Power distribution network uninterrupted operation safety early warning system and method

Abstract

The invention discloses a power distribution network uninterrupted operation safety early warning system and method, and relates to the technical field of power operation safety. According to the system, through deep cooperation of the data set construction module, the target detection module, the behavior analysis module and the edge deployment module, transformation from post-event alarm to pre-event early warning is realized. Wherein the data set construction module accurately extracts a protection appliance state change key frame by using a P2 branch of an improved INSP-YOLO model; the target detection module adopts a special model comprising an MS-RFPM module, a C2f-CAA module and a CSIB module; the behavior analysis module realizes behavior prediction through human body key point trajectory analysis; and the edge deployment module jointly optimizes and deploys the models to the mercuric chloride equipment. All the modules are mutually enhanced through a closed-loop feedback mechanism, the problems of early warning lag, insufficient small target detection precision and the like in the prior art are solved, and the timeliness and accuracy of safety control are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power system safety operation technology, and in particular to a safety early warning system and method for uninterrupted power distribution network operation. Background Technology

[0002] Live-line work on power distribution networks is a crucial means of ensuring power supply reliability. However, the on-site environment is complex and high-risk. Traditional safety supervision mainly relies on manual on-site supervision or ordinary video surveillance, which suffers from problems such as insufficient operation and maintenance personnel, inability to monitor and issue early warnings in real time, and delayed response to non-standard operations. While existing technologies have attempted to utilize computer vision technology for safety monitoring, they generally suffer from the following shortcomings:

[0003] Insufficient model accuracy: General target detection models (such as the original YOLO series) have low detection accuracy for small-sized safety protective equipment such as insulating gloves and goggles, and are difficult to adapt to complex power operation scenarios (such as obstruction and changes in light).

[0004] Data quality constraints: The lack of targeted, high-quality dedicated datasets leads to poor model training results and weak generalization ability.

[0005] Lagging behavior recognition: Most solutions can only identify unsafe behaviors that have already occurred (such as not wearing a safety helmet), and cannot identify dynamic and continuous behaviors (such as the process of removing gloves), and lack the ability to predict potential risks.

[0006] Real-time challenges: Complex models require a large amount of computation, making it difficult to achieve real-time inference on resource-constrained edge devices, resulting in delayed warnings.

[0007] Therefore, there is an urgent need for a systematic solution that integrates high-quality data construction, high-precision small target detection, unsafe behavior prediction, and low-latency edge deployment to achieve the transformation from passive alarm to proactive intervention. Summary of the Invention

[0008] The purpose of this invention is to provide a safety early warning system and method for live-line work in power distribution networks, aiming to solve existing technical problems.

[0009] To address the aforementioned problems, according to one aspect of this application, an embodiment of the present invention provides a power distribution network live-line working safety early warning system, comprising:

[0010] The dataset construction module is used to build an unsafe behavior monitoring dataset through multi-view acquisition, model-assisted annotation and verification, and keyframe extraction based on target detection and temporal analysis.

[0011] The target detection module includes an improved INSP-YOLO model, which is trained based on the unsafe behavior monitoring dataset to identify workers and safety equipment. In the backbone network of the improved INSP-YOLO model, two C2f modules are replaced with a multi-scale receptive field prior module MS-RFPM, and the latter two C2f modules are replaced with a cross-stage fusion anchor attention module C2f-CAA.

[0012] The behavior analysis module is used to identify and predict unsafe behaviors based on human keypoint sequences;

[0013] The edge deployment module is used to convert trained object detection models and behavior analysis models into a format suitable for edge computing devices and deploy them.

[0014] The system achieves early warning through the following deep collaboration mechanism:

[0015] The dataset construction module utilizes the high-precision detection capability of the P2 branch of the INSP-YOLO model for small targets to accurately extract key frames of changes in the state of protective equipment from the video stream;

[0016] The behavior analysis module receives the bounding box of the person detected by the INSP-YOLO model, and locates and crops the target region based on the bounding box, and inputs it into the lightweight pose estimation model.

[0017] The edge deployment module uses the INSP-YOLO model and the behavior analysis model as a joint inference engine, and performs unified conversion and optimization through the CANN platform.

[0018] New unsafe behavior samples identified on edge devices are triggered to be uploaded and fed back to the dataset building module for incremental learning of the model.

