A power operation safety intelligent early warning method and system based on 5G communication and artificial intelligence

Abstract

This invention belongs to the field of artificial intelligence, specifically a method and system for intelligent early warning of power operation safety based on 5G communication and artificial intelligence. It includes real-time collection of worker location and movement data via a 5G communication sensing network, preprocessing and trajectory feature extraction. The system comprises a 5G communication sensing network module, an edge computing processing module, a lightweight trajectory prediction module, a risk assessment and early warning module, and a linkage execution module. This application achieves end-to-end millisecond-level latency, meeting the real-time requirements of high-voltage live-line operation scenarios. It improves the prediction accuracy of collaborative operation scenarios by modeling the interaction relationships of multiple workers using graph attention networks, controlling the model parameter size to within 5MB and the inference latency to within 10ms to achieve real-time edge deployment, thus constructing a proactive intelligent safety defense closed loop.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically a method and system for intelligent early warning of power operation safety based on 5G communication and artificial intelligence. Background Technology

[0002] Safety management and control of power operations is a core guarantee for safe production in the power industry. With the continuous expansion of the power grid and the increasing complexity of on-site operations, the traditional manual passive monitoring mode is difficult to meet the real-time and high-reliability early warning requirements of complex operation scenarios. Integrating 5G communication and artificial intelligence algorithms and building an active intelligent early warning system based on personnel trajectory prediction is an important development direction for power operation safety management and control technology.

[0003] Existing power operation safety early warning solutions mainly use cloud-deployed trajectory prediction models to model and analyze the movement trajectory of workers, and rely on cloud computing power to complete trajectory reasoning and risk analysis, which can realize basic safety early warning functions for routine and simple operation scenarios.

[0004] Existing trajectory prediction models are complex in structure and computationally intensive, making them unsuitable for deployment at the edge. The high latency of cloud inference links makes it difficult to meet the millisecond-level early warning requirements for high-risk operations such as high-voltage live-line work. Furthermore, they do not consider the trajectory coupling and interaction relationships of multi-person collaborative operations, which can easily lead to prediction bias and low early warning accuracy. They suffer from dual technical shortcomings in real-time performance and accuracy, and cannot meet the high-precision proactive safety early warning needs of complex power operation scenarios.

[0005] Therefore, this invention provides a method and system for intelligent early warning of power operation safety based on 5G communication and artificial intelligence. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows: A power operation safety intelligent early warning method based on 5G communication and artificial intelligence, comprising the following specific steps:

[0008] S1: Real-time collection of location and movement data of workers through a 5G communication sensing network deployed at the power operation site;

[0009] S2: Preprocess and extract trajectory features from the collected location and motion data to build a trajectory feature library for the collaborative movement of multiple workers;

[0010] S3: Construct a lightweight interactive sensing trajectory prediction model based on the trajectory feature library, and deploy the lightweight interactive sensing trajectory prediction model on edge computing nodes;

[0011] S4: Utilize a lightweight interactive sensing trajectory prediction model deployed on edge computing nodes to predict the future movement trajectory of workers in real time, and conduct risk assessment based on the spatial relationship between the predicted trajectory and the preset safety boundary.

[0012] S5: When the predicted trajectory is about to enter the safety boundary, it automatically triggers multi-level early warning commands and performs defensive actions in conjunction with them;

[0013] S6: The warning instructions are sent to the on-site execution terminal with a latency of milliseconds through the 5G communication network, realizing end-to-end linkage between risk identification and defense response.

[0014] Preferably, in the step of collecting the location and movement data of workers in real time through a 5G communication sensing network deployed at the power operation site, the 5G communication sensing network includes ultra-densely deployed 5G base stations and distributed edge computing units. The 5G base stations are equipped with indoor positioning enhancement units, which achieve high-precision three-dimensional positioning of workers through time-of-arrival positioning technology and angle positioning technology. The positioning accuracy is better than 10 centimeters, and the positioning data update frequency is not less than 100 Hz. The distributed edge computing units are directly connected to the 5G base stations using a fronthaul network architecture to achieve local offloading processing of sensing data and reduce data transmission latency.

[0015] Preferably, in the steps of preprocessing and extracting trajectory features from the collected location data and motion state data, the preprocessing process includes performing Kalman filtering to denoise the original positioning data, removing abnormal drift points, and performing time synchronization calibration to ensure that the trajectory data of multiple targets are aligned under a unified time scale; the trajectory feature extraction process uses a sliding window mechanism to extract kinematic features such as the velocity vector, motion direction, displacement distance, and trajectory curvature of the workers within a preset time window, and simultaneously extracts interactive features such as the relative distance, relative speed, and motion trend similarity between multiple workers. The kinematic features and interactive features together constitute a multidimensional trajectory feature vector, providing a high-dimensional input representation for subsequent trajectory prediction.

[0016] Preferably, in the step of constructing a lightweight interactive perception trajectory prediction model based on the trajectory feature library, the lightweight interactive perception trajectory prediction model adopts an encoder-decoder architecture. The encoder part uses a graph attention network to explicitly model the interaction relationship between multiple workers, and the decoder part uses a long short-term memory network to generate multi-step prediction trajectories. The model lightweighting process includes replacing standard convolutional layers with depthwise separable convolutions to reduce the number of parameters, using knowledge distillation technology to transfer knowledge from large prediction models to small models, and using parameter quantization technology to compress 32-bit floating-point parameters to 8-bit fixed-point parameters. The number of parameters in the lightweighted model is no more than 5 megabytes, and the inference latency is no more than 10 milliseconds, meeting the real-time constraints of edge computing nodes.

[0017] Preferably, in the step of using a lightweight interactive sensing trajectory prediction model deployed on an edge computing node to predict the future movement trajectory of the worker in real time, the edge computing node is equipped with a hardware accelerator for neural networks, supporting parallel computing to meet the computational needs of simultaneous prediction of multiple targets; the prediction process takes the historical trajectory data of the most recent 3 seconds as input and outputs the predicted trajectory sequence within the next 5 seconds, with a prediction time step of 0.1 seconds; the risk assessment process generates a risk level score by calculating the area of ​​the overlapping region between the predicted trajectory and the preset safety boundary and the overlapping time window, and when the risk level score exceeds a preset threshold, it is determined to be a high-risk event.

[0018] Preferably, in the step of automatically triggering multi-level early warning instructions and linking defensive actions when the predicted trajectory is about to enter the safety boundary, the multi-level early warning mechanism includes three levels: yellow warning, orange warning, and red warning. A yellow warning corresponds to the predicted trajectory possibly touching the safety boundary within 10 seconds, an orange warning corresponds to the predicted trajectory possibly touching the safety boundary within 5 seconds, and a red warning corresponds to the predicted trajectory about to touch the safety boundary within 2 seconds. The linked defensive actions include triggering an audible and visual alarm to remind workers, linking the access control system to restrict workers from entering the danger zone, and linking a robotic arm or barrier to perform physical interception. The content of the early warning instruction includes the danger zone number, warning level, expected contact time, and recommended avoidance direction.

[0019] Preferably, in the step of sending the warning command to the on-site execution terminal via the 5G communication network with a millisecond-level latency, the 5G communication network is configured with network slicing technology to allocate dedicated network resource slices for the security warning service, ensuring priority transmission of the warning command in network congestion scenarios; the end-to-end latency includes edge computing node processing latency, core network transmission latency, and wireless air interface latency, with the total latency controlled within 20 milliseconds; the on-site execution terminal includes a smart safety helmet, wearable wristband, on-site audible and visual alarm, and automatic defense execution mechanism, with each terminal accessing the network through a 5G IoT gateway to receive and execute the warning command.

