Distribution line anomaly detection model training method, anomaly detection method and system

Abstract

The invention discloses a training method of a distribution line anomaly detection model, belongs to the technical field of distribution line anomaly detection, and solves the problems that an existing distribution line anomaly detection mode is low in detection precision and poor in recognition performance. The method comprises the steps that normal and abnormal historical distribution line images in different distribution line scenes are collected and preprocessed, the preprocessed distribution line sample images are labeled, and a distribution line sample image sample set is constructed; for each distribution line sample image in the distribution line sample image sample set, performing local and global attention feature extraction based on a graph model to obtain a relation feature vector of each distribution line sample image; and training the anomaly detection model based on the relation feature vector and the label of each distribution line sample image in the distribution line sample image sample set to obtain a trained anomaly detection model.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution line anomaly detection, and in particular to a power distribution line anomaly detection model training method, anomaly detection method and system. Background Art

[0002] Distribution line anomaly detection refers to the power supply company discovering and locating anomalies in the distribution line during the power production and maintenance process. Common anomalies include bird nests, foreign objects, discharge traces and background interference, etc., to ensure the safety of residential and commercial electricity use.

[0003] In the traditional mode, the inspection process of distribution lines mainly relies on the naked eye observation of power workers. However, in the actual inspection process of distribution lines, there are problems such as interference from foreign objects, excessive sight distance, and the risk of high-altitude operations. It is difficult to effectively meet the needs of actual power inspections by relying solely on manual observation.

[0004] With the development of information technology, researchers use machine learning to handle anomaly detection and classification tasks. They collect images of distribution lines through hardware (such as cameras, handheld terminals and other data devices), and use machine learning to train the corresponding collected data to achieve feature extraction, positioning and unknown anomaly prediction of abnormal targets in distribution lines. Although this type of method has a certain degree of improvement in intelligence compared to traditional manual methods, it requires a lot of manpower, material and financial resources. Therefore, in order to save costs and improve the accuracy and efficiency of intelligent inspection safety under distribution lines, the distribution line anomaly detection method based on drone vision has received more attention.

[0005] Different from the traditional machine vision method, the distribution line anomaly detection technology based on the drone perspective relies more on artificial intelligence, deep learning and other technologies. Through model training, optimization and deployment integration, it can achieve more automated anomaly detection and target positioning, and complete the prediction of unknown anomalies. At present, the anomaly detection method based on various visual features to obtain anomaly detection results has received widespread attention from industry insiders. However, for distribution line detection, due to the special characteristics of distribution lines such as single scene, large line of sight, and unknown abnormal targets, the existing anomaly detection methods cannot meet the needs of distribution line anomaly detection.

[0006] The existing anomaly detection methods based on drone perspective mainly include the following steps: (1) data collection; (2) feature extraction; (3) abnormal sample classification detection; (4) prediction of unknown anomalies. Through the above steps, the cost and time required in the manual inspection process are reduced to a certain extent. However, the disadvantage of this type of anomaly detection method is that in the process of feature extraction and model optimization, due to the characteristics of a wide variety of abnormal targets and a small number of samples, noise is easily generated, which in turn affects the training quality of the later model. In addition, another important reason why this type of anomaly detection method is prone to noise interference is that during the inspection of distribution lines by drones, factors such as differences in shooting angles, different lighting conditions, excessive line visibility, background interference, and risks of high-altitude operations may affect the detection accuracy and positioning of specific abnormal targets, thereby reducing the detection and recognition performance of distribution line anomalies. Summary of the invention

[0007] In view of the above analysis, the embodiments of the present invention aim to provide a distribution line anomaly detection model training method, anomaly detection method and system, so as to solve the problems of low detection accuracy and poor recognition performance of existing distribution line anomaly detection methods.

[0008] First, the present invention provides a method for training a distribution line anomaly detection model, the method comprising:

[0009] Collect normal and abnormal historical distribution line images under different distribution line scenarios and preprocess them, label the preprocessed distribution line sample images, and construct a distribution line sample image sample set;

[0010] For each distribution line sample image in the distribution line sample image sample set, local and global attention feature extraction based on the graph model is performed to obtain the relational feature vector of each distribution line sample image;

[0011] Based on the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained to obtain a trained anomaly detection model.