[0019] In some implementations, the dataset construction module includes:

[0020] The model-assisted labeling unit uses a pre-trained first object detection model to predict the initial dataset, compares and corrects the prediction results with human labels, and obtains the second basic dataset.

[0021] The keyframe extraction unit uses the trained second target detection model to identify video data in frames. When a key state change is detected, the video frames at that moment and the preset time before and after are extracted as keyframes to form the third basic dataset.

[0022] The unsafe behavior monitoring dataset is formed by merging the second basic dataset and the third basic dataset.

[0023] In some implementations, the multi-scale receptive field prior module MS-RFPM fuses multi-scale features and target prior information by using parallel multi-scale deep-dilated convolutions and strip convolutions. The strip convolutions include depthwise separable convolutions with a horizontal dimension of 1×(2k+1) and a vertical dimension of (2k+1)×1.

[0024] In some implementations, the cross-stage fusion anchor attention module C2f-CAA adopts a two-stream feature fusion architecture and applies a hierarchical awareness convolutional kernel scaling strategy. In the shallow feature extraction stage, a horizontal long kernel of 1×9 with an inflation rate of 4 is used, while in the deep feature extraction stage, a horizontal long kernel of 1×5 with an inflation rate of 2 is used.

[0025] In some implementations, the neck network of the improved INSP-YOLO model also includes a cross-scale interaction module (CSIB). The CSIB flattens and stitches the feature maps of different scales of the future autonomous backbone network through a linear mapping layer, and then achieves cross-scale feature interaction through multi-receptive field dilated convolution.

[0026] In some implementations, the improved INSP-YOLO model further includes a small target detection branch, which upsamples the maximum-scale feature map output by the neck network and then splices and fuses it with the output of the cross-scale interaction module CSIB to form a 160×160 resolution detection head.

[0027] In some implementations, the behavior analysis module achieves early warning through the following steps:

[0028] Extracting key points of the human body using the YOLOPOSE model;

[0029] The keypoint sequence is input into the spatiotemporal graph convolutional network ST-GCN for unsafe behavior classification;

[0030] By analyzing the spatial relationship between the movement trajectory of specific key points on the human body and the preset risk area, predictive warnings can be issued before unsafe actions are completed.

[0031] The specific human body key points include the wrist key points, and the preset risk area includes the head area. When the wrist key points are detected moving towards the head area and the speed exceeds the threshold, an early warning to remove protective equipment is triggered.

[0032] In some implementations, the edge deployment module is specifically used for:

[0033] Model training was completed on the Huawei Cloud ModelArts platform;

[0034] The model is converted to .om format suitable for Ascend edge devices using the ATC tool on the Huawei CANN platform;

[0035] The deployment process of model loading, data preprocessing, inference, and result postprocessing is completed on edge devices;

[0036] The edge device is the Atlas200DKA2 developer kit based on the Huawei Ascend 310 chip.

[0037] This invention also provides a method for safety early warning of power distribution network uninterrupted operation based on any of the above-described systems, comprising the following steps:

[0038] The dataset construction module is used to construct an unsafe behavior monitoring dataset, wherein the keyframe extraction step uses the P2 branch of the INSP-YOLO model to detect changes in the status of protective equipment.

[0039] The improved INSP-YOLO model was trained using the unsafe behavior monitoring dataset.

[0040] The trained model is deployed to edge devices via the edge deployment module.

[0041] Edge devices acquire real-time video streams of operations and perform real-time analysis and prediction using pre-deployed target detection and behavior analysis modules.

[0042] When an unsafe behavior is predicted, a warning signal is triggered before the action is completed;

[0043] New unsafe behavior samples identified on edge devices are fed back to the dataset building module to enable continuous model optimization.

[0044] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the method described above.

[0045] Compared with the prior art, the power distribution network live-line working safety early warning system and method of the present invention have at least the following beneficial effects:

[0046] High detection accuracy: The improved INSP-YOLO model, especially the introduction of MS-RFPM, C2f-CAA modules and P2 branch, significantly improves the detection accuracy of small targets and targets with specific shapes (mAP improvement of about 5-10%).

[0047] Advanced early warning: By combining attitude estimation and behavior prediction models, it has achieved a leap from "post-event alarm" to "pre-event early warning", truly preventing problems before they occur.

[0048] Strong system integration: It has formed a complete technical closed loop from data preparation and algorithm optimization to deployment, with deep adaptation of each link and excellent overall performance.

