Safety management system for electric power engineering construction

Abstract

The application relates to the technical field of construction safety management, and discloses a power engineering construction safety management system, which comprises a real-time risk assessment scheduling unit, a forward-looking human factor risk perception unit and a risk simulation mapping unit. The real-time risk assessment scheduling unit is used for generating a safety state characteristic map based on heterogeneous safety data of a construction site, identifying a risk event of the construction site through the safety state characteristic map, and generating an intervention instruction and a resource scheduling scheme to optimize the construction site. The application can comprehensively capture multi-dimensional information of the construction site by collecting personnel biological characteristics, equipment operation logs, environmental indexes and work video streams and other heterogeneous safety data through a multi-modal perception unit, breaks the limitation of single and one-sided data collection in the traditional management mode, and enables management personnel to obtain more rich and comprehensive on-site safety information, so that the overall safety condition of the construction site can be more clearly understood.

Description

Technical Field

[0001] This invention relates to the field of construction safety management technology, and more specifically, to a power engineering construction safety management system. Background Technology

[0002] During power engineering construction, the construction site environment is complex and ever-changing, involving various types of equipment, a large number of construction personnel, and diverse work scenarios, posing numerous challenges to safety management. Traditional power engineering construction safety management methods largely rely on manual inspections and paper records, which have significant limitations. Manual inspections are limited by the experience, energy, and frequency of the inspectors, making it difficult to achieve comprehensive and real-time monitoring of the construction site. Often, they can only be traced after an accident occurs, failing to identify potential risks in advance.

[0003] Paper records are prone to data omissions and errors, and data processing and analysis are inefficient, making it difficult to quickly identify safety hazards during construction. With the continuous expansion of power engineering scale, the increase in the types of construction equipment, and the frequent turnover of workers, traditional management methods can no longer meet the safety management needs of complex construction environments.

[0004] Existing safety management systems often focus on the collection and monitoring of single data, such as monitoring only equipment operation data or personnel location information. They lack the ability to integrate and process multi-source heterogeneous data such as personnel biometrics, equipment operation logs, environmental indicators, and operation video streams. Because they cannot achieve collaborative analysis of multiple data, it is difficult to build a comprehensive understanding of the safety status of the construction site, resulting in low accuracy in identifying abnormal patterns during construction and often failing to detect risk events in a timely and accurate manner.

[0005] When safety hazards are discovered, most existing systems can only issue simple warning signals and lack the ability to automatically generate graded intervention instructions and resource allocation plans based on the risk level. They are unable to quickly coordinate on-site security equipment, mechanical braking systems and personnel communication devices to respond in a coordinated manner, resulting in untimely and inadequate risk handling, making it difficult to effectively avoid safety accidents and seriously affecting the safety assurance of power engineering construction.

[0006] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0007] In response to the problems in related technologies, this invention proposes a power engineering construction safety management system to overcome the aforementioned technical problems existing in the existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows:

[0009] The power engineering construction safety management system includes: a real-time risk assessment and scheduling unit, a proactive human-cause risk perception unit, and a risk simulation and mapping unit. The real-time risk assessment and scheduling unit generates a safety status feature map based on heterogeneous safety data from the construction site, identifies risk events at the construction site using this map, and generates intervention instructions and resource scheduling schemes to optimize the construction site. The proactive human-cause risk perception unit, connected and interacting with the real-time risk assessment and scheduling unit, assesses the dynamic coupling risk entropy value of construction personnel during power engineering construction based on their physiological state vectors, and sends this value to the real-time risk assessment and scheduling unit for state deviation warnings. The risk simulation and mapping unit, also connected and interacting with the real-time risk assessment and scheduling unit, constructs a real-world model based on point cloud data of the power engineering construction site, generates parameter constraints based on the construction plan and equipment configuration management, determines the risk intensity value, and outputs the optimized power engineering construction plan and risk control measures to the real-time risk assessment and scheduling unit. The system achieves safety management of the power engineering construction site through the combination of these three units.

[0010] Preferably, the real-time risk assessment and scheduling unit includes: a multimodal perception unit for collecting heterogeneous safety data from the power engineering construction site, including personnel biometrics, equipment operation logs, environmental indicators, and operation video streams; a distributed computing unit for normalizing and extracting features from the heterogeneous safety data to generate a safety status feature map; a dynamic risk assessment unit for using the safety status feature map as input and employing a deep spatiotemporal network to identify abnormal patterns at the power engineering construction site and determine risk event descriptors; an adaptive response unit for matching the risk event descriptors with a power engineering safety strategy library to generate tiered intervention instructions and resource allocation schemes; and an instruction execution and resource scheduling unit for implementing scheduling processing at the power engineering construction site according to the tiered intervention instructions and resource allocation schemes, and combining the output results of the forward-looking human risk perception unit and the risk simulation mapping unit to provide early warning processing for the power engineering construction site.

[0011] Preferably, the dynamic risk assessment unit includes: a pattern learning module, used to learn the distinguishing boundary between normal and abnormal working modes in power engineering construction from a safety state feature map using an adversarial generative network; a probability inference module, used to analyze the matching probability between the power engineering construction site and the risk mode using a Bayesian neural network, with the distinguishing boundary between normal and abnormal working modes as a reference, and generate a risk probability distribution; a propagation analysis module, used to call a temporal inference network based on the risk probability distribution to predict the propagation path and impact range of risk events at the power engineering construction site, and determine the propagation analysis results; and a descriptor generation module, used to generate risk event descriptors containing risk category, intensity index, and spatiotemporal coordinates based on the risk probability distribution and the propagation analysis results.

[0012] Preferably, the adaptive response unit includes: a strategy matching module, used to perform syntax parsing on the risk event descriptor, extract target fields, and map and verify the target fields with the emergency plan execution concepts in the power engineering construction safety strategy library to determine the risk emergency strategy; a resource planning module, used to calculate the type, quantity, and response time requirements of the emergency resources required in the risk emergency strategy based on the risk event descriptor, and generate an emergency plan including resource allocation routes and resource arrival times; a decision optimization module, used to establish optimization rules with the objectives of risk containment effect, resource usage cost, and construction progress impact, and combine the emergency plan with the non-dominated solution set to find the technical output optimization decision plan; and an instruction synthesis module, used to convert the optimization decision plan into a textual instruction sequence, and use a knowledge parsing model to parse the textual instruction sequence to generate a hierarchical intervention instruction frame that conforms to the communication protocol specification.

[0013] Preferably, the forward-looking human risk perception unit includes: a physiological signal perception module, used to collect near-infrared brain functional imaging signals, eye movement trajectory signals, skin conductance response signals, and surface electromyography signals of construction workers during power engineering construction to characterize the physiological state vector of construction workers; a behavioral semantic analysis module, used to parse the work video stream, identify the standard of construction workers' work actions, the area of ​​their gaze, and the spatial interaction between construction workers and hazards, and generate a behavioral semantic vector; and a dynamic coupling risk entropy calculation module, used to receive the physiological state vector, behavioral semantic vector, and environmental indicators, and establish a nonlinear mapping model to evaluate the dynamic coupling risk entropy value of construction workers when performing power engineering construction, so as to generate a human state deviation warning and send it to the instruction execution resource scheduling unit.

[0014] Preferably, the behavioral semantic analysis module includes: a personnel action state recognition module, used to extract limb points of construction workers in the operation video stream through image extraction technology, and calculate the arm extension angle and body center of gravity offset of the construction workers based on the limb points to quantify the standard of the construction workers' operation actions; a personnel position state recognition module, used to locate the hazard source and the line of sight of the construction workers in the operation video stream, and quantify the Euclidean distance between the construction workers and the hazard source based on the angle between the line of sight of the construction workers and the hazard source to determine the line of sight attention area and spatial interaction relationship; and a behavior vector fusion determination module, used to fuse the standard of operation actions, line of sight attention area and spatial interaction relationship as a behavioral semantic vector.

[0015] Preferably, the extraction of limb points of construction workers in the operation video stream using image extraction technology includes: performing frame segmentation processing on the operation video stream to generate clear frames and occluded frames of construction worker actions; using deep learning technology to analyze the clear frames of construction worker actions and identify the clear limb points of the construction workers; setting a classification threshold based on digital refocusing technology and energy gradient function, and combining element image inpainting technology to repair and reconstruct the target objects in the occluded frames of construction worker actions, and outputting the repaired clear frames of construction worker actions based on the processing results; using deep learning technology to analyze the repaired clear frames of construction worker actions, identifying the repaired limb points of the construction workers, and combining them with the clear limb points to output the limb points of the construction workers in the operation video stream.

[0016] Preferably, a classification threshold is set based on digital refocusing technology and an energy gradient function, and element image inpainting technology is used to repair and reconstruct the target object in the occluded frames of construction worker actions. The output of the clear frames of the repaired construction worker actions includes: determining the light intensity and direction information from different perspectives in the construction worker action scene based on the occluded frames, obtaining the element image array at the corresponding perspective, and performing digital refocusing processing on the element image array along the axial depth; obtaining a sequence of clear and blurred images distributed along the depth based on the digital refocusing processing results, and analyzing the sum of gradient values ​​of all adjacent points in the image sequence to determine the focus evaluation result of the image sequence at the corresponding axial depth; and determining the focus evaluation result when the focus evaluation result meets the preset... After setting a standard value, the image sequence undergoes binary encoding transformation and cost aggregation to determine the classification threshold. Based on this threshold, the foreground occlusion region in the construction worker action occlusion frame is removed. After removing the occlusion region, a clear construction worker action frame running alongside the occluded frame is selected as a sample image. The disparity value between the sample image and adjacent elements in the occluded frame is calculated. The sample image is then shifted according to the corresponding disparity value, and the pixel values ​​at the corresponding positions are extracted and fused into the foreground occlusion region of the occluded frame to obtain the repaired construction worker action occlusion frame. Finally, the repaired frame undergoes 3D reconstruction to output the repaired clear construction worker action frame.

[0017] Preferably, after the focus evaluation result meets the preset standard value, the image sequence is subjected to binary encoding transformation and cost aggregation processing to determine the classification threshold. The foreground occlusion region in the construction worker's action occlusion frame is then removed based on the classification threshold. This includes: selecting a transformation window according to the target size; comparing the grayscale values ​​of pixels within the transformation window with the center pixel to generate a binary bitstream; using the binary bitstream to perform binary encoding transformation on the image sequence to complete the cost matching processing of the image sequence; after cost matching processing, linear interpolation processing is performed on the pixel intensity of the image sequence; the minimum disparity of each pixel in the image sequence is analyzed; and additional smoothing constraints are added to construct an energy function. Cost aggregation of the image sequence is completed by applying different penalty terms to adjacent disparity changes; the disparity value of each pixel in the image sequence at the minimum cost aggregation is obtained, and after left-right consistency check, disparity filling and filtering, the disparity map of each element image array is obtained; according to the transformation relationship between the image sequence and the disparity map, the classification threshold of the construction worker target object and the occluder object is determined, and according to the imaging principle that the disparity value of the occluder object is greater than the disparity value of the construction worker target object, the foreground occluded area in the element image array with the disparity value greater than the classification threshold is marked to complete the classification of the occluder object and the construction worker target object, so as to realize the removal of the foreground occluded area in the construction worker action occlusion frame.

[0018] Preferably, the risk simulation mapping unit includes: a digital mapping module, used to construct a 3D real-world model of the power engineering construction site based on point cloud data, and generate a benchmark virtual scene by combining physical rules and behavioral rules in a hybrid modeling technique to determine the risk intensity value; a data fusion module, used to acquire the construction schedule and the configuration of construction personnel and equipment management as multi-source external data, and extract action points and state change points after formatting the multi-source external data; convert the action points and state change points into an event stream recognizable by the benchmark virtual scene, and extract the parameterized constraints required for the benchmark virtual scene deduction from the multi-source external data, and add the event stream and parameterized constraints to the benchmark virtual scene in chronological order; and a knowledge evolution module, used to match the disposal strategy with the risk intensity value output by the benchmark virtual scene, verify the performance evaluation of the disposal strategy in different scenarios, generate a risk evolution path map, and output the optimized construction plan and risk control measure adjustment suggestions to the instruction execution resource scheduling unit.

[0019] Preferably, the method of constructing a 3D real-scene model of a power engineering construction site based on point cloud data and generating a benchmark virtual scene by combining physical rules and behavioral rules includes: acquiring point cloud data of the power engineering construction site using a laser scanning station, and uniformly generating a 3D real-scene scene after registering and fusing the point cloud data, which serves as the 3D real-scene model of the power engineering construction site; analyzing the motion and behavior of construction equipment under external forces using multibody dynamics and materials mechanics, and adding this as a physical rule to the 3D real-scene model; treating construction personnel as intelligent agents, assigning behavioral states to the agents using collaborative game theory, and converting the output of the behavioral states into a control rule base, which is then added to the 3D real-scene model; and constructing a benchmark virtual scene of the power engineering construction site for a future target time period based on the 3D real-scene model with added physical and behavioral rules, combined with current construction technology.

[0020] Preferably, the method of employing collaborative game theory to assign behavioral states to agents and converting the output of these behavioral states into a control rule base includes: configuring a working state for each agent and modeling agents involved in future power engineering construction plans as agents with state stochasticity; simulating the interaction between the agent configuration after working state configuration and the agents with state stochasticity during power engineering construction based on a 3D real-scene model; analyzing the state transition process of agents during work based on the interaction state, and evaluating the fuzziness and stochasticity of the 3D real-scene model in the simulation process based on the state transition process, and defining a standard evaluation set for agents; based on the standard evaluation set, using agents as game participants, constructing a multi-agent dynamic game function with the goals of operational safety and efficiency, analyzing the equilibrium strategies of agents under different working conditions and different scenario events, and converting the equilibrium strategies into cloud droplet distribution characteristics; encoding the control rules of agents in the construction process and state transition process based on the cloud droplet distribution characteristics, as a behavioral control rule base.

[0021] The beneficial effects of this invention are as follows:

[0022] 1. This invention collects heterogeneous safety data such as personnel biometrics, equipment operation logs, environmental indicators, and operation video streams at the construction site through a multimodal sensing unit. It can achieve comprehensive capture of multi-dimensional information at the construction site, breaking the limitations of single and one-sided data collection in traditional management methods. This allows managers to obtain richer and more comprehensive on-site safety information, thereby gaining a clearer understanding of the overall safety status of the construction site.

[0023] 2. This invention generates a safety status feature map with spatiotemporal correlation by normalizing and extracting features from heterogeneous safety data. This solves the problem of inconsistent formats and difficulty in collaborative analysis of multi-source heterogeneous data. Through data normalization, the differences between different types of data are eliminated, facilitating subsequent integration and analysis. The feature extraction process can extract key safety information from massive amounts of data. Combined with spatiotemporal correlation, a safety status feature map is constructed, which presents the safety status of the construction site in a more intuitive and organized way, helping managers to quickly grasp key safety information.

[0024] 3. This invention, based on a safety status feature map, identifies abnormal patterns in the construction process through a deep spatiotemporal network and outputs risk event descriptors with confidence scores, significantly improving the ability to identify abnormal situations during construction. The deep spatiotemporal network has powerful data analysis and pattern recognition capabilities, enabling it to accurately mine abnormal patterns from a safety status feature map with spatiotemporal correlation. Furthermore, the confidence scores allow managers to understand the severity of risk events, providing a clear basis for subsequent risk management and avoiding the omission of safety hazards due to inaccurate or untimely identification of abnormal situations.

[0025] 4. Based on risk event descriptors and combined with an engineering safety strategy library, this invention generates tiered intervention instructions and resource allocation schemes, enabling intelligent and differentiated handling of risk events. Different levels of risk events correspond to different intervention measures and resource allocation methods, avoiding the problems of single response methods, resource waste, or insufficient resources in traditional risk handling. Through tiered intervention instructions, appropriate response measures can be taken for risks of different severities, while the resource allocation scheme can ensure that the required human and material resources can be rationally allocated and quickly dispatched when handling risks, improving the efficiency and pertinence of risk handling.