[0012] Based on the above scheme, this scheme also makes the following improvements:

[0013] Furthermore, the relational feature vector of each distribution line sample image is obtained by executing:

[0014] The backbone network model is used to extract the feature map of the distribution line sample image;

[0015] The feature map of the distribution line sample image is processed into blocks to obtain multiple local feature blocks, and the corresponding local attention feature map is extracted from each local feature block;

[0016] Extracting a global attention feature map from the feature map of the distribution line sample image;

[0017] Using an undirected graph model based on edge regression, the nonlinear mapping enhancement of multi-channel features is performed on all local attention feature maps and global attention feature maps to obtain the relational feature vectors of distribution line sample images.

[0018] Furthermore, using an undirected graph model based on edge regression, all local attention feature maps and global attention feature maps are enhanced by nonlinear mapping of multi-channel features to obtain the relational feature vectors of the distribution line sample images, and then:

[0019] All local attention feature maps and global attention feature maps are used as vertices in the undirected graph model based on edge regression, and the correlation between the vertices is calculated;

[0020] The information of the association relationship between each vertex in the undirected graph model is integrated to obtain the relationship feature vector of the distribution line sample image.

[0021] Further, based on the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained to obtain a trained anomaly detection model; execute:

[0022] According to the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained until the loss function converges to obtain a trained anomaly detection model.

[0023] Furthermore, during the training process of the anomaly detection model, the loss function L is:

[0024] L=(1-λ)L ID +λL pre (1)

[0025] Among them, L ID is the distance penalty loss function, L pre is the distance constraint loss function, and λ is a hyperparameter that controls classification and prediction.

[0026] Furthermore, the collected historical distribution line images are preprocessed by performing the following operations:

[0027] The historical distribution line images are screened, cleaned, deduplicated and transformed to obtain distribution line sample images with consistent sizes.

[0028] Second, the present invention discloses a method for detecting anomalies in a distribution line. The method for detecting anomalies in a distribution line is implemented based on a trained anomaly detection model obtained by the training method for anomaly detection model of a distribution line described above; the method for detecting anomalies in a distribution line comprises:

[0029] Collect images of the power distribution lines to be inspected;

[0030] Extracting local and global attention features based on a graph model on the distribution line image to be detected to obtain a relational feature vector of the distribution line image to be detected;

[0031] The relational feature vector of the distribution line image to be detected is input into the trained anomaly detection model, and the anomaly detection model performs anomaly detection prediction and outputs the corresponding anomaly detection result.

[0032] Based on the above scheme, this scheme also makes the following improvements:

[0033] Furthermore, the anomaly detection method further includes:

[0034] After collecting the image of the distribution line to be detected, it is determined whether the size of the image of the distribution line to be detected is consistent with the size of the distribution line sample image. If not, the size of the image of the distribution line to be detected is transformed to obtain the image of the distribution line to be detected that is consistent with the size of the distribution line sample image.

[0035] Third, the present invention discloses an abnormality detection system for a power distribution line, the abnormality detection system comprising:

[0036] An image acquisition module, used for acquiring images of the power distribution lines to be inspected;

[0037] A relational feature vector extraction module, used for performing local and global attention feature extraction based on a graph model on the distribution line image to be detected, to obtain a relational feature vector of the distribution line image to be detected;

[0038] The anomaly detection module is used to input the relational feature vector of the distribution line image to be detected into the trained anomaly detection model, and the anomaly detection model performs anomaly detection and outputs the corresponding anomaly detection result.

[0039] Based on the above scheme, this scheme also makes the following improvements:

[0040] Furthermore, the anomaly detection system is transplanted to an unmanned aerial vehicle platform, so that the unmanned aerial vehicle platform can perform real-time detection on the captured images of the power distribution lines to be detected.

[0041] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0042] The present invention has the following positive effects and advantages:

[0043] 1) The present invention extracts local and global attention features based on the graph model, and can obtain the relational feature vector of each distribution line sample image. That is, the nonlinearity of the attention feature is increased through the graph model, the complexity of model training is guaranteed, and the diversity of training samples is increased from the feature level. Compared with the traditional method of increasing the diversity of target features at the image level, the final recognition effect is better. In general, the detection efficiency and detection accuracy are enhanced, the training efficiency of the model is increased, and it has good transferability. It solves the problems of low detection accuracy and poor recognition performance of the existing distribution line anomaly detection method.

