Method and system for recognizing wearing state of high-altitude operation protection equipment of electric power system

Abstract

The invention discloses a wearing state identification method for a high-altitude operation protection device of an electric power system. The method comprises the following steps: acquiring and preprocessing image data worn by the high-altitude operation protection device of the electric power system to construct an image data set; based on the double-branch neural network and the graph attention network, constructing a wearing state identification initial model of the electric power system high-altitude operation protection equipment, and training to obtain a wearing state identification model of the electric power system high-altitude operation protection equipment; and adopting the obtained wearing state identification model of the electric power system high-altitude operation protection equipment to carry out wearing state identification of the protection equipment in an actual electric power system high-altitude operation process. The invention also discloses a system for realizing the wearing state identification method of the high-altitude operation protection equipment of the electric power system. According to the method, the wearing state of the high-altitude operation protection equipment of the electric power system can be recognized, the reliability is higher, the accuracy is higher, and the recognition effect is better.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electrical automation, and particularly relates to a wearing state recognition method and system of high-altitude operation protective equipment of a power system. BACKGROUND

[0002] With the development of economy and technology and the improvement of people's living standards, electric energy has become an essential secondary energy in people's production and life, bringing endless convenience to people's production and life. Therefore, ensuring the stable and reliable supply of electric energy has become one of the most important tasks of the power system.

[0003] High-altitude operation of the power system is an important means in the process of maintenance, maintenance, patrol, construction, etc. of the power system. Therefore, ensuring the safety of high-altitude operation of the power system has become one of the important work of the power system. The wearing of high-altitude operation protective equipment can greatly reduce the risk of high-altitude operation of the power system and improve the safety of high-altitude operation personnel. Therefore, real-time monitoring of the wearing state of the protective equipment for high-altitude operation of the power system is of great significance.

[0004] At present, researchers have applied artificial intelligence models to the detection of protective equipment wearing in the process of high-altitude operation of the power system. For example, CN113537019A discloses a human key point recognition method based on an AlphaPose model, which is used to detect the wearing state of a safety helmet of a substation personnel; CN116311082A proposes a YOLO-KP detection model using a YOLOv7 network architecture, which combines an improved SURF algorithm to extract key part feature points for wearing detection; CN115937735A describes a method for monitoring the wearing specification of a safety belt for high-altitude operation by using a marker ball and color recognition technology, which uses a FASTER-RCNN algorithm to identify the operation personnel.

[0005] However, the existing monitoring and identification schemes have greatly reduced accuracy in complex dynamic environments and in the presence of occlusion. In particular, in the case of occlusion exceeding 50% or more, the identification accuracy of such schemes is significantly reduced. In a multi-person interaction scenario, the misidentification probability of such schemes is also high. SUMMARY

[0006] One of the purposes of the present application is to provide a wearing state recognition method of high-altitude operation protective equipment of a power system, which has high reliability, good accuracy and good effect.

[0007] The second purpose of the present application is to provide a system for implementing the wearing state recognition method of high-altitude operation protective equipment of a power system.

[0008] The method for recognizing the wearing status of protective equipment for high-altitude operations in power systems provided by this invention includes the following steps:

[0009] S1. Acquire image data of the wearing of protective equipment for high-altitude operations in the power system;

[0010] S2. Preprocess the image data obtained in step S1 to construct an image dataset;

[0011] S3. Based on a dual-branch neural network and a graph attention network, an initial model for recognizing the wearing status of protective equipment for high-altitude operations in power systems is constructed.

[0012] S4. Using the image dataset constructed in step S2, train the initial model for recognizing the wearing status of protective equipment for high-altitude operations in the power system constructed in step S3 to obtain the model for recognizing the wearing status of protective equipment for high-altitude operations in the power system.

[0013] S5. Using the power system high-altitude operation protective equipment wearing status recognition model obtained in step S4, the wearing status recognition of protective equipment during actual power system high-altitude operations is performed.

[0014] Step S2, which involves preprocessing the image data obtained in step S1, specifically includes the following steps:

[0015] The data preprocessing includes data labeling and data cleaning;

[0016] Data annotation, including annotation of acquired image data;

[0017] Data cleaning includes image data filtering, image data enhancement, and cropping image data to a set size.

