-level risk threshold interval is divided into a low risk interval, a medium risk interval, a high risk interval and an emergency risk interval, corresponding to a propagation potential value less than 0.3, greater than or equal to 0.3 and less than 0.6, greater than or equal to 0.6 and less than 0.9, and greater than or equal to 0.9. When the propagation potential value falls into the high risk interval, a yellow warning is triggered, and a risk transmission path topology is pushed to the project manager's mobile terminal. When the propagation potential value falls into the emergency risk interval, a red warning is triggered, and a synchronous on-site sound and light alarm is started, an evacuation instruction is sent to the smart safety helmet worn by the relevant operating personnel, a speed reduction or shutdown instruction is sent to the heavy machinery controller, and an increased pumping flow instruction is sent to the foundation dewatering system.

[0047] The risk suppression action sequence includes a first-level suppression action, a second-level suppression action and a third-level suppression action. The first-level suppression action is executed when the yellow warning is triggered, and the content is to broadcast voice prompts to the relevant area, highlight the risk transmission path on the monitoring large screen, and push the inspection task to the safety officer's handheld terminal. The second-level suppression action is executed within 5s after the red warning is triggered, and the content is to cut off unnecessary power supply in the relevant area, start the emergency lighting system, open all safety passage access control, and send an emergency braking instruction to the tower crane and elevator. The third-level suppression action is executed within 15s after the red warning is triggered, and the content is to start the automatic spraying dust suppression system around the foundation pit, activate the prestressed tension device of the structural support system, send a linkage request to the fire control center, and send the estimated number and type of injury information to the hospital emergency center.

[0048] In the warning decision and execution stage, first, the propagation potential prediction value is compared with the four-level threshold interval to determine the highest matching level. According to the matching level, a warning message containing the warning level, trigger time, risk node number and potential value is generated. All transmission paths from the source risk node to the current node are traced back, the path cumulative weight is calculated and sorted, and the three paths with the highest weight are extracted as the key transmission links. The warning message and key transmission link information are packaged into a structured instruction data packet, with an execution priority and timeliness label. The structured instruction data packet is disassembled, and the warning level and key transmission link information are extracted. According to the type of risk node in the key transmission link, it is mapped to the address list of the corresponding on-site execution device. According to the warning level, the preset action sequence is matched to generate a device control instruction queue, which is sorted by priority and distributed to each execution device. The execution confirmation signals returned by each device are received, and an execution status report is generated after being summarized and returned to the man-machine interaction display unit.

[0049] In the human-computer interaction display stage, the propagation potential values of the risk nodes are mapped into a color gradient, which is superimposed on the construction plan to form a risk heat map. The key transmission links are dynamically displayed in the form of arrow flow, and the thickness of the arrow is proportional to the path weight, forming a transmission path animation. All early warning events are arranged in reverse chronological order, and can be filtered by level, region and type to form a list of early warning information. The online status, instruction receiving status and action execution progress of each linked device are displayed in real time to form a device status board.

[0050] The construction safety risk monitoring system comprises: A multi-modal perception network deployment unit is configured to arrange various types of environmental, personnel, equipment and structural monitoring sensors at a preset density and location on the construction site. A space-time data fusion processing unit is configured to perform time synchronization, space mapping, noise filtering and feature extraction on multi-source heterogeneous perception data to generate a standardized risk factor time series matrix. A risk transmission map construction unit is configured to load a preset risk transmission rule library and a historical accident case library, and dynamically generate and update the correlation map between risk factors. A risk propagation potential prediction unit is configured to call a graph neural network model to perform propagation path deduction and potential value calculation on the current risk situation. A multi-level early warning decision unit is configured to compare the propagation potential value with a preset threshold interval, determine the early warning level, and generate corresponding early warning signals and blocking node suggestions. A risk suppression execution unit is configured to receive early warning signals and execute risk intervention measures on various security, mechanical and electrical and emergency equipment on site according to a preset action sequence. A human-computer interaction display unit is configured to display the risk heat map, transmission path animation, early warning information list and device linkage status on the central control room large screen and mobile terminal simultaneously.