[0049] Good real-time performance: The lightweight design of the model and the optimized deployment for edge devices ensure that the stringent requirements of real-time monitoring can still be met in resource-constrained environments.

[0050] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0053] Figure 2 This is a flowchart of the dataset construction method;

[0054] Figure 3 This is a structural diagram of the improved INSP-YOLO model.

[0055] Figure 4 This is a schematic diagram of behavior recognition and prediction (such as helmet removal warning);

[0056] Figure 5 This is a flowchart of the edge deployment module. Detailed Implementation

[0057] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments, structures, features, and effects according to the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "an embodiment" or "an embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0058] In the description of this invention, it should be clearly stated that the terms "first," "second," etc., in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence; the terms "vertical," "lateral," "longitudinal," "front," "rear," "left," "right," "up," "down," "horizontal," etc., indicate orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, and are merely for the convenience of describing this invention, and do not mean that the device or element referred to must have a specific orientation or position, and therefore should not be construed as a limitation of this invention.

[0059] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0060] like Figure 1-5 As shown, this embodiment of the invention provides a power distribution network live-line working safety early warning system, including:

[0061] The dataset construction module is used to build an unsafe behavior monitoring dataset through multi-view acquisition, model-assisted annotation and verification, and keyframe extraction based on target detection and temporal analysis.

[0062] The target detection module includes an improved INSP-YOLO model, which is trained based on the unsafe behavior monitoring dataset to identify workers and safety equipment. In the backbone network of the improved INSP-YOLO model, two C2f modules are replaced with a multi-scale receptive field prior module MS-RFPM, and the latter two C2f modules are replaced with a cross-stage fusion anchor attention module C2f-CAA.

[0063] The behavior analysis module is used to identify and predict unsafe behaviors based on human keypoint sequences;

[0064] The edge deployment module is used to convert trained object detection models and behavior analysis models into a format suitable for edge computing devices and deploy them.

[0065] The system achieves early warning through the following deep collaboration mechanism:

[0066] The dataset construction module utilizes the high-precision detection capability of the P2 branch of the INSP-YOLO model for small targets to accurately extract key frames of changes in the state of protective equipment from the video stream;

[0067] The behavior analysis module receives the bounding box of the person detected by the INSP-YOLO model, and locates and crops the target region based on the bounding box, and inputs it into the lightweight pose estimation model.

[0068] The edge deployment module uses the INSP-YOLO model and the behavior analysis model as a joint inference engine, and performs unified conversion and optimization through the CANN platform.

[0069] New unsafe behavior samples identified on edge devices are triggered to be uploaded and fed back to the dataset building module for incremental learning of the model.

[0070] In this embodiment, the system achieves pre-operational safety warnings for uninterrupted power distribution network operations through the deep collaborative operation of a dataset construction module, a target detection module, a behavior analysis module, and an edge deployment module. The dataset construction module first collects video data from the work site from multiple perspectives, uses a pre-trained target detection model for auxiliary annotation and verification, compares and corrects the prediction results with manually labeled data, and leverages the high-precision detection capability of the P2 branch of the INSP-YOLO model for small targets to accurately extract keyframes of changes in the state of protective equipment from the video stream. The two types of data are then merged to form a high-quality dataset for monitoring unsafe behaviors. The improved INSP-YOLO model in the target detection module replaces two C2f modules with MS-RFPM in its backbone network, replaces the latter two C2f modules with C2f-CAA, adds a CSIB module to the neck network, and adds a small target detection branch. After training based on the constructed dataset, it can accurately identify workers and safety protective equipment such as insulating gloves and goggles. The behavior analysis module receives bounding boxes of personnel detected by the INSP-YOLO model, locates and crops the target area, inputs it into a lightweight pose estimation model, extracts the sequence of human key points, and analyzes the spatial relationship between the key point motion trajectory and the preset risk area through ST-GCN to predict unsafe behaviors. The edge deployment module uses the target detection model and behavior analysis model as a joint inference engine, and deploys them to Ascend edge devices after unified conversion and optimization by the CANN platform. New unsafe behavior samples identified by the edge devices are fed back to the dataset construction module to achieve incremental model learning. The closed-loop feedback and mutual reinforcement of the modules in the entire system solve the problems of delayed early warning and insufficient accuracy of small target detection in existing technologies, greatly improving the timeliness and accuracy of safety management and realizing the transformation from "post-event alarm" to "pre-event warning", adapting to complex power operation scenarios and reducing operational safety risks.