[0018] Step S3, which involves constructing an initial model for recognizing the wearing status of protective equipment for high-altitude operations in power systems based on a dual-branch neural network and a graph attention network, specifically includes the following steps:

[0019] The initial model for identifying the wearing status of protective equipment for high-altitude operations in power systems consists of a mask / heatmap generation module, a graph construction module, a graph attention module, and a feature fusion module connected in series.

[0020] A mask / heatmap generation module is constructed based on a dual-branch neural network; the mask / heatmap generation module is used to simultaneously generate target segmentation masks and structural key point heatmaps of input image data;

[0021] The graph construction module is used to construct a scene relationship graph based on the target segmentation mask and structural key point heatmap of the obtained input image data.

[0022] A graph attention module is constructed based on a graph attention network; the graph attention module is used to process the constructed scene relationship graph and update node features;

[0023] The feature fusion module is used to perform multimodal feature fusion based on the output of the graph attention module, and generate the final identification result of the wearing status of the high-altitude operation protective equipment in the power system.

[0024] The mask / heatmap generation module based on a dual-branch neural network specifically includes the following steps:

[0025] The mask / heatmap generation module includes a feature extraction module, a segmentation module, and a key point module;

[0026] A ResNet-50 network is used as the feature extraction module. The feature extraction module is used to extract global features of the input image and inputs the obtained global features into both the segmentation module and the key point module.

[0027] The segmentation module consists of three convolutional layers connected in sequence; the segmentation module is used to generate a target segmentation mask M for the input global features;

[0028] The keypoint module consists of three convolutional layers connected in sequence; the keypoint module is used to generate a structural keypoint heatmap H of the input global features.

[0029] The graph construction module's processing steps include the following:

[0030] The output of the mask / heatmap generation module includes a segmentation mask M and a key point heatmap H;

[0031] The key point heatmap is smoothed using a Gaussian kernel to better characterize the spatial distribution of protective equipment; the smoothing process is represented as follows:

[0032] In the formula Heatmap of key points after smoothing; The pixel coordinates on the image plane; Let i be the coordinates of the i-th key point; The standard deviation of the Gaussian kernel;

[0033] Based on the segmentation mask M and the smoothed keypoint heatmap Construct a scene relationship diagram , where node set V represents the detected key points and edge set E represents the spatial relationship between nodes;

[0034] Node features Composed of visual features and spatial coordinates, represented as ;in Key point coordinates; The visual feature extraction process includes: After passing through several convolutional layers, the feature mapping tensor FFF is obtained through a non-linear activation function, and then a region sampling function is applied. Processing is performed to complete the visual feature extraction process. For the region sampling function;

[0035] Edge features Represented as ,in This represents the Euclidean distance between node i and node j; Let be the angle between the line connecting node i and node j and the reference direction; It is a feature mapping function, and , For the set nonlinear activation function, As the first weight parameter, is the second weighting parameter, and b is the bias term.

[0036] The processing steps of the graph attention module include the following:

[0037] A graph attention network is used as the graph attention module;

[0038] A graph attention network is used to process the scene relationship graph and update node features; represented as:

[0039] In the formula For the first Layer node characteristics; The activation function is set; For nodes The set of neighboring nodes; For the first Attention weights of layers; For the first The parameter matrix to be learned for the layer;

[0040] The following formula is used to calculate the first... layer attention weights :

[0041]

[0042] In the formula It is a linear rectified function with leakage parameters; The attention vector to be learned; This is for splicing operations.

[0043] The feature fusion module's processing procedure includes the following steps:

[0044] The processing procedure of the feature fusion module is represented by the following formula:

[0045] In the formula This is the output of the feature fusion module; The output of the graph attention module; For fusion function, , This is the first weight matrix. This is the second weight matrix. This is a bias term.

[0046] The training described in step S4 specifically includes the following steps:

[0047] The following formula is used as the total loss function for the training process:

[0048] In the formula This represents the total loss function value. This is the first weight value; The value of the Focal Loss function; As the second weight; This represents the mean squared error loss function value. This is the third weight value; This represents the value of the cross-entropy loss function.

