Distribution network operation anomaly detection and safety protection multi-scale feature recognition method based on weak supervision

Abstract

The invention discloses a distribution network operation anomaly detection and safety protection multi-scale feature recognition method based on weak supervision, and aims to solve the problems of non-standard wearing, high detection cost, poor environmental adaptability, waste of computing resources and poor real-time performance of safety protection equipment on a distribution network operation site. According to the invention, an anomaly detection model based on multi-instance learning is established, and video spatio-temporal features are extracted by using a 3D convolutional network, so that abnormal picture monitoring under weak supervision is realized; constructing an image feature enhancement processing model, and improving image details through gamma transformation and graying processing; designing a multi-scale feature extraction and convolution recognition model, and recognizing the wearing state of the safety protection equipment in combination with CNN and GraphSAGE algorithms of mixed channel attention; according to the method, the data labeling cost is reduced, the environmental adaptability and the recognition efficiency are improved, comprehensive, efficient and intelligent safety protection management on a power distribution network operation site is realized, potential safety hazards are found and corrected in time, and the safety and the efficiency of power operation are improved.

Description

Technical Field

[0001] The present invention relates to the field of distribution network and artificial intelligence technology, and in particular to a multi-scale feature recognition method for distribution network operation anomaly detection and safety protection based on weak supervision. Background Art

[0002] The distribution network operation environment is complex, facilities are densely packed, and the operation scope is narrow. These characteristics bring great challenges to the safety of on-site workers. During the non-stop operation, high-intensity electric fields, transient electric shocks and steady-state electric shocks, as well as air gap discharges that may be caused by misoperation during the operation, all seriously threaten the safety of workers. Therefore, safety protection for non-stop operation becomes the key to ensure the smooth implementation of the operation.

[0003] At present, a variety of safety protection equipment is widely used in the power industry, such as insulating helmets, insulating protective clothing, insulating protective gloves, etc., to ensure the safety of workers. However, in the actual operation process, there are still cases where workers do not wear safety protection equipment properly or do not wear them, which can easily lead to safety accidents. In order to strengthen the safety control of the operation site and improve the level of production safety, traditional safety supervision methods mainly rely on manual inspections and on-site supervision, but this method has problems such as high labor costs, low supervision efficiency, and slow response speed, which is difficult to meet the needs of modern power operation safety management.

[0004] With the rapid development of artificial intelligence technology, target detection methods have gradually been introduced into the safety management of power operation sites, providing new ideas and technical routes for the safety management of non-stop operation sites. Using target detection technology, intelligent supervision of the entire operation process can be achieved, and irregular behaviors of operators can be discovered and corrected in a timely manner, thereby effectively ensuring the safety of operators and the entire operation task. However, the existing methods for identifying power safety protection equipment still have some shortcomings: High data labeling cost: Training recognition models requires a large number of labeled data sets. Collecting images that meet the requirements and labeling and classifying them consumes a lot of labor costs and time.

[0005] Poor environmental adaptability: Power operations are usually carried out outdoors. Affected by factors such as natural light brightness and weather conditions, the images recorded by photographic equipment may not be clear enough. In addition, there are many pieces of equipment and personnel at the work site, and the scene is complex, making it difficult for the recognition model to run stably.

[0006] Waste of computing resources: Existing recognition models usually process all elements in the entire image, but in actual operation videos, only a few key frames contain abnormal information. This results in a large amount of computing resources being wasted on processing redundant information, reducing the operating efficiency of the model.

[0007] Poor real-time performance: Power operations have high requirements for real-time performance, but existing recognition models usually adopt a larger model structure to meet accuracy requirements. As a result, when embedded in smart devices such as drones, the hardware performance cannot support the real-time operation of the model.

[0008] In response to the above problems, the present invention proposes a distribution network operation anomaly detection and safety protection multi-scale feature recognition method based on weak supervision, aiming to improve the safety management and control level of power operation sites, reduce data annotation costs, enhance environmental adaptability, reduce the waste of computing resources, and improve the real-time performance of the model. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a distribution network operation anomaly detection and safety protection multi-scale feature recognition method based on weak supervision, so as to solve the technical problems of high detection cost, poor environmental adaptability, waste of computing resources and poor real-time performance of improperly wearing safety protection equipment at the distribution network operation site.

[0010] In order to solve the above technical problems, the technical solution adopted by the present invention is: a distribution network operation anomaly detection and safety protection multi-scale feature recognition method based on weak supervision, comprising the following steps: Step 1: Establish an abnormal image recognition model; Step 2: Build an image feature enhancement processing model; Step 3: Build a multi-scale feature extraction and convolutional recognition model.

[0011] In a preferred embodiment, Step 1 comprises the following steps: Step 1.1: The operation video shot by the monitoring system is used as the detection target. The video is divided into positive and negative packets according to whether the protective gear is not worn or worn incorrectly. The positive packet contains abnormal clips, while the negative packet does not contain abnormal clips. Step 1.2: During the training phase, the video is divided into a fixed number of segments, each segment generates an instance, and the video set is defined as an instance bag. Each bag contains several instances and is assigned a label. The positive bag label is abnormal and the negative bag label is normal. Step 1.3: Extract the spatiotemporal feature information of the instance through the 3D convolutional network, use the loss function to optimize the model, and realize abnormal image monitoring based on weak supervision.

[0012] In the preferred solution, the 3D convolutional network in Step 1.3 includes an input layer, a 3D convolutional layer, a 3D pooling layer, a fully connected layer and an output layer. Multiple continuous frames are used as input information, and spatiotemporal features are extracted through the 3D convolutional layer. The 3D pooling layer reduces the time and space dimensions, and the fully connected layer and the output layer output anomaly scores.

[0013] In a preferred solution, the expression of the 3D convolution is as follows: (1) In the formula, The convolution kernel of the 3D convolution layer is; is the depth of the convolution kernel in the time dimension; is the height and width of the convolution kernel; is the number of input channels; is the number of output channels; is the input video tensor; for time of Channel output characteristics, is the image time series; is the vertical position of the image; is the horizontal position of the image; is the vertical position of the convolution kernel; is the lateral position of the convolution kernel; is the time position of the convolution kernel; Keep the key part of the timing information and perform average pooling on the output features: (2) is the average pooling feature map of the output; is the input feature map; is the pooling window size in time, height, and width for each instance Assign a score After the training is completed, the instance score is judged according to the instance score. The instance with a higher score is considered to be an abnormal instance. , all instances have low scores; for positive packets , at least one instance score is larger, and the maximum instance score in the positive bag is higher than that in the negative bag, that is: (3) In the formula, is the set of positive packets containing abnormal instances, is a set of negative packets that do not contain abnormal instances, is the fraction of instances in the positive bag; is the fraction of instances in the negative bag; is a function to be learned that maps the anomaly score of a fragment, ranging from 0 to 1. In order to maximize the match between the bag label and the instance score, the following loss function is used : (4) In the formula, For video tags; Representation instance score; is the sigmoid function, which is used to convert scores into probabilities; Indicates the highest instance score in the bag; the MIL model is used to monitor the video clips for incorrect or non-worn safety protection and output the instances in image format.