[0051] The multi-modal perception network deployment unit includes a sensor selection sub-module, an installation positioning sub-module and a communication networking sub-module. The sensor selection sub-module automatically matches the required sensor types and quantities according to the engineering type and risk level, and outputs the device list and technical parameter table. The installation positioning sub-module generates an optimal sensor distribution scheme based on the building information model, including three-dimensional coordinates, installation height, orientation angle and fixing method. The communication networking sub-module uses an industrial wireless mesh network protocol to build a self-healing data transmission network, with a maximum hop count of no more than five hops between network nodes, and an end-to-end transmission delay of less than 50 ms.

[0052] The spatio-temporal data fusion processing unit comprises a time synchronization submodule, a spatial registration submodule, a feature extraction submodule, and a matrix generation submodule. The time synchronization submodule obtains a time signal from a reference clock source through a network time protocol, adds a time stamp to each sensor data packet, and performs interpolation alignment. The spatial registration submodule converts sensor physical coordinates to a unified construction coordinate system using a laser scanning point cloud and video image feature point matching algorithm. The feature extraction submodule calls corresponding signal processing algorithm libraries according to sensor types and outputs standardized feature vectors. The matrix generation submodule aligns all feature vectors according to time steps, sorts them according to risk factor categories, and splices them into a two-dimensional matrix structure.

[0053] The risk transmission graph construction unit comprises a rule loading submodule, a case matching submodule, a graph initialization submodule, and a weight updating submodule. The rule loading submodule reads the latest version of the risk transmission rule library from local storage or a cloud server, parses the source nodes, target nodes, transmission conditions, and initial weights in the rules. The case matching submodule performs similarity calculation on the current engineering features and the metadata in the historical accident case library, and selects the three cases with the highest similarity as weight calibration samples. The graph initialization submodule generates an initial directed weighted graph according to the loaded rules, with nodes as risk factors and edges as transmission paths. The weight updating submodule performs Bayesian posterior probability correction on the initial weights according to the actual accident evolution path in the calibration samples, and outputs the dynamically adjusted risk transmission graph.

[0054] The risk transmission graph construction unit comprises a rule loading submodule, a case matching submodule, a graph initialization submodule, and a weight updating submodule. The rule loading submodule reads the latest version of the risk transmission rule library from local storage or a cloud server, parses the source nodes, target nodes, transmission conditions, and initial weights in the rules. The case matching submodule performs similarity calculation on the current engineering features and the metadata in the historical accident case library, and selects the three cases with the highest similarity as weight calibration samples. The graph initialization submodule generates an initial directed weighted graph according to the loaded rules, with nodes as risk factors and edges as transmission paths. The weight updating submodule performs Bayesian posterior probability correction on the initial weights according to the actual accident evolution path in the calibration samples, and outputs the dynamically adjusted risk transmission graph.

[0055] The multi-stage early warning decision unit comprises a threshold comparison submodule, an early warning generation submodule, a path analysis submodule, and an instruction packaging submodule. The threshold comparison submodule compares the propagation potential prediction value with the four-level threshold interval to determine the highest matching level. The early warning generation submodule generates an early warning message containing the early warning level, trigger time, risk node number, and potential value according to the matching level. The path analysis submodule traces all the transmission paths from the source risk node to the current node, calculates the path cumulative weight and sorts it, and extracts the top three paths with the highest weight as the key transmission links. The instruction packaging submodule packages the early warning message and key transmission link information into a structured instruction data packet, and adds an execution priority and timeliness label.

[0056] The risk suppression execution unit comprises an instruction analysis submodule, a device addressing submodule, an action sequence scheduling submodule, and an execution state feedback submodule. The instruction analysis submodule disassembles the structured instruction data packet to extract the early warning level and key transmission link information. The device addressing submodule maps the risk node type in the key transmission link to the corresponding on-site execution device address list. The action sequence scheduling submodule matches the preset action sequence according to the early warning level to generate a device control instruction queue, which is sorted by priority and then distributed to each execution device. The execution state feedback submodule receives the execution confirmation signals returned by each device, generates an execution state report after summarizing, and returns it to the human-computer interaction display unit.