[0049] This invention also provides a system for recognizing the wearing status of protective equipment for high-altitude operations in power systems, comprising a data acquisition module, a data processing module, a model building module, a model training module, and a status recognition module; the data acquisition module, data processing module, model building module, model training module, and status recognition module are connected in series; the data acquisition module is used to acquire image data of the wearing of protective equipment for high-altitude operations in power systems and upload the data information to the data processing module; the data processing module is used to preprocess the acquired image data according to the received data information to construct an image dataset and upload the data information to the model building module; the model building module is used to preprocess the acquired image data according to the received data information. Based on a dual-branch neural network and a graph attention network, an initial model for recognizing the wearing status of protective equipment for high-altitude operations in power systems is constructed, and the data information is uploaded to the model training module. The model training module is used to train the initial model for recognizing the wearing status of protective equipment for high-altitude operations in power systems using the constructed image dataset based on the received data information, thereby obtaining a model for recognizing the wearing status of protective equipment for high-altitude operations in power systems, and uploading the data information to the status recognition module. The status recognition module is used to recognize the wearing status of protective equipment during actual high-altitude operations in power systems based on the received data information and the obtained model for recognizing the wearing status of protective equipment for high-altitude operations in power systems.

[0050] The present invention provides a method and system for recognizing the wearing status of protective equipment for high-altitude operations in power systems. By acquiring and processing images of the wearing status of protective equipment for high-altitude operations in power systems to construct a dataset, and constructing and training a wearing status recognition network based on a dual-branch neural network and a graph attention network, the present invention can not only recognize the wearing status of protective equipment for high-altitude operations in power systems, but also has higher reliability, higher accuracy, and better recognition effect. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0052] Figure 2 This is a schematic diagram comparing performance indicators under occlusion conditions in an embodiment of the method of the present invention.

[0053] Figure 3 This is a schematic diagram illustrating the accuracy variation trend under different degrees of occlusion in an embodiment of the method of the present invention.

[0054] Figure 4 This is a schematic diagram illustrating the processing time variation trend under different degrees of occlusion in an embodiment of the method of the present invention.

[0055] Figure 5 This is a schematic diagram comparing the positioning errors of key structural points under different degrees of occlusion in an embodiment of the method of the present invention.

[0056] Figure 6 This is a schematic diagram comparing the error growth rate under different degrees of occlusion in an embodiment of the method of the present invention.

[0057] Figure 7 This is a schematic diagram comparing the cumulative probability under different degrees of occlusion in an embodiment of the method of the present invention.

[0058] Figure 8 This is a schematic diagram comparing errors under different degrees of occlusion in an embodiment of the method of the present invention.

[0059] Figure 9 This is a functional diagram of the system of the present invention. Detailed Implementation

[0060] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The method for identifying the wearing status of protective equipment for high-altitude operations in power systems disclosed in this invention includes the following steps:

[0061] S1. Acquire image data of the wearing of protective equipment for high-altitude operations in the power system;

[0062] S2. Preprocess the image data obtained in step S1 to construct an image dataset; specifically, this includes the following steps:

[0063] The data preprocessing includes data labeling and data cleaning;

[0064] Data annotation, including annotation of acquired image data;

[0065] Data cleaning includes image data filtering, image data enhancement, and cropping image data to a set size;

[0066] S3. Based on a dual-branch neural network and a graph attention network, construct an initial model for recognizing the wearing status of protective equipment for high-altitude operations in power systems; specifically including the following steps:

[0067] The initial model for identifying the wearing status of protective equipment for high-altitude operations in power systems consists of a mask / heatmap generation module, a graph construction module, a graph attention module, and a feature fusion module connected in series.

[0068] A mask / heatmap generation module is constructed based on a dual-branch neural network; the mask / heatmap generation module is used to simultaneously generate target segmentation masks and structural key point heatmaps of input image data;

[0069] The graph construction module is used to construct a scene relationship graph based on the target segmentation mask and structural key point heatmap of the obtained input image data.