[0071] In some implementations, the dataset construction module includes:

[0072] The model-assisted labeling unit uses a pre-trained first object detection model to predict the initial dataset, compares and corrects the prediction results with human labels, and obtains the second basic dataset.

[0073] The keyframe extraction unit uses the trained second target detection model to identify video data in frames. When a key state change is detected, the video frames at that moment and the preset time before and after are extracted as keyframes to form the third basic dataset.

[0074] The unsafe behavior monitoring dataset is formed by merging the second basic dataset and the third basic dataset.

[0075] In this embodiment, the dataset construction module efficiently constructs a high-quality dataset for monitoring unsafe behaviors through the collaborative work of the model-assisted annotation unit and the keyframe extraction unit. The model-assisted annotation unit first calls a pre-trained first target detection model to predict the initial dataset collected from multiple perspectives, generating preliminary annotation results. Then, the predicted results are compared one by one with manually annotated labels to correct mislabeling and omissions in the model predictions, eliminate invalid data, and form a second basic dataset with accurate and reliable annotations. This reduces the workload of manual annotation and improves annotation accuracy. The keyframe extraction unit uses the trained and optimized second target detection model to perform frame-by-frame recognition on continuous video data, monitoring in real time the changes in the status of workers' safety protective equipment and key turning points in their work behavior within the video stream. When key status changes such as the protective equipment changing from being worn to being removed, or from being intact to being damaged, are detected, video frames within a preset time range before and after that moment are automatically extracted as keyframes. These keyframes collectively embody the core characteristics of unsafe behaviors, constituting a highly representative third basic dataset. Finally, the second and third basic datasets are merged to form a comprehensive and focused dataset for monitoring unsafe behaviors. This provides high-quality data support for the training of subsequent target detection and behavior analysis models, solves the problems of high cost and poor sample validity in traditional dataset construction, and improves model training effect and generalization ability.

[0076] In some implementations, the multi-scale receptive field prior module MS-RFPM fuses multi-scale features and target prior information by using parallel multi-scale deep-dilated convolutions and strip convolutions. The strip convolutions include depthwise separable convolutions with a horizontal dimension of 1×(2k+1) and a vertical dimension of (2k+1)×1.

[0077] In this embodiment, MS-RFPM effectively fuses multi-scale features and prior target information through a unique parallel convolutional structure, enhancing the model's ability to detect targets of different scales. This module uses parallel multi-scale deep-dilated convolutions and strip convolutions. The strip convolutions include depthwise separable convolutions with a horizontal dimension of 1×(2k+1) and a vertical dimension of (2k+1)×1. Multi-scale deep-dilated convolutions expand the model's receptive field by setting different dilation rates, enabling the capture of target features of different sizes in the work environment, effectively perceiving both large worker bodies and small safety equipment. The strip convolutions target long, narrow targets such as insulating rods in power operations. Through horizontal and vertical strip convolution kernels, they accurately extract the morphological features of these targets, avoiding detection omissions or misjudgments due to the special shape of the target. The two convolutional structures work in parallel, extracting features from each and then fusing them together. This allows the model to fully acquire key information about multi-scale and multi-shaped targets during the feature extraction stage, providing rich feature support for subsequent target detection. This significantly improves the model's detection accuracy for various targets in complex power operation scenarios, especially enhancing the ability to recognize small and elongated targets, and adapting to complex environments such as occlusion and changes in lighting.

[0078] In some implementations, the cross-stage fusion anchor attention module C2f-CAA adopts a two-stream feature fusion architecture and applies a hierarchical awareness convolutional kernel scaling strategy. In the shallow feature extraction stage, a horizontal long kernel of 1×9 with an inflation rate of 4 is used, while in the deep feature extraction stage, a horizontal long kernel of 1×5 with an inflation rate of 2 is used.