[0026] 5. Based on tiered intervention instructions, this invention coordinates the collaborative operation of on-site security equipment, mechanical braking systems, and personnel communication devices, enabling rapid implementation of risk management. Upon receiving a tiered intervention instruction, the instruction execution resource scheduling unit can quickly link various on-site equipment and devices to form a unified response, avoiding delays caused by poor coordination between different devices. For example, when a high-risk abnormal operation of equipment is identified, the instruction execution resource scheduling unit can promptly coordinate the mechanical braking system to stop the equipment, simultaneously notify nearby workers to evacuate via personnel communication devices, and activate on-site security equipment for alert, comprehensively ensuring the safety of personnel and equipment at the construction site and effectively reducing the possibility of safety accidents. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of a power engineering construction safety management system according to an embodiment of the present invention;

[0029] Figure 2 This is a flowchart of a real-time risk assessment and scheduling unit in a power engineering construction safety management system according to an embodiment of the present invention;

[0030] Figure 3 This is a flowchart of dynamic risk assessment in the power engineering construction safety management system according to an embodiment of the present invention;

[0031] Figure 4 This is a sequence diagram of the power engineering construction safety management system according to an embodiment of the present invention;

[0032] Figure 5This is a flowchart illustrating the extraction of limb points of construction workers from a work video stream within a power engineering construction safety management system according to an embodiment of the present invention.

[0033] In the picture:

[0034] 1. Real-time risk assessment and scheduling unit; 2. Proactive human factor risk perception unit; 3. Risk simulation and mapping unit. Detailed Implementation

[0035] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0036] According to an embodiment of the present invention, a power engineering construction safety management system is provided.

[0037] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figures 1 to 5 As shown, the power engineering construction safety management system according to an embodiment of the present invention can be found here. Figure 1 This embodiment provides a power engineering construction safety management system, including a real-time risk assessment and scheduling unit 1, a forward-looking human factor risk perception unit 2, and a risk simulation and mapping unit 3;

[0038] The real-time risk assessment and scheduling unit 1 is used to generate a safety status feature map based on heterogeneous safety data at the construction site, and to identify risk events at the construction site through the safety status feature map in order to generate intervention instructions and resource scheduling schemes to optimize the construction site.

[0039] In one embodiment, the real-time risk assessment and scheduling unit 1 includes a multimodal perception unit for collecting heterogeneous safety data from the power engineering construction site, including personnel biometrics, equipment operation logs, environmental indicators, and operation video streams; a distributed computing unit for normalizing and extracting features from the heterogeneous safety data to generate a safety status feature map; a dynamic risk assessment unit for using the safety status feature map as input and employing a deep spatiotemporal network to identify abnormal patterns at the power engineering construction site and determine risk event descriptors; an adaptive response unit for matching the risk event descriptors with a power engineering safety strategy library to generate tiered intervention instructions and resource allocation schemes; and an instruction execution and resource scheduling unit for implementing scheduling processing at the power engineering construction site according to the tiered intervention instructions and resource allocation schemes, and combining the output results of the forward-looking human risk perception unit 2 and the risk simulation mapping unit 3 to provide early warning processing for the power engineering construction site.

[0040] It should be explained that the real-time risk assessment and scheduling unit 1 includes a multimodal perception unit, a distributed computing unit, a dynamic risk assessment unit, an adaptive response unit, an instruction execution resource scheduling unit, and a forward-looking human factor risk perception unit 2. In this embodiment, the multimodal perception unit is deployed at the construction site to collect heterogeneous safety data, including personnel biometrics, equipment operation logs, environmental indicators, and work video streams. The heterogeneous safety data is transmitted to the distributed computing unit, which performs normalization processing and feature extraction on the heterogeneous safety data to generate a safety status feature map with spatiotemporal correlation. The dynamic risk assessment unit receives the safety status feature map and analyzes the dynamic patterns in the construction process through a built-in deep spatiotemporal network model to identify abnormal patterns that deviate from the normal state. The system then outputs risk event descriptors with confidence scores. The adaptive response unit queries the engineering safety strategy library based on the input risk event descriptors, performs strategy matching and decision optimization, and finally generates graded intervention instructions and resource allocation plans. The instruction execution resource scheduling unit receives and parses the graded intervention instructions, coordinates and controls on-site security equipment, mechanical braking systems, and personnel communication devices, and executes specific intervention actions. The proactive human risk perception unit 2 operates independently, collects and integrates multimodal physiological signals and work behavior data of construction personnel in real time, calculates the dynamic coupling risk entropy of human-machine-environment state, predicts potential human errors before they occur, generates early warnings of human state deviations, and sends them to the adaptive response unit, thus forming a complete closed-loop management system of perception, assessment, decision-making, and execution.

[0041] See Figure 2The multimodal sensing unit acts as the nerve ending, and its high-precision positioning module constructs a three-dimensional spatial sensing network at the construction site. This module, through ultra-wideband positioning base stations deployed at high points in the construction area, such as the tops of tower cranes or temporary light poles, forms a complementary positioning system with the satellite differential positioning module on mobile devices. The ultra-wideband base station receives pulse signals from personnel safety helmet tags and machinery tags at a frequency of hundreds of times per second, and calculates two-dimensional plane coordinates using a time difference of arrival algorithm. Meanwhile, the satellite differential module receives correction data from ground reference stations, fuses the calculated elevation coordinates with the plane coordinates, and finally outputs millimeter-level three-dimensional position information containing latitude, longitude, and altitude. This fusion positioning method effectively overcomes the signal attenuation problem of single technologies in underground spaces or environments obstructed by large equipment. The elevation coordinates are calculated based on the correction data from the ground reference station using satellite differential positioning technologies such as real-time dynamic carrier phase differential technology. By continuously receiving satellite navigation signals through a ground reference station fixed at a known coordinate point, and using its known precise coordinates to calculate the error correction amount of the satellite signal, the correction data is transmitted in real time via a data link. The mobile station at the construction site uses this correction data to perform high-precision processing on its own carrier phase observations, thereby calculating the three-dimensional coordinate increment relative to the reference station, including the elevation component. Specifically, the mobile station performs differential processing on the received reference station correction data and its own carrier phase observations to eliminate common errors such as satellite clock bias, receiver clock bias, and atmospheric delay. By solving the integer ambiguity of the carrier phase, the clean phase observations are converted into high-precision station-satellite geometric distances. Using these geometric distances from multiple satellites, combined with the known precise coordinates of the reference station, the three-dimensional coordinate increment of the mobile station relative to the reference station, including the elevation direction, is calculated. The fusion of elevation data and planar coordinates, and the output of millimeter-level three-dimensional position information, are achieved through a sensor fusion algorithm. This algorithm uses the absolute three-dimensional coordinates including elevation calculated by satellite differential technology as a reference, supplemented by stable and continuous relative planar displacement data provided by ultra-wideband technology, to perform real-time data fusion and optimization estimation. This process simultaneously optimizes the planar and elevation coordinates, ultimately outputting fused, unified millimeter-level three-dimensional position information including latitude, longitude, and altitude.

[0042] The equipment condition monitoring module focuses on sensing the health status of construction machinery. Nine-axis inertial measurement units (IMUs) are installed on key components such as the tower crane slewing mechanism and crawler crane hydraulic pumps to continuously collect acceleration, angular velocity, and geomagnetic field data in three directions. The spatial attitude angles of the mechanical components are calculated in real time using a sensor fusion algorithm. The process of calculating spatial attitude angles using the sensor fusion algorithm within the equipment condition monitoring unit is as follows: The multi-axis IMU continuously collects acceleration, angular velocity, and magnetic field data of the mechanical components in three directions. First, the raw data undergoes preprocessing such as filtering and noise reduction. The processed multi-source data is then input into a fusion core, such as a Kalman filter. This core utilizes the characteristic of accelerometer data reflecting the direction of gravity in a static state to correct pitch and roll angles, and uses magnetometer data to correct yaw angles. This constrains and corrects the attitude change predictions obtained by integrating gyroscope data. Through real-time iterative optimization, the final output is... The system generates high-precision spatial attitude angles for mechanical components and installs arrayed acoustic sensors on the engine casing, gearbox, and other parts. These sensors feature a high-temperature resistant and shock-resistant design, enabling them to capture wide-band sound and vibration signals generated during equipment operation. After preprocessing, the raw signals undergo spectrum analysis via an edge computing gateway to identify characteristic frequency components related to equipment faults. Preprocessing primarily involves filtering the acoustic and vibration signals, such as bandpass filtering to retain fault characteristic frequency bands, detrending, and segmented windowing to remove noise and standardize the signal. Spectrum analysis via the edge computing gateway involves performing a Fast Fourier Transform on the preprocessed signal segments, converting the time-domain signal into a frequency-domain spectrum. The gateway then identifies prominent characteristic frequency components in the spectrum and matches them against a pre-stored database of equipment fault characteristic frequencies to identify frequency components related to specific faults.

[0043] In the high-precision positioning module, the ultra-wideband positioning base station calculates the tag's two-dimensional plane coordinates using the time difference of arrival algorithm as follows: Multiple base stations deployed at known locations receive pulse signals emitted by the same tag, accurately measuring the time difference between the arrival times of the signals at different base stations. Since the propagation speed of radio waves is constant, each time difference corresponds to a distance difference. The set of points where the distance difference between the tag and two base stations is constant forms a hyperbola. The tag must lie on this hyperbola. By solving for the intersection of two or more such hyperbolas, the tag's two-dimensional plane coordinates can be determined. Let the tag coordinates be (x, y). The calculation formula of this algorithm can be expressed as:

[0044] ;

[0045] Wherein, the coordinates of base stations A and B are respectively ( , ), ( , ), This represents the time difference between the measured signal arriving at base station A and base station B from the tag, where the speed of light is... The equation defines a hyperbola with foci A and B. Two such hyperbolic equations are obtained using at least three base stations. By solving the simultaneous equations and finding their intersection, the tag coordinates can be determined. , ).

[0046] The environmental parameter acquisition module is responsible for building the ecological environment perception capability of the construction site. The laser scattering particulate matter monitor uses a constant flow sampling pump to draw in air samples and measures the intensity of scattered light generated by the laser beam irradiating particulate matter in the cavity. It then calculates the real-time concentration values ​​of PM2.5 and PM10 using the Mie scattering model. The multi-gas analyzer integrates an electrochemical sensor and a photoionization detector to continuously monitor eight types of gases, including oxygen, carbon monoxide, and hydrogen sulfide. All sensor data are transmitted to the data acquisition unit via the Modbus protocol, with temperature and humidity compensation parameters to ensure measurement accuracy. The Mie scattering model used in the laser scattering particulate matter monitor is a well-known principle in the field of optical particulate matter concentration measurement. Its working principle is as follows: laser irradiation of particulate matter in the air produces scattering, and the intensity of the scattered light is related to the surface area or volume of the particulate matter. The monitor measures the intensity of the scattered light at a specific angle and, based on the physical relationship between light intensity and particle size distribution and concentration established by the Mie scattering theory, calculates the real-time mass concentration values ​​of PM2.5 and PM10.

[0047] The panoramic vision module adopts a multi-sensor fusion vision acquisition scheme. A multispectral imager deployed at a high point simultaneously acquires visible light, near-infrared, and thermal infrared images. Visible light images are used for target recognition, near-infrared images enhance the ability to distinguish vegetation from materials, and thermal infrared images identify equipment overheating anomalies through temperature distribution differences. A binocular stereo vision camera deployed on the ground generates depth point cloud data of the work area by calculating the parallax information of the left and right images. This visual data is compressed using H.265 encoding and then transmitted to the edge server through a gigabit industrial ring network.

[0048] In one embodiment, the dynamic risk assessment unit includes a pattern learning module, used to learn the distinguishing boundary between normal and abnormal working modes of power engineering construction from a safety state feature map using a generative adversarial network; a probability inference module, used to analyze the matching probability between the power engineering construction site and the risk mode using a Bayesian neural network with reference to the distinguishing boundary between normal and abnormal working modes, and generate a risk probability distribution; a propagation analysis module, used to call a temporal inference network based on the risk probability distribution to predict the propagation path and impact range of risk events at the power engineering construction site, and determine the propagation analysis results; and a descriptor generation module, used to generate risk event descriptors containing risk category, intensity index, and spatiotemporal coordinates based on the risk probability distribution and the propagation analysis results.

[0049] In one embodiment, the adaptive response unit includes: a strategy matching module, used to perform syntactic parsing on the risk event descriptor, extract target fields, and map and verify the target fields with the emergency plan execution concepts in the power engineering construction safety strategy library to determine the risk emergency strategy; a resource planning module, used to calculate the type, quantity, and response time requirements of the emergency resources required in the risk emergency strategy based on the risk event descriptor, and generate an emergency plan including resource allocation routes and resource arrival times; a decision optimization module, used to establish optimization rules with the objectives of risk containment effect, resource usage cost, and construction progress impact, and combine the emergency plan with the non-dominated solution set to find the technical output optimization decision plan; and an instruction synthesis module, used to convert the optimization decision plan into a textual instruction sequence, and use a knowledge parsing model to parse the textual instruction sequence to generate a hierarchical intervention instruction frame that conforms to the communication protocol specification.

[0050] It should be further explained that the distributed computing unit undertakes the preprocessing and feature extraction of massive heterogeneous data. Its data standardization module establishes a unified data specification framework, uses coordinate transformation algorithms to convert WGS84 geographic coordinates into the construction local coordinate system for positioning data, performs minimum-maximum normalization processing on vibration data to make it fall within the zero-to-one range, and performs sliding window mean filtering on environmental monitoring data to eliminate instantaneous fluctuation interference. It also includes a data quality assessment module, which automatically identifies abnormal data points and triggers a data retransmission mechanism by setting threshold ranges and continuous consistency checks. The threshold range set by the data quality assessment module in its data standardization module refers to a reasonable numerical range for judging whether the data is abnormal. The specific value of this range is not fixed and needs to be calibrated and set during system deployment based on the specifications of the specific sensors, historical normal data of the on-site installation environment, and engineering experience. For example, it can be set as the upper and lower limits of the historical normal value fluctuation range.

[0051] The feature extraction module employs a spatiotemporal graph convolutional network to mine deep features of the construction scene. This network divides the construction site into a topologically related grid graph structure. Each grid node contains feature vectors such as equipment status, personnel density, and environmental parameters. Graph convolution operations capture spatial dependencies by aggregating feature information from adjacent nodes, such as the dynamic changes in personnel density within the tower crane's operating radius. Gated loop units process feature sequences of continuous time steps, learning the periodic patterns of construction activities and the temporal patterns of equipment operation, such as the phased features of concrete pouring operations. The spatiotemporal graph convolutional network in the feature extraction module follows these steps: constructing a dynamic construction scene graph based on the planar layout of the construction area and the logic of the work process; dividing the construction site into regular grids, with each grid serving as a graph node, and its initial feature vector incorporating normalized personnel density and equipment status within that grid. The network incorporates state codes and environmental parameter values; the connecting edges between nodes include both spatial edges representing physical adjacencies and logical edges defined based on operational process dependencies and the scope of equipment functional influence; the network aggregates the features of each node and its neighboring nodes through graph convolution operations to capture the spatial dependencies of the construction area; to further learn collaborative patterns over a wider range, a graph sampling and aggregation mechanism is adopted, enabling nodes to aggregate information from their higher-order neighboring nodes; after spatial convolution, each node obtains features containing local spatial context, and these node features are organized according to time series and input into a gating loop module. This module independently processes the time series data of each node, and through its internal gating mechanism, learns temporal evolution patterns such as periodic personnel flow and standard equipment operating trajectories. The output of the gating loop module at the end of the sequence is used as the feature representation of the node that integrates spatiotemporal information for subsequent graph construction.