[0070] A graph attention module is constructed based on a graph attention network; the graph attention module is used to process the constructed scene relationship graph and update node features;

[0071] The feature fusion module is used to perform multimodal feature fusion based on the output of the graph attention module and generate the final identification result of the wearing status of the high-altitude operation protective equipment in the power system.

[0072] In specific implementation, the mask / heatmap generation module based on a dual-branch neural network includes the following steps:

[0073] The mask / heatmap generation module includes a feature extraction module, a segmentation module, and a key point module;

[0074] A ResNet-50 network is used as the feature extraction module. The feature extraction module is used to extract global features of the input image and inputs the obtained global features into both the segmentation module and the key point module.

[0075] The segmentation module consists of three convolutional layers connected in sequence; the segmentation module is used to generate a target segmentation mask M for the input global features;

[0076] The keypoint module consists of three convolutional layers connected in sequence; the keypoint module is used to generate a structural keypoint heatmap H of the input global features;

[0077] In practice, the graph construction module's processing steps include the following:

[0078] The output of the mask / heatmap generation module includes a segmentation mask M and a key point heatmap H;

[0079] The key point heatmap is smoothed using a Gaussian kernel to better characterize the spatial distribution of protective equipment; the smoothing process is represented as follows:

[0080] In the formula Heatmap of key points after smoothing; The pixel coordinates on the image plane; Let i be the coordinates of the i-th key point; The standard deviation of the Gaussian kernel;

[0081] Based on the segmentation mask M and the smoothed keypoint heatmap Construct a scene relationship diagram , where node set V represents the detected key points and edge set E represents the spatial relationship between nodes;

[0082] Node features Composed of visual features and spatial coordinates, represented as ;in Key point coordinates; The visual feature extraction process includes: After passing through several convolutional layers, the feature mapping tensor FFF is obtained through a non-linear activation function, and then a region sampling function is applied. Processing is performed to complete the visual feature extraction process. This is a region sampling function; it can effectively capture the texture, edge, and semantic features of the neighborhood of key points, providing a high-quality visual description for subsequent node feature construction and graph attention update.

[0083] Edge features Represented as ,in Let be the Euclidean distance between node i and node j; Let be the angle between the line connecting node i and node j and the reference direction; It is a feature mapping function, and , For the set nonlinear activation function, As the first weight parameter, Here, b is the second weight parameter, and b is the bias term. This function maps spatial distance and angular direction information into low-dimensional edge feature representations, thereby effectively capturing the geometric relationships between keypoints and providing input for feature updates in the graph attention module.

[0084] In practice, the graph attention module's processing steps include the following:

[0085] A graph attention network is used as the graph attention module;

[0086] The graph attention network adopts the network structure proposed by Velickovic et al. in the paper "Graph Attention Networks" (ICLR2018);

[0087] A graph attention network is used to process the scene relationship graph and update node features; represented as:

[0088] In the formula For the first Layer node characteristics; The activation function is set; For nodes The set of neighboring nodes; For the first Attention weights of layers; For the first The parameter matrix to be learned for the layer;

[0089] The following formula is used to calculate the first... layer attention weights :

[0090] In the formula It is a linear rectified function with leakage parameters; The attention vector to be learned; This is for splicing operations.

[0091] The feature fusion module's processing procedure includes the following steps:

[0092] The processing procedure of the feature fusion module is represented by the following formula:

[0093] In the formula This is the output of the feature fusion module; The output of the graph attention module; For fusion function, , This is the first weight matrix. This is the second weight matrix. The bias term is used; through this fusion function, visual features and graph structure features can be weighted and combined in the same feature space to generate high-level fusion features that combine local texture information and global topological relationships of the image, effectively improving the robustness and accuracy of protective equipment wearing status recognition.