[0057] The human-computer interaction display unit comprises a risk heat map rendering submodule, a transmission path animation submodule, an early warning information list submodule, and a device state dashboard submodule. The risk heat map rendering submodule maps the propagation potential value of each risk node to a color gradient and displays it on the construction plan. The transmission path animation submodule dynamically displays the key transmission links in the form of flowing arrows, with the arrow thickness proportional to the path weight. The early warning information list submodule arranges all early warning events in reverse chronological order, and supports filtering by level, region, and type. The device state dashboard submodule displays the online status, instruction reception status, and action execution progress of each linked device in real time.

[0058] The embodiment realizes the dead angle-free collection of all-factor risk factors in the construction site by constructing a multi-modal perception network covering the environment, personnel, equipment and structure. Through spatio-temporal data fusion processing technology, heterogeneous data is unified to a standard spatio-temporal framework to provide high-quality input for risk modeling. By introducing dynamic risk transmission atlas and graph neural network, the risk spatio-temporal transmission path is quantitatively deduced and the propagation potential is predicted for the first time, breaking through the limitation of traditional methods which can only identify isolated risk points. By setting multi-level early warning thresholds and hierarchical inhibition action sequences, the safety control measures are accurately matched with the risk severity, avoiding excessive intervention or insufficient response. Through the automatic linkage execution of the whole system, the time delay from risk identification to intervention measures landing is greatly shortened, and the accident prevention window period is improved from minutes to seconds. Finally, the construction safety risk is fundamentally changed from post-tracing to pre-judging, from point alarm to link blocking, and from manual decision to intelligent linkage, reducing the probability of occurrence and potential loss of major safety accidents.

[0059] It should be noted that, in this text, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0060] Although the embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A construction safety risk monitoring method characterized by, The application relates to a construction site risk prediction and control method based on a multi-modal perception network. A multi-modal perception network is deployed in the whole construction site to synchronously collect physical environment parameters, personnel behavior trajectories, equipment operation states and structure deformation data. The collected multi-source heterogeneous data is subjected to space-time alignment and feature normalization processing to construct a risk factor time sequence matrix in a unified space-time coordinate system. Based on a preset risk transmission rule library and a historical accident case library, a dynamic correlation graph between risk factors is constructed, and the correlation graph comprises risk source nodes, transmission path edges and path weight coefficients. A graph neural network is used to jointly embed and learn the risk factor time sequence matrix and the dynamic correlation graph to generate a potential risk propagation potential value of each risk node at the current time. The propagation potential value is compared with a preset multi-level risk threshold interval, and when the propagation potential value of any risk node falls into a high-risk interval, a corresponding level of early warning signal is triggered, and a visual topological structure of a risk transmission path and a key blocking node suggestion are synchronously output. The early warning signal and the blocking suggestion are pushed to a site control terminal, and a sound and light alarm device, a personnel positioning system and an equipment emergency stop interface are connected to execute a preset risk suppression action sequence.

2. The construction safety risk monitoring method according to claim 1, characterized by, The collected multi-source heterogeneous data is subjected to space-time alignment and feature normalization processing to construct a risk factor time sequence matrix in a unified space-time coordinate system. The coordinated universal time output by a global satellite navigation system timing module is used as a reference time source to stamp a unified time stamp on all sensor data, and the time synchronization error is less than 10 ms. A two-dimensional Cartesian coordinate system established based on a total site plan is used as a reference space framework, and the space positions of all sensor collection points are mapped to a unified coordinate plane through joint calibration of a laser range finder and a video monitoring camera, and the space positioning error is less than 5 cm. A sliding window mean filtering algorithm is used to eliminate instantaneous fluctuations of temperature and humidity data, and the window length is 30 s; an exponential weighted moving average method is used to perform trend smoothing on dust concentration data, and the weight coefficient is 0.85; noise data are integrated and calculated according to the equivalent continuous sound level, and the integration period is 1 min; a background difference method is used to extract the motion target contour of video image data, and the personnel identity and behavior intention are bound in combination with radio frequency identification tag data; a regional maximum temperature extraction method is used to obtain the hotspot temperature value of infrared thermal imaging data; a fast Fourier transform is used to extract the main frequency component and amplitude feature of the inclination and vibration data; a temperature compensation algorithm is used to eliminate the influence of environmental temperature drift on the structure strain data, and the compensation coefficient is obtained through laboratory calibration; All processed feature vectors are spliced into a unified format risk factor time sequence matrix according to a preset dimension, and the number of rows corresponds to the time step, and the number of columns corresponds to the total number of risk factor categories.