[0052] The feature fusion module employs a cross-modal attention mechanism to organically integrate multi-source information. This mechanism assigns weight coefficients based on the contribution of different data sources to the current risk assessment task. The core of this cross-modal attention mechanism lies in its use of a trainable small neural network to evaluate the relevance of each data source's feature vector to the current risk assessment task and output an initial score. The initial scores from all data sources are then normalized, for example, using a softmax function, so that the sum of the final weight coefficients for each data source is 1. Furthermore, the magnitude of these coefficients directly reflects their contribution to the current risk judgment, thus achieving multi-source feature fusion. The adaptive fusion of features, for example, when environmental sensors detect an increase in combustible gas concentration, automatically increases the attention weight of thermal imaging camera data while reducing the processing priority of irrelevant area video data. This dynamic weight allocation strategy is implemented through a trainable attention network, which adaptively adjusts the fusion strategy according to the real-time scene context. The specific construction and working process of the spatiotemporal graph convolutional network in the feature extraction module is as follows: A dynamic construction scene graph is constructed based on the planar layout and operational logic of the construction area. The construction site is divided into regular grids, with each grid serving as a graph node. The initial feature vector of the node fuses the normalized personnel density, equipment status code, and ring information within that grid. The parameters of the environment are defined, and the edges between nodes include not only spatial edges representing physical adjacency but also logical edges defined based on work process dependencies and functional relationships. Graph convolution operations are performed on this graph, learning the spatial dependencies of the construction area by aggregating the features of each node and its first-order neighbors. To further capture a wider range of collaborative patterns, a graph sampling and aggregation mechanism is adopted, enabling nodes to aggregate information from their higher-order neighbors. After spatial convolution, each node obtains a feature representation containing its local spatial context. These node features are organized according to time series and input into a gated recurrent module network. This module independently processes the time series of each node, using update gates and reset gates. The mechanism learns and memorizes temporal evolution patterns such as the periodic migration of personnel from living areas to work areas and the movement of equipment along fixed paths. The features output by the gating loop module at the final time step are used as the final representation of the node, which integrates spatiotemporal information, to construct a safety status feature map. The map construction module organizes the processed features into a structured three-dimensional map. The temporal dimension of the map records the history of feature changes with a timestamp sequence, the spatial dimension corresponds to the absolute coordinate system of the construction site, and the feature dimension contains normalized multimodal data feature values. This map structure not only fully preserves the spatiotemporal correlation of the data, but also supports complex queries and real-time updates through the indexing mechanism of the graph database.

[0053] In the collaborative operation of the multimodal sensing unit and the distributed computing unit, data flow follows strict temporal logic. The raw data collected by the sensing terminal first enters the edge computing layer for preliminary filtering and timestamp alignment, and then is transmitted via the 5G private network to the distributed computing unit in the regional data center. Within the nodes, a pipeline architecture is used to process data standardization, feature extraction, and fusion tasks in parallel. The resulting security status feature map is published to the data bus via a message middleware for subsequent subscription by the risk assessment module. The generation of the security status feature map specifically includes the following steps: the high-precision positioning module of the multimodal sensing unit provides real-time, millimeter-level spatial coordinates for personnel and equipment using ultra-wideband and satellite differential technology; the feature extraction module of the distributed computing unit uses a spatiotemporal graph convolutional network and a gated recurrent module to extract deep features representing personnel flow patterns and equipment trajectory characteristics from heterogeneous data containing the aforementioned coordinates; the map construction module combines the extracted deep feature vector of each location point with its corresponding spatial... Coordinates are bound and each frame of data is assigned a timestamp. These spatial coordinates, feature vectors, and timestamp triplets are organized into a structured 3D map, namely the security status feature map. The entire processing flow adopts a microservice architecture design, and each processing unit module can independently expand resources. Containerization technology ensures high service availability. The system also establishes a data lineage tracing mechanism to record the complete transformation path from raw data to the feature map, meeting the audit requirements of security supervision. During implementation, the adaptability design to the field environment must be carefully considered. The sensing terminal adopts an IP67 protection-rated shell to resist wind and rain corrosion, and the internal circuit is treated to prevent electromagnetic interference to ensure signal stability. The distributed computing unit is deployed in a shockproof cabinet, and redundant power supply and network design are used to ensure continuous operation capability. A comprehensive calibration and maintenance mechanism is also established. The positioning base station is regularly calibrated, the gas sensor is calibrated with standard gas according to plan, and the vision equipment is equipped with an automatic cleaning device to maintain lens transparency. These measures work together to ensure the long-term reliability of the system.

[0054] The multimodal perception unit further integrates a physiological signal perception module. This module is specifically manifested as an intelligent safety device with a built-in multi-channel near-infrared brain functional imaging sensor, a miniature eye tracker, a skin conductance response sensor, and a surface electromyography sensor. The near-infrared brain functional imaging sensor reflects the cognitive load and attention level of the worker by collecting dynamic changes in the blood oxygen concentration in the prefrontal cortex. The miniature eye tracker tracks the pupil movement trajectory, fixation point, and blink frequency, and analyzes whether the worker's gaze and attention allocation matches the current task by combining the scene video stream provided by the panoramic vision module. The skin conductance response sensor monitors skin electrical conductance activity as an indicator of emotional arousal level. The surface electromyography sensor is deployed in the muscle groups of the main operating limbs to capture muscle electrical signal activation patterns before fine manipulation.

[0055] The physiological feature analysis module corresponding to the distributed computing unit performs synchronization, noise reduction, and artifact removal preprocessing on multi-source asynchronous physiological signals. It employs a deep learning network based on attention mechanisms to extract features from the preprocessed temporal data of brain functional signals, quantifying changes in connection strength between different functional networks in the brain. For eye-tracking signals, it extracts indicators representing attention stability, such as fixation entropy and saccade speed. By fusing activation timing and pattern features from electromyographic signals, it predicts the standardization and potential error tendencies of operational actions. This module outputs a multi-dimensional human-caused state feature vector, which, along with features such as personnel location and behavioral trajectory, is input into the feature fusion module. In the feature fusion module, physiological status is dynamically assessed through a cross-modal attention mechanism. The correlation weights between human factors characteristics and external behaviors and environmental characteristics in risk assessment enable a more comprehensive and forward-looking perception of risks in human-machine-environment systems. The steps are as follows: First, the multi-dimensional human factor state feature vector output from the physiological feature analysis module, the behavioral semantic vector output from the behavioral semantic analysis module, and the environmental equipment state vectors obtained in real-time from the environmental parameter acquisition module and equipment status monitoring module are aligned in timestamps and used as input to a cross-modal attention mechanism. Then, a trainable correlation evaluation network is constructed: the core of this mechanism is a small neural network specifically designed for human factor risk assessment in power construction. This network receives a fused vector obtained by concatenating the above three types of feature vectors. Its structure includes a fully connected layer and a soft layer. The softmax output layer, a fully connected layer, maps the high-dimensional fusion vector to a low-dimensional intermediate representation. Its weight parameters are learned during model training. Dynamic weight calculation and fusion: The softmax output layer calculates and outputs three scalar weight coefficients based on the content of the current fusion vector. These coefficients correspond to the contributions of physiological state, behavioral semantics, and environmental equipment status features to the overall risk assessment at the current moment. The sum of these weight coefficients is 1, and their values ​​change dynamically. For example, when the system detects a sudden increase in environmental wind speed, the weight of the environmental equipment status features automatically increases; when it detects abnormally dispersed eye-tracking trajectories, the weight of the behavioral semantic features increases accordingly. Weighted feature generation: The calculated dynamic weights are multiplied by their corresponding original feature vectors. The data are then summed to generate a weighted comprehensive feature vector. This vector reflects the system's dynamic assessment of the importance of multi-source information and is input into the subsequent dynamically coupled risk entropy calculation module for the final quantification of human-caused state deviation risks. When the system identifies significant fatigue, inattention, or high emotional stress characteristics, the dynamic risk assessment unit will generate a risk event descriptor containing categories of human-caused risk precursors. The adaptive response unit then matches personalized pre-intervention strategies accordingly. For example, it can provide visual guidance to specific workers focusing on key task areas through the augmented reality device of the personnel guidance module, or send soothing voice prompts through a directional sound wave device, while simultaneously pushing warning information to the smart terminal of the on-site safety officer.

[0056] See Figure 3 The physical carrier of the dynamic risk assessment unit is deployed on high-performance computing nodes of a cloud computing platform. These nodes are equipped with professional graphics processors and tensor computing modules, enabling real-time inference operations for complex neural network models. The core software architecture adopts a microservice design pattern, encapsulating functions such as pattern learning, probability inference, propagation analysis, and descriptor generation into independent services. Data interaction is achieved through remote procedure call protocols. The application architecture of deep spatiotemporal networks is reflected in the workflow of the dynamic risk assessment unit. This core takes a safety state feature map as input. Its internal pattern learning module uses a generative adversarial network to learn from this map to establish a benchmark for the normal construction mode. The Bayesian neural network in the probability inference unit uses the aforementioned map and the learned patterns as references to calculate the risk probability and uncertainty of the current state deviating from the normal mode. The propagation analysis module can call a temporal inference network to analyze the temporal correlations in the map and predict risk transmission. The core of the entire assessment process lies in using a deep network model to perform hierarchical processing and inference on the feature map containing spatiotemporal correlations. Finally, the descriptor generation module integrates and outputs risk event descriptors.

[0057] The pattern learning module employs a generative adversarial network framework to construct a dynamic feature learning mechanism. The generator network consists of fully connected layers and deconvolutional layers. Its input layer receives random noise vectors that conform to a normal distribution and outputs simulated safety state feature maps through multi-layer neural network transformation. The discriminator network adopts a convolutional neural network structure. Its input layer receives mixed data of real feature maps and generated maps. After extracting spatial features through sliding scan of convolutional kernels, it outputs the true / false discrimination result. During the adversarial training process of the pattern learning unit, the real feature map is the core safety state feature map generated and input by the distributed computing unit. It represents the true safety state of the construction site.

[0058] The generated map is simulated by the generator network using random noise as input and its internal network structure. The discriminator network drives adversarial learning by continuously distinguishing between mixed input data, such as the real security status feature map and the generated map, thus enabling the generator to learn to approximate the data distribution of the real security status. Therefore, the security status feature map is the real training data source and target, while the generated map is the simulated data used for comparative learning during training. Both serve the pattern learning unit to learn the boundary of the normal operation pattern. In the pattern learning module, the generator network consists of fully connected layers and deconvolutional layers to learn the data distribution of the normal security status feature map; the discriminator network uses a convolutional neural network structure to distinguish the input feature maps. Whether the spectrum comes from real data or a generator, the two are trained adversarially so that the feature map generated by the generator infinitely approximates the distribution of real normal operation data, thereby implicitly learning the boundary between normal and abnormal patterns. The Bayesian neural network built by the probability inference module maintains random dropout operation during inference and performs multiple forward propagations on the same input to obtain a series of risk probability outputs. The mean of these outputs is used as the final risk matching probability, and the standard deviation is the confidence score of the output. The deep spatiotemporal network here is the feature extraction network composed of the spatiotemporal graph convolutional network and the gated recurrent module used to construct the safety status feature map. Together with the generative adversarial network in the pattern learning module and the Bayesian neural network in the probability inference module, it constitutes the complete model system of the dynamic risk assessment unit. The feature extraction network is responsible for extracting high-level feature maps with spatiotemporal correlations from the raw data; the generative adversarial network uses these maps to learn the boundaries of normal patterns; and the Bayesian neural network uses the feature maps to make quantitative inferences on risk probabilities and confidence levels. The two networks form a dynamic game relationship during training. The generator continuously optimizes the realism of its generated maps, and the discriminator continuously improves its ability to identify real and fake maps. This adversarial training process enables the system to gradually master the feature distribution boundaries of normal construction activities. When the real-time input feature maps deviate significantly from the learned normal patterns, the system can keenly capture abnormal signals.

[0059] The probabilistic inference module constructs an uncertainty quantification model based on a Bayesian deep learning framework. Its network structure introduces a random dropout layer during forward propagation. By maintaining the dropout activation state during inference, it samples the posterior distribution of model parameters. Each inference process is equivalent to drawing samples from the parameter distribution for forward computation, obtaining the probability distribution of the output result through multiple samplings. The unit input layer receives dimensionality-reduced feature vectors from the feature extraction layer. After nonlinear transformation through three hidden layers, the output layer generates probability values ​​corresponding to various risk events. These probability values ​​not only include the most likely risk type judgment but also obtain confidence estimates by calculating the standard deviation of multiple sampling results. The training process of the probabilistic inference module uses a variational inference method to approximate a complex posterior distribution. By minimizing the KL divergence between the variational distribution and the true posterior, the network parameters are optimized. This design enables the system to reasonably assess the uncertainty of its judgments and automatically reduce the confidence score when the data quality is poor or the scenario is complex.

[0060] In the dynamic risk assessment unit, the deep spatiotemporal network refers to the feature extraction model composed of the spatiotemporal graph convolutional network and gated recurrent units used to extract safety status feature maps. The feature maps extracted by this network are the input basis for the pattern learning module and the probability inference module. The generative adversarial network in the pattern learning module uses these feature maps as training data and learns the data distribution boundary of the normal operation mode through adversarial game between the generator and the discriminator. The Bayesian neural network in the probability inference module uses the aforementioned feature maps as input and outputs the risk matching probability and its uncertainty measure through multiple forward propagation inferences with a random drop-out mechanism. The deep spatiotemporal network, generative adversarial network and Bayesian neural network together constitute a hierarchical risk assessment model system, which collaboratively completes the calculation of risk probability and confidence from raw data.

[0061] The propagation analysis module focuses on studying the spatiotemporal transmission patterns of risk events. The temporal reasoning network constructed by this module integrates causal convolution and self-attention mechanisms. The causal convolution module ensures that the output of each time step relies only on current and historical information and does not use future data by restricting the receptive field direction of the convolution kernel, thus meeting the requirements of real-time prediction. The self-attention mechanism dynamically calculates the correlation weights of different time steps in the feature sequence and identifies the historical moment with the most influence on the current risk propagation. The network output layer contains multiple prediction heads, corresponding to the probability of risk propagation, the expansion speed of the scope of influence, and the identification of key propagation paths, respectively. The propagation analysis module also integrates the physical constraint knowledge of the construction scenario. For example, by importing the equipment layout diagram and operation process relationship diagram of the construction site, it constructs structured prior information on risk propagation. This prior information is embedded in the message passing mechanism of the neural network in the form of a graph structure, making the propagation prediction more in line with engineering reality.

[0062] The descriptor generation module, serving as the output interface of the dynamic risk assessment unit, is responsible for integrating various analysis results and generating standardized risk descriptions. This module receives risk probability distributions from the probability inference module, propagation prediction results from the propagation analysis module, and anomaly scores from the pattern learning module. It then generates structured risk event descriptors through a decision fusion algorithm. The three-dimensional spatial coordinates contained in these descriptors refer to millimeter-level three-dimensional location information, including latitude, longitude, and altitude, acquired and output by the high-precision positioning module of the multimodal perception unit. This coordinate information, as part of the original heterogeneous safety data, is processed and integrated into a safety status feature map. The dynamic risk assessment unit identifies risks based on this map; therefore, the spatial coordinates in its output descriptors are of the same origin, used to accurately locate the physical location of risk events. The descriptors are grouped using JSON format. The risk assessment descriptor consists of three main parts: metadata, risk assessment, and spatiotemporal information. The metadata section records the timestamp of risk identification, data source version number, and processing log ID. The risk assessment section details the identified risk type code, risk intensity quantification value, confidence score, and possible derivative risk types. The spatiotemporal information section includes the spatial coordinates of the risk source, the estimated radius of influence, the predicted duration of the risk, and a list of key nodes in the transmission path. The generation of each field undergoes rigorous logical verification. For example, the risk intensity value is calculated by weighting multiple factors such as the deviation of abnormal characteristics from the standard pattern and the spread speed of the risk event. The confidence score is calibrated by comprehensively assessing the uncertainty of the model itself and data quality indicators. The generation of risk type and intensity indicators in the risk event descriptor relies on in-depth analysis of the correlation patterns between multiple nodes in the safety status feature map.