[0044] 2) The present invention obtains multi-channel attention features and constructs a graph model, and sets two loss functions for classification and prediction to restrain each other to achieve the best effect. According to the acquisition of multi-channel attention features, a top-down hierarchical feature acquisition method is adopted to increase the discriminability of abnormal features, increase the robustness of the model, and greatly reduce the impact of noise labels on model training.

[0045] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0047] Figure 1 A flow chart of a method for training a distribution line anomaly detection model provided in Example 1 of the present invention;

[0048] Figure 2 A framework diagram of power distribution line anomaly detection provided by Embodiment 2 of the present invention;

[0049] Figure 3 A flow chart of a method for detecting anomalies in a power distribution line provided in Embodiment 2 of the present invention;

[0050] Figure 4 This is a schematic diagram of the structure of the abnormality detection system for distribution lines provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0051] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0052] Specific embodiment 1 of the present invention provides a method for training a distribution line anomaly detection model, the flow chart of which is as follows: Figure 1 shown.

[0053] Step S1: collect and preprocess normal and abnormal historical distribution line images under different distribution line scenarios, label the preprocessed distribution line sample images, and construct a distribution line sample image sample set.

[0054] In this embodiment, in each distribution line scenario, normal and abnormal historical distribution line images can be directly collected through drone cameras, handheld terminals and other sensor devices; or, distribution line videos can be collected through drone cameras, handheld terminals and other sensor devices to extract normal and abnormal historical distribution line images. It should be noted that in order to improve the abnormality detection effect of the distribution line, in each distribution line scenario, as many historical distribution line images of different abnormal types as possible are collected. It should be noted that since the quality of the collected images directly affects the accuracy of subsequent abnormality analysis and detection, it is necessary to ensure that the collected images are clear and the information is sufficient, so that the details of the surface of the transmission line can be captured in time.

[0055] Preferably, the collected historical distribution line images can be preprocessed in the following manner: the historical distribution line images are screened, cleaned, deduplicated and image transformed to obtain distribution line sample images of the same size. In the specific implementation process, by screening and cleaning the collected historical distribution line images, low-quality images or video sequences can be removed and high-quality images can be retained. At the same time, images of the same type with too high similarity are deduplicated to avoid category deviation caused by too many similar samples of the same type, while increasing the diversity of images and improving the robustness of the model. In addition, image transformation operations can be performed on the collected historical distribution line images, including random cropping, scaling, enhancement and twisting, so as to obtain distribution line sample images of the same size, so as to facilitate subsequent feature extraction and matching, and enhance the robustness of the data and the generalization ability of the model. Afterwards, each distribution line sample image is labeled separately, and the label of each distribution line sample image is normal or a specific abnormal type, so as to construct a distribution line sample image sample set.

[0056] Step S2: for each distribution line sample image in the distribution line sample image sample set, local and global attention feature extraction based on the graph model is performed to obtain a relational feature vector of each distribution line sample image.

[0057] The specific implementation process of step S2 is described below.

[0058] Step S21: extracting a feature map of a distribution line sample image using a backbone network model.

[0059] Preferably, in this embodiment, the backbone network model can be implemented using a convolutional neural network, such as a ResNet / Transformer model. The feature map f0∈R output by the backbone network model H×W×C , where H, W, and C represent the height, width, and number of channels of the feature map, respectively.

[0060] Step S22: performing block processing on the feature map of the distribution line sample image to obtain multiple local feature blocks, and extracting the corresponding local attention feature map from each local feature block.

[0061] In the specific implementation process, the feature map of the distribution line sample image can be horizontally segmented to obtain multiple local feature blocks. Next, the corresponding local attention feature map can be extracted from each local feature block, that is, local attention features are extracted and reshaped for each local feature block to extract the corresponding local attention feature map.

[0062] Assume that the tth local feature block is f t express, Where T represents the number of local feature blocks, t∈[1,T].

[0063] For example, for the tth local feature block f t Perform local attention feature extraction to obtain the corresponding local attention feature, and use f′ t It means that f′ t ∈R h×W×C , where h represents the height of the local attention feature. Next, for the local attention feature f′ t Reshape the local attention feature f′ t is a three-dimensional vector, i.e. h×W×C. Below, we respectively calculate the three-dimensional vector f′ t The first two dimensions h×W and the last two dimensions W×C are reshaped and reduced in dimension. Among them, the first two dimensions are reshaped to obtain the first two-dimensional feature vector Q∈R K×C , where K = h × W, and then for f′ t Perform transposition to obtain the eigenvector K∈R C×K , where K = h × W, and then the eigenvectors Q and K are Operation (matrix multiplication), get L∈R C×C , and then the original local feature f tThen perform a similar matrix multiplication operation with L, and finally obtain the local attention feature map corresponding to the t-th local feature block, using f" t Indicates f″ t ∈R h×W×C .