[0079] In this embodiment, C2f-CAA employs a dual-stream feature fusion architecture, combined with a hierarchical kernel scaling strategy, to optimize feature extraction and enhance the model's ability to capture target features. The dual-stream feature fusion architecture can simultaneously process feature information from different sources, achieving complementary feature fusion, enhancing the richness of feature representation, and allowing the model to perceive target features more comprehensively. This module applies a hierarchical kernel scaling strategy, adjusting the convolutional kernel parameters according to different stages of feature extraction. In the shallow feature extraction stage, a horizontal kernel of 1×9 and an inflation rate of 4 are used, at which point the model focuses on capturing the overall contour and large-scale features of the target, expanding the receptive field to cover more target areas. In the deep feature extraction stage, the configuration is switched to a horizontal kernel of 1×5 and an inflation rate of 2, focusing on the detailed features and key local information of the target, improving the accuracy of feature extraction. This dynamic adjustment strategy allows the model to focus on different aspects at different feature extraction stages, ensuring both overall perception of the target and accurate capture of detailed features. This effectively improves the ability to identify the features of workers and safety equipment, especially in complex work scenarios where target features are not obvious and are easily interfered with, thus enhancing the model's anti-interference ability and detection accuracy.

[0080] In some implementations, the neck network of the improved INSP-YOLO model also includes a cross-scale interaction module (CSIB). The CSIB flattens and stitches the feature maps of different scales of the future autonomous backbone network through a linear mapping layer, and then achieves cross-scale feature interaction through multi-receptive field dilated convolution.

[0081] In this embodiment, the CSIB module enhances the information flow and fusion between feature maps of different scales through a unique feature processing flow, thereby improving the model's feature representation capability. The module first flattens the feature maps of the future autonomous backbone network at different scales using a linear mapping layer, eliminating dimensional differences between the feature maps. Then, the flattened feature maps at each scale are stitched together to integrate multi-scale feature information. The stitched feature information is then input into a multi-receptive-field dilated convolutional layer, where features are further extracted through dilated convolutions with different dilation rates, achieving deep interaction and fusion of cross-scale features. This process breaks down the information barriers between feature maps of different scales, allowing shallow, fine-grained features to fully combine with deep, abstract features, enabling the model to simultaneously utilize the feature advantages of different scales and accurately identify targets of different sizes. For small-sized safety protective equipment and larger workers' bodies in live-line work on power distribution networks, efficient detection can be achieved through the fused features, solving the problem of insufficient detection accuracy caused by inadequate feature fusion at different scales in traditional models, and improving the model's target detection performance in complex scenarios.

[0082] In some implementations, the improved INSP-YOLO model further includes a small target detection branch, which upsamples the maximum-scale feature map output by the neck network and then splices and fuses it with the output of the cross-scale interaction module CSIB to form a 160×160 resolution detection head.

[0083] In this embodiment, the improved INSP-YOLO model adds a small target detection branch specifically designed for detecting small-sized safety protective gear such as insulating gloves and goggles, significantly improving the accuracy of small target detection. This branch first upsamples the maximum-scale feature map output by the neck network to increase the resolution of the feature map, retaining more detailed features of small targets and avoiding the loss of small target information due to insufficient feature map resolution. Then, the upsampled feature map is stitched and fused with the output of the cross-scale interaction module CSIB, integrating the feature information from multi-scale interaction with high-resolution feature information to form a 160×160 resolution detection head. The high-resolution detection head can more accurately locate the position of small targets, and combined with the rich fused feature information, effectively distinguishes small targets from background noise, reducing missed and false detections of small targets. In live-line work on power distribution networks, the wearing status of small-sized safety protective gear directly affects work safety. The design of this small target detection branch specifically addresses the problem of low accuracy in small target detection by general target detection models, ensuring that the model can accurately identify the status of small-sized protective gear, providing a reliable basis for subsequent unsafe behavior prediction, and further improving the accuracy of the system's safety warnings.

[0084] In some implementations, the behavior analysis module achieves early warning through the following steps:

[0085] Extracting key points of the human body using the YOLOPOSE model;

[0086] The keypoint sequence is input into the spatiotemporal graph convolutional network ST-GCN for unsafe behavior classification;

[0087] By analyzing the spatial relationship between the movement trajectory of specific key points on the human body and the preset risk area, predictive warnings can be issued before unsafe actions are completed.

[0088] The specific human body key points include the wrist key points, and the preset risk area includes the head area. When the wrist key points are detected moving towards the head area and the speed exceeds the threshold, an early warning to remove protective equipment is triggered.