[0020] Preferably, in the encoder of the lightweight interactive perception trajectory prediction model, the graph attention network calculates the attention weight between each worker node and other worker nodes around it. The attention weight reflects the degree of coupling of the motion trajectories of multiple workers during collaborative operations. When multiple workers perform cooperative operations such as collaborative hoisting or equipment traction, the encoder can automatically capture the motion constraint relationship between workers and encode it into the prediction features, thereby improving the trajectory prediction accuracy in collaborative operation scenarios.

[0021] Preferably, during the risk assessment process, the shape of the safety boundary can be set to a circle, rectangle, or irregular polygon according to the degree of danger of the actual work area. The safety boundary supports dynamic scaling to adapt to the risk control needs of different work stages. The dynamic scaling rule of the safety boundary is linked to the working voltage level. The boundary radius of the high-voltage live work area is automatically expanded to 1.5 times that of the regular area to ensure sufficient safety margin in high-risk work scenarios.

[0022] Preferably, the multi-level early warning mechanism also supports an early warning escalation strategy. When a low-level early warning is triggered, the system continuously monitors the risk situation. If the degree of overlap between the predicted trajectory and the safety boundary does not improve within a preset time window, the early warning level is automatically upgraded. During the early warning escalation process, SMS and telephone notifications are sent to on-site safety management personnel to ensure that management personnel can keep abreast of risk dynamics in complex operating scenarios.

[0023] Preferably, the system further includes an adaptive update module for the prediction model. The adaptive update module corrects the parameters of the prediction model in reverse according to the actual occurrence of the warning event. When the deviation between the actual trajectory and the predicted trajectory exceeds the preset tolerance, the system automatically records the environmental context information for that period and updates the prediction model through incremental learning, so that the model can adapt to the movement habits of specific workers and the trajectory distribution characteristics under specific environments.

[0024] Preferably, the system further includes a three-dimensional digital twin module for the work area. The three-dimensional digital twin module uses multi-source sensing data transmitted back from the 5G network to construct a dynamic three-dimensional scene of the work site in real time. The spatial relationship between the predicted trajectory, safety boundary and warning area is superimposed and displayed in the three-dimensional digital twin scene, providing safety supervisors with an intuitive risk situation awareness interface that supports multi-angle rotation and zoom operations.

[0025] The present invention also provides an intelligent early warning system for power operation safety based on 5G communication and artificial intelligence. The system includes a 5G communication sensing network module, an edge computing processing module, a lightweight trajectory prediction module, a risk assessment and early warning module, and a linkage execution module.

[0026] Preferably, the 5G communication sensing network module is configured to build an ultra-low latency integrated communication and sensing network at the power operation site. This module includes a densely deployed 5G small base station cluster, a distributed edge computing unit, and an indoor positioning enhancement unit. All base stations and edge computing units are directly connected via fronthaul fiber optic cables to achieve local processing and offloading of sensing data. The 5G communication sensing network module provides the system with a high-bandwidth, low-latency data transmission channel and also undertakes the high-precision indoor positioning function for operators. Positioning data and video data are transmitted through independent network slices without interference, ensuring the reliable delivery of critical early warning data.

[0027] Preferably, the edge computing processing module is electrically connected to the 5G communication sensing network module. This module includes a multi-core central processing unit, a field-programmable gate array, and a dedicated neural network accelerator. Each processing unit is interconnected via a high-speed bus. The edge computing processing module is responsible for performing data preprocessing, feature extraction, and trajectory prediction inference tasks. All computing tasks are completed in a closed loop at the edge, without relying on cloud computing power. This module supports containerized deployment and can dynamically load or unload prediction model instances according to the scale of the operation, realizing elastic scaling of computing resources.

[0028] Preferably, the lightweight trajectory prediction module is integrated within the edge computing processing module. This module includes an interaction-aware encoding submodule and a trajectory decoding submodule. The interaction-aware encoding submodule uses a graph attention network architecture to encode features of a spatial relationship graph composed of multiple workers, outputting a trajectory feature representation containing interaction relationships. The trajectory decoding submodule uses a lightweight long short-term memory network to generate a multi-step predicted trajectory sequence based on the encoded features. The model parameters of the lightweight trajectory prediction module are deeply compressed, enabling millisecond-level inference on resource-constrained edge devices, meeting the computational requirements for real-time prediction.

[0029] Preferably, the risk assessment and early warning module is connected to the lightweight trajectory prediction module and is configured to perform spatial conflict detection between the predicted trajectory and the safety boundary and generate multi-level early warnings. The module internally maintains a safety boundary database and a safety level rule base for the work area, supporting various boundary types such as circles, rectangles, and irregular polygons. The risk assessment and early warning module calculates the overlapping area between the predicted trajectory and the safety boundary using geometric algorithms, and outputs a risk level score based on the overlapping area and the overlapping time window. When the risk level score reaches a preset threshold, the module automatically generates an early warning instruction of the corresponding level and sends it to the linkage execution terminal via the 5G network.

[0030] Preferably, the linkage execution module is communicatively connected to the risk assessment and early warning module. This module includes an early warning information distribution submodule and an execution equipment scheduling submodule. The early warning information distribution submodule formats and encapsulates the early warning instructions and pushes them to the field execution terminal via a 5G network, supporting multiple notification methods such as SMS, telephone, APP push, and audible and visual alarms. The execution equipment scheduling submodule intelligently schedules automatic defense execution mechanisms based on the early warning level and the location of the dangerous area, including triggering audible and visual alarms, linking access control locks, and activating physical interception devices. The linkage execution module has millisecond-level instruction forwarding capability, ensuring that the end-to-end latency from risk identification to the completion of defense actions meets the safety response requirements for high-voltage live-line work.

[0031] Preferably, the system further includes a predictive model adaptive update module, which is electrically connected to the lightweight trajectory prediction module. This module continuously monitors the deviation between the predicted trajectory and the actual trajectory. When the deviation exceeds a preset tolerance, it automatically triggers a model parameter update process. The update process includes collecting environmental context data during the deviation period, labeling and correcting samples, performing incremental learning training, and hot updating of model parameters to ensure that the predictive model can continuously adapt to the specific working environment and the movement characteristics of the workers.

[0032] Preferably, the system further includes a 3D digital twin display module, which establishes data connections with both the 5G communication sensing network module and the risk assessment and early warning module. This module receives multi-source sensing data transmitted from the site and renders a dynamic 3D scene of the work site in real time. The 3D scene overlays and displays the real-time location of the workers, predicted trajectory curves, safety boundary areas, and early warning information, allowing safety supervisors to intuitively grasp the overall risk situation through a visual interface. The 3D digital twin display module supports remote access, allowing safety supervisors to view the safety status of the work site in real time via a web or mobile device.

[0033] The beneficial effects of this invention are as follows:

[0034] 1. The present invention provides a method and system for intelligent early warning of power operation safety based on 5G communication and artificial intelligence. By deploying a lightweight trajectory prediction model on edge computing nodes and combining it with a 5G communication network to achieve millisecond-level end-to-end transmission, it completely solves the bottleneck of hundreds of milliseconds of latency caused by cloud inference in the prior art. This enables the early warning response speed to meet the stringent real-time requirements of high-risk scenarios such as high-voltage live-line work, and allows for proactive intervention before dangerous behavior actually occurs.