[0064] During the specific implementation process, the execution process of step S22 can be realized based on the local attention model.

[0065] Step S23: extracting a global attention feature map from the feature map of the distribution line sample image.

[0066] In the specific implementation process, due to the particularity and unknown type of abnormal targets in the distribution line, in addition to some local information, all significant features need to be learned. In the specific implementation process, the feature map of the distribution line sample image output by the backbone network model can be extracted and reshaped by global attention features. By taking similar operations as above, the corresponding global attention features can be obtained, represented by F″, F″∈R H×W×C .

[0067] During the specific implementation process, the execution process of step S23 can be implemented based on the global attention model.

[0068] Step S24: using an edge regression-based undirected graph model, perform nonlinear mapping enhancement of multi-channel features on all local attention feature maps and global attention feature maps to obtain a relational feature vector of the distribution line sample image.

[0069] During the specific implementation process, all local attention feature maps and global attention feature maps are respectively used as vertices in an undirected graph model based on edge regression, and the correlation between each vertices is calculated (for example, the correlation relationship is calculated using similarity metrics such as Euclidean distance or Mahalanobis distance); the correlation between each vertices in the undirected graph model is integrated to obtain the relationship feature vector of the distribution line sample image.

[0070] For example, assume that the undirected graph model G Q =(G, P(Q)}, where G represents a graph, Q represents a vertex, and P(Q) represents the relationship between vertices in the graph. G=(V, E), V, E represent the set of vertices and edges in the graph, respectively. In the present embodiment, V={1, 2, ...., t, t+1}, there are t+1 vertices in total, and the number of vertices is the sum of all local attention feature maps and global attention feature maps. E∈V×V, this method can complete the nonlinear mapping of multi-channel features, integrate the information of the output multi-attention feature maps, and realize hierarchical feature mapping from coarse to fine and from top to bottom. In the specific implementation process, an undirected graph model based on edge regression is constructed according to an undirected graph model with local attention features and global attention feature maps as vertices.

[0071] In the specific implementation process, the graph model is used to perform nonlinear mapping enhancement on all local attention feature maps and global attention feature maps, and the relationship feature vector of the distribution line sample image can be obtained. That is, this embodiment completes the nonlinear mapping enhancement of the corresponding attention features based on the graph model, and its purpose is to use the graph model to map the image space relationship. Specifically, each local attention feature map and global attention feature map are used as vertices of the graph model, and the relationship between attention (the relationship between the local attention mechanism itself and the global attention) (the relative position information between the local parts of the image) is used as the edge to construct the structure of the graph model, and the characteristics of the graph structure are used to complete the coordinated association relationship between the attention features, thereby improving the nonlinear expression ability of the data features.

[0072] Step S3: Based on the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained until the loss function converges to obtain a trained anomaly detection model.

[0073] Preferably, in this embodiment, the anomaly detection model is implemented based on a regression model. In the process of training the regression model, the loss function is optimized to achieve classification detection of distribution line images and prediction of unknown anomalies.

[0074] According to the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model and the loss function therein are trained to obtain a trained anomaly detection model.

[0075] Preferably, during the training process of the anomaly detection model, the loss function L is:

[0076] L=(1-λ)L ID +λL pre (1)

[0077] Among them, L ID is the distance penalty loss function (classification loss function), L pre is the distance constraint loss function (prediction discriminant loss function), and λ is a hyperparameter that controls classification and prediction.

[0078] This embodiment uses anomaly adversarial generation loss function L pre Complete the prediction and identification of abnormal samples, expressed as:

[0079]

[0080] Among them, y, denote the true label and the predicted label respectively, It represents the overlapping area between the anomaly algorithm detection box and the real anomaly area. gt and pre represent the size of the real anomaly area and the size of the algorithm detection box, respectively.

[0081] The framework diagram of the power distribution line abnormality detection provided by the embodiment of the present invention is as follows Figure 2 shown.