[0063] By analyzing feature maps, abnormal co-occurrence or coordinated changes of personnel biometric nodes and equipment operation trajectory nodes in time and space are identified, thereby determining the composite risk of personnel approaching dangerous operating equipment under stress. The confidence score is mainly derived from the uncertainty measure of the risk probability distribution calculated by the Bayesian neural network in the probability inference module, that is, the degree of dispersion of the probability distribution, and is comprehensively calibrated in combination with the real-time data quality assessment results output by the data standardization unit. The more concentrated the probability distribution and the higher the data quality, the higher the confidence score.

[0064] During the internal data flow of the dynamic risk assessment unit, modules exchange data efficiently through a shared memory mechanism. The feature distribution parameters extracted by the pattern learning module are stored in a distributed cache for use by the probability inference module. The propagation pattern map generated by the propagation analysis module is directly mapped to the memory space of the descriptor generation module. A pipelined parallel architecture is adopted to improve processing efficiency. When a new pattern learning task is executed in the background, the real-time risk assessment process in the foreground can continue. This design ensures that the system can continuously optimize its risk identification capabilities without interrupting service. The core also establishes a comprehensive quality control mechanism, including integrity checks on the input feature map, verification of the rationality of the output results of each module, and logical consistency checks on the final descriptor. Any data found to be inaccurate at any stage will be addressed. Any anomalies will trigger a reprocessing procedure or a degradation strategy. The operation and maintenance of the dynamic risk assessment unit requires a professional technical support system. The model update team regularly collects new construction scenario data to incrementally train the network, the algorithm optimization team continuously tracks the latest research results to improve the model architecture, and the operation and maintenance engineers monitor the resource consumption and response latency indicators of the core operation. A model version management mechanism has also been established. Each model update retains a backup of the historical version. When the performance of the new version degrades, it can be quickly rolled back to a stable version. This design ensures the continuity and reliability of the risk assessment service. The interface between the core and other modules of the system adopts a standardized design. Through clearly defined application programming interfaces and data exchange formats, it is ensured that the risk assessment capabilities can be flexibly called and integrated by upper-layer applications.

[0065] In terms of handling abnormal situations, the dynamic risk assessment unit is designed with a multi-level degradation strategy. When a certain analysis unit fails, it can automatically switch to a simplified analysis mode. For example, when the propagation analysis unit is unavailable, the propagation prediction function can be suspended based solely on the characteristics at the current moment. The core also includes a self-monitoring mechanism that tracks the computational load and output quality of each analysis unit in real time. When performance degradation or abnormal output is detected, alarms are automatically triggered and detailed logs are recorded, providing data support for subsequent troubleshooting and system optimization. The entire risk assessment process establishes a complete audit trail chain, with detailed logs recorded for each processing step from feature input to descriptor output, meeting the compliance requirements of engineering safety supervision.

[0066] The adaptive response unit and the instruction execution resource scheduling unit together constitute a closed-loop control system from risk identification to actual intervention. The adaptive response unit is deployed on the server cluster in the construction site command center. Its hardware configuration adopts a high-availability architecture, equipped with dual power supply redundancy and hot backup nodes. The software adopts a service-oriented architecture design, encapsulating functional modules such as strategy matching, resource planning, decision optimization, and instruction synthesis into independently scalable microservices. The engineering safety policy library built into the strategy matching module is stored in a graph database. Each emergency plan exists in the form of a node. The node attributes include fields such as plan number, applicable risk type, trigger condition threshold, response level, and specific handling steps. The nodes are connected by edges to represent the derivative or mutually exclusive relationships between plans. When a risk event descriptor is input into the strategy matching module, the module first... The descriptor undergoes syntax parsing to extract key fields such as risk type code, intensity index, and spatial coordinates. The strategy matching unit's syntax parsing of the risk event descriptor includes the following steps: verifying that the descriptor conforms to a predefined structured data format, such as JSON-LD; extracting top-level fields such as risk type, confidence level, and spatial coordinates; mapping and verifying the extracted field values ​​with the built-in power engineering construction safety domain ontology to ensure semantic validity and logical consistency, such as verifying the legality of the risk type code and whether the spatial coordinates are within the valid construction range; and converting the verified structured information into an internal query representation that facilitates efficient pattern matching in a graph database. By introducing the domain ontology for semantic-level verification and enrichment, this goes beyond simple format parsing and improves the accuracy and reliability of strategy matching.

[0067] Pattern matching queries are performed in the graph database. The matching process uses a rule-based inference algorithm, which comprehensively considers the similarity of risk features and the degree of satisfaction of the contingency plan triggering conditions. For example, for the risk event of overload of a high-altitude work platform, the system will simultaneously match the equipment overload handling plan and the special plan for high-altitude work, and judge the applicability of the plan based on whether the risk coordinates are within the high-altitude work area. When generating risk event descriptors, the descriptor generation module determines the risk category and intensity index based on the comprehensive analysis of the correlation patterns between multiple nodes in the safety status feature map. For example, by analyzing the spatiotemporal coordinated abnormal changes of personnel biometric nodes and adjacent equipment operation nodes in the map, composite risks such as personnel approaching dangerous equipment under stress can be identified. The confidence score in the descriptor mainly comes from the uncertainty of the risk probability distribution output by the Bayesian neural network in the probability inference module, that is, the degree of dispersion of the probability distribution, and is combined with the real-time evaluation results of the quality of the heterogeneous data currently used by the data standardization module for comprehensive calibration. The more concentrated the probability distribution and the higher the data quality, the higher the generated confidence score.

[0068] The resource planning module connects to the real-time data interface of the construction management system. These interfaces provide dynamic resource data such as personnel attendance information, equipment operating status, and material inventory. The module establishes a resource visualization model, which abstracts large equipment such as tower cranes, pump trucks, and generators on the construction site into resource objects with location attributes, functional attributes, and status attributes. It also abstracts human resources such as construction teams, safety officers, and technicians into intelligent agents with skill levels, current locations, and task loads. The resource planning algorithm calculates the type, quantity, and response time requirements of necessary emergency resources based on the impact range estimate and intensity index in the risk event descriptor. For example, for the identified risk of excessive displacement of the foundation pit slope, the algorithm automatically selects available large-scale pumping equipment, sandbag reserves, and work teams with foundation pit emergency rescue qualifications, generating a detailed plan that includes resource allocation routes, estimated arrival times, and collaborative operation requirements. The decision optimization module uses a multi-objective programming method to solve for the optimal response strategy. The optimization model established by this module comprehensively considers three core objectives: risk containment effect, resource usage cost, and construction progress impact. The risk containment effect is measured by the rate of risk intensity reduction and the degree of control over the impact range. Resource usage cost includes direct economic indicators such as equipment scheduling costs and personnel overtime costs. The construction progress impact is evaluated by the critical path delay time. The optimization process uses an iterative algorithm to find the non-dominated solution set and outputs a comprehensive decision including response level classification, action time series, and resource allocation plan. During the decision generation process, the effect data of historical disposal cases are referenced to perform a weighted evaluation of the actual effects of different strategies in similar scenarios.

[0069] The instruction synthesis module undertakes the crucial task of transforming abstract decisions into executable instructions. This unit internally establishes an equipment instruction template library and a personnel operation guide library. The template library stores the control protocol formats for various field devices; for example, the emergency stop instruction for a tower crane uses a specific function code and register address from the Modbus RTU protocol, and the closing instruction for a smart gate conforms to the data frame structure of the ISO / IEC 15118 standard. The instruction synthesis process first performs semantic parsing on the optimization decision, identifying the sequence of equipment actions and personnel operations to be executed. The semantic parsing of the optimization decision by the instruction synthesis unit employs an analysis method combined with an operational knowledge graph. Specifically, the structured content of the optimization decision is converted into an enhanced textual instruction sequence, utilizing a domain-specific... The knowledge parsing model identifies the core operational actions, operational objects, operational parameters, and constraints in the instructions. These identified elements are mapped and bound to standard equipment control modes and personnel guidance procedures in the operation knowledge graph, thereby determining the specific, executable sequence of equipment actions and personnel operation items corresponding to each abstract operational intent. This method achieves a deep understanding and decomposition of complex operational intents through the rich semantic associations provided by the knowledge graph. Based on the equipment type, the corresponding instruction template is called from the template library, and specific parameters are filled into the template to generate binary control instructions. At the same time, personnel guides of different levels of detail are generated according to the complexity of the operation items. Simple operations generate voice prompts and text, while complex processes generate illustrated electronic checklists.

[0070] The instruction synthesis module parses optimization decisions into executable instructions as follows: Semantic parsing is performed on the optimization decision to identify atomic operation objects, target states, and constraints. The module accesses the device instruction template library, which pre-stores control instruction protocol templates for various field devices. These templates define the instruction frame structure, function code, address mapping, and parameter format. The semantically parsed parameters are filled into the corresponding device's instruction template, generating a control instruction frame conforming to the device's communication protocol specifications, either in raw binary or a specific encoding format. For example, generating an instruction to stop the main hoist of a specific tower crane might include binary fields such as the target device address, stop function code, and checksum. Natural language-based text or voice guidance is generated for personnel. The instruction synthesis module also includes a conflict detection mechanism to check the logical consistency of multiple simultaneously generated device instructions, preventing contradictory control commands.

[0071] The instruction distribution module of the instruction execution resource scheduling unit uses message middleware to achieve reliable transmission of instructions. This module is designed with hierarchical message topics corresponding to instructions of different urgency levels. Urgent instructions are pushed in real time through high-priority topics, while regular instructions are delivered at least once through persistent topics. The equipment control module establishes a communication connection with the field programmable logic controller (PLC) via industrial Ethernet, supporting multiple industrial protocols such as PROFINET and EtherCAT. The module internally performs protocol conversion, transforming a unified internal command format into control messages recognizable by specific devices. Message transmission employs retry mechanisms and timeout controls to ensure reliable delivery of critical commands. The personnel guidance module integrates augmented reality and spatial sound field technologies. The augmented reality system uses positioning data and scene recognition algorithms to overlay virtual information such as danger zone boundaries, avoidance path arrows, and equipment status indicators onto the safety glasses' display screen. The spatial sound field system utilizes beamforming technology to achieve directional audio transmission, clearly delivering voice warnings to construction personnel in specific areas even in noisy environments. The area management module connects to on-site intelligent gates, warning light poles, and drone base stations via an IoT gateway. The unit internally implements geofencing management, dynamically adjusting the electronic fence range based on risk coordinates. It automatically triggers warnings when personnel or equipment approach a danger zone. The drone patrol system automatically plans patrol routes based on risk levels and transmits high-definition video streams in real time for monitoring by the command center.

[0072] The protocol conversion module's function is achieved collaboratively by the instruction distribution module and the equipment control module within the instruction execution resource scheduling unit. Its working steps are as follows: The instruction distribution module receives hierarchical intervention instructions from the adaptive response unit and parses them into specific equipment action sequences, personnel warning messages, and area management strategies. The equipment control module, based on the equipment type in the action sequence, calls a pre-set equipment instruction template library to convert the abstract actions into standardized control messages conforming to specific industrial communication protocols such as PROFINET and EtherCAT. These control messages are then sent via industrial Ethernet to the corresponding programmable logic controllers or devices in the field, driving security equipment, mechanical braking systems, etc., to execute specific actions, completing the closed loop from instruction to control signal parsing and execution.

[0073] In the collaborative operation of the adaptive response unit and the instruction execution resource scheduling unit, an event-driven architecture is adopted to ensure real-time response. Upon arrival of a risk event descriptor, a multi-threaded parallel processing flow is immediately initiated. Strategy matching, resource planning, and decision optimization are executed in a pipelined manner, overlapping each other. A complete feedback mechanism is established throughout the response process, transmitting the instruction execution status back to the response engine in real time. The engine dynamically adjusts subsequent strategies based on the execution results. For example, if the equipment braking command fails to effectively reduce the risk, it will automatically escalate to stronger intervention measures such as regional power outages. A degradation processing strategy is also designed: when a failure occurs in an execution unit, a backup control channel is automatically activated or manual handling mode is switched to ensure safe intervention. To ensure continuity, the system needs to be integrated with existing construction management processes during implementation. Adaptive response units should establish data exchange interfaces with project schedule management systems and quality acceptance systems to ensure consistency between safety decisions and construction plans. The deployment of command execution resource scheduling units should fully consider adaptability to the site environment. Equipment control modules should be dustproof and waterproof. Personnel guidance equipment should undergo ergonomic testing to ensure comfort during extended wear. Area management equipment should meet the rapid deployment requirements of temporary construction sites. The system also establishes a comprehensive training mechanism, developing specialized operation training plans for personnel in different positions, and conducting emergency response drills through a virtual simulation platform to ensure that relevant personnel are proficient in system operation procedures.

[0074] The forward-looking human risk perception unit 2 is connected and interacts with the real-time risk assessment and scheduling unit 1. It is used to assess the dynamic coupling risk entropy value of the construction personnel when performing power engineering construction based on the physiological state vector of the construction personnel, and send the dynamic coupling risk entropy value to the real-time risk assessment and scheduling unit 1 for state deviation warning.

[0075] In one embodiment, the forward-looking human risk perception unit 2 includes: a physiological signal perception module, used to collect near-infrared brain functional imaging signals, eye movement trajectory signals, skin conductance response signals, and surface electromyography signals of construction workers during power engineering construction to characterize the physiological state vector of construction workers; a behavioral semantic analysis module, used to parse the work video stream, identify the standard of the construction workers' work actions, the area of ​​their gaze, and the spatial interaction between the construction workers and the hazard source, and generate a behavioral semantic vector; and a dynamic coupling risk entropy calculation module, used to receive the physiological state vector, the behavioral semantic vector, and environmental indicators, and establish a nonlinear mapping model to evaluate the dynamic coupling risk entropy value of construction workers when performing power engineering construction, so as to generate a human state deviation warning and send it to the instruction execution resource scheduling unit.

[0076] In one embodiment, the behavioral semantic analysis module includes: a personnel action state recognition module, used to extract limb points of construction workers in the operation video stream through image extraction technology, and calculate the arm extension angle and body center of gravity offset of the construction workers based on the limb points to quantify the standard of the construction workers' operation actions; a personnel position state recognition module, used to locate the hazard source and the line of sight of the construction workers in the operation video stream, and quantify the Euclidean distance between the construction workers and the hazard source based on the angle between the line of sight of the construction workers and the hazard source to determine the line of sight attention area and spatial interaction relationship; and a behavior vector fusion determination module, used to fuse the standard of operation actions, line of sight attention area and spatial interaction relationship as a behavioral semantic vector.