[0014] In a preferred embodiment, Step 2 comprises the following steps: Step 2.1: Use the gamma function to adjust the brightness and contrast of the image to be identified and enhance the image details; Step 2.2: Grayscale the image to reduce data processing and storage space and extract image structure information.

[0015] In the preferred solution, the mathematical formula of the gamma function transformation in Step 2.1 is: (5) In the formula, is the original image at position The gray value of the pixel at ; is the pixel gray value at the same position in the transformed image, is a constant used to adjust the overall brightness; is the gamma index; after gamma transformation, the bright parts of the image will be compressed, the dark parts will be enhanced, and the details will become more obvious in the light and dark areas.

[0016] In the preferred solution, the grayscale processing of the image in Step 2.2 is to map the three-dimensional color data to a one-dimensional grayscale value. In the grayscale image, the value of each pixel is a scalar, which represents its brightness or grayscale value, reducing the storage space required for data processing and extracting the image structure information. In order not to affect the original features of the image, the weighted average method is used to calculate the grayscale value. The standard grayscale The formula is as follows: (6) In the formula, is the value of the red channel, is the value of the green channel, is the value of the blue channel.

[0017] In a preferred solution, Step 3 comprises the following steps: Step 3.1: Use a convolutional neural network with mixed channel attention as a multi-scale feature extraction model, and use convolution kernels of different scales to perform convolution operations on the input feature map to obtain a multi-scale feature map; Step 3.2: Introduce the channel attention mechanism to dynamically adjust the weights of different channels of the feature map; Step 3.3: Input the feature graph output by the convolutional neural network into the GraphSAGE algorithm to construct the graph structure, and obtain the final embedding vector of each node by aggregating and updating the features of neighboring nodes; Step 3.4: According to the final embedded vector, identify the wearing status of safety protection equipment such as hard hats, insulating protective clothing, insulating capes, and insulating sleeves.

[0018] In the preferred solution, the multi-scale feature extraction in Step 3.1 is to perform convolution operations on the input feature map at different scales to obtain feature maps of multiple scales. The convolution process is as follows: (7) In the formula, is the input image, is the convolution kernel, It is the feature map formed after convolution. The input image is The pixel value of the position, Represents the convolution kernel and the input image Relative position relationship; Convolution operations are performed using convolution kernels of (3×3), (5×5), and (7×7) to obtain feature maps of different scales: (8) In the formula, Indicates the use The convolution kernel is used for convolution. Indicates the use The convolution kernel is used for convolution. Indicates the use The convolution kernel is used for convolution. , , The output feature maps are spliced ​​and integrated to form a comprehensive feature map containing multi-scale information. : (9) Then perform maximum pooling on the feature map to get the final feature map : (10).

[0019] In a preferred solution, the specific steps of Step 3.2 include: The input feature map Perform global average pooling to compress the spatial dimension. For each channel, calculate the vector after global average pooling. : (11) In the formula, , and Represent the height, width and number of channels respectively, is the vertical position of the image, is the horizontal position of the image, is the specific position in the feature map of the input channel c; finally, a C-dimensional vector is obtained , contains the feature information of each channel; after two layers of fully connected layer networks and activation functions, the attention weight of each channel is generated : (12) in , is the weight matrix of the fully connected layer, It is the Sigmoid activation function, which is used to map the weight value to [0, 1]; It is a nonlinear activation function. When the input value is ≤0, the output is 0. When the output value is >0, the input value is output. After being processed by the channel attention mechanism module, the channel weight vector is obtained. , and apply it to the original feature map On the top, we get the weighted feature map : (13).

[0020] In the preferred solution, the GraphSAGE algorithm in Step 3.3 obtains the final embedding vector of each node by aggregating and updating the features of neighboring nodes, and uses it as the input of the classification task to complete the identification of safety protection equipment.

[0021] In the preferred solution, the specific steps of Step 3.3 are as follows: After CNN processes the original image, its output feature map is input into the GNN network and converted into a graph structure. For each node in the graph GraphSAGE will select from its neighbor set Randomly sample K subsets from the set to form a new set , the node features are updated as follows: (14) Is a node In the The feature representation of the layer, Is a node exist The feature representation of the layer, is the learnable weight matrix, is the activation function, For Node The set of neighbors, is an aggregation function. After the collection is completed, mean aggregation is performed based on the feature information of neighbor nodes: (15) in is the number of iteration steps, Neighbor node Features, is the aggregated neighbor feature, For Node No. The node features are combined with the aggregated neighbor features, and the node features are updated through linear transformation and activation function. The update formula is: (16) in, is the learnable weight matrix of the current layer, is the activation function, It means that the node's own features and the features of neighboring nodes are spliced ​​together. Each layer (or iteration step) of GraphSAGE can obtain higher-order neighborhood information of the node. The features of each node contain information from its neighboring nodes. After multiple layers of aggregation and updating, the final embedding vector of each node is obtained. , and finally use it as the input of the classification task to complete the identification of safety protection equipment such as safety helmets, insulating capes, protective clothing, insulating sleeves, etc.