3. The construction safety risk monitoring method according to claim 2, wherein, Based on a preset risk transmission rule library and a historical accident case library, a dynamic correlation graph between risk factors is constructed, and the correlation graph comprises risk source nodes, transmission path edges and path weight coefficients. The preset risk transmission rule library is loaded, and the risk transmission rule library comprises four basic rules of physical transmission rules, behavior transmission rules, equipment coupling rules and structure failure rules. Similarity calculation is performed between the current engineering characteristics and the metadata in the historical accident case library, and the three cases with the highest similarity are selected as the weight calibration samples; An initial directed weighted graph is generated according to the loaded rules, with nodes being risk factors and edges being transmission paths; The initial weights are modified according to the actual accident evolution path in the calibration sample, and the dynamically adjusted risk transmission graph is output.

4. The construction safety risk monitoring method according to claim 3, characterized by, The risk factor time series matrix and the dynamic correlation graph are jointly embedded by using a graph neural network to generate the potential risk propagation potential value of each risk node at the current time, including: A pre-trained spatio-temporal graph convolution network model is called, the input layer of which receives the risk factor time series matrix and the dynamic correlation graph, the hidden layer of which contains three graph convolution modules, and the neighborhood aggregation range of each graph convolution module is first-order adjacent nodes, second-order adjacent nodes and third-order adjacent nodes, and the activation function uses a rectified linear unit; Three-layer graph convolution operations are performed, and batch normalization and residual connection are performed after each operation. The final node embedding vector is transformed by full connection, and the propagation potential prediction value of each risk node at the next time step is output, which is equal to the current node risk intensity value multiplied by the sum of path weight coefficients and then multiplied by a time decay factor, which is the negative time step power of a natural constant.

5. The construction safety risk monitoring method according to claim 4, wherein According to the comparison between the propagation potential value and the preset multi-level risk threshold interval, when the propagation potential value of any risk node falls into the high-risk interval, the corresponding level of early warning signal is triggered, including: The propagation potential prediction value is compared with the four-level threshold interval, which is divided into low-risk interval, medium-risk interval, high-risk interval and emergency risk interval, corresponding to propagation potential value less than 0.3, greater than or equal to 0.3 and less than 0.6, greater than or equal to 0.6 and less than 0.9, greater than or equal to 0.9; When the propagation potential value falls into the high-risk interval, a yellow warning is triggered, and the risk transmission path topological graph is pushed to the project manager's mobile terminal; When the propagation potential value falls into the emergency risk interval, a red warning is triggered, and the on-site sound and light alarm is started synchronously, the evacuation instruction is sent to the smart safety helmet worn by the relevant operating personnel, the speed reduction or shutdown instruction is sent to the heavy machinery controller, and the increased pumping flow instruction is sent to the foundation dewatering system.

6. The construction safety risk monitoring method according to claim 5, wherein A preset risk suppression action sequence is executed, including: When the yellow warning is triggered, the first-level suppression action is executed, which includes broadcasting voice prompts to the relevant area, highlighting the risk transmission path on the monitoring screen, and pushing the inspection task to the safety officer's handheld terminal; Within 5s after the red warning is triggered, the second-level suppression action is executed, which includes cutting off unnecessary power supply in the relevant area, starting the emergency lighting system, opening all safety passage access control, and sending emergency braking instructions to the tower crane and elevator; Within 15s after the red warning is triggered, the third-level suppression action is executed, which includes starting the automatic spraying dust suppression system around the foundation pit, activating the prestressed tensioning device of the structural support system, sending a linkage request to the fire control center, and sending the estimated number and type of injured personnel to the hospital emergency center.