[0035] 2. The present invention provides a method and system for intelligent early warning of power operation safety based on 5G communication and artificial intelligence. By introducing a graph attention network to explicitly model the interaction relationship between multiple operators, it solves the problem that existing trajectory prediction technology only predicts a single target independently and ignores the motion coupling relationship in collaborative operation scenarios. This significantly improves the trajectory prediction accuracy in multi-person collaborative operation scenarios and reduces the risk of missed and false alarms caused by prediction deviations.

[0036] 3. The intelligent early warning method and system for power operation safety based on 5G communication and artificial intelligence described in this invention, through deep lightweight compression processing of the prediction model, including deep separable convolutional replacement, knowledge distillation and transfer, and parameter quantization, controls the model parameter size to within 5 megabytes and the inference latency to within 10 milliseconds, realizing the real-time deployment of complex prediction models on resource-constrained edge computing nodes, and providing a feasible technical path for edge intelligence.

[0037] 4. The intelligent early warning method and system for power operation safety based on 5G communication and artificial intelligence described in this invention realizes full-link automation from risk identification, early warning generation, instruction issuance to defense execution by constructing a multi-level early warning mechanism and a linkage defense system. This enables the prediction results to be directly linked with the physical safety boundary, driving the execution of automatic defense actions. It fundamentally changes the passive situation in the prior art where early warning and defense are separated, and constructs an active intelligent safety defense closed loop. Attached Figure Description

[0038] The invention will now be further described with reference to the accompanying drawings.

[0039] Figure 1 This is a structural block diagram of an intelligent early warning system for power operation safety based on 5G communication and artificial intelligence, as described in this invention.

[0040] Figure 2 This is a flowchart of a method for intelligent early warning of power operation safety based on 5G communication and artificial intelligence in this invention. Detailed Implementation

[0041] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0042] Example 1: This application provides an intelligent early warning system for power operation safety based on 5G communication and artificial intelligence. It includes a 5G communication sensing network module, an edge computing processing module, a lightweight trajectory prediction module, a risk assessment and early warning module, a coordinated execution module, a predictive model adaptive update module, and a 3D digital twin display module. The 5G communication sensing network module is used to collect real-time location and motion status data of the workers. The edge computing processing module is used to preprocess and extract trajectory features from the collected location and motion status data. The lightweight trajectory prediction module is used to construct a lightweight interactive sensing trajectory prediction model and predict the future movement trajectory of the workers in real time. The risk assessment and early warning module is used for risk assessment and multi-level early warning command generation. The coordinated execution module is used to send early warning commands to the on-site execution terminal and dispatch automatic defense execution mechanisms. The predictive model adaptive update module is used to correct the parameters of the prediction model in reverse according to the actual occurrence of the early warning event, achieving continuous model optimization. The 3D digital twin display module is used to receive multi-source sensing data transmitted from the site and render a dynamic 3D scene of the work site in real time, providing safety supervisors with an intuitive risk situation awareness interface.

[0043] like Figure 1 As shown, the intelligent early warning system for power operation safety constructed in this embodiment includes five core functional modules. The modules are tightly coupled through standardized data interfaces, forming a complete closed-loop link from sensing and data acquisition to coordinated execution.

[0044] The 5G communication sensing network module is located at the core of the power operation site. This module includes a densely deployed 5G small base station cluster, a distributed edge computing unit, and an indoor positioning enhancement unit. The densely deployed 5G small base station cluster consists of twelve 5G small base stations arranged in a rectangular grid with a three-meter interval, covering the entire operation area and ensuring seamless 5G signal coverage within the work space. The distributed edge computing unit is located near each 5G small base station and is directly connected to the corresponding small base station via fronthaul fiber to achieve local offloading and processing of sensing data. The indoor positioning enhancement unit is integrated inside each 5G small base station. This unit is equipped with a high-precision time difference of arrival (TDOA) measurement module and an angle measurement antenna array. It achieves high-precision three-dimensional positioning of workers through TDOA positioning technology and angle positioning technology. In this embodiment, the positioning accuracy is better than ten centimeters, and the positioning data update frequency reaches one hundred hertz, that is, the latest position coordinates are output every ten milliseconds. The 5G communication sensing network module is also equipped with an independent network slice. This slice allocates dedicated communication resources for safety early warning services, ensuring the priority and reliable delivery of early warning data during network transmission.

[0045] The edge computing processing module is directly connected to the 5G communication sensing network module via a 10 Gigabit Ethernet interface. This module includes a multi-core CPU, a field-programmable gate array (FPGA), and a dedicated neural network accelerator. The multi-core CPU uses an eight-core architecture with a clock speed of 2.5 GHz, responsible for performing general computing tasks such as data preprocessing, feature extraction, and task scheduling. The FPGA uses Xilinx KU115 series chips and is equipped with massively parallel logic units (MLLUs) for implementing customized calculations such as high-speed data filtering and coordinate transformation. The dedicated neural network accelerator integrates dedicated matrix operation units and tensor processing units, enabling extremely high-performance computations. The energy efficiency ratio enables the inference operations of the deep learning model. The processing units are interconnected via a high-speed bus with a bandwidth of no less than 256 gigabits per second, ensuring rapid data flow between the processing units. The edge computing processing module supports a containerized deployment environment built on the Kubernetes container orchestration framework, which can dynamically load or unload prediction model instances according to the current job scale, realizing elastic scaling of computing resources. In this embodiment, the edge computing processing module runs three lightweight trajectory prediction model instances simultaneously, corresponding to three different job partitions, to achieve comprehensive coverage of the entire job site.

[0046] The lightweight trajectory prediction module is integrated into the neural network accelerator of the edge computing processing module. This module includes an interaction-aware encoding submodule and a trajectory decoding submodule. The interaction-aware encoding submodule is implemented using a graph attention network architecture. It maintains a dynamically updated spatial relationship graph, where nodes correspond to individual workers, and edge weights correspond to motion correlation scores between workers. The graph attention network calculates the attention weight between each worker node and its surrounding worker nodes. The calculation of the attention weight considers three factors: relative distance, relative speed, and similarity of historical motion trajectories between workers. When multiple workers perform cooperative operations such as hoisting or equipment traction, the graph attention network can automatically capture the relationships between workers. The motion constraints are encoded into the prediction features. The trajectory decoding submodule is implemented using a lightweight long short-term memory network. This network contains two stacked long short-term memory units, each containing fifty hidden state nodes. It generates multi-step predicted trajectory sequences through stepwise prediction. The model parameters of the lightweight trajectory prediction module undergo three stages of lightweighting. In the first stage, depthwise separable convolutions are used to replace standard convolutional layers to reduce the number of parameters. In the second stage, knowledge distillation technology is used to transfer knowledge from large prediction models to small models. In the third stage, parameter quantization technology is used to compress 32-bit floating-point parameters into 8-bit fixed-point parameters. After lightweighting, the model parameter size is 4.8 megabytes and the inference latency is 8 milliseconds, which meets the real-time constraints of edge computing nodes.