[0022] The weakly supervised distribution network operation anomaly detection and safety protection multi-scale feature recognition method provided by the present invention has the following beneficial effects: 1. The present invention solves the technical problems of improper wearing of safety protection equipment, high detection cost, poor environmental adaptability, waste of computing resources and poor real-time performance at the distribution network operation site; 2. The weakly supervised learning adopted by the present invention reduces the data annotation cost and improves the training efficiency and scalability of the model by using weakly labeled data for training; 3. The image feature enhancement adopted by the present invention uses the gamma function to adjust the image brightness and contrast, as well as the grayscale processing, which solves the influence of the light and dark environment and complex scenes of the power operation environment on the recognition accuracy, and enhances the recognition accuracy in low brightness and complex scenes; 4. The multi-scale feature extraction adopted by the present invention performs convolution operations on the input feature map at different scales to obtain feature maps of multiple scales, thereby improving the model's ability to capture details; 5. The present invention introduces a channel attention mechanism to dynamically adjust the weights of different channels of the feature map, so that the model pays more attention to important features and reduces the waste of computing resources; 6. The GraphSAGE feature recognition of the present invention further processes the extracted features using graph neural networks, thereby improving the recognition accuracy of safety protection equipment; 7. The present invention uses lightweight model design to accurately identify abnormal images, concentrates computing resources on images that need to be identified, reduces computing and storage pressure, improves model recognition efficiency, and meets real-time requirements; 8. Through weakly supervised learning and multi-scale feature extraction, the present invention can more accurately detect abnormal images at the work site and improve the accuracy of abnormality detection; 9. The introduction of image feature enhancement and channel attention mechanism in the present invention enables the model to maintain a high recognition accuracy in low brightness and complex scenes, thereby enhancing the robustness of the model; 10. The multi-scale feature extraction and channel attention mechanism of the present invention reduces the waste of computing resources, improves the operation efficiency of the model, and reduces the computing cost; 11. The method of the present invention provides an intelligent management tool for the distribution network operation site, which can timely discover and correct potential safety hazards, improve the safety of operators, and realize intelligent management; 12. The present invention uses weakly supervised learning, image processing, deep learning, and graph neural network technologies to achieve anomaly detection in power operations and multi-scale feature recognition of safety protection equipment. Through intelligent means, the safety and efficiency of power operations are improved, and the risk of accidents caused by irregular operations or safety hazards is reduced; 13. The present invention adopts weak supervision recognition. By marking the video itself and a small number of images containing anomalies, the C3D network convolution mines the abnormal video features to achieve the function of identifying abnormal instances, effectively reducing the cost of manual image annotation, reducing the cost of data annotation, and achieving high efficiency and high scalability; 14. The present invention enhances the image features of the image files to be identified and detected, uses the gamma function to adjust the brightness and contrast of the image to be identified, highlights the details and features, and performs nonlinear transformation on the grayscale value of the image, thereby solving the problem of difficulty in identification when the brightness of the power operation scene is low, reducing the influence of the environment light and the complex operation scene on the accuracy of the recognition model, and overcoming the influence of the environment light and the complex scene on the recognition accuracy; 15. The present invention introduces a channel attention mechanism to dynamically adjust the weights of different channels of the feature map, so that the model pays more attention to important features and reduces the waste of computing resources. During feature recognition, the node's own features are merged with the aggregated neighbor features, and the node features are updated through linear transformation and activation function to achieve feature splicing, further improving the robustness of the model. 16. The method of the present invention also incorporates edge computing technology, optimizes the allocation of computing resources, ensures real-time monitoring, and achieves comprehensive, efficient and intelligent safety, protection and management of distribution network operation sites through the collaborative work of the cloud and the edge. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be further described below in conjunction with the accompanying drawings and implementation examples: Figure 1 It is a technical flow chart of the present invention; Figure 2 It is a schematic diagram of the structure of the multi-scale feature extraction model of the present invention; Figure 3 This is a schematic diagram of the GraphSAGE feature recognition model of the present invention. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments: Example 1 The weakly supervised distribution network operation anomaly detection and safety protection multi-scale feature recognition method includes the following steps: Step 1: Establish an abnormal image recognition model; Step 2: Build an image feature enhancement processing model; Step 3: Build a multi-scale feature extraction and convolutional recognition model.

[0025] In this embodiment, Step 1 includes the following steps: Step 1.1: The operation video shot by the monitoring system is used as the detection target. The video is divided into positive and negative packets according to whether the protective gear is not worn or worn incorrectly. The positive packet contains abnormal clips, while the negative packet does not contain abnormal clips. Step 1.2: During the training phase, the video is divided into a fixed number of segments, each segment generates an instance, and the video set is defined as an instance bag. Each bag contains several instances and is assigned a label. The positive bag label is abnormal and the negative bag label is normal. Step 1.3: Extract the spatiotemporal feature information of the instance through the 3D (Three Dimensions) convolutional network, and use the loss function to optimize the model to achieve abnormal image monitoring based on weak supervision.

[0026] Furthermore, the 3D convolutional network in Step 1.3 includes an input layer, a 3D convolutional layer, a 3D pooling layer, a fully connected layer and an output layer. Multiple continuous frames are used as input information, and spatiotemporal features are extracted through the 3D convolutional layer. The 3D pooling layer reduces the time and space dimensions, and the fully connected layer and the output layer output anomaly scores.

[0027] Furthermore, the expression of the 3D convolution is as follows: (1) In the formula, The convolution kernel of the 3D convolution layer is; is the depth of the convolution kernel in the time dimension; is the height and width of the convolution kernel; is the number of input channels; is the number of output channels; is the input video tensor; for time of Channel output characteristics, is the image time series; is the vertical position of the image; is the horizontal position of the image; is the vertical position of the convolution kernel; is the lateral position of the convolution kernel; is the time position of the convolution kernel; Keep the key part of the timing information and perform average pooling on the output features: (2) is the average pooling feature map of the output; is the input feature map; is the pooling window size in time, height, and width for each instance Assign a score After the training is completed, the instance score is judged according to the instance score. The instance with a higher score is considered to be an abnormal instance. , all instances have low scores; for positive packets , at least one instance score is larger, and the maximum instance score in the positive bag is higher than that in the negative bag, that is: (3) In the formula, is the set of positive packets containing abnormal instances, is a set of negative packets that do not contain abnormal instances, is the fraction of instances in the positive bag; is the fraction of instances in the negative bag; is a function to be learned that maps the anomaly score of a fragment, ranging from 0 to 1. In order to maximize the match between the bag label and the instance score, the following loss function is used : (4) In the formula, For video tags; Representation instance score; is the sigmoid function, which is used to convert scores into probabilities; Indicates the highest instance score in the bag; the MIL (Multiple Instance Learning) model is used to monitor video clips for incorrect or non-worn safety protection and output the instances in image format.

[0028] Furthermore, the Step 2 includes the following steps: Step 2.1: Use the gamma function to adjust the brightness and contrast of the image to be identified and enhance the image details; Step 2.2: Grayscale the image to reduce data processing and storage space and extract image structure information.

[0029] Furthermore, the mathematical formula of the gamma function transformation in Step 2.1 is: (5) In the formula, is the original image at position The grayscale value of the pixel at is usually in the range of [0, 255]. is the pixel gray value at the same position in the transformed image, Is a constant used to adjust the overall brightness. This invention sets it to 1 (it can be adjusted according to the overall brightness requirements). It is the gamma index, which is set to 1.5 in this embodiment. After the gamma transformation, the bright part of the image will be compressed, the dark part will be enhanced, and the details will become more obvious in the light and dark areas.