[0094] S4. Using the image dataset constructed in step S2, train the initial model for recognizing the wearing status of protective equipment for high-altitude operations in the power system constructed in step S3 to obtain the model for recognizing the wearing status of protective equipment for high-altitude operations in the power system; specifically including the following steps:

[0095] The following formula is used as the total loss function for the training process:

[0096] In the formula This represents the total loss function value. This is the first weight value; The value of the Focal Loss function; As the second weight; This represents the mean squared error loss function value. This is the third weight value; This represents the value of the cross-entropy loss function;

[0097] in, and The aim is to enable the model to accurately segment targets and locate key points from images; This is used to drive the model to learn and internalize the physical and geometric prior knowledge that high-altitude work protective equipment should follow, so that it can still make reliable judgments that meet the constraints when there are obstructions and changes in the viewing angle, and achieve the overall optimization goal.

[0098] S5. Using the power system high-altitude operation protective equipment wearing status recognition model obtained in step S4, the wearing status recognition of protective equipment during actual power system high-altitude operations is carried out; in specific applications, the power system high-altitude operation protective equipment wearing status recognition model obtained in step S4 can be deployed in the monitoring room of the construction site to realize real-time monitoring and recognition of the wearing status of high-altitude operation protective equipment.

[0099] The method of the present invention will be further described below with reference to an embodiment:

[0100] A comprehensive simulation environment was constructed: the simulation platform is based on MATLAB R2021b and PyTorch 1.9.0, and the hardware environment includes an Intel Xeon Gold 6248 processor, 256 GB DDR4 RAM, and 2 TB NVMe SSD storage. The model design fully considers the dynamic characteristics and occlusion factors of high-altitude operation scenarios, and the parameter settings strictly adhere to the technical specifications in the fields of computer vision and safety monitoring. Specifically, the dual-branch neural network uses ResNet-50 as the backbone network, with the segmentation branch and keypoint branch outputting 2-channel and 17-channel feature maps, respectively. The node feature dimension of the scene relationship graph construction module is set to 256, and the edge feature dimension is 128. The graph attention network is configured with 8 attention heads and a hidden layer dimension of 512. These parameter settings ensure that the model can effectively capture the complex topological relationships and occlusion situations in high-altitude operation scenarios.

[0101] like Figure 2 As shown, the performance metrics of the three methods were compared under 50% occlusion conditions. Figure 3 The generation was implemented using MATLAB code, and the reproducibility of the results was ensured by setting a random number seed. Three methods and three performance metrics were defined in the data preparation stage. Grouped bar charts were used in the plotting process, and error bars were added to reflect the dispersion of the data. The results show that the keypoint and graph network fusion method proposed in this application performs best in all metrics. Specifically, the recognition accuracy of this method reaches 95.32%, significantly higher than the 72.18% of the traditional segmentation method and the 68.43% of the keypoint detection method. Meanwhile, its false positive rate and false negative rate are 4.68% and 2.87%, respectively, far lower than the other two methods. This result fully demonstrates the significant advantages of the proposed method in the occlusion-resistant recognition task and verifies the effectiveness of fusing keypoint detection and graph networks.

[0102] Figure 3 and Figure 4 This further illustrates the comparison of accuracy and processing time for the three detection methods under different levels of occlusion. The figure was generated using MATLAB code and employs a dual y-axis design to simultaneously display the trends in accuracy and processing time. From... Figure 3 As can be seen, the accuracy of the proposed method remains superior as the occlusion intensity gradually increases from 0% to 80%, especially under high occlusion conditions. For example, at 80% occlusion, the accuracy of the proposed method remains at 71.28%, while the accuracy of the other two methods drops to 56.78% and 50.86%, respectively. This indicates that the proposed method effectively alleviates the information loss caused by occlusion by fusing keypoint information with graph structure features, thereby improving the model's ability to identify occluded targets. Figure 4 The results show that the method presented in this application also has a significant advantage in processing time. As the degree of occlusion increases, the processing time decreases from the initial 23.45 milliseconds to 15.05 milliseconds. This efficiency is mainly attributed to the optimized design of the method in this application in feature extraction and graph structure reasoning, which reduces unnecessary computational redundancy.