[0047] The risk assessment and early warning module is connected to the output of the lightweight trajectory prediction module. Internally, this module maintains a safety boundary database and a safety level rule base for the work area. The safety boundary database stores the safety boundary parameters for each work area, including the boundary's geometric shape, center coordinates, and radius or side length. This database supports three boundary types: circular, rectangular, and irregular polygonal. The safety level rule base stores trigger rules for multi-level early warning mechanisms. A yellow warning corresponds to a predicted trajectory that may touch the safety boundary within ten seconds; an orange warning corresponds to a predicted trajectory that may touch the safety boundary within five seconds; and a red warning corresponds to a predicted trajectory that may touch the safety boundary within... Within two seconds of approaching the safety boundary, the risk assessment and early warning module uses geometric algorithms to calculate the overlap area between the predicted trajectory and the safety boundary. Based on the overlap area and overlap time window, it outputs a risk level score. When the risk level score reaches a preset threshold, the module automatically generates a corresponding level of early warning instruction. The early warning instruction includes the danger zone number, warning level, expected contact time, and recommended avoidance direction. The safety boundary supports dynamic scaling, and the scaling rule of this function is linked to the working voltage level. The boundary radius of the high-voltage live-line working area is automatically expanded to 1.5 times that of the regular area to ensure sufficient safety margin in high-risk working scenarios.

[0048] The linkage execution module and the risk assessment and early warning module establish a communication connection through a 5G network. This module includes an early warning information distribution submodule and an execution equipment scheduling submodule. The early warning information distribution submodule formats and encapsulates the early warning instructions and pushes them to the on-site execution terminal through the 5G network, supporting four notification methods: SMS, telephone, application push, and audible and visual alarms. The execution equipment scheduling submodule intelligently schedules automatic defense execution mechanisms according to the early warning level and the location of the dangerous area, including triggering audible and visual alarms to remind workers, linking the access control system to lock the entrances and exits of the corresponding area, and activating physical interception devices. In this embodiment, the physical interception device adopts a hydraulically driven high-speed baffle. The response time of the baffle rising from below the ground to the fully deployed position does not exceed 300 milliseconds. The linkage execution module has millisecond-level instruction forwarding capability, and the end-to-end latency from receiving the early warning instruction to completing the instruction issuance operation is controlled within 5 milliseconds.

[0049] The predictive model adaptive update module establishes a data feedback connection with the lightweight trajectory prediction module. This module continuously monitors the deviation between the predicted trajectory and the actual trajectory. The predictive model adaptive update module has a deviation threshold judge. When the deviation value exceeds the preset tolerance, the model parameter update process is automatically triggered. The update process includes four stages: collecting environmental context data for the deviation period, labeling correction samples, performing incremental learning training, and hot updating of model parameters. Environmental context data includes factors such as the type of work, number of personnel, weather conditions, and light intensity at that time. The labeling of correction samples is automatically completed by the labeling system. This system generates supervision signals based on the difference between the actual trajectory and the predicted trajectory during the deviation period. Incremental learning training is performed in the background of the edge computing processing module. A gradient-based online learning algorithm is used to gradually adjust the model parameters. The hot updating of model parameters adopts a seamless switching strategy. After the new model parameters are trained and verified in the background, they immediately replace the old model parameters without interrupting the real-time prediction function of the system.

[0050] The 3D digital twin display module establishes data connections with the 5G communication sensing network module and the risk assessment and early warning module. This module receives multi-source sensing data transmitted from the site and performs real-time rendering processing. The rendering engine of the 3D digital twin display module is built on the Unity3D engine and can render dynamic 3D scenes of the work site at a refresh rate of 60 frames per second. The rendered 3D scene overlays and displays the real-time position of the workers, predicted trajectory curves, safety boundary areas, and early warning information, allowing safety supervisors to intuitively grasp the overall risk situation through a visual interface. The 3D digital twin display module provides a remote access interface, allowing safety supervisors to view the safety status of the work site in real time through a web browser or mobile application. This interface uses the WebSocket protocol to achieve bidirectional real-time data transmission.

[0051] like Figure 2 As shown, the intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to the present invention includes the following specific steps:

[0052] Step 1: Data Acquisition and Transmission. The data acquisition module collects the location and movement data of the workers in real time through a 5G communication sensing network deployed at the work site. A super-dense network composed of twelve 5G small base stations continuously transmits positioning reference signals. The smart safety helmets worn by the workers have built-in positioning signal receiving modules and safety communication modules. After receiving positioning reference signals from at least three 5G small base stations, the smart safety helmet measures the time difference of arrival of each signal through the time difference of arrival measurement module and uploads the measurement results to the corresponding distributed edge computing unit through the safety communication module. The distributed edge computing unit calculates the three-dimensional coordinates of the workers based on the time difference of arrival and angle positioning technology, with a positioning accuracy better than ten centimeters. The update frequency of the positioning data is 100 Hz, that is, the system refreshes the position coordinates of all workers every ten milliseconds. The collected position data is transmitted to the edge computing processing module through the fronthaul fiber. At the same time, the original positioning reference signal measurement data is also backed up and stored in the form of original data packets for subsequent data verification and quality assessment.

[0053] Step Two: Data Preprocessing and Trajectory Feature Extraction. After receiving the location data stream from the 5G communication sensing network module, the multi-core central processing unit of the edge computing processing module performs data preprocessing. The preprocessing process includes three sequentially executed sub-steps. First, the raw positioning data is subjected to Kalman filtering for noise reduction. The state vector of the Kalman filter contains three dimensions: position coordinates, velocity, and acceleration. The filter performs optimal estimation and fusion based on the state estimate from the previous moment and the observation from the current moment, effectively removing random noise and abnormal drift points from the positioning data. In this embodiment, the process noise parameters of the Kalman filter are... The noise parameters are set to 0.01 meters per second squared, 0.5 meters per second cubicd, and 0.1 meters per observation. Next is outlier removal, which detects abnormal drift points by setting upper limits for velocity and acceleration. When an instantaneous velocity exceeding 8 meters per second or an instantaneous acceleration exceeding 3 meters per second squared is detected for a positioning data point, the data point is marked as an outlier and replaced with an interpolated result. Finally, there is time synchronization calibration, which aligns multi-target trajectory data under a unified time scale to ensure that positioning data from different operators have the same timestamp reference.

[0054] After preprocessing, the edge computing processing module performs trajectory feature extraction. The trajectory feature extraction process is implemented through a sliding window mechanism. The sliding window duration is set to three seconds, and the window sliding step size is set to 0.1 seconds. Within each sliding window, the system extracts two types of feature vectors. The first type is kinematic features, including the average velocity, average direction of movement, average displacement distance, and trajectory curvature of the workers within the window. The average velocity is calculated by removing the displacement within the window and dividing by the time interval. The average direction of movement is obtained by calculating the direction angle of the displacement vector using the arctangent function. The trajectory curvature is obtained by dividing the second derivative of the displacement vector by the magnitude of the first derivative of the velocity vector. The second type is interaction features, including the relative distance, relative velocity, and similarity of movement trends among multiple workers. The relative distance is calculated using the Euclidean distance formula. The relative velocity is obtained by the magnitude of the difference between the velocity vectors of two workers. The similarity of movement trends is obtained by calculating the cosine of the angle between the two velocity vectors. The kinematic features and interaction features are concatenated to form a 128-dimensional multidimensional trajectory feature vector, which serves as the standard input representation for the subsequent trajectory prediction model.