[0082] Specific embodiment 2 of the present invention provides a method for detecting abnormality of a power distribution line, the flow chart of which is as follows: Figure 3 As shown, it is implemented based on the anomaly detection model trained in Example 1. The anomaly detection method of the distribution line includes:

[0083] Step S1: collecting images of the power distribution lines to be inspected;

[0084] Step S2: extracting local and global attention features based on a graph model for the distribution line image to be detected, and obtaining a relational feature vector of the distribution line image to be detected;

[0085] Step S3: Input the relational feature vector of the distribution line image to be detected into the trained anomaly detection model, and the anomaly detection model performs anomaly detection and outputs a corresponding anomaly detection result. The anomaly detection result is normal, or a corresponding anomaly type.

[0086] Preferably, the anomaly detection method further comprises, after step S1, adding the following steps:

[0087] After collecting the image of the distribution line to be detected, it is determined whether the size of the image of the distribution line to be detected is consistent with the size of the distribution line sample image. If not, the size of the image of the distribution line to be detected is transformed to obtain the image of the distribution line to be detected that is consistent with the size of the distribution line sample image.

[0088] In the specific implementation process, since this embodiment is based on a supervised target anomaly detection task, it is essentially a target classification problem, in which each data sample has corresponding category information. For the randomly selected batch samples, the above-mentioned local and global attention features are extracted, and after the nonlinear mapping is completed by the graph model, it is used as the input of the loss function to complete the optimization of the entire framework. In this embodiment, the corresponding classification loss function L can be set ID , to achieve classification and identification of abnormal samples. In addition, for the acquisition of local and global attention features of the above abnormal samples, in order to obtain the spatial location information of the abnormal samples, this embodiment also designs a corresponding prediction loss function L pre , in order to achieve the prediction of unknown anomalies. The specific training process can be found in the subsequent implementation steps.

[0089] After completing the optimization training of the entire framework above, the trained framework is used to test the input samples. For each abnormal sample, its feature vector can be stored in the database. When classification and recognition are required, the feature vector of the query image is compared with the features of all samples in the database to find several similar images as the classification results. In addition, for unknown abnormal samples, that is, abnormal targets that are not in the library images, a prediction loss function is used to distinguish the differences between the remaining normal samples and predict the unknown abnormal targets. The specific prediction process is shown in the subsequent implementation steps.

[0090] In addition, after the network model is trained, the trained model is used to extract the feature vector of the abnormal sample image. For each abnormal sample, its feature vector can also be stored in the database. When classification and recognition are required, the feature vector of the query image is compared with all sample features in the database to find several similar images as the classification results.

[0091] Preferably, in this embodiment, in order to facilitate subsequent model training, the sample set of distribution line sample images can be divided into multiple batches of sample subsets, and the number of distribution line sample images in each batch of sample subsets is consistent (such as 500). During the model training process, the data sampling operation will randomly select a number of product sample images from multiple batches, and then use the selected distribution line sample images and their labels to perform model training, so that the model can learn more different product sample features to improve the generalization ability of the model. At the same time, this method can also avoid the model from overfitting certain abnormal features in a single sample subset, thereby reducing the risk of overfitting during the model optimization process.

[0092] A specific embodiment 3 of the present invention further provides an abnormality detection system for a power distribution line, the structural diagram of which is shown in FIG. Figure 4 As shown, the anomaly detection system includes an image acquisition module, a relationship feature vector extraction module and an anomaly detection module; wherein,

[0093] An image acquisition module, used for acquiring images of the power distribution lines to be inspected;

[0094] A relational feature vector extraction module, used for performing local and global attention feature extraction based on a graph model on the distribution line image to be detected, to obtain a relational feature vector of the distribution line image to be detected;

[0095] The anomaly detection module is used to input the relational feature vector of the distribution line image to be detected into the trained anomaly detection model, and the anomaly detection model performs anomaly detection and outputs the corresponding anomaly detection result.

[0096] During the specific implementation process, the anomaly detection system of the distribution line can be transplanted to the corresponding UAV platform to realize real-time detection of anomalies in the distribution line and improve the efficiency and accuracy of anomaly detection in the distribution line.

[0097] The specific implementation process of the embodiment of the present invention can be found in the above method embodiment, which will not be described in detail in this embodiment.

[0098] Since the principles of this embodiment are the same as those of the above method embodiment, the system also has the corresponding technical effects of the above method embodiment.