[0103] To more comprehensively evaluate the performance of the method in this application, Figures 5-8 This paper presents a multi-dimensional comparative analysis of the location errors of structural key points under different degrees of occlusion. The figure, generated using MATLAB code, contains four sub-figures, each illustrating the error characteristics of the three methods from a different perspective. Figure 5 The results demonstrate the trend of average keypoint localization error as the degree of occlusion increases. The results show that the method described in this application maintains the lowest localization error across the entire occlusion range. For example, at 90% occlusion, the error of the method described in this application is 12.26 pixels, while the other two methods are 25.18 pixels and 24.11 pixels, respectively. Figure 6 Further analysis of the error growth rate revealed that the method in this application exhibited the most gradual growth trend, with a maximum growth rate of approximately 1.18 pixels / 5%, which is significantly lower than other methods. Figure 7 and Figure 8 The superiority of the proposed method was statistically verified by using the cumulative error distribution and box plots, demonstrating that it is more concentrated and stable in overall performance and has better predictive consistency and anti-interference ability.

[0104] like Figure 9The diagram shows the functional modules of the system of the present invention: The system disclosed in this invention for recognizing the wearing status of protective equipment for high-altitude operations in power systems includes a data acquisition module, a data processing module, a model building module, a model training module, and a status recognition module; these modules are connected in series. The data acquisition module acquires image data of the wearing of protective equipment for high-altitude operations in power systems and uploads the data to the data processing module. The data processing module preprocesses the acquired image data according to the received data to construct an image dataset and uploads the data to the model building module. The model building module preprocesses the acquired image data according to the received data to construct an image dataset and uploads the data to the model building module. The model building module is used to... The received data is used to construct an initial model for recognizing the wearing status of protective equipment for high-altitude operations in power systems, based on a dual-branch neural network and a graph attention network. This data is then uploaded to the model training module. The model training module trains the initial model using the constructed image dataset, based on the received data, to obtain a model for recognizing the wearing status of protective equipment for high-altitude operations in power systems. This model is then uploaded to the status recognition module. The status recognition module uses the obtained model to recognize the wearing status of protective equipment during actual high-altitude operations in power systems, based on the received data.

Claims

1. A method for identifying the wearing state of power system aerial work protective equipment, comprising the following steps: S1. Obtaining image data of the wearing of power system aerial work protective equipment; S2. Preprocessing the image data obtained in step S1 to construct an image data set; S3. Constructing an initial model for identifying the wearing state of power system aerial work protective equipment based on a double-branch neural network and a graph attention network; S4. Training the initial model for identifying the wearing state of power system aerial work protective equipment constructed in step S3 using the image data set constructed in step S2 to obtain a model for identifying the wearing state of power system aerial work protective equipment; S5. Using the model for identifying the wearing state of power system aerial work protective equipment obtained in step S4 to identify the wearing state of protective equipment during actual power system aerial work.

2. The power system work clothes state recognition method according to claim 1, characterized in that The preprocessing of the image data obtained in step S1 in step S2 comprises the following steps: The data preprocessing comprises data labeling and data cleaning; The data labeling comprises labeling the obtained image data; The data cleaning comprises filtering the image data, enhancing the image data, and cropping the image data to a set size.

3. The power system work clothes state recognition method according to claim 2, characterized by The construction of the initial model for identifying the wearing state of power system aerial work protective equipment based on a double-branch neural network and a graph attention network in step S3 comprises the following steps: The initial model for identifying the wearing state of power system aerial work protective equipment comprises a mask / heat map generation module, a graph construction module, a graph attention module, and a feature fusion module connected in sequence; The mask / heat map generation module is constructed based on a double-branch neural network and is used to simultaneously generate a target segmentation mask and a structure key point heat map of the input image data; The graph construction module is used to construct a scene relationship graph based on the target segmentation mask and the structure key point heat map of the input image data obtained; The graph attention module is constructed based on a graph attention network and is used to process the constructed scene relationship graph and update the node features; The feature fusion module is used to perform multi-modal feature fusion based on the output of the graph attention module and generate the final identification result of the wearing state of power system aerial work protective equipment.