[0055] Step 3: Model Construction and Deployment. The edge computing processing module constructs a lightweight interactive perception trajectory prediction model based on the trajectory feature vectors extracted in Step 2. The model adopts an encoder-decoder architecture. The encoder part uses a graph attention network, and the decoder part uses a long short-term memory network. The input of the graph attention network is a spatial relationship graph composed of multiple workers. The node features of the graph are the trajectory feature vectors of each worker. The edge weights are dynamically calculated according to the distance between workers. The computation process of the graph attention network is divided into three stages. The first stage is feature projection, where the original node features are mapped to a high-dimensional feature space through a fully connected layer. The second stage is attention weight calculation, where for each node, the attention coefficient with all its neighboring nodes is calculated. The attention coefficient is obtained by fusing node features, edge weights, and position encoding through a multilayer perceptron. The third stage is feature aggregation, where the output feature of each node is the weighted sum of the features of all its neighboring nodes, and the weights are the corresponding attention coefficients. The input of the long short-term memory network is the high-level semantic features output by the encoder, and the output is the predicted trajectory sequence within the next five seconds. The long short-term memory network contains a two-layer structure, with fifty hidden units in each layer. The network generates prediction results step by step with a time step of 0.1 seconds.

[0056] The lightweight model processing includes three sequential compression steps. The first step is to replace the standard convolutional layer with depthwise separable convolution. Depthwise separable convolution decomposes the standard convolution into two steps: channel-wise convolution and pointwise convolution, which significantly reduces the number of parameters and computational complexity. The second step is to train a lightweight student model using knowledge distillation technology. The teacher model is a standard trajectory prediction model, and the student model is a lightweight model modified by depthwise separable convolution. The loss function of knowledge distillation integrates the cross-entropy loss of hard labels and the knowledge transfer loss of soft labels. The third step is to compress the 32-bit floating-point parameters into 8-bit fixed-point parameters using parameter quantization technology. The quantization process adopts a symmetric quantization scheme, using the maximum absolute value of the parameter as the quantization scale factor. The compressed model has 4.8 megabytes of parameters and an inference latency of 8 milliseconds, which can run in real time on the neural network accelerator of the edge computing processing module. The completed lightweight model is loaded into the edge computing processing module through containerized deployment. Each model instance occupies 30% of the computing resources and 2 gigabytes of memory resources of the neural network accelerator.

[0057] Step 4: Trajectory Prediction and Risk Assessment. The lightweight trajectory prediction module takes the historical trajectory data of the last three seconds as input and outputs the predicted trajectory sequence for the next five seconds. The specific prediction process is as follows: The interactive perception encoding submodule first constructs a spatial relationship graph at the current moment. The number of nodes in the graph is equal to the total number of workers in the current work area. The node features are the trajectory feature vector sequence of the past three seconds. The graph attention network encodes the spatial relationship graph and outputs a trajectory feature representation containing interactive relationships. The dimension of this feature representation is 256. After receiving the encoded features, the trajectory decoding submodule generates the predicted trajectory step by step through the long short-term memory network. The decoding process is executed in 50 steps with a step size of 0.1 seconds, and outputs the coordinate sequence of 50 predicted points, which constitutes the predicted trajectory for the next five seconds.

[0058] After receiving the predicted trajectory output by the lightweight trajectory prediction module, the risk assessment and early warning module performs spatial conflict detection and multi-level early warning generation operations. The specific process of spatial conflict detection is as follows: The risk assessment and early warning module reads the safety boundary parameters of the current work area from the safety boundary database. In this embodiment, the geometric shape of the safety boundary is circular, and the circular safety boundary is uniquely defined by the coordinates of the center and the radius. The risk assessment and early warning module calculates the distance from each point on the predicted trajectory to the center of the circle. When the distance from a predicted point to the center of the circle is less than the radius of the safety boundary, the point is determined to be in a dangerous area. The risk assessment and early warning module further calculates the area of ​​the overlapping region between the predicted trajectory and the safety boundary. The overlapping time window and the area calculation adopt a numerical integration method. The predicted trajectory is discretized into one hundred sampling points and then the overlapping area is accumulated. The time window is from the current time to the time when the first overlapping point of the predicted trajectory appears. The risk level score is calculated using a weighted comprehensive score method. The score formula is described as multiplying the overlapping area by the weight coefficient A and adding the overlapping time window by the weight coefficient B to obtain the comprehensive score. The weight coefficient A is set to ten and the weight coefficient B is set to five. When the comprehensive score exceeds the yellow warning threshold, a yellow warning instruction is generated. When the comprehensive score exceeds the orange warning threshold, an orange warning instruction is generated. When the comprehensive score exceeds the red warning threshold, a red warning instruction is generated.

[0059] Step 5: Early Warning Command Generation and Defense Action Linkage. When the risk assessment and early warning module determines that an early warning needs to be triggered, it automatically generates multi-level early warning commands of corresponding levels. The content of the yellow early warning command includes: the current danger zone number is Zone A, the warning level is yellow, the estimated contact time is 12 seconds, and the recommended avoidance direction is to move 3 meters to the left. The content of the orange early warning command includes: the danger zone number is Zone A, the warning level is orange, the estimated contact time is 6 seconds, and the recommended avoidance direction is to immediately retreat to the safety line. The content of the red early warning command includes: the danger zone number is Zone A, the warning level is red, the estimated contact time is 1.5 seconds, and the recommended avoidance direction is to remain still and wait for physical interception.

[0060] After receiving the warning command, the linkage execution module performs linkage defense actions. For a yellow warning, the linkage execution module pushes an alarm message to the smart safety helmet worn by the worker through the warning information distribution submodule. The smart safety helmet emits a voice prompt "Pay attention to safety, please stay away from the danger area" through its built-in speaker. At the same time, the warning light on the top of the safety helmet flashes yellow. For an orange warning, in addition to pushing an alarm message to the smart safety helmet, the linkage execution module also links the access control system to lock the entrance and exit of the danger area to prevent other workers from accidentally entering the danger area. At the same time, the sound and light alarm installed on site is activated in orange alarm mode. For a red warning, the linkage execution module immediately triggers the physical interception device. The hydraulically driven high-speed baffle rises from below the ground. It takes 280 milliseconds for the baffle to rise to the fully deployed position. Once fully deployed, the baffle can effectively prevent workers from entering the danger area. At the same time, the linkage execution module sends SMS and telephone notifications to the on-site safety management personnel to ensure that the management personnel can keep abreast of the risk dynamics in complex work scenarios.

[0061] Step Six: Issuance of Warning Commands and End-to-End Linkage. The linkage execution module issues warning commands to the on-site execution terminal via the 5G communication network with a millisecond-level latency. The 5G communication network is configured with a dedicated network slice for safety warning services. This network slice reserves dedicated wireless resource blocks and dedicated core network transmission channels to ensure that warning commands can still be transmitted with priority even in network congestion scenarios. The end-to-end latency consists of three parts: the first part is the edge computing node processing latency, which is the processing time from the risk assessment and warning module generating the warning command to the linkage execution module receiving the command, and this latency is two milliseconds; the second part is the core network transmission latency, which is the transmission time from the edge computing node through the core network to the 5G base station, and this latency is three milliseconds; the third part is the wireless air interface latency, which is the transmission time from the 5G base station through the wireless air interface to the on-site execution terminal, and this latency is five milliseconds. The sum of the three parts of the latency results in a total end-to-end latency of ten milliseconds, which meets the requirements for safety response timeliness in high-risk scenarios such as high-voltage live-line work.