[0030] Furthermore, the grayscale processing of the image in Step 2.2 is to map the three-dimensional color data to a one-dimensional grayscale value. In the grayscale image, the value of each pixel is a scalar, which represents its brightness or grayscale value, reducing the storage space required for data processing and extracting image structure information. In order not to affect the original features of the image, the weighted average method is used to calculate the grayscale value. The standard grayscale The formula is as follows: (6) In the formula, is the value of the red channel, is the value of the green channel, is the value of the blue channel.

[0031] Furthermore, the Step 3 includes the following steps: Step 3.1: Use a convolutional neural network with mixed channel attention as a multi-scale feature extraction model, and use convolution kernels of different scales to perform convolution operations on the input feature map to obtain a multi-scale feature map; Step 3.2: Introduce the channel attention mechanism to dynamically adjust the weights of different channels of the feature map; Step 3.3: Input the feature graph output by the convolutional neural network into the GraphSAGE (Graph Sampling and Aggregation) algorithm to construct the graph structure. By aggregating and updating the features of neighboring nodes, the final embedding vector of each node is obtained. Step 3.4: According to the final embedded vector, identify the wearing status of safety protection equipment such as hard hats, insulating protective clothing, insulating capes, and insulating sleeves.

[0032] Furthermore, the convolutional neural network of Step 3.1 uses (3×3), (5×5), and (7×7) convolution kernels to perform convolution operations to obtain feature maps at different scales, and generates channel attention weights through global average pooling and two-layer fully connected layer networks.

[0033] Furthermore, the GraphSAGE algorithm in Step 3.3 obtains the final embedding vector of each node by aggregating and updating the features of neighboring nodes, and uses it as the input of the classification task to complete the identification of safety protection equipment.

[0034] Example 2 In another preferred embodiment, based on the above-mentioned embodiment 1, this embodiment elaborates on the anomaly detection based on weakly supervised learning in the distribution network operation anomaly detection and safety protection multi-scale feature recognition method based on weak supervision, which specifically includes the following steps: 1. Data preparation Surveillance videos of power operation sites are selected as data sources. These videos contain normal operations and abnormal events (such as workers not wearing safety helmets, illegal operations, etc.). The videos are cut into multiple video instances by time period, and a video-level label is assigned to each video instance to indicate whether the video instance contains abnormal events.

[0035] 2. 3D convolution feature extraction The C3D network (Convolutional 3D) is used to extract spatiotemporal features of each video instance. The C3D network contains multiple three-dimensional convolutional layers, pooling layers, and fully connected layers, which can capture dynamic information in the video. After the video instance is input into the C3D network, the feature vector of each video instance is obtained.

[0036] 3. Construction of multi-instance learning framework A multi-instance learning framework is constructed, and the feature vector of each video instance is used as the input of the MIL framework; the MIL framework includes a bag generation module and a classifier module; the bag generation module combines the feature vectors of multiple video instances into a bag, and the classifier module classifies the bag to determine whether the bag contains abnormal events.

[0037] 4. Model training and optimization The MIL framework is trained using a dataset with video-level labels. During the training process, the parameters of the classifier are optimized so that the MIL framework can accurately determine whether a package contains abnormal events. After the training is completed, an anomaly detection model based on weakly supervised learning is obtained.

[0038] 5. Anomaly Detection The new power operation site monitoring video is input into the trained anomaly detection model, and the model will output the judgment result of whether each video instance contains abnormal events; based on these judgment results, anomaly detection of the power operation site can be achieved.

[0039] Example 3 In another preferred embodiment, based on the above-mentioned embodiment 1, this embodiment elaborates on the multi-scale feature recognition in the multi-scale feature recognition method for abnormal detection and safety protection of distribution network operation based on weak supervision, which specifically includes the following steps: 1. Image preprocessing The monitoring video of the power operation site is processed frame by frame to obtain a series of image frames; each frame of the image is subjected to gamma transformation and grayscale processing to enhance the image features.

[0040] 2. Multi-scale feature extraction A convolutional neural network (CNN) with mixed channel attention is used to extract multi-scale features from preprocessed image frames; the CNN network contains multiple convolutional layers, pooling layers, and fully connected layers, where convolutional layers use convolution kernels of different sizes to capture features of different scales. Through the channel attention mechanism, the output of the convolutional layer is weighted to highlight important features.

[0041] 3. Graph Neural Network Feature Recognition The multi-scale feature map extracted by CNN is converted into a graph structure, in which each feature point is regarded as a node of the graph. The GraphSAGE algorithm is used to perform feature convolution recognition on the graph, and the final feature representation of each node is obtained by aggregating and updating the node features.

[0042] 4. Security protection identification According to the final feature representation of the node, a classifier (such as support vector machine, neural network, etc.) is used to classify the node to determine whether it belongs to safety protection related features (such as safety helmets, safety belts, etc.); based on the classification results, safety protection identification of the power operation site can be achieved.

[0043] Example 4 In another preferred embodiment, based on the above embodiments 1, 2, and 3, this embodiment elaborates on the two-stage recognition process in the multi-scale feature recognition method for abnormal detection and safety protection of distribution network operations based on weak supervision, which specifically includes the following steps: 1. Anomaly detection stage Use the anomaly detection model based on weakly supervised learning in Example 3 to perform anomaly detection on the surveillance video of the power operation site; cut the video into multiple video instances and input them into the model for judgment; if the model determines that a video instance contains an abnormal event, mark the video instance as a suspected abnormality.

[0044] 2. Fine Identification Stage The video instances marked as suspected abnormalities are further processed; the multi-scale feature recognition method in Example 4 is used to perform security protection recognition on each frame of the suspected abnormal video instance; based on the recognition results, the specific type and location of the abnormal event can be determined.

[0045] 3. Result output and alarm The recognition results are displayed to the user in a visual manner, such as marking the location and type of abnormal events; at the same time, when serious safety hazards are identified, the alarm mechanism is triggered and relevant personnel are notified in a timely manner for processing.

[0046] Example 5 In another preferred embodiment, based on the above-mentioned embodiments 1, 2, 3, and 4, this embodiment provides a specific application of a method for abnormal detection of power operations and multi-scale feature recognition for safety protection based on weakly supervised learning.