[0099] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0100] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A training method for a distribution line anomaly detection model, characterized in that: The method comprises: Collect normal and abnormal historical distribution line images under different distribution line scenarios and preprocess them, label the preprocessed distribution line sample images, and construct a distribution line sample image sample set; For each distribution line sample image in the distribution line sample image sample set, local and global attention feature extraction based on the graph model is performed to obtain the relational feature vector of each distribution line sample image; Based on the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained to obtain a trained anomaly detection model.

2. The training method of the distribution line anomaly detection model according to claim 1, characterized in that: To obtain the relational feature vector for each distribution line sample image, execute: The backbone network model is used to extract the feature map of the distribution line sample image; The feature map of the distribution line sample image is processed into blocks to obtain multiple local feature blocks, and the corresponding local attention feature map is extracted from each local feature block; Extracting a global attention feature map from the feature map of the distribution line sample image; Using an undirected graph model based on edge regression, the nonlinear mapping enhancement of multi-channel features is performed on all local attention feature maps and global attention feature maps to obtain the relational feature vectors of distribution line sample images.

3. The training method of the distribution line anomaly detection model according to claim 2, characterized in that: Using an undirected graph model based on edge regression, all local attention feature maps and global attention feature maps are enhanced by nonlinear mapping of multi-channel features to obtain the relational feature vectors of the distribution line sample images. Execute: All local attention feature maps and global attention feature maps are used as vertices in the undirected graph model based on edge regression, and the correlation between the vertices is calculated; The information of the association relationship between each vertex in the undirected graph model is integrated to obtain the relationship feature vector of the distribution line sample image.

4. The training method of the distribution line anomaly detection model according to claim 3, characterized in that: Based on the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained to obtain a trained anomaly detection model; implement: According to the relational feature vector and label of each distribution line sample image in the distribution line sample image sample set, the anomaly detection model is trained until the loss function converges to obtain a trained anomaly detection model.

5. The training method of the distribution line anomaly detection model according to claim 4, characterized in that: During the training process of the anomaly detection model, the loss function L is: L=(1-λ)L ID +λL pre (1) Among them, L ID is the distance penalty loss function, L pre is the distance constraint loss function, and λ is a hyperparameter that controls classification and prediction.

6. The training method of the distribution line anomaly detection model according to any one of claims 1 to 5, characterized in that: The collected historical distribution line images are preprocessed by performing the following operations: The historical distribution line images are screened, cleaned, deduplicated and transformed to obtain distribution line sample images with consistent sizes.

7. A method for detecting abnormality of a power distribution line, characterized in that: The distribution line anomaly detection method is implemented based on a trained anomaly detection model obtained by the training method of the distribution line anomaly detection model according to any one of claims 1 to 6; the distribution line anomaly detection method comprises: Collect images of the power distribution lines to be inspected; Extracting local and global attention features based on a graph model on the distribution line image to be detected to obtain a relational feature vector of the distribution line image to be detected; The relational feature vector of the distribution line image to be detected is input into the trained anomaly detection model, and the anomaly detection model performs anomaly detection prediction and outputs the corresponding anomaly detection result.

8. The method for detecting abnormality of a power distribution line according to claim 7, characterized in that: The anomaly detection method further includes: After collecting the image of the distribution line to be detected, it is determined whether the size of the image of the distribution line to be detected is consistent with the size of the distribution line sample image. If not, the size of the image of the distribution line to be detected is transformed to obtain the image of the distribution line to be detected that is consistent with the size of the distribution line sample image.

9. The abnormality detection system for power distribution lines according to claim 8, characterized in that: The anomaly detection system comprises: An image acquisition module, used for acquiring images of the power distribution lines to be inspected; A relational feature vector extraction module, used for performing local and global attention feature extraction based on a graph model on the distribution line image to be detected, to obtain a relational feature vector of the distribution line image to be detected; The anomaly detection module is used to input the relational feature vector of the distribution line image to be detected into the trained anomaly detection model, and the anomaly detection model performs anomaly detection and outputs the corresponding anomaly detection result.

10. The abnormality detection system for power distribution lines according to claim 9, characterized in that: The anomaly detection system is transplanted to an unmanned aerial vehicle platform, so that the unmanned aerial vehicle platform can perform real-time detection on the captured images of the distribution line to be detected.

Images