[0089] In this embodiment, the behavior analysis module uses human keypoint extraction, behavior classification, and trajectory analysis to achieve predictive early warning of unsafe behaviors and mitigate operational risks in advance. The module first calls the YOLOPOSE model to accurately extract human keypoints from the personnel target area output by the target detection module, including the coordinate information of key parts such as the wrist, head, and torso, forming a continuous keypoint sequence. This keypoint sequence is then input into the spatiotemporal graph convolutional network ST-GCN. ST-GCN analyzes the temporal and spatial variation patterns of the keypoints to classify the worker's behavior and identify any unsafe behaviors that have occurred. Simultaneously, the module continuously analyzes the spatial relationship between the movement trajectory of specific human keypoints, including wrist keypoints, and the preset risk areas, including the head area. When the movement of a wrist keypoint towards the head area is detected and the speed exceeds a set threshold, it is determined that the worker intends to remove protective equipment such as a helmet and goggles, triggering an early warning signal before this unsafe action is completed. This key point trajectory prediction method breaks through the limitation of traditional behavior recognition, which can only detect behaviors that have already occurred. It enables early prediction of potential unsafe behaviors, giving workers and supervisors more reaction time and effectively reducing the incidence of safety accidents caused by unsafe behaviors.

[0090] In some implementations, the edge deployment module is specifically used for:

[0091] Model training was completed on the Huawei Cloud ModelArts platform;

[0092] The model is converted to .om format suitable for Ascend edge devices using the ATC tool on the Huawei CANN platform;

[0093] The deployment process of model loading, data preprocessing, inference, and result postprocessing is completed on edge devices;

[0094] The edge device is the Atlas200DKA2 developer kit based on the Huawei Ascend 310 chip.

[0095] In this embodiment, the edge deployment module achieves efficient model deployment on edge devices through a standardized process, ensuring the system's real-time performance and usability. This module first trains the object detection and behavior analysis models on the Huawei Cloud ModelArts platform, leveraging the platform's powerful computing resources to improve training efficiency and effectiveness. After training, the trained models are converted to the .om format suitable for Ascend edge devices using the ATC tool on the Huawei CANN platform. This optimized format fully adapts to the edge device's hardware architecture, improving model inference speed. Subsequently, the complete deployment process—model loading, data preprocessing, inference, and result post-processing—is completed on the edge device. The edge device utilizes the Atlas200DKA2 developer kit based on the Huawei Ascend 310 chip, which boasts ample computing power and strong resource adaptability. After deployment, the edge device can receive real-time video streams from the field, quickly perform preprocessing such as video decoding and scaling, and then feed them into the model for inference analysis, instantly outputting detection results and early warning signals. This edge deployment approach avoids the latency caused by data transmission to the cloud, solves the problem of real-time inference of complex models on resource-constrained devices, ensures timely output of early warning signals, meets the needs of real-time monitoring of power distribution network uninterrupted operation, and improves the response speed of on-site safety management.

[0096] This invention also provides a method for safety early warning of power distribution network uninterrupted operation based on any of the above-described systems, comprising the following steps:

[0097] The dataset construction module is used to construct an unsafe behavior monitoring dataset, wherein the keyframe extraction step uses the P2 branch of the INSP-YOLO model to detect changes in the status of protective equipment.

[0098] The improved INSP-YOLO model was trained using the unsafe behavior monitoring dataset.

[0099] The trained model is deployed to edge devices via the edge deployment module.

[0100] Edge devices acquire real-time video streams of operations and perform real-time analysis and prediction using pre-deployed target detection and behavior analysis modules.

[0101] When an unsafe behavior is predicted, a warning signal is triggered before the action is completed;

[0102] New unsafe behavior samples identified on edge devices are fed back to the dataset building module to enable continuous model optimization.

[0103] In this embodiment, the early warning method follows a closed-loop process of "data construction - model training - deployment and application - continuous optimization" to achieve full-process safety early warning for live-line work in power distribution networks. First, an unsafe behavior monitoring dataset is constructed using a dataset construction module. Work videos are collected from multiple perspectives, and high-quality basic data is obtained through model-assisted annotation and verification. Then, keyframes of changes in the state of protective equipment are extracted using the P2 branch of the INSP-YOLO model, and these are merged to form a complete dataset. Subsequently, an improved INSP-YOLO model is trained using this dataset, optimizing model parameters and improving the model's detection accuracy for workers and safety equipment. After training, the model is converted to a format suitable for edge devices using an edge deployment module and deployed to Ascend edge devices. The edge devices collect the work video stream in real time, calling the deployed target detection and behavior analysis modules to perform real-time analysis of the video stream, extracting key human points and predicting unsafe behaviors. When a potential unsafe behavior is detected, an early warning signal is triggered immediately before the action is completed, reminding workers and supervisors to intervene in a timely manner. Meanwhile, edge devices automatically feed back newly identified unsafe behavior samples to the dataset construction module, providing new data for incremental model learning and continuously optimizing model performance. This method forms a complete technical closed loop from data to model to application, combining high-precision detection, early warning, and real-time response capabilities. It effectively solves problems such as insufficient personnel and delayed response in traditional supervision methods, and comprehensively improves the safety management level of live-line work in power distribution networks.