[0062] In this embodiment, the on-site execution terminals include four types of terminals: smart safety helmets, wearable wristbands, on-site audible and visual alarms, and automatic defense execution mechanisms. The smart safety helmets and wearable wristbands are connected to the network through a 5G IoT gateway, receive warning commands from the linkage execution module, and execute corresponding reminder actions. The on-site audible and visual alarms are directly connected to the 5G small base station via a wired connection, and trigger audible and visual alarms after receiving warning commands. The automatic defense execution mechanisms include the aforementioned hydraulically driven high-speed baffle and access control actuator, which are also connected to the 5G small base station via a wired connection. After receiving warning commands, they perform physical interception or access control locking operations. After receiving warning commands, all types of on-site execution terminals can complete command parsing and action execution within fifteen milliseconds, ensuring that the end-to-end latency from risk identification to the completion of defense actions does not exceed twenty milliseconds.

[0063] The intelligent early warning system for power operation safety constructed in this embodiment achieves millisecond-level end-to-end transmission by deploying a lightweight trajectory prediction model on edge computing nodes and combining it with a 5G communication network. This enables the early warning response speed to meet the stringent real-time requirements of high-risk scenarios such as high-voltage live-line work. The system introduces a graph attention network to explicitly model the interaction relationships between multiple operators, solving the problem of capturing motion coupling relationships in collaborative operation scenarios and significantly improving the trajectory prediction accuracy in multi-person collaborative operation scenarios. Through deep lightweight compression processing, the system controls the model parameter size to within 4.8 megabytes and the inference latency to within 8 milliseconds, enabling the real-time deployment of complex prediction models on resource-constrained edge computing nodes. By constructing a multi-level early warning mechanism and a linked defense system, the system achieves full-link automation from risk identification, early warning generation, instruction issuance to defense execution, thus constructing a proactive intelligent safety defense closed loop.

[0064] Example 2: This example is applied to the equipment maintenance operation scenario of a power substation. The operation area in this scenario is divided into multiple functional blocks. The safety boundary settings between different blocks are significantly different, and the operators need to move frequently between different blocks, which puts forward higher requirements for the dynamic adaptability of the early warning system.

[0065] This application provides a power operation safety intelligent early warning system based on 5G communication and artificial intelligence, which is applied to substation equipment maintenance scenarios. It includes a 5G communication sensing network module, an edge computing processing module, a lightweight trajectory prediction module, a risk assessment and early warning module, a linkage execution module, a prediction model adaptive update module, and a three-dimensional digital twin display module. The functional definitions of each module are consistent with those in Embodiment 1.

[0066] In this embodiment, the configuration of the 5G communication sensing network module has been specifically optimized. The ultra-dense 5G small base station cluster consists of eighteen 5G small base stations. These small base stations are deployed differently according to the equipment layout within the substation. They are densely deployed at two-meter intervals in the high-voltage equipment area to ensure positioning accuracy, and sparsely deployed at five-meter intervals in low-risk areas such as the central control room to reduce deployment costs. The number of distributed edge computing units has been increased to six. Each distributed edge computing unit manages the data access and data preprocessing tasks of three 5G small base stations. The indoor positioning enhancement unit has added a multipath signal recognition algorithm. This algorithm analyzes the multipath component characteristics of the positioning reference signal, identifies and eliminates non-line-of-sight signals caused by reflections from metal equipment, and significantly improves the positioning accuracy in the high-voltage equipment area. In this embodiment, the positioning accuracy in the high-voltage equipment area is better than eight centimeters, and the positioning accuracy in areas such as the central control room is better than fifteen centimeters.

[0067] In this embodiment, the edge computing processing module adopts a distributed deployment architecture. Six distributed edge computing units are deployed at the core of each work zone. The units are interconnected through a fiber optic ring network to form redundancy. When a distributed edge computing unit fails, its computing tasks can be automatically switched to an adjacent backup unit within 50 milliseconds to ensure the continuous operation of the system. The edge computing processing module supports dynamic resource scheduling. This function dynamically adjusts the resource quota of the prediction model instance according to the real-time load of each work zone. During peak maintenance periods, more computing resources are allocated to the high-voltage equipment area, and idle resources are used for model pre-training and parameter optimization during work breaks.

[0068] In this embodiment, the lightweight trajectory prediction module has undergone adaptive training and optimization for the substation scenario. The node features of the graph attention network have been enhanced with equipment association features, which encode the distance and relative positional relationship between the operator and the surrounding live equipment. This enhances the prediction model's ability to understand equipment contact risks. The output sequence length of the trajectory decoding submodule has been extended from five seconds to eight seconds to adapt to the characteristics of slow operator movement and long warning response windows in the substation scenario. The lightweight model processing adopts the same compression scheme as in Embodiment 1. The compressed model parameter size is 4.7 megabytes, and the inference latency is 7 milliseconds.

[0069] In this embodiment, the risk assessment and early warning module is configured with a dynamic safety boundary management function. The shape of the safety boundary is set as an irregular polygon to adapt to the irregular equipment layout and energized area boundaries within the substation. The dynamic scaling rules of the safety boundary are linked to the working voltage level. When the working voltage level is 500 kV, the safety boundary radius automatically expands to 1.8 times that of the regular area; when the working voltage level is 220 kV, the safety boundary radius automatically expands to 1.5 times that of the regular area; and when the working voltage level is 110 kV, the safety boundary radius automatically expands to 1.3 times that of the regular area. The safety boundary database supports real-time updates. When the maintenance work plan changes, safety supervisors can dynamically adjust the safety boundary through the editing interface of the 3D digital twin display module. The adjusted safety boundary parameters are synchronized to the risk assessment and early warning module within ten seconds.

[0070] In this embodiment, the adaptive update module of the early warning model is configured with a scenario-based learning function. This module establishes independent model parameter fine-tuning strategies for different work scenarios. When the system switches to a new work scenario, the adaptive update module automatically loads the fine-tuning parameter set corresponding to the scenario, without having to start model training from scratch. The scenario-based learning function effectively shortens the model's adaptation cycle to the new work scenario. The trajectory prediction accuracy under the new scenario can reach a stable level within five minutes after the model is deployed. In addition, the adaptive update module is also configured with a personalized learning function for workers. This function adjusts the model parameters according to the movement habits of specific workers, enabling the prediction model to more accurately predict the movement intentions of the workers.

[0071] In this embodiment, the 3D digital twin display module has optimized 3D modeling for substation scenarios. The module pre-imports the substation's design drawings and equipment models to construct a high-precision 3D static scene base. During real-time rendering, the module overlays and displays the real-time location of the workers, predicted trajectory curves, safety boundary areas, and early warning information. The module also integrates equipment status monitoring functions, overlaying real-time current and voltage data of energized equipment onto the corresponding equipment models, providing safety supervisors with a more comprehensive operational situation awareness. The module supports multi-terminal synchronous access, allowing up to twenty clients to simultaneously view the same 3D scene of the work site online, with each client maintaining a refresh rate of over 45 frames per second.