[0047] First, the weakly supervised learning model is trained. A small number of labeled images of power operation sites are selected as the training set, and a large number of unlabeled images are collected as auxiliary training data. A weakly supervised learning algorithm, such as semi-supervised learning or self-supervised learning, is used to preliminarily classify the training set, and the classification ability of the model is continuously improved by mining the potential information in the unlabeled data.

[0048] Next, multi-scale feature extraction is performed. For the input power operation site images, MSCNN (Multi-Scale Convolutional Neural Network) is used for feature extraction; MSCNN contains multiple convolutional layers of different scales, and each convolutional layer can extract feature information of different scales; through MSCNN, we can obtain the feature representation of the image at different scales, thereby capturing the detailed information in the image more comprehensively.

[0049] Then, the graph neural network is combined for feature fusion and classification. The multi-scale features extracted by MSCNN are input into GNN (Graph Neural Network), and the nodes and edges of GNN are used to represent the relationship between features and perform feature fusion. Finally, the fused features are classified by the GNN classifier to determine whether there are any abnormalities at the power operation site and whether the safety protection equipment is standardized.

[0050] In the specific implementation, we can adjust the network structure and parameter settings of MSCNN and GNN according to actual needs to obtain better recognition effects; in addition, we can also use data enhancement, transfer learning and other technologies to further improve the generalization ability and recognition accuracy of the model.

[0051] Example 6 In another preferred embodiment, based on embodiments 1, 2, 3, 4, and 5, this embodiment optimizes the multi-scale feature extraction part.

[0052] In Example 5, we used MSCNN for multi-scale feature extraction; in this example, we introduced a more advanced multi-scale feature extraction method - Pyramid Convolutional Neural Network (PCNN).

[0053] PCNN realizes multi-scale feature extraction of images by constructing multiple convolutional layers of different scales and combining them in the form of a pyramid. Compared with MSCNN, PCNN can more effectively capture multi-scale features in images and improve the accuracy and robustness of feature extraction.

[0054] In the specific implementation, we replace the MSCNN in Example 5 with PCNN, and keep other parts unchanged; through experimental verification, the recognition effect of this embodiment is better than that of Example 5, which further proves the effectiveness and innovation of the present invention.

[0055] Compared with the closest existing technology, the main difference of the present invention lies in the use of a method that combines weakly supervised learning algorithms and multi-scale feature extraction technology to achieve abnormal detection of power operations and multi-scale feature recognition of safety protection under limited labeled data; among them, weakly supervised learning algorithms can reduce dependence on labeled data and improve the generalization ability of the model; and multi-scale feature extraction technology can more comprehensively capture the feature information in the image and improve the recognition accuracy; in addition, the present invention also combines graph neural networks for feature fusion and classification, further enhancing the recognition ability of the model.

[0056] Example 7 In another preferred embodiment, based on embodiments 1, 2, 3, 4, 5, and 6, this embodiment further illustrates the technical solution of the multi-scale feature recognition method for abnormal detection and safety protection of distribution network operations based on weak supervision of the present invention: 1. Establish an abnormal image recognition model: The anomaly detection method based on multi-instance learning is a strategy that uses weak labels (i.e. partial labels) to perform anomaly detection. In multi-instance learning, labels are usually assigned to a bag of instances rather than individual instances (data points). Usually, the entire bag is labeled as abnormal (positive) or normal (negative), but the labels of the specific data points in the bag are unknown. This learning framework is suitable for handling anomaly detection tasks. Abnormal data is usually small and relatively discrete, and only a few instances in the bag may show abnormal behavior.

[0057] The work videos shot by the monitoring system are used as detection targets. The videos are divided into positive packets (including abnormal segments) and negative packets (excluding abnormal segments) according to whether the protective gear is not worn or worn incorrectly. In the training phase, the videos are divided into a fixed number of segments, and each segment generates an instance. In MIL, the video set is defined as an instance packet. Each package It consists of several instances, denoted as ,in Yes Package The number of instances in each package. Contains a tag (Indicates package is marked as abnormal, i.e., a positive packet), (Indicates package is marked as normal, i.e., negative bag), there is at least one instance in the positive bag is abnormal.

[0058] The spatiotemporal feature information of the instance is extracted through a 3D convolutional network. The C3D network structure includes an input layer, a 3D convolutional layer, a 3D pooling layer, a fully connected layer, and an output layer. It takes multiple consecutive frames as input information and has a shape of (T, H, W, C), where T is the number of time frames of the video (time dimension), H and W are the height and width of each frame of the video (spatial dimension), and C is the number of channels per frame; the convolution kernel of the 3D convolutional layer is , is the depth of the convolution kernel in the time dimension, is the height and width of the convolution kernel, is the number of input channels, is the number of output channels, and the 3D convolution is as follows: (1) In order to reduce the time and space dimensions and retain the key parts of the timing information, the output features are average pooled: (2) is the pooling window size in time, height, and width for each instance Assign a score After the training is completed, the instance score can be judged according to the instance score. Usually, the instance with a higher score is considered to be an abnormal instance. , because all the instances are normal fragments, the scores of all the instances are low. , since the positive package contains at least one abnormal instance, there is at least one instance with a larger score. Usually, the maximum instance score in the positive package is higher than that in the negative package, that is: (3) In the formula, is a function to be learned that maps the anomaly score of a fragment, ranging from 0 to 1. In order to maximize the match between the bag label and the instance score, the following loss function is used : (4) The MIL model is used to monitor video clips for situations where safety protection is worn incorrectly or not worn, and the instances are output in image format.

[0059] 2. Construct an image feature enhancement processing model: Use the gamma function to adjust the brightness and contrast of the image to be identified, highlight the details and features, and perform nonlinear transformation on the grayscale value of the image to solve the problem of difficulty in identification when the brightness of the power operation scene is low. The mathematical formula of the gamma transformation is as follows: (5) In the formula, is the original image at position The grayscale value of the pixel at is usually in the range of [0, 255]. is the pixel gray value at the same position in the transformed image, Is a constant used to adjust the overall brightness. This example is set to 1 (it can be adjusted according to the overall brightness requirements). It is the gamma index, which is set to 1.5 in this embodiment. After the gamma transformation, the bright part of the image will be compressed, the dark part will be enhanced, and the details will become more obvious in the light and dark areas.

[0060] In the processed image, each pixel still has three color channels: red (R, Red), green (G, Green) and blue (B, Blue). Image grayscale is used to map the three-dimensional color data to a one-dimensional grayscale value. In a grayscale image, the value of each pixel is usually a scalar, indicating its brightness or grayscale value, which can greatly reduce the storage space required for data processing and extract image structure information. In order not to affect the original features of the image, the weighted average method is used to calculate the grayscale value. The standard grayscale formula is as follows: (6).