[0077] It needs to be explained that the forward-looking human risk perception unit 2 predicts deviations in human condition through the coordinated operation of a physiological signal perception module, a behavioral semantic analysis module, and a dynamically coupled risk entropy calculation module. In specific implementation, the physiological signal perception module is integrated into personnel intelligent safety equipment such as safety helmets and wearable vests. The physiological signal perception unit synchronously collects near-infrared brain functional imaging signals, eye movement trajectory signals, skin conductance response signals, and surface electromyography signals from construction workers during high-altitude tower installation operations. The signals are wirelessly transmitted to the edge computing gateway at a sampling frequency of 100 times per second. The eye movement trajectory signals include pupil diameter change sequences and fixation point coordinate sequences. The near-infrared brain functional imaging signals provide data on changes in blood oxygen concentration in the prefrontal cortex. The surface electromyography signals come from the trapezius and biceps brachii muscles. These real-time physiological signals constitute a physiological state vector characterizing the cognitive load and fatigue state of personnel. The behavioral semantic analysis module receives signals from the tower... The panoramic vision unit of the base receives the operational video stream, and the behavioral semantic analysis module analyzes the video stream frame by frame. In specific implementation, the behavioral semantic analysis module uses a deep learning skeleton extraction algorithm to identify key points of the personnel's limbs, calculates the personnel's arm extension angle and body center of gravity offset as standardized indicators of operational actions, and uses an object detection network to locate hazards in the video, such as exposed high-voltage lines, and calculates the angle between the personnel's line of sight direction vector and the bounding box of the hazard. Combined with the Euclidean distance between the personnel and the hazard, it outputs a behavioral semantic vector describing the area of ​​focus of the line of sight and the spatial interaction relationship. It can be understood that in the example scenario of climbing the tower, the behavioral semantic analysis module recognizes that the personnel's line of sight continuously deviates from the foot spike position and turns to an irrelevant area, and the arm extension angle exceeds the safety specification threshold. These behavioral semantic data are compared with the normal operation benchmark data, in which the line of sight should be focused on the foot spike and the arm angle should remain stable.

[0078] The dynamic coupling risk entropy calculation module receives physiological state vectors, behavioral semantic vectors, and environmental equipment state vectors. The environmental equipment state vectors include data from environmental parameter acquisition units and equipment status monitoring units, such as wind speed and equipment vibration values. In implementation, the module establishes a nonlinear mapping model that fuses the three vectors into a joint probability distribution within the context of the tower installation task. This joint probability distribution is compared with a preset ideal safety benchmark distribution, which is derived from historical safety operation data statistics. The module quantifies the dynamic coupling risk entropy value in real time by calculating the Jensen-Shannon divergence, expressed by the following formula:

[0079] ;

[0080] in Represents the dynamic coupling risk entropy. The joint probability distribution of the representative personnel's physiological state vector, behavioral semantic vector, and environmental equipment state vector. Represents the ideal safety baseline distribution. yes and The mean distribution is , Representing the Kullback-Leibler divergence, in data comparison, when fatigue leads to increased skin conductance response signals and scattered eye movement trajectory signals in the physiological state vector, while the behavioral semantic vector shows a decrease in action standardization, at this time... Distribution deviation Distribution, calculated The value increased from the normal range of 0.05 to 0.38, where the adaptive threshold is dynamically adjusted according to the work stage. For example, the threshold for high-altitude work is set to 0.30. When the value exceeds 0.30, the dynamic coupling risk entropy calculation module generates a risk event descriptor. The risk event descriptor includes specific personnel identification such as employee number, risk type labeled as human-caused state deviation, and predicted deviation direction as inattention and muscle fatigue. The risk event descriptor is immediately sent to the adaptive response unit. It can be understood that the dynamic coupling risk entropy calculation module achieves a quantitative assessment of the dynamic coupling risk of human, machine, and environment by continuously comparing the real-time joint probability distribution with the ideal safety benchmark distribution, thereby providing early warning before personnel make operational errors.

[0081] In one embodiment, extracting limb points of construction workers from a work video stream using image extraction technology includes: performing frame segmentation on the work video stream to generate clear frames and occluded frames of construction worker actions; using deep learning technology to analyze the clear frames of construction worker actions and identify clear limb points of the construction workers; setting a classification threshold based on digital refocusing technology and energy gradient function, and combining element image inpainting technology to repair and reconstruct target objects in the occluded frames of construction worker actions, and outputting repaired clear frames of construction worker actions based on the processing results; using deep learning technology to analyze the repaired clear frames of construction worker actions, identifying the repaired limb points of the construction workers, and combining them with the clear limb points to output the limb points of the construction workers in the work video stream.

[0082] In one embodiment, a classification threshold is set based on digital refocusing technology and an energy gradient function, and element image inpainting technology is combined to repair and reconstruct the target object from occluded frames of construction worker actions. The output of a clear frame of the repaired construction worker action includes: determining the light intensity and direction information from different viewpoints in the construction worker action scene based on the occluded frame; acquiring an element image array from the corresponding viewpoint; and performing digital refocusing processing on the element image array along the axial depth. Based on the digital refocusing processing results, a sequence of clear and blurred images distributed along the depth is obtained, and the sum of gradient values ​​of all adjacent points in the image sequence is analyzed to determine the focus evaluation result of the image sequence at the corresponding axial depth. The focus evaluation result satisfies... After setting a standard value, the image sequence is subjected to binary encoding transformation and cost aggregation to determine the classification threshold. The foreground occlusion area in the construction worker action occlusion frame is removed based on the classification threshold. After the occlusion area of ​​the construction worker action occlusion frame is removed, a clear construction worker action frame in the same row as the occlusion frame is selected as a sample image, and the disparity value of the adjacent elements in the sample image and the occlusion frame is calculated. The sample image is shifted according to the corresponding disparity value, and the pixel value of the corresponding position is extracted and fused into the foreground occlusion area of ​​the construction worker action occlusion frame to obtain the repaired construction worker action occlusion frame. The repaired construction worker action occlusion frame is then subjected to three-dimensional reconstruction processing to output the repaired clear construction worker action frame.

[0083] In one embodiment, after the focus evaluation result meets the preset standard value, binary encoding transformation and cost aggregation processing are performed on the image sequence to determine the classification threshold. The foreground occlusion region in the construction worker's action occlusion frame is then removed based on the classification threshold. This includes: selecting a transformation window according to the target size; comparing the grayscale values ​​of pixels within the transformation window with the center pixel to generate a binary bitstream; using the binary bitstream to perform binary encoding transformation on the image sequence to complete the cost matching processing of the image sequence; after cost matching processing, linear interpolation processing is performed on the pixel intensity of the image sequence; the minimum disparity of each pixel in the image sequence is analyzed; and additional smoothing constraints are added to construct an energy function. The system performs cost aggregation of image sequences by applying different penalty terms to adjacent disparity changes. It obtains the disparity value of each pixel in the image sequence at the minimum cost aggregation, and after left-right consistency check, disparity filling and filtering, it obtains the disparity map of each element image array. Based on the conversion relationship between the image sequence and the disparity map, it determines the classification threshold between construction worker targets and occluders. Based on the imaging principle that the disparity value of the occluder is greater than that of the construction worker target, it marks the foreground occluded areas in the element image array whose disparity value is greater than the classification threshold, thus completing the classification of occluders and construction worker targets and realizing the removal of foreground occluded areas in the construction worker action occlusion frame.

[0084] It needs to be explained that the essence of extracting limb points of construction workers from a work video stream using image extraction technology is to perform frame segmentation processing on the input work video stream. This involves breaking down the continuous video stream into individual static image frames according to a fixed time step or frame interval. Simultaneously, pre-classification is performed based on the completeness of the construction worker's actions within the frame, the occlusion ratio, and pixel clarity, dividing all image frames into clear frames and occluded frames. The purpose is to decompose the continuous video stream into the smallest independently processable image elements, while pre-processing the two types of frames separately. This allows clear frames to directly enter the deep learning recognition stage, while occluded frames enter a dedicated repair and reconstruction path, achieving precise allocation of computing resources and parallel acceleration of the processing flow, meeting the low-latency requirements of real-time monitoring at construction sites. For the clear frames of construction worker actions obtained through the segmentation, a convolutional neural network model optimized based on human pose estimation (such as HRNet or OpenPose industrial optimization version) is directly used for analysis. Pre-trained convolutional kernels are used to extract the human body contours and other features of the construction workers within the clear frames. Key features such as joint texture and limb edges are accurately located and output in 2D / 3D limb key points that conform to the topology of the human skeleton. For the occluded frames of construction workers' actions obtained by splitting, digital refocusing preprocessing based on the principle of light field imaging is performed. Specifically, this is to restore the light intensity and propagation direction information of the construction worker's action scene from different perspectives based on the occluded frame, and generate an element image array for the corresponding perspective. This element image array is a set of sub-images corresponding to different microlens perspectives and different depth planes. Each element image corresponds to the light sampling result of a specific depth and specific perspective in the scene. Digital refocusing processing is performed on the element image array along the axial depth of the camera's optical axis, that is, the element image array is integrated and superimposed at different depth layers along the depth direction. The camera's focus plane is readjusted in digital space, and the image originally focused on the foreground occluder is refocused on the construction worker target in the mid-to-background. Thus, the image information of different depth planes within the occluded frame can be restored from the light field dimension, and the imaging information of the foreground occluder and the background construction worker in the depth dimension can be separated.

[0085] After digital refocusing, a sequence of images is generated, distributed sequentially along the depth of the camera's optical axis. Images whose focus depth matches the depth of the construction worker's target are in sharp focus, while those whose focus depth does not match are in blurry, out-of-focus. The focus effect of this image sequence is quantitatively evaluated based on the energy gradient function. Specifically, this is achieved by utilizing the core principle that sharp images have sharp pixel grayscale changes in the edge contour areas and larger gradient values ​​between adjacent pixels, while blurry images have gradual pixel grayscale changes and smaller gradient values ​​between adjacent pixels. The sum of the gradient values ​​of all adjacent pixels in each image in the image sequence is calculated to obtain the focus evaluation value corresponding to that image. The larger the gradient sum value, the higher the degree of focus and the better the sharpness of the image; conversely, the lower the degree of focus and the blurrier the image. Finally, by traversing the focus evaluation values ​​of the entire depth sequence, the focus evaluation result for the corresponding axial depth is determined, locating the optimal focused image at the depth of the construction worker's target. This transforms the subjective image sharpness / blurriness into a quantifiable and comparable numerical focus evaluation result, achieving automated and precise positioning of the optimal focus depth during digital refocusing.

[0086] When the focus evaluation result reaches the preset standard value based on the focus evaluation value of clear frames at the construction site, it indicates that the image at that depth has achieved clear focus on the construction worker target, possessing the image quality foundation for separating occlusions from the target object. The process then proceeds to the binary encoding transformation and cost aggregation processing stage, where binary encoding transformation is performed. Specifically, a rectangular sliding window (such as a 3×3 or 5×5 neighborhood window) adapted to the human body size of the construction worker is used as the transformation window. This window slides pixel by pixel across the image. For each pixel within the window, its grayscale value is compared with the grayscale value of the center pixel of the window. If the grayscale value is greater than or equal to the center pixel's grayscale value, the grayscale value is calculated. The grayscale value of the center pixel is recorded as 1, and anything less is recorded as 0. Finally, all pixels within the window generate a fixed-length binary code stream as the feature code for that center pixel. After all pixels in the entire image sequence have been encoded, the binary encoding transformation and cost matching processing of the image sequence is complete. This transforms the originally continuous pixel grayscale values ​​into binary feature codes with strong anti-interference capabilities, completing pixel-level feature matching. After cost matching, the cost aggregation stage begins. Specifically, this involves first performing linear interpolation on the pixel intensity of the image sequence to achieve sub-pixel-level disparity accuracy improvement, enabling disparity calculation... The precision breaks through the limitation of integer pixels. The essence of disparity is the pixel position offset of the same target point in the same scene across element images from different viewpoints. The disparity value is directly related to the axial depth of the target point; the closer the depth, the larger the disparity value of foreground occluders, and the farther the depth, the smaller the disparity value of mid-to-background construction workers. Subsequently, an additional smoothing constraint is added to construct an energy function. This energy function contains a data term representing the disparity matching cost for each pixel, and a smoothing term that penalizes disparity changes in adjacent pixels. When the disparity change between adjacent pixels is small, a smaller penalty term is applied to ensure the disparity continuity of the same target area. When the disparity between adjacent pixels changes abruptly, a large penalty term is applied to the edge region between the target and the background to avoid edge blurring caused by excessive smoothing. By minimizing this energy function, the cost aggregation of the entire image sequence is completed, and the optimal, globally consistent disparity value of each pixel is obtained. The initial, single-pixel independent disparity matching cost can be transformed into a globally optimized disparity result that conforms to the scene spatial topology, eliminating the interference of mismatched points and noise points on disparity calculation. At the same time, sub-pixel level disparity accuracy is improved through linear interpolation, providing a high-precision disparity data foundation for the accurate classification of occluded objects and targets.

[0087] After cost aggregation, the optimal disparity value of each pixel in the image sequence at the minimum cost aggregation is obtained. Disparity map post-processing and classification threshold determination are then performed. Specifically, the initial disparity values ​​are first checked for left-right consistency. The disparity maps calculated from the left-view element images and the right-view element images are cross-validated, retaining only pixels with consistent disparity values ​​across the two views. Inconsistent pixels are marked as occluded or invalid regions, eliminating a large number of mismatched points. Disparity filling is then performed. For invalid regions and occluded holes marked by the left-right consistency check, weighted interpolation of effective neighborhood disparity values ​​or background filling algorithms are used to fill disparity holes, ensuring the integrity of the disparity map. Next, filtering is performed, using median filtering or bilateral filtering to remove noise points from the disparity map while preserving edge information. Finally, a complete, high-precision, hole-free disparity map is obtained. Each pixel value in the disparity map represents the optimal disparity value for each pixel. Based on the disparity value and axial depth of the pixel in the original image, and using the statistical characteristics of the disparity map, the Otsu method or K-means clustering algorithm is used to divide the pixel values ​​in the disparity map into a low disparity value cluster corresponding to the construction worker target and a high disparity value cluster corresponding to the foreground occlusion. The boundary value between the two clusters is the classification threshold for the construction worker target and the occlusion. Finally, based on the imaging principle that the closer the object distance, the greater the disparity value, the pixel areas in the disparity map with disparity values ​​greater than the classification threshold are marked as foreground occlusion areas, and the pixel areas with disparity values ​​less than or equal to the classification threshold are marked as construction worker target areas and background areas, respectively. This completes the accurate classification of occlusion and construction worker target, and removes the foreground occlusion area in the construction worker action occlusion frame. Then, through disparity statistical analysis, an adaptive classification threshold is determined to achieve automated, pixel-level accurate segmentation and removal of the foreground occlusion area and the construction worker target area.