4. The power system work clothes state recognition method according to claim 3, characterized by The construction of the mask / heat map generation module based on a double-branch neural network comprises the following steps: The mask / heat map generation module comprises a feature extraction module, a segmentation module, and a key point module; ResNet-50 network is used as the feature extraction module, which is used to extract the global features of the input image and simultaneously input the obtained global features to the segmentation module and the key point module; The segmentation module comprises three layers of convolutional layers connected in sequence, which is used to generate a target segmentation mask M of the input global features; The key point module comprises three layers of convolutional layers connected in sequence, which is used to generate a structure key point heat map H of the input global features.

5. The power system work clothes state recognition method according to claim 4, characterized in that The processing process of the graph construction module comprises the following steps: The output of the mask / heat map generation module comprises a segmentation mask M and a key point heat map H; The key point heat map is smoothed by a Gaussian kernel to better represent the spatial distribution of the protective equipment. The smoothing process is represented as: In the formula The key point heat map after smoothing processing; is the pixel coordinate position on the image plane; is the coordinate of the i-th key point; is the standard deviation of the Gaussian kernel; According to the segmentation mask M and the smoothed key point heat map , a scene relation graph is constructed , wherein a node set V represents the detected key points, and an edge set E represents the spatial relationship between the nodes. Node feature consisting of visual features and spatial coordinates, denoted as ; wherein is a key point coordinate; denotes a visual feature extraction process, specifically comprising: is sequentially processed through a plurality of convolutional layers, and then processed through a nonlinear activation function to obtain a feature mapping tensor FFF, and then processed through a region sampling function to complete the visual feature extraction process, is a region sampling function; edge feature is represented as where denotes the Euclidean distance between node i and node j; is the angle between the edge between node i and node j and the reference direction; is a feature mapping function, and , is a set non-linear activation function, is a first weight parameter, is a second weight parameter, and b is a bias term.

6. The power system work clothes state recognition method according to claim 5, characterized by The processing procedure of the graph attention module, specifically comprising the following steps: The graph attention network is adopted as the graph attention module; The graph attention network is adopted to process the scene relation graph and update the node features, which is represented as: In the formula is the first layer node feature; is the set activation function; is the node neighbor node set; is the first layer attention weight; is the first layer to-be-learned parameter matrix; The attention weight of the layer is calculated using the following equation layer : wherein is a linear rectifying function with a leakage parameter; is an attention vector to be learned; is a concatenation operation. The processing procedure of the feature fusion module, specifically comprising the following steps: The processing procedure of the feature fusion module is represented by the following formula: wherein is the output of the feature fusion module; is the output of the graph attention module; is a fusion function, , is a first weight matrix, is a second weight matrix, is a bias term.

7. The power system work clothes state recognition method according to claim 6, characterized in that The training of step S4 specifically comprises the following steps: The following formula is adopted as the total loss function of the training procedure: In the formula is a total loss function value; is a first weight value; is a Focal Loss loss function value; is a second weight; is a mean square error loss function value; is a third weight value; is a cross-entropy loss function value.

8. A system for implementing the method for recognizing the wearing state of the electric power system aerial work protective equipment according to any one of claims 1 to 7, characterized by The system comprises a data acquisition module, a data processing module, a model construction module, a model training module and a state recognition module. The data acquisition module, the data processing module, the model construction module, the model training module and the state recognition module are sequentially connected. The data acquisition module is configured to acquire image data of the protective equipment for high-altitude work of a power system, and upload the data information to the data processing module. The data processing module is configured to pre-process the acquired image data according to the received data information, to construct an image data set, and upload the data information to the model construction module. The model construction module is configured to construct an initial model for recognizing the wearing state of the protective equipment for high-altitude work of the power system based on the double-branch neural network and the graph attention network according to the received data information, and upload the data information to the model training module. The model training module is configured to train the constructed initial model for recognizing the wearing state of the protective equipment for high-altitude work of the power system by using the constructed image data set, to obtain a model for recognizing the wearing state of the protective equipment for high-altitude work of the power system, and upload the data information to the state recognition module. The state recognition module is configured to recognize the wearing state of the protective equipment in the actual high-altitude work process of the power system by using the obtained model for recognizing the wearing state of the protective equipment for high-altitude work of the power system.

Images