[0061] 3. Construct a multi-scale feature extraction and convolutional recognition model: Use the convolutional neural network with mixed channel attention as the multi-scale feature extraction model after image feature enhancement, and use the GraphSAGE algorithm as the feature convolutional recognition model.

[0062] The overall structure of CNN includes: a feature information extraction module consisting of an input layer, a convolution layer, and a pooling layer, and a classification module consisting of a fully connected layer and an output layer; multi-scale feature extraction is to perform convolution operations on the input feature map at different scales to obtain feature maps of multiple scales. The convolution process is as follows: (7) In the formula, is the input image, is the convolution kernel, It is the feature map formed after convolution. The input image is The pixel value of the position, Represents the convolution kernel and the input image The relative position relationship of the nodes; this embodiment uses convolution kernels of (3×3), (5×5), and (7×7) to perform convolution operations to obtain feature maps at different scales: (8) In the formula, Indicates the use The convolution kernel is used for convolution. Indicates the use The convolution kernel is used for convolution. Indicates the use The convolution kernel is used for convolution. , , The output feature maps are spliced ​​and integrated to form a comprehensive feature map containing multi-scale information. : (9) Then perform maximum pooling on the feature map to get the final feature map : (10) However, the original image contains noise and redundant information, and CNN will use the same weight to convolve the whole image, resulting in a waste of resources. The channel attention mechanism is introduced to dynamically adjust the weights of different channels of the feature map, so that the model pays more attention to important features. Perform global average pooling to compress the spatial dimension. For each channel, calculate the vector after global average pooling. : (11) In the formula, , and Represent the height, width and number of channels respectively, is the vertical position of the image, is the horizontal position of the image, is the specific position in the feature map of the input channel c; finally, a C-dimensional vector is obtained , contains the feature information of each channel; after two layers of fully connected layer networks and activation functions, the attention weight of each channel is generated : (12) in , is the weight matrix of the fully connected layer, It is the Sigmoid activation function, which is used to map the weight value to [0, 1]; It is a nonlinear activation function. When the input value is ≤0, the output is 0. When the output value is >0, the input value is output. After being processed by the channel attention mechanism module, the channel weight vector is obtained. , and apply it to the original feature map On the top, we get the weighted feature map : (13) After CNN processes the original image, its output feature map is input into the GNN (Graph Neural Networks) network and converted into a graph structure. Each spatial position or pixel of the feature map is defined as a node. The size of the feature map is , each Position as a node, The dimensional feature vector is used as the initial feature of the node, and the edge relationship is established by setting the neighbor pixel position connection method within the 8-neighborhood range, and finally forming a structural graph ,in is a collection of nodes, is a set of edges.

[0063] For each node in the graph GraphSAGE will select from its neighbor set Randomly sample K subsets from the set to form a new set , the node characteristics are updated as follows: (14) Is a node In the The feature representation of the layer, Is a node exist The feature representation of the layer, is the learnable weight matrix, is the activation function, For Node The set of neighbors, is an aggregation function. After the collection is completed, mean aggregation is performed based on the feature information of neighbor nodes: (15) in is the number of iteration steps, Neighbor node Features, is the aggregated neighbor feature, For Node No. The node features are combined with the aggregated neighbor features, and the node features are updated through linear transformation and activation function. The update formula is: (16) in, is the learnable weight matrix of the current layer, is the activation function, Indicates the concatenation of the node's own features and the features of its neighboring nodes; each layer (or iteration step) of GraphSAGE can obtain higher-order neighborhood information of the node, and the features of each node include information from its neighboring nodes; after multiple layers of aggregation and updating, the final embedding vector of each node is obtained , and finally use it as the input of the classification task to complete the identification of safety protection equipment such as safety helmets, insulating capes, protective clothing, insulating sleeves, etc.

[0064] 4. Training of anomaly detection and recognition models: 1) Anomaly detection model training: Collect operation videos such as operation site monitoring or drone cameras, divide them according to whether there are anomalies, and label the videos with anomalies as 1, and the videos without anomalies as 0; use the abnormal operation clips as input to perform feature recognition training on the C3D module embedded in the MIL network to improve the accuracy of abnormal network recognition; 2) Multi-scale feature extraction and convolutional recognition model: After using LabelImg to annotate various protective equipment, perform image feature enhancement and grayscale conversion, use the CNN multi-scale feature extraction model with channel attention mechanism for training, and input the output feature information into the GraphSAGE network for training.

[0065] Example 8 In another preferred embodiment, based on embodiments 1, 2, 3, 4, 5, 6, and 7, this embodiment further illustrates the technical solution of the multi-scale feature recognition method for abnormal detection and safety protection of distribution network operations based on weak supervision in combination with the accompanying drawings: The flowchart of the anomaly detection and security protection multi-scale feature recognition method based on weak supervision is shown in the attached figure. Figure 1 As shown in the figure; firstly, the operation video recorded by the monitoring equipment or camera equipment such as drones arranged on site is collected, the video is divided into instances and then feature C3D convolution recognition is performed, and the scoring is performed based on the matching of the detected picture features with the marked abnormal picture features to determine whether there is an abnormality. If there is no abnormality, the detection of the instance is terminated; if there is an abnormality, the instance image is input into the image feature enhancement model, and the gamma transform and grayscale conversion are performed to improve the image details and reduce the influence of ambient light and complex scenes on the recognition accuracy; the processed image is input into the multi-scale feature extraction model of the trained fusion channel attention mechanism to obtain a weighted feature map, and the GraphSAGE feature recognition model recognizes and divides the safety protection equipment into safety helmets, insulating protective clothing, insulating shawls, and insulating sleeves according to the feature map structure information; detects whether there is a wearing abnormality, and if there is no abnormality, the recognition of this instance is terminated; if there is an abnormality, a warning prompt is issued, and then the recognition of this instance is terminated.

[0066] The structure of the multi-scale feature extraction model is shown in the attached figure. Figure 2As shown in the figure, the abnormal instances identified by MIL are subjected to (3×3), (5×5), and (7×7) feature convolutions to obtain multi-scale feature maps, and the feature maps are subjected to global average pooling to compress the spatial dimensions. For each channel, the vector after global average pooling is calculated; after two layers of fully connected layer networks and activation functions, the attention weights of each channel are generated and applied to the original feature map F to obtain a weighted feature map.