[0072] In this embodiment, the linkage execution module expands the types and number of execution terminals for substation scenarios. The newly added execution terminals include transformer oil temperature monitoring terminals, switchgear status monitoring terminals, and grounding switch control terminals. These terminals are connected to the 5G network through industrial IoT gateways. Based on the warning level and the location of the dangerous area, the linkage execution module not only dispatches the original audible and visual alarms and access control systems, but also links the corresponding equipment monitoring terminals to perform protective operations. For example, when it detects that a worker is about to accidentally enter a live compartment, the linkage execution module, in addition to triggering an audible and visual alarm, also sends an emergency trip command to the switchgear control terminal of that compartment to cut off the compartment's power supply and eliminate the risk of electric shock.

[0073] Example 3: This example is applied to a scenario of collaborative operation of live-line working vehicles in power distribution networks. This scenario involves the coordinated cooperation of multiple live-line working vehicles and multiple ground workers. The working space covers a wide area and there are a large number of mobile devices, which puts forward higher requirements for the multi-target tracking and cross-device coordination capabilities of the early warning system.

[0074] This application provides an intelligent early warning system for power operation safety based on 5G communication and artificial intelligence, which is applied to the collaborative operation scenario of live-line working vehicles in power distribution networks. The system includes a 5G communication sensing network module, an edge computing processing module, a lightweight trajectory prediction module, a risk assessment and early warning module, a linkage execution module, a prediction model adaptive update module, and a three-dimensional digital twin display module.

[0075] In this embodiment, the 5G communication sensing network module adopts a mobile deployment scheme. The system is equipped with two mobile 5G base stations, which are installed on the operation command vehicle and can be quickly deployed according to changes in the location of the operation site. The two mobile 5G base stations maintain a distance of 150 meters between them, jointly covering the operation area of ​​the operation vehicle. The mobile 5G base stations are equipped with vehicle-mounted positioning enhancement units. These units receive Beidou satellite navigation signals and differential correction data from ground reference stations to achieve precise positioning of the mobile base stations, ensuring the accuracy of the base station coordinates. In addition, the operation vehicle itself is equipped with a vehicle-mounted positioning module, which outputs the real-time position and attitude information of the operation vehicle at a frequency of 10 Hz, with a positioning accuracy better than 20 centimeters.

[0076] In this embodiment, the edge computing processing module adopts a cloud-edge collaborative architecture. The edge side consists of two vehicle-mounted edge computing units, each managing the personnel data accessed by a mobile 5G base station. The cloud side consists of a computing cluster of a regional data center, which undertakes the tasks of trajectory correlation analysis and global risk assessment across work areas. The vehicle-mounted edge computing units and the cloud computing cluster are connected through a dedicated enterprise line of the 5G network, with a transmission bandwidth of 500 megabits per second and a transmission latency of 15 milliseconds. The cloud-edge collaborative architecture enables the system to achieve global situational awareness across work areas while maintaining local real-time response.

[0077] In this embodiment, the lightweight trajectory prediction module extends the multi-target collaborative prediction capability. The graph attention network of the interactive perception encoding submodule not only models the interaction relationship between operators, but also the interaction relationship between operators and the work vehicle. The work vehicle, as a special node in the graph, participates in the calculation of attention weights. The feature vector of the work vehicle node contains information such as the position, speed, boom extension angle, and working radius of the work vehicle. This information is collected and uploaded in real time by on-board sensors. When there is a motion constraint relationship between the operator and the work vehicle, such as when the operator needs to follow the boom of the work vehicle, the graph attention network can automatically capture this collaborative motion mode and encode it into the prediction features. The trajectory decoding submodule adds multimodal output capability. Each operator outputs three candidate prediction trajectories, corresponding to three possible motion intentions: normal operation, emergency avoidance, and collaborative cooperation, providing a more comprehensive decision-making basis for risk assessment.

[0078] In this embodiment, the risk assessment and early warning module adds a safety boundary management function for the operating area of ​​the work vehicle. The operating radius of the work vehicle is defined as a circular area with the center of the work vehicle as the center and the maximum extension length of the working arm as the radius. This area is dynamically updated according to the real-time extension angle of the working arm. The risk assessment and early warning module overlays and analyzes the operating area of ​​the work vehicle with the preset safety boundary. When the predicted trajectory overlaps with both the operating area of ​​the work vehicle and the preset safety boundary, it is judged as a high-priority risk event and the early warning level is automatically upgraded by one level. In addition, the risk assessment and early warning module is also equipped with a work vehicle collision warning function. This function calculates the overlap of the predicted trajectories between each work vehicle. When there is a collision risk between the predicted trajectories of two work vehicles, a dedicated work vehicle avoidance warning is triggered.

[0079] In this embodiment, the linkage execution module adds a direct linkage function with the work vehicle control system. The linkage execution module establishes a safe communication link with the work vehicle control system through the vehicle communication interface. When a red warning command is received, in addition to performing the conventional audible and visual alarm and access control locking operations, it also sends an emergency stop command to the work vehicle control system. After receiving the emergency stop command, the work vehicle control system immediately locks the braking system and cuts off the power output of the boom to ensure that the work vehicle enters a safe parking state. The linkage execution module also supports group coordination with multiple work vehicles. When a collision risk is detected between multiple work vehicles, the linkage execution module simultaneously sends an avoidance path planning command to the relevant work vehicles to guide each work vehicle to orderly evacuate the danger zone according to the preset avoidance route.

[0080] Comparative Example: Comparative Example 1 adopts a traditional cloud-based centralized trajectory prediction scheme. This scheme deploys the trajectory prediction model on a remote cloud server. The positioning data collected by the 5G base station is uploaded to the cloud server via the core network for inference calculation. The prediction result is then transmitted back to the field execution terminal via the core network. In this scheme, the end-to-end latency of a single trajectory prediction includes a data upload latency of 100 milliseconds, a cloud inference latency of 200 milliseconds, and a result transmission latency of 100 milliseconds, for a total latency of 400 milliseconds. Due to the influence of network congestion and server load fluctuations, the end-to-end latency fluctuates between 200 and 600 milliseconds, which cannot meet the real-time requirements of early warning response in high-voltage live-line working scenarios.

[0081] Comparative Example 2 uses a standalone trajectory prediction scheme. This scheme deploys a simplified trajectory prediction model on a general-purpose server on-site, without relying on a 5G communication network. The positioning data is transmitted to the server for processing via wired connection. In this scheme, due to the lack of support from an ultra-dense 5G base station network, the positioning accuracy is only 50 centimeters, which cannot meet the high-precision positioning requirements of the work site. In addition, the simplified model ignores the interaction between multiple workers and only predicts the trajectory independently for each worker, resulting in a prediction accuracy of only 72% in collaborative work scenarios.

[0082] Comparative Example 3 adopts a local model-free threshold triggering scheme. This scheme does not use machine learning models for trajectory prediction. It only determines whether to trigger an alarm based on the real-time distance between the worker and the safety boundary. When the distance between the worker and the safety boundary is less than the alarm threshold, an alarm is triggered immediately. Otherwise, no alarm is triggered. In this scheme, due to the lack of trajectory prediction capability, the system can only issue an alarm when the worker has approached or entered the safety boundary. The reserved safety margin time is less than two seconds, which cannot provide sufficient response time for the execution of linkage defense actions. The alarm effect is limited to passive reminders rather than active intervention.

[0083] Comparative Example 4 uses a standard convolutional neural network for trajectory prediction but without lightweight processing. This scheme deploys the standard trajectory prediction model on edge computing nodes, with a model parameter size of 120 megabytes and an inference latency of 80 milliseconds. Although the prediction accuracy of this scheme is comparable to that of this invention, the inference latency far exceeds the real-time constraints of edge computing nodes, making it unable to support continuous trajectory prediction updates at 100 Hz, and the timeliness of the prediction results cannot be guaranteed.