[0088] After removing the occluded area, the next step is element image inpainting and 3D reconstruction. Specifically, this involves first selecting clear, unobstructed frames within the same video sequence, camera position, and adjacent time step as sample images. These sample images exhibit high spatiotemporal continuity with the occluded frame's construction scene, shooting angle, and personnel movements, making them optimal inpainting samples. Next, the disparity value between adjacent element images in the sample image and the occluded frame is calculated. This disparity represents the pixel position offset of the same scene target point in the sample image and the occluded frame. This disparity corresponds to the pixel displacement between the two frames caused by shooting time and minor personnel movements. Finally, the sample image is reconstructed according to the calculated disparity value. Subpixel-level displacement processing is performed on the difference to precisely align the features of the construction workers' limbs and contours in the sample image with the unoccluded areas of the construction workers in the occluded frame. Pixel values ​​corresponding to the foreground occluded area in the occluded frame are extracted from the aligned sample image and seamlessly integrated into the foreground occluded area of ​​the occluded frame to complete pixel filling of the occluded area, resulting in a preliminarily repaired occluded frame. Based on the previously obtained disparity map and element image array, 3D reconstruction processing is performed on the preliminarily repaired occluded frame to restore the 3D human skeleton model of the construction workers and the 3D structure of the scene. 3D spatial optimization is performed on the pixel values, contours, and motion continuity of the repaired area. The correction process ensures that the limb movements, human proportions, and texture details in the repaired area are completely consistent with the unoccluded areas, without any stitching marks, movement distortion, or proportional imbalance. The final output is a clear frame of the repaired construction worker's movements that conforms to the topology of human movement, is unoccluded, and distortion-free. This allows for pixel-level precise repair of occluded areas based on spatiotemporally continuous clear sample frames. Furthermore, spatial topology correction of the repaired content is achieved through 3D reconstruction. The 3D reconstruction correction process ensures the continuity and rationality of the repaired construction worker's limb movements in three-dimensional space, ensuring that the repaired frame has the same recognition accuracy as the original clear frame. This process is applied to the output of the construction worker... The motion restoration of clear frames uses the same deep learning human pose estimation model as the clear frames for analysis, accurately locating and outputting the limb points of the construction workers in the restored frames. Finally, the clear limb points identified from the clear frames and the restored limb points identified from the clear frames are merged and temporally smoothed according to the time sequence of the video stream. The final output is a continuous and complete sequence of limb points of the construction workers throughout the entire operation video stream, without blind spots, to complete the limb point recognition of the restored frames. The recognition results of the two types of frames are temporally merged to fill the gaps in limb point data during occlusion periods, achieving a complete restoration of the limb movements of the construction workers in the entire operation video stream.

[0089] Risk simulation mapping unit 3 is connected and interacts with real-time risk assessment and scheduling unit 1. It is used to build a real-scene model based on point cloud data of power engineering construction site, and combine construction plan and equipment configuration management to generate parameter constraints, determine risk intensity value, and output optimized power engineering construction plan and risk control measures to real-time risk assessment and scheduling unit 1.

[0090] Safety management of power engineering construction sites is achieved by combining real-time risk assessment and scheduling unit 1, forward-looking human factor risk perception unit 2, and risk simulation and mapping unit 3.

[0091] In one embodiment, the risk simulation mapping unit 3 includes:

[0092] The digital mapping module is used to construct a 3D real-scene model of the power engineering construction site based on point cloud data, and generate a benchmark virtual scene by combining physical rules and behavioral rules in a hybrid modeling technology to determine the risk intensity value.

[0093] The data fusion module is used to acquire construction schedule plans and construction personnel and equipment management configurations as multi-source external data. After formatting the multi-source external data, it extracts action points and state change points. It converts the action points and state change points into event streams that can be recognized by the benchmark virtual scene, and extracts the parameterized constraints required for the benchmark virtual scene derivation from the multi-source external data. It adds the event streams and parameterized constraints to the benchmark virtual scene in chronological order.

[0094] The knowledge evolution module is used to match the risk intensity value output by the benchmark virtual scenario with the disposal strategy, verify the performance evaluation of the disposal strategy in different scenarios, generate a risk evolution path map, and output the optimized construction plan and risk control measure adjustment suggestions to the instruction execution resource scheduling unit.

[0095] In one embodiment, constructing a 3D real-scene model of a power engineering construction site based on point cloud data and generating a benchmark virtual scene using a hybrid modeling technique combining physical and behavioral rules includes: acquiring point cloud data of the power engineering construction site using a laser scanning station, and uniformly generating a 3D real-scene scene after registering and fusing the point cloud data, serving as the 3D real-scene model of the power engineering construction site; analyzing the motion and behavioral state of construction equipment under external forces using multibody dynamics and materials mechanics, and adding this as a physical rule to the 3D real-scene model; treating construction personnel as intelligent agents, assigning behavioral states to the agents using collaborative game theory, and converting the output of the behavioral states into a control rule base, obtaining behavioral rules that are added to the 3D real-scene model; and constructing a benchmark virtual scene of the power engineering construction site for a future target time period based on the 3D real-scene model with added physical and behavioral rules, combined with current construction technology.

[0096] In one embodiment, employing collaborative game theory to assign behavioral states to agents and converting the output of these behavioral states into a control rule base includes: configuring a working state for each agent and modeling agents involved in future power engineering construction plans as agents with state stochasticity; simulating the interaction between the agent configuration after working state configuration and the agents with state stochasticity during power engineering construction based on a 3D real-scene model; analyzing the state transition process of agents during work based on the interaction state, and evaluating the fuzziness and stochasticity of the 3D real-scene model during the simulation based on the state transition process, defining a standard evaluation set for agents; and constructing a multi-agent dynamic game function based on the standard evaluation set, using agents as game participants, with the goals of operational safety and efficiency, analyzing the equilibrium strategies of agents under different working conditions and different scenario events, and converting the equilibrium strategies into cloud droplet distribution characteristics. Based on the cloud droplet distribution characteristics, encoding the control rules of agents during the construction process and state transition process, as a behavioral control rule base.

[0097] It needs to be explained that in the process of using collaborative game theory to assign behavioral states to intelligent agents and converting the output of these behavioral states into a control rule base, the scope of intelligent agents in the power engineering construction scenario is first clarified. This includes all digitally managed entities participating in the construction process, such as construction machinery intelligent agents and worker intelligent agents. Standardized discrete working states are then configured for each real-time manageable intelligent agent, specifically including core states such as idle, waiting, in operation, and early warning. Simultaneously, corresponding quantitative attribute parameters are bound to each working state, including operational capacity thresholds, energy consumption limits, safety constraint boundaries, response latency, resource utilization, and state transition triggering conditions. This completes the basic digital modeling of deterministic and controllable intelligent agents. Then, the future power... The intelligent agents involved in the engineering construction plan that have not yet entered the site or have uncertainties are modeled as intelligent agents with state stochasticity based on historical power construction data, equipment failure probability, environmental impact parameters, and personnel operation deviation statistics. Specifically, discrete-time Markov chains are used to represent the stochastic characteristics of their state transitions, clarifying the probability distribution of transitions between different states, the random fluctuation range, and the probability of triggering events. This allows construction entities with different types, functions, and management modes in the physical world to be uniformly transformed into standardized intelligent agents that can be simulated and interacted with in the digital space. At the same time, deterministic and controllable intelligent agents are distinguished from stochastic intelligent agents in the future plan, and uncertainties throughout the entire construction cycle are incorporated into the modeling system in advance.

[0098] After completing the agent modeling, a multi-agent full-cycle interactive simulation was conducted based on the 3D real-scene digital twin model of power engineering construction. During the simulation, deterministic agents with completed work state configurations and future-planning agents with stochastic states were first precisely deployed to their corresponding spatial locations and time nodes in the 3D real-scene model according to the construction organization design's time sequence plan, work zoning, process logic, and spatial constraints. Then, using the Monte Carlo sampling simulation method, thousands of repeated sampling simulations were conducted to address the state transitions of stochastic agents, random environmental events, and operating condition fluctuations, fully reproducing the entire construction process from construction preparation, foundation construction, tower erection, line erection, live-line work to final acceptance. During the construction cycle, the full range of interaction states between multiple agents are recorded. Specifically, this includes interaction in the spatial dimension such as overlapping and obstacle avoidance of work areas, interaction in the process dimension such as sequential dependence and time-series coordination, interaction in the resource dimension such as equipment, interaction in the safety dimension such as safe distance constraints for live / non-live work, and emergency response in fault scenarios. At the same time, the system collects the status data, interaction behavior data, spatial location data, and safety and efficiency index data of all agents in real time during the simulation process, forming a complete multi-agent interaction state dataset. Then, in a digital twin space with zero on-site trial and error costs, the system fully reproduces the dynamic interaction behavior of multiple agents in the entire process of power construction and quantifies the coupling relationship between different agents.

[0099] Based on the interactive state dataset, this study analyzes the state transition process of intelligent agents, quantifies simulation uncertainties, and defines a standard evaluation set. First, it analyzes the state transition process, extracting the temporal chain of state changes for each agent throughout the entire construction cycle from the interactive state dataset. It traces all transition paths from the initial state to the final state for each agent, clarifying the triggering events, preconditions, transition delays, and probabilities of each state transition, as well as the impact of each transition on surrounding related agents, overall construction progress, and safety. Simultaneously, it uses a Markov state transition matrix to quantify the state transition patterns of each agent, constructing a state transition topology network for the entire intelligent agent system, and clarifying the coupling relationships between the state transitions of different agents. Based on the quantification results of the state transition process, the fuzziness and randomness in the simulation process of the 3D real-scene model are evaluated. Fuzziness refers to the boundary fuzziness of evaluation indicators such as safety, efficiency, and compliance in the power construction scenario. That is, safety and insecurity, high efficiency and low efficiency are not black-and-white binary boundaries, but have continuous transition intervals. The membership function in fuzzy mathematics is used to quantify the fuzziness, and the fuzzy intervals and membership calculation methods of different evaluation indicators are clarified. Randomness refers to the random fluctuations of the agent's state transition, the random occurrence of interactive events, and the random changes of environmental parameters. The variance, coefficient of variation, and information entropy in probability statistics are used to quantify the randomness, and the probability distribution and fluctuation range of different random events are clarified.

[0100] Based on the quantitative results of fuzziness and randomness, and combined with the industry norms, safety management standards, and project schedule and efficiency requirements of power engineering construction, a standard evaluation set for intelligent agents is defined. This evaluation set takes operational safety and operational efficiency as the two core dimensions, and each dimension is divided into 5 levels of standardized evaluations. The safety dimension is extremely safe, safe, critically safe, unsafe, and extremely unsafe, while the efficiency dimension is extremely high efficiency, high efficiency, compliant efficiency, low efficiency, and extremely low efficiency. At the same time, each evaluation level is bound to a corresponding quantitative indicator range, membership degree range, state transition constraints, and interaction behavior boundaries, so that each evaluation level has a clear and calculable quantitative judgment standard, without subjective fuzzy definitions. This quantifies the inherent laws and coupling relationships of state transitions of multiple intelligent agents, and clarifies the triggering logic and scope of influence of state changes.

[0101] After defining the standard evaluation set, multi-agent dynamic collaborative game modeling and equilibrium strategy solving are carried out. First, all modeled agents are defined as game participants. Each agent possesses an independent strategy space, action set, payoff function, and information set. The strategy space refers to the set of all behavioral strategies an agent can choose during construction, including work sequence adjustment, work space avoidance, resource allocation priority adjustment, fault emergency response actions, state transition trigger threshold adjustment, and collaborative linkage rules. Each strategy corresponds to a different level in the standard evaluation set. The action set is the set of specific actions that can be executed in real time within the strategy space. The information set is all information that the agent can obtain regarding its own state, the states of surrounding agents, the working environment, and construction progress. The payoff function is constructed entirely based on the standard evaluation set, with work safety and work efficiency as the dual core optimization objectives. The evaluation membership degrees of the safety dimension and the efficiency dimension are weighted and summed using preset weight coefficients for each project to form the comprehensive payoff function for each agent. Simultaneously, the weight coefficients can be dynamically adjusted according to the project type; for example, increasing the safety weight for live-line work projects and increasing the efficiency weight for emergency repair projects. Subsequently, based on the electrical... To address the temporal progression characteristics of power construction, a multi-agent dynamic collaborative game function is constructed. This function fully encompasses the strategy space, payoff function, state transition equation, and information set of all game participants. It also incorporates hard constraints such as power industry safety regulations, project schedule requirements, total resource limits, and safety boundaries for live-line work, clarifying that this game is a collaborative game aimed at achieving optimal global overall payoff for all agents. Based on the constructed dynamic collaborative game function, a combination of backward induction and deep reinforcement learning algorithms is used to solve for perfect equilibrium strategies in subgames under different construction conditions and scenario events. The construction conditions cover all typical power construction scenarios, including live-line work, power outage work, and high-altitude work. The scenario events cover all typical scenarios, including normal operation, equipment failure, sudden weather changes, and safety warnings. The obtained equilibrium strategies are Pareto optimal, meaning that no agent can improve their individual payoff without reducing the overall safety and efficiency payoff by changing their own strategy, ensuring the global optimality of the strategy. Furthermore, each equilibrium strategy is bound to a corresponding combination of working conditions, scenarios, agent states, and payoff quantification results, forming a complete equilibrium strategy library.

[0102] After solving the equilibrium strategy, the strategy is transformed into a cloud model and the control rules are encoded, ultimately forming a behavior control rule base. First, cloud model theory is used to transform the solved equilibrium strategy into cloud droplet distribution characteristics. The cloud model is a mathematical model specifically designed to handle the uncertainty transformation between qualitative concepts and quantitative values. It can simultaneously characterize the fuzziness and randomness of evaluation indicators, perfectly matching the uncertain characteristics of the construction process in the early stages of quantification. Each equilibrium strategy corresponds to a normal cloud model, which contains three core numerical features: expectation Ex, entropy En, and hyperentropy He. The expectation Ex is the median of the cloud droplet distribution. The core value corresponds to the optimal core control parameter of the equilibrium strategy, that is, the optimal action value that best meets the overall safety and efficiency goals. Entropy En corresponds to the fuzziness of the qualitative concept, that is, the reasonable fluctuation range of the control parameter, matching the membership interval corresponding to the standard evaluation set. Hyperentropy He corresponds to the randomness of entropy, that is, the random dispersion of the fluctuation range, matching the randomness of the construction process quantified in the early stage. Through this transformation, each abstract equilibrium strategy is transformed into a numerical model with clear cloud droplet distribution characteristics that can adapt to the uncertainties on site. Each cloud droplet is a specific feasible implementation sample of the strategy in the field working conditions.

[0103] Based on the cloud droplet distribution characteristics, the control rules of the intelligent agent in the entire construction process and state transition process are encoded. Specifically, a standardized IF-THEN production rule structure is adopted to bind and encode the triggering conditions and the control actions of the equilibrium strategy in the construction process. The IF part is the precondition of the rule, which includes the current working state of the intelligent agent, the construction condition, the triggering scene event, the interaction state of surrounding intelligent agents, and the membership degree of safety and efficiency indicators. The THEN part is the execution action of the rule, which includes the behavioral action instructions of the intelligent agent, the state transition rules, the collaborative linkage requirements, and the cloud droplet distribution characteristics of the control parameters. The optimal value, fluctuation range and random adaptation interval of the control parameters are clearly defined to ensure that the rules not only conform to the globally optimal equilibrium strategy, but also adapt to the fuzzy boundaries and random conditions on site, rather than rigid fixed thresholds. Finally, all the encoded control rules are classified, verified and stored according to intelligent agent type, construction procedure, working condition type and scene event to form a complete behavior control rule library that can be directly implemented and executed. This rule library can be directly distributed to the edge controllers of each intelligent agent to realize real-time on-site management and control, or it can be deployed on the digital twin platform to realize online pre-rehearsal and dynamic management and control of the entire construction process.

[0104] It should be explained that the digital mapping module adopts a technical approach of fusing high-precision scanned point clouds with building information models. It acquires millimeter-level precision point cloud data through laser scanning stations set up at the construction site. Each scanning station generates hundreds of millions of spatial coordinate points per hour. This point cloud data is registered and fused using an iterative nearest-point algorithm to generate a 3D reality model of the entire construction site. The model building process employs hierarchical detailing technology, using simplified models for areas far from the viewpoint and retaining complete geometric details for near-field areas. This design balances rendering efficiency and visualization accuracy requirements. The module integrates a real-time data-driven mechanism, accessing readings from on-site sensors via industrial IoT protocols. For example, it maps tower crane stress monitoring data to the corresponding positions in the virtual tower crane model and converts personnel positioning coordinates into avatar motion trajectories in the virtual scene, achieving bidirectional mapping between physical entities and virtual models. The application of the digital mapping module in risk management is reflected in its dynamic simulation function. When the system identifies a risk of excessive wind speed during the hoisting of large equipment, it first recreates the accident scenario in the virtual scene: parameters such as the crane boom angle, the mass distribution of the hoisted object, and wind speed and direction data are precisely assigned to the virtual model. The simulation engine calculates the risk development process based on physical laws, considering the amplification effect of wind loads on the swaying of the hoisted object, the possibility of crane structural resonance, and the chain reaction of the hoisted object's fall trajectory. The simulation process employs multi-threaded calculations, simultaneously simulating the effects of multiple response strategies. For example, strategy A immediately stops hoisting and activates wind-resistant anchoring; strategy B slowly lowers the hoisted object to a safe area; and strategy C urgently adjusts the boom angle to avoid wind direction. The simulation results for each strategy generate quantitative indicators, forming a comparative analysis table of response strategies as shown in Table 1 below.