[0067] The GraphSAGE feature recognition model is shown in the attached Figure 3 As shown in the figure, after CNN processes the original image, its output feature map is converted into a graph structure, and mean aggregation is performed according to the feature information of neighbor nodes; the node's own features are merged with the aggregated neighbor features, and the node features are updated through linear transformation and activation function. After multiple layers of aggregation and updating, the final embedding vector of each node is obtained, which is used as the input of the classification task to complete the recognition of safety protection equipment such as safety helmets, insulating capes, protective clothing, and insulating sleeves.

[0068] Through the detailed description of the above embodiments, it is proved that the use of weakly supervised learning methods for anomaly detection reduces the cost of data annotation and improves the practicability of the model; the combination of three-dimensional convolutional neural networks and convolutional neural networks with mixed channel attention for multi-scale feature extraction improves the recognition accuracy in low light and complex scenes; the feature maps are further processed using graph neural networks to enhance the model's recognition ability for complex features; a two-stage recognition process is designed, with anomaly detection followed by fine recognition, which reduces the consumption of computing resources and improves recognition efficiency.

[0069] The significant effect of the present invention is that it improves the accuracy and robustness of power operation safety detection, while reducing the dependence on labeled data; through the combination of weakly supervised learning algorithm and multi-scale feature extraction technology, the present invention can more effectively utilize limited labeled data and extract richer feature information, thereby improving the recognition accuracy; in addition, the application of graph neural network also further enhances the model's ability to handle complex backgrounds and occlusion situations; this conclusion has been fully verified by experimental data, showing that the present invention has significant advantages and application prospects in the field of power operation safety detection.

[0070] In the preferred scheme, the 3D convolutional network in Step 1.3 includes an input layer, a 3D convolutional layer, a 3D pooling layer, a fully connected layer and an output layer. Multiple continuous frames are used as input information, and spatiotemporal features are extracted through the 3D convolutional layer. The 3D pooling layer reduces the time and space dimensions, and the fully connected layer and the output layer output anomaly scores. The above settings enable the model to fully capture the dynamic change information in the video data, effectively distinguish between normal and abnormal behavior patterns, improve the accuracy and robustness of abnormal behavior detection, and provide strong support for the practical application of intelligent monitoring systems.

[0071] In the preferred solution, the GraphSAGE algorithm in Step 3.3 obtains the final embedding vector of each node by aggregating and updating the features of neighboring nodes, and uses it as the input of the classification task to complete the identification of safety protection equipment; the above settings significantly improve the recognition accuracy and efficiency; in addition, the attention mechanism is introduced to optimize the node feature weight distribution, further enhancing the model's ability to capture key features, ensuring stable identification of various types of safety protection equipment even in complex scenarios.

[0072] In summary, the weakly supervised distribution network operation anomaly detection and safety protection multi-scale feature recognition method proposed in the present invention solves the technical problems of high detection cost, poor environmental adaptability, waste of computing resources and poor real-time performance of improperly worn safety protection equipment at the distribution network operation site; the present invention adopts a multiple instance learning (MIL) framework and uses weakly labeled data for training, which reduces the cost of data annotation and improves the training efficiency and scalability of the model; in view of the low brightness and complex environment of the power operation scene, the scheme introduces the gamma function to adjust the image brightness and contrast, as well as the grayscale processing technology, which effectively enhances the details and features of the image and improves the accuracy of the recognition model; the scheme of the present invention combines the convolutional neural network (CNN) with mixed channel attention and the multi-scale feature extraction technology, which can capture image features at different scales and improve the model's recognition ability for safety protection equipment. At the same time, the GraphSAGE algorithm is introduced to further process the features, which enhances the feature representation ability of the model; the technical scheme of the present invention creatively applies the weakly supervised learning method to power operation anomaly detection, by using video packet-level labels instead of single frames The proposed method uses labels of different levels to train, which effectively solves the problem of scarce abnormal data and difficult labeling in power operation monitoring videos, and improves the efficiency and accuracy of anomaly detection. It combines image feature enhancement technology and multi-scale feature extraction methods, first improves image quality through image enhancement technology, and then captures key information in the image through multi-scale feature extraction technology. This collaborative optimization method significantly improves the performance of the recognition model in low-light and complex scenes. It also creatively introduces the GraphSAGE algorithm to perform convolution recognition on the extracted features, fully utilizes the information representation ability of graph structure data, improves the recognition accuracy of safety protection equipment, and provides a new technical means for power operation safety detection. In addition, the scheme also integrates an adaptive learning rate adjustment strategy, dynamically adjusts the learning rate according to the change of loss during training, accelerates model convergence and improves training stability. Combining these innovations, the present invention has brought revolutionary improvements to power operation safety monitoring. In practical applications, the scheme has been successfully applied to multiple power operation safety monitoring projects, significantly reducing the safety accident rate, improving operation efficiency, and providing strong support for the intelligent transformation and safe development of the power industry.

Claims

1. A weakly supervised distribution network operation anomaly detection and safety protection multi-scale feature recognition method, characterized in that: The method comprises the following steps: Step 1: Establish an abnormal image recognition model; Step 2: Build an image feature enhancement processing model; Step 3: Build a multi-scale feature extraction and convolutional recognition model.

2. According to the weakly supervised distribution network operation anomaly detection and safety protection multi-scale feature recognition method according to claim 1, it is characterized in that: Step 1 includes the following steps: Step 1.1: The operation video shot by the monitoring system is used as the detection target. The video is divided into positive and negative packets according to whether the protective gear is not worn or worn incorrectly. The positive packet contains abnormal clips, while the negative packet does not contain abnormal clips. Step 1.2: During the training phase, the video is divided into a fixed number of segments, each segment generates an instance, and the video set is defined as an instance bag. Each bag contains several instances and is assigned a label. The positive bag label is abnormal and the negative bag label is normal. Step 1.3: Extract the spatiotemporal feature information of the instance through the 3D convolutional network, use the loss function to optimize the model, and realize abnormal image monitoring based on weak supervision.

3. According to claim 2, the method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision is characterized in that: The 3D convolutional network in Step 1.3 includes an input layer, a 3D convolutional layer, a 3D pooling layer, a fully connected layer and an output layer. It takes multiple consecutive frames as input information, extracts spatiotemporal features through the 3D convolutional layer, reduces the time and space dimensions through the 3D pooling layer, and outputs anomaly scores through the fully connected layer and the output layer.