7. The construction safety risk monitoring method according to claim 6, wherein The multi-modal perception network comprises a distributed temperature and humidity sensor array, a dust concentration detection unit, a noise monitoring probe, a high-definition video monitoring camera, an infrared thermal imager, a laser range finder, an inclination sensor, a vibration accelerometer, a personnel radio frequency identification tag reader, a heavy machinery operation state acquisition module, and a structural stress and strain monitoring sheet; The distributed temperature and humidity sensor array is arranged at every 50m 2 Lay out a collection point; the dust concentration detection unit is arranged at 3m in the upwind direction and downwind direction of the main dust operation area; the noise monitoring probe is installed on the surface of the high noise equipment shell and within 1m radius of the ear side of the operating personnel; the high-definition video monitoring camera covers all the main operation surfaces and channel intersections, and the lens focal length can be adjusted in the range of 35mm to 150mm; the infrared thermal imager is aimed at the core heat dissipation components of large mechanical equipment and the surface of electrical cabinet, and the temperature measurement range is -20℃ to 500℃; the laser range finder is installed at the end of the tower crane jib and the top of the base pit edge protection rail column, and the measurement accuracy is ±1mm; the inclination sensor is fixed at the top of the scaffold upright rod and the main beam node of the formwork support system, and the measuring range is ±15°; the vibration accelerometer is pasted on the handle of the concrete vibrator and the base of the pile foundation construction hammering equipment, and the sampling frequency is greater than 1KHz; the personnel radio frequency identification tag reader is arranged at the entrance of each operation area and the dangerous operation permission issuing point, and the reading distance is adjustable in the range of 0.5m to 3m; the heavy machinery running state acquisition module accesses the engine control unit and hydraulic system pressure sensor of the excavator, crane and concrete pump truck through the controller area network bus interface; the structure stress and strain monitoring sheet is pasted on the surface of the main reinforcement of the deep foundation pit support pile and the middle part of the key compression member of the large-span formwork support system, and the strain measurement range is ±5000με.

8. The construction safety risk monitoring method according to claim 7, wherein, Trace all the conduction paths from the source risk node to the current node, calculate the path cumulative weight and sort it, and extract the top three paths with the highest weight as the key conduction links; Pack the early warning message and key conduction link information into a structured instruction data packet, and attach the execution priority and timeliness label; Disassemble the structured instruction data packet and extract the early warning level and key conduction link information; According to the type of risk node in the key conduction link, map it to the corresponding list of field execution device addresses; Match the preset action sequence according to the early warning level, generate a device control instruction queue, and distribute it to each execution device according to the priority; Receive the execution confirmation signals returned by each device, summarize them to generate an execution status report, and return it to the man-machine interaction display unit.

9. A construction safety risk monitoring system, characterized by, It includes: A multi-modal perception network deployment unit is used to deploy various types of environmental, personnel, equipment and structural monitoring sensors at the construction site according to the preset density and position; A spatio-temporal data fusion processing unit is used to perform time synchronization, spatial mapping, noise filtering and feature extraction on multi-source heterogeneous perception data to generate a standardized risk factor time series matrix; A risk conduction map construction unit is used to load a preset risk conduction rule library and a historical accident case library to dynamically generate and update the correlation map between risk factors; A risk propagation potential prediction unit is used to call a graph neural network model to perform propagation path deduction and potential value calculation on the current risk situation; A multi-level early warning decision unit is used to compare the propagation potential value with the preset threshold interval, determine the early warning level, and generate the corresponding early warning signal and block node suggestion; A risk suppression execution unit is used to receive the early warning signal, and execute risk intervention measures on various security, mechanical and electrical and emergency equipment in the field according to the preset action sequence; A man-machine interaction display unit is used to display the risk heat map, conduction path animation, early warning information list and device linkage state on the central control room large screen and mobile terminal simultaneously.

10. The construction safety risk monitoring system of claim 9, wherein, The spatio-temporal data fusion processing unit includes a time synchronization submodule, a spatial registration submodule, a feature extraction submodule, and a matrix generation submodule; The time synchronization submodule obtains a time signal from a reference clock source through a network time protocol, adds a time stamp to each sensor data packet, and performs interpolation alignment; The spatial registration submodule converts the sensor physical coordinates to a unified construction coordinate system using a laser scanning point cloud and video image feature point matching algorithm; The feature extraction submodule calls the corresponding signal processing algorithm library according to the sensor type, and outputs the standardized feature vector; The matrix generation submodule aligns all feature vectors by time step, sorts them by risk factor category, and concatenates them into a two-dimensional matrix structure.

Images