[0105] Table 1: Comparative Analysis of Lifting Risk Management Strategies

[0106] The simulation process pays special attention to the second and third-order effects of the response measures. For example, while strategy A can quickly control the risk, it may cause the hoisted object to remain suspended in the air for an extended period, increasing uncertainty. Strategy C has the least impact on the schedule but requires operators to have high emergency response skills. These simulation results are presented to managers through a visual interface, supporting drag-and-drop parameter adjustments and real-time re-simulation. Managers can modify wind speed thresholds, try different landing paths, and observe the changing trends of the simulation results. The knowledge evolution module uses a reinforcement learning mechanism to continuously accumulate system experience. The module is designed with a state-action-reward triplet learning framework. The state space contains 128-dimensional feature vectors such as risk characteristics, environmental parameters, and resource status. The action space corresponds to a set of executable response strategies. The reward function comprehensively evaluates multiple objectives such as risk control effectiveness, resource utilization efficiency, and the degree of impact on schedule. After each risk response is completed, a complete response trajectory is recorded, including the initial state, the sequence of actions taken, and the final result. This data is anonymized and stored in the system. The system includes a replay buffer; a distributed asynchronous optimization architecture where multiple learning agents explore different strategy selection paths in parallel, continuously optimizing strategy network parameters through value function approximation; and a knowledge evolution module update mechanism that dynamically adjusts decision parameter systems, maintaining a strategy weight matrix within the module to record the historical performance evaluation of each treatment strategy in different scenarios. For example, for risks such as foundation pit seepage, historical data analysis reveals that the success rate of using a combination strategy of adding drainage equipment and slope reinforcement is significantly higher than that of a single drainage strategy, and this experience is encoded into strategy selection preferences. Furthermore, a strategy innovation mechanism is established: when encountering new risk patterns or when traditional strategies are ineffective, the system combines and mutates existing strategies to generate new treatment plans and evaluate their potential value.

[0107] The synergy between the digital mapping module and the knowledge evolution module is reflected in the closed-loop optimization process. The advantageous strategies recommended by the knowledge module are first simulated and verified in the digital mapping environment. The simulation results are fed back to the knowledge evolution module as a basis for strategy evaluation. This virtual-real training method significantly reduces the cost of trial and error in the field. For example, the system tries out various new support schemes in a virtual environment, quickly transforming successful experiences into executable strategies. The data interaction between the two modules uses a unified timestamp alignment mechanism to ensure the time consistency between virtual simulations and actual records, facilitating comparative analysis of effects. A strict quality control system is established during module implementation. The digital mapping module regularly verifies model accuracy, correcting model errors by comparing virtual measurements with actual survey data. The knowledge evolution module sets up a strategy evaluation committee to manually review important strategies generated by the system, preventing decisions that do not conform to engineering realities. Version management adopts a gray-scale release mechanism; newly optimized risk assessment models are first piloted in select construction areas, and gradually rolled out to the entire site after confirming good results.

[0108] The digital mapping module, through the data fusion module, receives real-time access to the four-level construction schedule from the engineering management platform, refined early warning data from the meteorological service interface, and future configuration plans from the personnel and equipment management system. This multi-dimensional dynamic data is parsed and transformed into a time-series event stream and parameterized constraints that drive the virtual scene. The specific steps are as follows: First, multi-source data access and standardization: The system fusion module receives external data in real-time, including the four-level construction schedule from the engineering management platform (structured data), refined early warning data from the meteorological service (time-series data), and future configuration plans from the personnel and equipment management system (list data). The data parsing unit of the digital mapping module first performs format parsing and time alignment on these heterogeneous data, uniformly converting them into event prototypes with standard timestamps. The second step is the extraction of time-series event streams: the parsing unit extracts key actions and state change points from the data and converts them into events recognizable by the virtual scene engine; for example, the start of hoisting operations in the construction plan is converted into a hoisting start event, the effective time of gust warnings in meteorological data is converted into a wind speed level change event, and the departure of welder A in the personnel plan is converted into a key position absence event. Each event includes a type, trigger timestamp, affected entity identifier, and event parameters. The third step is the generation of parameterized constraints: the boundary conditions and rule parameters required for the deduction are extracted from the data; for example, the maximum allowable wind speed for hoisting operations is extracted from the construction plan as a safety threshold parameter; the skill level and location of backup welder B are extracted from the equipment management system as resource replacement constraints; and the spatial clearance on the hoisting path is extracted from the point cloud data as a geometric constraint parameter. These parameters are organized into key-value pairs and injected into the simulation engine in sync with the event stream. The fourth step is to inject events and parameters into the virtual scene: the generated event stream and parameterized constraints are parsed and injected into the baseline virtual scene constructed by the digital mapping module in chronological order. The event stream drives the scene state to evolve according to the plan and time, while the parameterized constraints serve as inputs to the physical and behavioral rule models in the simulation engine, dynamically adjusting the boundary conditions of the simulation to ensure that the virtual simulation is highly consistent with the future actual construction plan. When conducting forward-looking simulations, the digital mapping module first constructs a baseline virtual scene for the next three days based on the current construction plan.

[0109] The simulation engine employs a hybrid modeling approach that combines physical and behavioral rules: on the one hand, it calculates the sway of hoisted objects under strong winds based on a built-in mechanical model; on the other hand, based on multi-agent simulation technology, it models personnel and equipment planned to enter the site as agents that follow safety procedures but possess a certain degree of state randomness, simulating their interactions under complex working conditions. Specifically, this includes the following steps: First, physical rule modeling: for physical objects in construction (such as tower cranes, hoisted objects, and scaffolding), the simulation engine has a built-in physical model based on multibody dynamics and materials mechanics. These models are defined by a series of differential equations used to calculate the motion and deformation of entities under external forces. For example, the swing model of a hoisted object is described by a formula, where represents the swing angle of the hoisted object, represents the gravitational acceleration, represents the boom length, represents the damping coefficient, and represents a function of wind disturbance that varies with time. This model takes real-time wind speed, boom length, and hoisted object mass as inputs and calculates the spatial position and swing trajectory of the hoisted object. Secondly, behavioral rule modeling: For active entities such as construction workers and work teams, the inference engine uses multi-agent simulation technology for modeling. Each agent is given a behavioral model based on a finite state machine, and its state transitions (such as waiting, working, moving, and avoiding danger) are controlled by a predefined rule base. The rule base encodes safety procedures, work processes, and simple decision logic. For example, if the hoisted object enters the area below, the state changes from operation to avoidance and the object moves along a preset path. The agent determines its behavior based on events in the virtual scene (such as the actions of other equipment and warning signals) and its own state. Finally, dual-model coupling and collaborative inference: during the inference process, the physical model and the behavioral model are bidirectionally coupled through the shared virtual scene state. The equipment state (such as the position of the boom) calculated by the physical model is perceived by the agent in the behavioral model, thus affecting its decision. Conversely, the agent's behavior (such as personnel entering a certain area) will affect the simulation environment of the physical model as an event or boundary condition.

[0110] After the simulation was initiated, the system simulated a combined extreme working condition during construction work across an elevated railway line on the afternoon of the second day, encountering a level 6 gust of wind and a key welder being absent due to an unforeseen situation. The engine dynamically extrapolated the potential risk evolution path under this condition through multi-threaded parallel computation. For example, in the simulated combined extreme condition, wind first caused slight swaying of the work platform; this swaying further affected the welding operation. Due to insufficient skills of on-site operators or the need for substitutes, welding efficiency decreased; decreased efficiency led to overall delays in the process; and these delays forced subsequent hoisting operations, originally planned for the daytime, to be rescheduled for the relatively higher safety risk of the nighttime period. Through the simulation, the system ultimately generated a risk evolution path map presented as a superposition of a Gantt chart and a risk heatmap. This map visually displays the key nodes of risk transmission between different stages and their probability of occurrence. The specific steps are as follows: Step 1, simulation sequence data acquisition and feature extraction: After completing a forward-looking hybrid simulation in the digital mapping module, the engine recorded each discrete time step within the entire simulation cycle. The scene state snapshot is used to extract two types of features for each time step: task time sequence features, including the set of ongoing construction tasks and their progress percentages, for subsequent Gantt chart generation; and scene risk features, which are encoded into a multi-dimensional risk feature vector by a lightweight risk encoder network, which encodes the physical state parameters of the current time step, such as equipment stress values, personnel density, environmental indicators, and agent alarm status. Step two involves constructing a spatiotemporal causal graph and probabilistic reasoning. Using the overall layout topology of the construction scenario as the spatial graph structure, nodes represent key construction areas or equipment, and edges represent physical connectivity or operational logic relationships. The risk feature vector for each time step is then used. Values ​​are assigned to nodes at corresponding spatial locations, forming a spatiotemporal graph sequence. This sequence is then input into a pre-trained causal graph convolutional network. This network, through stacked causal spatiotemporal convolutional layers, learns the propagation pattern of risk features on the graph over time, under the constraint of prohibiting the transmission of future information to the past. The network output is the risk propagation probability distribution of each node at each time step. Step 3: Identify which upstream nodes the risk originates from and their contribution probability; Step 4: Generate an overlay map. The map generator receives two inputs: one is the task time-series data extracted from Step 1, used to draw a baseline Gantt chart, with time on the horizontal axis and task on the vertical axis; the other is the spatiotemporal node risk propagation probability obtained from Step 2. The generator first depends on Calculate the comprehensive risk intensity value of each node at each time step. The calculation formula is:

[0111] ;

[0112] in, Represents the current node. Represents its upstream neighbor nodes, The norm of the node's own risk feature vector. It is the self-risk contribution coefficient. yes The neighbor set of the generator, on the time task plane of the Gantt chart, is based on the spatial location of the task, and combines different spatial nodes at the same time. Values ​​are mapped to color depth and precisely overlaid on the spatiotemporal coordinates of the corresponding task strip in the form of a semi-transparent thermal layer. High-risk areas are covered with warm color depth, thus simultaneously displaying the origin, transmission path and intensity evolution of task progress plans and risks in spatiotemporal space on a single graph.

[0113] The decision optimization module of the adaptive response unit intervenes, trying various pre-intervention strategies in the digital mapping module based on the map. For example, strategy one involves advancing high-risk construction to the morning of the second day when winds are lower; strategy two involves immediately dispatching backup welders and installing windproof netting on the same day. The digital mapping module quickly simulates the implementation effects of each strategy, quantitatively assessing its impact on the overall project duration, safety risk entropy, and resource costs. Based on the simulation results, the decision optimization module outputs an optimized construction plan and risk control measure adjustment suggestions, and simultaneously updates the engineering safety strategy library and the contingency plan of the resource planning unit. When the actual construction progress approaches the simulated risk window, the system can activate corresponding early warnings and resource preparations in advance, proactively mitigating risks.

[0114] The system's learning effectiveness is visualized through a knowledge graph. Each risk type and its corresponding handling strategy forms a node, with edges between nodes representing the strength of the association and the frequency of application. Managers can retrieve handling experience for similar cases using the graph. The module also includes a knowledge transfer function, allowing effective strategies accumulated in one project to be adaptively modified and applied to new projects with similar characteristics, accelerating the development of safety management capabilities for new projects. The entire knowledge evolution process establishes a complete traceability chain, recording the data foundation and reasoning process for each strategy optimization, meeting the data reliability requirements of engineering quality management. The digital mapping module's hardware configuration adopts a distributed rendering architecture. The main server handles physical computation and data management tasks, while multiple rendering nodes process visualization output from different perspectives, supporting concurrent access from multiple terminals. The knowledge evolution module is deployed on a high-performance computing cluster, utilizing graphics processors to accelerate the training process of deep learning models. The system includes a periodic snapshot function to save model states at different points in time for easy performance comparison and rollback operations. Both modules adopt a modular design, exchanging data with other parts of the system through standard interfaces, ensuring functional independence and system scalability.

[0115] It should be noted that the power engineering construction safety management system also includes a human-computer interaction module. The operating interface is built using a 3D visualization engine based on WebGL technology. The engine supports smooth rendering of large construction scenes containing millions of polygons in ordinary web browsers. The interface layout adopts a configurable dashboard design. The main view area displays a real-world 3D model of the construction site. The model surface is rendered with different colors according to the real-time risk level. High-risk areas are displayed in dark red with a pulsed warning halo, while safe areas are displayed in light green. The right side of the interface is equipped with a collapsible function panel. The upper part of the panel displays a Gantt chart showing the handling progress of currently active risk events, the middle part displays a visual chart of resource allocation status, and the lower part provides tools for historical data query and report generation. Managers can rotate, zoom, and pan the 3D scene using multi-touch gestures. When clicking on a specific device or area, a detailed information card pops up. The card displays structured data such as device operating parameters, maintenance records, and related risk events in layers. The collaborative decision-making function of the human-computer interaction module supports real-time collaboration among multiple users. When the system identifies a major risk event, the handheld terminals of relevant managers will receive a meeting invitation simultaneously. After entering the virtual decision-making room, the perspective operations and annotation actions of each user will be displayed synchronously in real time. During collaboration, users can use 3D annotation tools to draw evacuation routes, mark key areas of concern, or add text annotations in a virtual scene. All annotation data includes timestamps and creator information. The system automatically records voice-to-text transcripts and key decision points during discussions. The historical data backtracking function provides multi-dimensional drill-down analysis capabilities. Users can filter historical events by combining multiple conditions such as time range, risk type, and responsible unit. The system generates interactive time-series charts to show the changing trends of risk occurrence frequency and response effectiveness. The report generation engine adopts a template-based design, with preset standard templates such as safety inspection reports, accident analysis reports, and monthly safety assessments. After the user selects a time range, the system automatically fills in the data and generates printable documents.

[0116] The data fusion module adopts an enterprise service bus architecture to achieve data integration with external systems. The bus is configured with multiple adapters to support different integration methods such as direct database connection, web service calls, and file transfer. Integration with the engineering management platform is achieved through a two-way data synchronization mechanism. The safety management system receives construction progress plans and work arrangement information, and simultaneously feeds back identified risk events and handling suggestions to the engineering management system. Integration with the personnel information system uses a real-time interface call method. When it is necessary to obtain construction personnel qualification information, the system queries the personnel database through the LDAP protocol to verify whether the operator is qualified to handle specific risks. Integration with the equipment management system achieves asynchronous data exchange through message queues. The safety management system subscribes to equipment status change messages and simultaneously publishes equipment control commands and maintenance suggestions. The data consistency guarantee mechanism of the data fusion module adopts a distributed transaction processing mode. Critical business operations, such as the issuance of risk handling commands, use a two-phase commit protocol to ensure that commands either take effect simultaneously in all relevant systems or are completely rolled back. The data lineage analysis function records the source system, transformation process, and usage trajectory of each data item, forming a complete data flow map. When data anomalies are detected, the source of the problem can be quickly located. The module also implements service level monitoring, which monitors the connection status and data exchange quality with various external systems in real time. When an interface abnormality is detected, it automatically triggers an alarm and starts a backup communication channel.