4. The method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision according to claim 3 is characterized in that: The specific steps and expressions of the 3D convolution are as follows: （1）； In the formula, The convolution kernel of the 3D convolution layer is; is the depth of the convolution kernel in the time dimension; is the height and width of the convolution kernel; is the number of input channels; is the number of output channels; is the input video tensor; for time of Channel output characteristics, is the image time series; is the vertical position of the image; is the horizontal position of the image; is the vertical position of the convolution kernel; is the lateral position of the convolution kernel; is the time position of the convolution kernel; Keep the key part of the timing information and perform average pooling on the output features: （2）； is the average pooling feature map of the output; is the input feature map; is the pooling window size in time, height, and width for each instance Assign a score After the training is completed, the instance score is judged according to the instance score. The instance with a higher score is considered to be an abnormal instance. , all instances have low scores; for positive packets , at least one instance score is larger, and the maximum instance score in the positive bag is higher than that in the negative bag, that is: （3）； In the formula, is the set of positive packets containing abnormal instances, is a set of negative packets that do not contain abnormal instances, is the fraction of instances in the positive bag; is the fraction of instances in the negative bag; is a function to be learned that maps the anomaly score of a fragment, ranging from 0 to 1. In order to maximize the match between the bag label and the instance score, the following loss function is used : （4）； In the formula, For video tags; Representation instance score; is the sigmoid function, which is used to convert scores into probabilities; Indicates the highest instance score in the bag; the MIL model is used to monitor the video clips for incorrect or non-worn safety protection and output the instances in image format.

5. According to the weakly supervised distribution network operation anomaly detection and safety protection multi-scale feature recognition method of claim 1, it is characterized in that: The Step 2 includes the following steps: Step 2.1: Use the gamma function to adjust the brightness and contrast of the image to be identified and enhance the image details. The mathematical formula of the gamma function transformation is: （5）； In the formula, is the original image at position The gray value of the pixel at is the pixel gray value at the same position in the transformed image, is a constant used to adjust the overall brightness. It is the gamma index. After gamma transformation, the bright parts of the image will be compressed, the dark parts will be enhanced, and the details will become more obvious in the light and dark areas; Step 2.2: Grayscale the image to reduce data processing and storage space and extract image structure information.

6. The method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision according to claim 5 is characterized in that: The grayscale processing of the image in Step 2.2 is to map the three-dimensional color data to a one-dimensional grayscale value. In a grayscale image, the value of each pixel is a scalar, which represents its brightness or grayscale value, reducing the storage space required for data processing and extracting image structure information. In order not to affect the original features of the image, the weighted average method is used to calculate the grayscale value. Standard grayscale The formula is as follows: （6）； In the formula, is the value of the red channel, is the value of the green channel, is the value of the blue channel.

7. The method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision according to claim 1 is characterized in that: Step 3 includes the following steps: Step 3.1: Use a convolutional neural network with mixed channel attention as a multi-scale feature extraction model, and use convolution kernels of different scales to perform convolution operations on the input feature map to obtain a multi-scale feature map; Step 3.2: Introduce the channel attention mechanism to dynamically adjust the weights of different channels of the feature map; Step 3.3: Input the feature graph output by the convolutional neural network into the GraphSAGE algorithm to construct the graph structure, and obtain the final embedding vector of each node by aggregating and updating the features of neighboring nodes; Step 3.4: According to the final embedded vector, identify the wearing status of safety protection equipment such as hard hats, insulating protective clothing, insulating capes, and insulating sleeves.

8. The method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision according to claim 7 is characterized in that: The multi-scale feature extraction in Step 3.1 is to perform convolution operations on the input feature map at different scales to obtain feature maps of multiple scales. The convolution process is as follows: （7）； In the formula, is the input image, is the convolution kernel, It is the feature map formed after convolution. The input image is The pixel value of the position, Represents the convolution kernel and the input image Relative position relationship; Convolution operations are performed using convolution kernels of (3×3), (5×5), and (7×7) to obtain feature maps of different scales: （8）； In the formula, Indicates the use The convolution kernel is used for convolution. Indicates the use The convolution kernel is used for convolution. Indicates the use The convolution kernel is used for convolution. , , The output feature maps are spliced ​​and integrated to form a comprehensive feature map containing multi-scale information. : （9）； Then perform maximum pooling on the feature map to get the final feature map : （10）。 9. The method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision according to claim 8 is characterized in that: The specific steps of Step 3.2 include: The input feature map Perform global average pooling to compress the spatial dimension. For each channel, calculate the vector after global average pooling. : （11）； In the formula, , and Represent the height, width and number of channels respectively, is the vertical position of the image, is the horizontal position of the image, is the specific position in the feature map of the input channel c; finally, a C-dimensional vector is obtained , contains the feature information of each channel; after two layers of fully connected layer networks and activation functions, the attention weight of each channel is generated : （12）； in , is the weight matrix of the fully connected layer, It is the Sigmoid activation function, which is used to map the weight value to [0, 1]; It is a nonlinear activation function. When the input value is ≤0, the output is 0. When the output value is >0, the input value is output. After being processed by the channel attention mechanism module, the channel weight vector is obtained. , and apply it to the original feature map On the top, we get the weighted feature map : （13）。 10. The method for detecting abnormalities in distribution network operations and identifying multi-scale features for safety protection based on weak supervision according to claim 9 is characterized in that: The GraphSAGE algorithm in Step 3.3 obtains the final embedding vector of each node by aggregating and updating the features of neighboring nodes, and uses it as the input of the classification task to complete the identification of safety protection equipment. The specific steps of Step 3.3 are as follows: The feature map output by the convolutional neural network is input into the GNN network and converted into a graph structure. For each node in the graph GraphSAGE will select from its neighbor set Randomly sample K subsets from the set to form a new set , the node characteristics are updated as follows: （14）； Is a node In the The feature representation of the layer, Is a node exist The feature representation of the layer, is the learnable weight matrix, is the activation function, For Node The set of neighbors, is an aggregation function. After the collection is completed, mean aggregation is performed based on the feature information of neighbor nodes: （15）； in is the number of iteration steps, Neighbor node Features, is the aggregated neighbor feature, For Node No. The node features are combined with the aggregated neighbor features, and the node features are updated through linear transformation and activation function. The update formula is: （16）； in, is the learnable weight matrix of the current layer, is the activation function, It means that the node's own features and the features of neighboring nodes are concatenated. Each layer or iteration of GraphSAGE obtains higher-order neighborhood information of the node. The features of each node include information from its neighboring nodes. After multiple layers of aggregation and updating, the final embedding vector of each node is obtained. .

Images