[0104] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the method described above.

[0105] In this embodiment, the electronic device ensures the efficient execution of the safety early warning method for uninterrupted power distribution network operations through the coordinated operation of hardware and software. The electronic device includes a memory and a processor. The memory stores computer programs and various types of data, including multi-view acquired video data of the operation, a constructed dataset for monitoring unsafe behaviors, and parameters of the trained target detection model and behavior analysis model, providing data support for the processor's computation. The processor, as the core computing unit, runs the computer program stored in the memory, sequentially executing each step of the early warning method. The processor first controls the dataset construction program to complete data acquisition, annotation verification, and keyframe extraction, constructing a high-quality dataset; then it calls the model training program, using the dataset to optimize and improve the INSP-YOLO model; subsequently, it starts the deployment program, converting the model to an adapted format and deploying it to the edge device; finally, it runs the analysis program in real time, receiving video streams acquired by the edge device, driving the target detection module and behavior analysis module to perform real-time inference, predicting unsafe behaviors, and triggering early warnings. The processor's high-efficiency computing power ensures the rapid execution of various programs, while the large-capacity memory guarantees the stable storage of data and programs. The two work together to enable electronic devices to stably and efficiently achieve safety early warning for uninterrupted power distribution network operations, providing reliable technical support for operational safety and adapting to the real-time monitoring needs of complex operational scenarios.

[0106] Two specific embodiments are listed below.

[0107] Example 1: Dataset Construction Taking the construction of a dataset of unsafe behaviors such as "removing insulating gloves" as an example:

[0108] Multi-view acquisition: At the 10kV distribution network live-line work site, ground cameras (looking up and at eye level) and cameras on the boom truck (looking down and at eye level) are set up to record videos of normal operations by workers and simulated unsafe operations.

[0109] Model-assisted annotation: A YOLOv8 model pre-trained on a general dataset is used to predict the initial images obtained from video frame extraction, generating preliminary labels. Professionals compare the model predictions with the real-world data, correcting mislabeling and omissions, and forming a high-quality second foundational dataset.

[0110] Keyframe extraction: A more accurate YOLOv8 model is trained using the dataset obtained in step two. The model is used to detect each frame of the video. When the state of "wearing insulated gloves" changes to "not wearing insulated gloves" in a continuous frame sequence, the system automatically extracts the state change point and the frames within 1 second before and after it, and stores them as keyframes in the third basic dataset.

[0111] Dataset merging: The second and third base datasets are merged to form the final dataset used to train the behavior monitoring model.

[0112] Example 2: Model Training and Edge Deployment

[0113] Training the INSP-YOLO model: Using the dataset constructed in Example 1, the improved INSP-YOLO model was trained on the Huawei Cloud ModelArts platform. The CIoU loss function was used during training.

[0114] Model conversion: First, convert the trained .pt model to .onnx format, and then use the ATC tool of Huawei CANN platform to convert it to .om format suitable for Atlas200IDKA2 edge devices.

[0115] Edge inference: Deploy the transformed model on edge devices. The video stream captured by the device's camera is preprocessed, such as decoding and scaling, and then fed into the model for inference. It can identify workers and safety equipment in real time and provide early warnings for unsafe behaviors or risks.