[0117] The performance optimization of the human-computer interaction module adopts a hierarchical data loading strategy. Initially, only the basic scene framework and high-risk area data are loaded. As the user's viewpoint approaches a specific area, the detailed model and real-time data for that area are dynamically loaded. Interface response speed is evaluated using the following optimization formula:

[0118] ;

[0119] in: This indicates the total response latency of the interface operation. This represents the number of subtasks contained in an interactive task. It is the weight coefficient of the i-th subtask in the user experience. Corresponding to the actual time spent on this sub-task, This represents the inherent delay constant of network transmission. Weighting coefficient User research determined that, for example, view rotation has a higher weight than clicking on UI elements because the former is more sensitive to the smoothness of the interaction. The system monitors UI performance metrics based on this formula, and automatically initiates optimization measures when the total response latency exceeds a threshold, such as reducing the rendering precision of non-focus areas or preloading data from potentially accessible adjacent areas.

[0120] The data fusion module employs a role-based access mechanism for security control. Each interface call requires verification of digital certificates and permission tags, and administrators at different levels are granted differentiated data access permissions. The operation audit function records detailed logs of all cross-system data access, including request time, requester identity, accessed data range, and processing results. These logs are digitally signed and stored in a secure storage area to meet compliance audit requirements. The module also implements a service degradation strategy, automatically switching to local cached data to provide services when an external system service interruption is detected, ensuring the continuous operation of core security functions. During implementation, the human-computer interaction module focuses on continuous improvement of user experience, regularly collecting user operation habit data to optimize the interface layout and adjusting the priority of function entry points based on user feedback. The data fusion module establishes an interface change management process, updating adapter configurations promptly when external systems are upgraded to ensure integration stability. Both modules adopt containerized deployment, supporting rapid expansion and failover. Automated testing and deployment are achieved through continuous integration pipelines, ensuring the efficiency and reliability of system updates. A complete user training system was established throughout the implementation process, creating specialized operation manuals and instructional videos for different roles, and using simulated operating environments to help users quickly master system usage.

[0121] Before key processes such as the hoisting of large main transformers, the construction manager submits a work plan through the human-computer interaction module. The system automatically calls the digital mapping module and, based on the plan, the equipment's 3D model, the day's weather forecast, and site cloud data, constructs a high-fidelity virtual scenario of the entire process from crane entry, outriggers, lifting, slewing, positioning to hook release and removal. When the simulation begins, the system uses a simulation engine based on multibody dynamics and discrete event coupling to accelerate the simulation of the entire work process. An independent safety rule reasoning engine works simultaneously, which is loaded with a digital safety red line rule library specifically defined for this hoisting, such as the distance between the boom and the surrounding 110kV overhead lines must not be less than 8.5 meters at any simulation time, no personnel models are allowed to enter the simulated area under the hoisted object, and the outrigger reaction force must not exceed the foundation bearing capacity design value, etc. If the virtual simulation completes successfully and all red-line rules are not triggered, the adaptive response unit will issue an encrypted digital work permit. This permit specifies the unique code for this lifting operation, the allowed operation time window, the crane number, maximum wind speed, and slewing angle limits, and is then distributed to the designated crane, relevant personnel, and positioning equipment on site. The equipment control unit of the instruction execution resource scheduling unit will only unlock the crane's relevant operating permissions after verifying this permit on the crane controller.

[0122] In actual operation, the system continuously compares the real-time distance between the boom and the power line obtained by the high-precision positioning module, the personnel intrusion status identified by the panoramic vision unit, and the boundary conditions in the digital operation permit. Once the real-time distance approaches the 8.5-meter red line or someone enters the restricted area, the adaptive response unit immediately sends a red line violation signal to the equipment control module. The preset safety override loop in the equipment control module will be executed unconditionally, directly triggering the crane's emergency stop system through a hard-wired connection, cutting off the main lifting power, and issuing the highest-level audible and visual alarm from the personnel guidance module, thereby forcibly interrupting the risky operation. The operation can only resume after the risk has been eliminated and reassessed.

[0123] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A safety management system for electrical power engineering construction, characterized in that include: Real-time risk assessment and scheduling unit, forward-looking human factor risk perception unit, and risk simulation and mapping unit; The real-time risk assessment and scheduling unit is used to generate a safety status feature map based on heterogeneous safety data at the construction site, and to identify risk events at the construction site through the safety status feature map, so as to generate intervention instructions and resource scheduling schemes to optimize the construction site. The forward-looking human risk perception unit is connected and interacts with the real-time risk assessment and scheduling unit. It is used to assess the dynamic coupling risk entropy value of the construction personnel when performing power engineering construction based on the physiological state vector of the construction personnel, and send the dynamic coupling risk entropy value to the real-time risk assessment and scheduling unit for state deviation warning. The risk simulation mapping unit is connected and interacts with the real-time risk assessment and scheduling unit. It is used to build a real-scene model based on the point cloud data of the power engineering construction site, and generate parameter constraints in combination with the construction plan and equipment configuration management to determine the risk intensity value, so as to output the optimized power engineering construction plan and risk control measures to the real-time risk assessment and scheduling unit. Safety management of power engineering construction sites is achieved by combining the real-time risk assessment and scheduling unit, the forward-looking human factor risk perception unit, and the risk simulation and mapping unit. The real-time risk assessment and scheduling unit includes: A multimodal sensing unit is used to collect heterogeneous safety data at the construction site of power engineering projects. The heterogeneous safety data includes personnel biometrics, equipment operation logs, environmental indicators, and operation video streams. Distributed computing units are used to normalize and extract features from heterogeneous security data to generate security status feature maps. The dynamic risk assessment unit is used to take the safety status feature map as input, use deep spatiotemporal networks to identify abnormal patterns at the power engineering construction site, and determine risk event descriptors. The adaptive response unit is used to match risk event descriptors with the power engineering safety strategy library to generate graded intervention instructions and resource allocation plans. The instruction execution resource scheduling unit is used to implement scheduling processing on the power engineering construction site according to the hierarchical intervention instructions and resource allocation plan, and to conduct early warning processing on the power engineering construction site in combination with the output results of the forward-looking human factor risk perception unit and the risk simulation mapping unit. The forward-looking human-cause risk perception unit includes: The physiological signal sensing module is used to collect near-infrared brain functional imaging signals, eye movement trajectory signals, skin conductance response signals and surface electromyography signals of construction workers during power engineering construction, so as to characterize the physiological state vector of construction workers. The behavioral semantic analysis module is used to parse the operation video stream, identify the standard of the construction workers' operation actions, the area of ​​their line of sight, and the spatial interaction between the construction workers and the hazard source, and generate behavioral semantic vectors. The dynamic coupling risk entropy calculation module is used to receive physiological state vectors, behavioral semantic vectors and environmental indicators, and establish a nonlinear mapping model to evaluate the dynamic coupling risk entropy value of construction personnel when performing power engineering construction, so as to generate early warning of human-caused state deviation and send it to the instruction execution resource scheduling unit. The risk simulation mapping unit includes: The digital mapping module is used to construct a 3D real-scene model of the power engineering construction site based on point cloud data, and generate a benchmark virtual scene by combining physical rules and behavioral rules in a hybrid modeling technology to determine the risk intensity value. The data fusion module is used to acquire construction schedule plans and construction personnel and equipment management configurations as multi-source external data. After formatting the multi-source external data, it extracts action points and state change points. It converts the action points and state change points into event streams that can be recognized by the benchmark virtual scene, and extracts the parameterized constraints required for the benchmark virtual scene derivation from the multi-source external data. It adds the event streams and parameterized constraints to the benchmark virtual scene in chronological order. The knowledge evolution module is used to match the risk intensity value output by the benchmark virtual scenario with the disposal strategy, verify the performance evaluation of the disposal strategy in different scenarios, generate a risk evolution path map, and output the optimized construction plan and risk control measure adjustment suggestions to the instruction execution resource scheduling unit.

2. The power engineering construction safety management system according to claim 1, characterized in that, The dynamic risk assessment unit includes: The pattern learning module is used to learn the distinguishing boundary between normal and abnormal working modes in power engineering construction from the feature map of the safe state using an adversarial generative network. The probability inference module is used to analyze the matching probability between the power engineering construction site and the risk mode using a Bayesian neural network, with the boundary between the normal working mode and the abnormal working mode as a reference, and to generate a risk probability distribution. The propagation analysis module is used to invoke a time-series inference network based on the risk probability distribution to predict the propagation path and impact range of risk events at power engineering construction sites, and to determine the propagation analysis results. The descriptor generation module is used to generate risk event descriptors that include risk category, intensity index and spatiotemporal coordinates based on the risk probability distribution and propagation analysis results.

3. The power engineering construction safety management system according to claim 2, characterized in that, The adaptive response unit includes: The strategy matching module is used to perform syntax parsing on risk event descriptors, extract target fields, and map and verify the target fields with the emergency plan execution concepts in the power engineering construction safety strategy library to determine the risk emergency strategy. The resource planning module is used to calculate the type, quantity, and response time requirements of emergency resources needed in the risk emergency strategy based on the risk event descriptor, and generate an emergency plan that includes resource allocation routes and resource arrival times. The decision optimization module is used to establish optimization rules with the objectives of risk mitigation effectiveness, resource usage costs, and construction progress impact, and to find technical output optimization decision schemes by combining emergency plans and non-dominated solution sets; The instruction synthesis module is used to convert optimization decision schemes into textual instruction sequences, and to parse the textual instruction sequences using a knowledge parsing model to generate hierarchical intervention instruction frames that conform to communication protocol specifications.

4. The power engineering construction safety management system according to claim 1, characterized in that, The behavioral semantic analysis module includes: The personnel action state recognition module is used to extract the limb points of construction workers in the operation video stream through image extraction technology, and calculate the arm extension angle and body center of gravity offset of the construction workers based on the limb points to quantify the standard of the construction workers' operation actions. The personnel position and status recognition module is used to locate the hazard source and the line of sight of the construction personnel in the operation video stream, and quantify the Euclidean distance between the construction personnel and the hazard source based on the angle between the line of sight of the construction personnel and the hazard source, and determine the line of sight attention area and spatial interaction relationship. The behavior vector fusion and determination module is used to fuse the standardization of work actions, the area of ​​visual attention, and spatial interaction relationships as behavior semantic vectors. The extraction of limb points of construction workers from the work video stream using image extraction technology includes: The video stream of the operation is processed by frame segmentation to generate clear frames of construction workers' actions and frames of construction workers' actions that are occluded. Deep learning technology is then used to analyze the clear frames of construction workers' actions and identify clear limb points of the construction workers. Based on digital refocusing technology and energy gradient function to set classification threshold, and combined with element image restoration technology, the occluded frames of construction workers' actions are repaired and the target objects are reconstructed. Based on the processing results, clear frames of construction workers' actions are output. Deep learning technology is used to analyze clear frames of construction workers' movements to identify the repaired limb points of the construction workers, and combine them with clear limb points to output the limb points of the construction workers in the operation video stream.

5. The power engineering construction safety management system according to claim 4, characterized in that, The process involves setting a classification threshold based on digital refocusing technology and an energy gradient function, and combining elemental image inpainting technology to repair and reconstruct target objects from occluded frames of construction worker actions. The resulting output includes clear frames of repaired construction worker actions. Based on the occlusion frames of construction workers' actions, the light intensity and direction information of different viewpoints in the construction worker's action scene are determined, the element image array under the corresponding viewpoint is obtained, and the element image array is digitally refocused along the axial depth. Based on the digital focusing processing results, image sequences with sharp focus and blurred defocus along the depth distribution are obtained, and the sum of gradient values ​​of all adjacent points of all pixels in the image sequence is analyzed to determine the focusing evaluation result of the image sequence at the corresponding axial depth. After the evaluation results meet the preset standard value, the image sequence is subjected to binary encoding transformation and cost aggregation to determine the classification threshold, and the foreground occlusion area in the construction worker action occlusion frame is removed based on the classification threshold. After removing the occlusion area of ​​the construction worker's action occlusion frame, select the clear construction worker action frame that is in the same row as the construction worker's action occlusion frame as the sample image, and calculate the disparity value between the sample image and the adjacent elements in the construction worker's action occlusion frame. The sample image is shifted according to the corresponding disparity value, and the pixel value at the corresponding position is extracted and fused into the foreground occlusion area of ​​the construction worker's action occlusion frame to obtain the repaired construction worker's action occlusion frame. The repaired construction worker's action occlusion frame is then subjected to three-dimensional reconstruction processing to output the clear frame of the repaired construction worker's action.

6. The power engineering construction safety management system according to claim 5, characterized in that, After the focus evaluation result meets the preset standard value, the process of performing binary encoding transformation and cost aggregation on the image sequence to determine the classification threshold, and removing the foreground occlusion region in the construction worker action occlusion frame based on the classification threshold includes: A transformation window is selected according to the target size. The gray values ​​of the pixels within the transformation window are compared with those of the center pixel to generate a binary code stream. The binary code stream is then used to perform binary encoding transformation on the image sequence to complete the cost matching processing of the image sequence. After cost matching, the pixel intensity of the image sequence is linearly interpolated to analyze the minimum disparity of each pixel in the image sequence. An additional smoothing constraint is added to construct an energy function. Cost aggregation of the image sequence is completed by applying different penalty terms to adjacent disparity changes. Obtain the disparity value of each pixel in the image sequence when it is aggregated at the minimum cost, and after left-right consistency check, disparity filling and filtering, obtain the disparity map of each element image array; Based on the conversion relationship between image sequence and disparity map, the classification thresholds for construction worker targets and occluders are determined. Based on the imaging principle that the disparity value of occluders is greater than that of construction worker targets, the foreground occluded areas in the element image array with disparity values ​​greater than the classification threshold are marked to complete the classification of occluders and construction worker targets, so as to remove the foreground occluded areas in the construction worker action occlusion frame.

7. The power engineering construction safety management system according to claim 1, characterized in that, The process of constructing a 3D reality model of the power engineering construction site based on point cloud data, and generating a benchmark virtual scene by combining physical rules and behavioral rules in a hybrid modeling technique, includes: Point cloud data of the power engineering construction site is obtained by using a laser scanning station, and after registration and fusion of the point cloud data, a unified three-dimensional real scene is generated as a three-dimensional real scene model of the power engineering construction site. Based on the construction equipment in the process of power engineering construction, the motion and behavior of the construction equipment under the action of external forces are analyzed by multibody dynamics and mechanics of materials, and added as physical rules to the three-dimensional real scene model. The construction workers in the power engineering construction process are regarded as intelligent agents. The collaborative game technology is used to assign behavioral states to the intelligent agents, and the output of the behavioral states is converted into a control rule base. The resulting behavioral rules are added to the three-dimensional real scene model. Based on the 3D real-world model with added physical and behavioral rules, and combined with current construction technology, a benchmark virtual scene of the power engineering construction site within the target future time period is constructed.

8. The power engineering construction safety management system according to claim 7, characterized in that, The step of assigning behavioral states to the agent using collaborative game theory and converting the output of the behavioral states into a control rule base includes: Each agent is configured with a working state, and the agents involved in the future power engineering construction plan are modeled as agents with state stochasticity; the interaction between the agent configuration after the working state is configured and the agents with state stochasticity during power engineering construction is simulated based on a 3D real scene model. Based on the interaction state analysis, the state transition process of the intelligent agent during operation is analyzed, and the fuzziness and randomness of the 3D real scene model in the simulation process are evaluated based on the state transition process, and a standard evaluation set of the intelligent agent is defined. Based on a standard evaluation set, agents are used as game participants. With the goals of work safety and efficiency, a multi-agent dynamic game function is constructed to analyze the equilibrium strategies of agents under different working conditions and different scenario events. The equilibrium strategies are then transformed into cloud droplet distribution features. Based on the cloud droplet distribution features, the control rules of agents in the construction process and state transition process are encoded as a behavior control rule library.

Images