[0084] Table 1: Comparison Table of Embodiments and Comparative Examples of the Present Invention

[0085] Comparison indicators Example 1 Example 2 Example 3 Comparative Example 1 Comparative Example 2 Comparative Example 3 Comparative Example 4 End-to-end latency (milliseconds) 10 12 15 400 250 8 95 Positioning accuracy (cm) 10 8 20 10 50 10 10 Trajectory prediction accuracy (%) 94 96 93 91 72 not applicable 93 Model parameter count (megabytes) 4.8 4.7 4.9 85 12 0 120 Inference delay (milliseconds) 8 7 9 200 45 0 80 Multi-objective collaborative prediction support support support Not supported Not supported Not supported Not supported Dynamic security boundary support support Not supported Not supported support Not supported Not supported Physical interception linkage support support support Not supported Not supported Not supported Not supported

[0086] As shown in the comparison table, the three embodiments of the present invention significantly outperform the comparative schemes in core performance indicators such as end-to-end latency, positioning accuracy, trajectory prediction accuracy, model parameter count, and inference latency. The present invention achieves millisecond-level end-to-end early warning response by deploying a lightweight trajectory prediction model on edge computing nodes and combining it with a 5G communication network; it significantly improves the trajectory prediction accuracy in multi-target collaborative operation scenarios by introducing graph attention networks for interaction relationship modeling; it enables the large prediction model to run in real time on resource-constrained edge devices through deep lightweight compression processing; and it realizes an active intelligent security defense closed loop by constructing a multi-level early warning mechanism and a linkage defense system. The technical solution of the present invention can effectively solve the shortcomings of existing technologies in terms of real-time performance, accuracy, and intelligence, and provides reliable technical support for power operation safety early warning.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent early warning of power operation safety based on 5G communication and artificial intelligence, characterized in that, Includes the following steps: S1: Real-time collection of location and movement data of workers through a 5G communication sensing network deployed at the power operation site; S2: Preprocess and extract trajectory features from the collected location and motion data to build a trajectory feature library for the collaborative movement of multiple workers; S3: Construct a lightweight interactive sensing trajectory prediction model based on the trajectory feature library, and deploy the lightweight interactive sensing trajectory prediction model on edge computing nodes; S4: Utilize a lightweight interactive sensing trajectory prediction model deployed on edge computing nodes to predict the future movement trajectory of workers in real time, and conduct risk assessment based on the spatial relationship between the predicted trajectory and the preset safety boundary. S5: When the predicted trajectory is about to enter the safety boundary, it automatically triggers multi-level early warning commands and performs defensive actions in conjunction with them; S6: The warning command is sent to the on-site execution terminal through the 5G communication network to realize end-to-end linkage between risk identification and defense response.

2. The intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to claim 1, characterized in that: The 5G communication sensing network includes ultra-densely deployed 5G base stations and distributed edge computing units. The 5G base stations are equipped with indoor positioning enhancement units, which achieve high-precision three-dimensional positioning of workers through time difference of arrival positioning technology and angle positioning technology.

3. The intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to claim 1, characterized in that: The preprocessing includes Kalman filtering to denoise the original positioning data, removing abnormal drift points, and performing time synchronization calibration; the trajectory feature extraction process uses a sliding window mechanism to extract kinematic features such as the velocity vector, direction of movement, displacement distance, and trajectory curvature of the workers within a preset time window, while also extracting interactive features such as the relative distance, relative speed, and similarity of movement trends among multiple workers.

4. The intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to claim 1, characterized in that: The lightweight interactive perception trajectory prediction model adopts an encoder-decoder architecture. The encoder part uses a graph attention network to explicitly model the interaction relationship between multiple workers, and the decoder part uses a long short-term memory network to generate multi-step prediction trajectories. The model lightweighting process includes replacing standard convolutional layers with depthwise separable convolutions to reduce the number of parameters, using knowledge distillation technology to transfer knowledge from large prediction models to small models, and using parameter quantization technology to compress floating-point parameters to fixed-point parameters.

5. The intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to claim 1, characterized in that: The prediction process uses historical trajectory data from the most recent 3 seconds as input and outputs a predicted trajectory sequence for the next 5 seconds. The risk assessment process generates a risk level score by calculating the area of ​​the overlap between the predicted trajectory and the preset safety boundary and the overlap time window. When the risk level score exceeds a preset threshold, it is determined to be a high-risk event.

6. The intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to claim 1, characterized in that: The multi-level early warning mechanism includes three levels: yellow, orange, and red. A yellow warning corresponds to a predicted trajectory that may reach the safety boundary within 10 seconds, an orange warning corresponds to a predicted trajectory that may reach the safety boundary within 5 seconds, and a red warning corresponds to a predicted trajectory that will reach the safety boundary within 2 seconds. The linked defense actions include triggering an audible and visual alarm to remind workers, linking the access control system to restrict workers from entering the danger zone, and linking a robotic arm or barrier to perform physical interception.

7. The intelligent early warning method for power operation safety based on 5G communication and artificial intelligence according to claim 1, characterized in that: The safety boundary is set to a circle, rectangle, or irregular polygon shape, and the safety boundary supports dynamic scaling to adapt to the risk management needs of different operation stages; the dynamic scaling rules of the safety boundary are linked to the operation voltage level.

8. A power operation safety intelligent early warning system based on 5G communication and artificial intelligence, applicable to the power operation safety intelligent early warning method based on 5G communication and artificial intelligence as described in any one of claims 1-7, characterized in that, include: The 5G communication sensing network module is configured to build an integrated communication and sensing network at the power operation site to collect the location and movement data of the workers. The edge computing processing module, connected to the 5G communication sensing network module, is configured to perform data preprocessing and feature extraction tasks. The lightweight trajectory prediction module, integrated within the edge computing processing module, is configured to encode the interaction relationships of multiple workers based on a graph attention network and generate predicted trajectories using a long short-term memory network. The risk assessment and early warning module, connected to the lightweight trajectory prediction module, is configured to perform spatial conflict detection between the predicted trajectory and the safety boundary and generate multi-level early warnings. The linkage execution module is communicatively connected to the risk assessment and early warning module and is configured to send early warning instructions to the field execution terminal and intelligently dispatch automatic defense execution mechanisms.

9. The intelligent early warning system for power operation safety according to claim 8, characterized in that: The lightweight trajectory prediction module includes an interactive perception encoding submodule and a trajectory decoding submodule; the interactive perception encoding submodule adopts a graph attention network architecture to encode features of a spatial relationship graph composed of multiple workers. The trajectory decoding submodule employs a lightweight long short-term memory network to generate a multi-step predicted trajectory sequence based on the encoding features.

10. The intelligent early warning system for power operation safety according to claim 8, characterized in that: The system also includes a predictive model adaptive update module and a 3D digital twin display module. The predictive model adaptive update module is connected to the lightweight trajectory prediction module and continuously monitors the deviation between the predicted trajectory and the actual trajectory. When the deviation exceeds a preset tolerance, the model parameter update process is automatically triggered. The 3D digital twin display module establishes a data connection with the 5G communication sensing network module and the risk assessment and early warning module to render and generate a dynamic 3D scene of the work site in real time.

Images