[0116] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A power distribution network live-line work safety early warning system, characterized in that, The system comprises: a dataset construction module for constructing an unsafe behavior monitoring dataset through multi-view acquisition, model-assisted annotation verification, and key frame extraction based on target detection and time series analysis; a target detection module comprising an improved INSP-YOLO model for training based on the unsafe behavior monitoring dataset to identify workers and safety protection tools; in the backbone network of the improved INSP-YOLO model, two C2f modules are replaced with a multi-scale receptive field prior module MS-RFPM, and the last two C2f modules are replaced with a cross-stage fusion anchor attention module C2f-CAA; a behavior analysis module for identifying and predicting unsafe behaviors based on human key point sequences; an edge deployment module for converting the trained target detection model and behavior analysis model into a format suitable for edge computing devices and deploying them; wherein the system realizes pre-warning through the following deep collaboration mechanism: the dataset construction module uses the high-precision detection capability of the P2 branch of the INSP-YOLO model for small targets to accurately extract key frames of protection tool state changes from video streams; the behavior analysis module receives the personnel bounding box detected by the INSP-YOLO model, locates and crops the target region based on the bounding box, and inputs it into a lightweight pose estimation model; the edge deployment module converts and optimizes the INSP-YOLO model and the behavior analysis model as a joint inference engine through the CANN platform; new unsafe behavior samples identified on the edge device are triggered to upload and flow back to the dataset construction module for incremental learning of the model.

2. The system of claim 1, wherein, The dataset construction module comprises: a model-assisted annotation unit that uses a pre-trained first target detection model to predict an initial dataset, compares the prediction results with manual labels, and corrects them to obtain a second basic dataset; a key frame extraction unit that uses a trained second target detection model to frame and identify video data, extracts video frames at a preset time before and after a key state change as key frames when a key state change is detected, and constitutes a third basic dataset; wherein the unsafe behavior monitoring dataset is combined from the second basic dataset and the third basic dataset.

3. The system of claim 1, wherein, The multi-scale receptive field prior module MS-RFPM fuses multi-scale features and target prior information through parallelly arranged multi-scale deep hollow convolution and strip convolution, and the strip convolution includes depth separable convolution with a horizontal size of 1×(2k+1) and a vertical size of (2k+1)×1.

4. The system of claim 1, wherein, The cross-stage fusion anchor attention module C2f-CAA adopts a dual-flow feature fusion architecture and applies a hierarchical perception convolution kernel scaling strategy, wherein a horizontal long kernel 1×9 with an expansion rate of 4 is used in the shallow feature extraction stage, and a horizontal long kernel 1×5 with an expansion rate of 2 is used in the deep feature extraction stage.

5. The system of claim 1, wherein, The improved INSP-YOLO model further comprises a small target detection branch, which is used for sampling the maximum scale feature map output by the neck network, and then performing splicing and fusion with the output of the cross-scale interaction module CSIB to form a detection head with a resolution of 160x160.

6. The system of claim 1, wherein, The behavior analysis module realizes the early warning by the following steps:

7. The system of claim 1, wherein, extracting human key points by using the YOLOPOSE model; inputting the key point sequence into a spatio-temporal graph convolution network ST-GCN for unsafe behavior classification; predictive warning is performed before the unsafe action is completed by analyzing the spatial relationship between the motion trajectory of a specific human key point and a preset risk area; wherein the specific human key point includes a wrist key point, and the preset risk area includes a head area, and when the wrist key point is detected to move towards the head area and the speed exceeds a threshold, a warning of removing protective equipment is triggered. The edge deployment module is specifically used for:

8. The system of claim 1, wherein, completing model training on the Huawei Cloud ModelArts platform; converting the model into an.om format suitable for Ascend edge devices by using the ATC tool of the Huawei CANN platform; completing the deployment process of model loading, data preprocessing, inference and result post-processing on the edge device; wherein the edge device is an Atlas 200DKA2 developer kit based on the Huawei Ascend 310 chip. The method comprises the following steps:

9. A safe pre-warning method for power distribution network operation without power cut based on the system of any one of claims 1-8, characterized in that, constructing an unsafe behavior monitoring dataset by using the dataset construction module, wherein the key frame extraction step uses the P2 branch of the INSP-YOLO model to detect changes in protective equipment state; training the improved INSP-YOLO model by using the unsafe behavior monitoring dataset; deploying the trained model to an edge device by using the edge deployment module; the edge device collects job video streams in real time, and uses the deployed target detection module and behavior analysis module for real-time analysis and prediction; when an unsafe behavior is predicted, a warning signal is triggered before the action is completed; the new unsafe behavior samples recognized on the edge device are fed back to the dataset construction module to realize continuous optimization of the model. The processor executes the program to realize the method of claim 9.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, ​

Images