Intelligent insulator detection method and system in power supply system

Abstract

The invention discloses an insulator intelligent detection method and system in a power supply system, relates to the technical field of artificial intelligence, and is used for simultaneously completing defect classification and damage area positioning of an insulator device under the condition that only a small number of defect area labeled samples exist. According to the method, the insulator feature analysis sub-network is used for extracting insulator features and performing defect classification, the expert knowledge analysis sub-network is used for positioning a damage area and extracting expert knowledge features, positive samples are generated for the same image, and a feedback loop of insulator device image features and expert knowledge features is formed; according to the method, the samples of the unmarked damage area are effectively utilized, mutual increment of the two features in the training process is achieved, the attention degree on the damage area is improved in a defect classification task and a damage positioning task, and meanwhile defect classification and damage positioning are achieved in a high-precision mode.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent detection method and system for insulators in a power supply system. Background Art

[0002] As mechanical support and electrical insulation components in transmission lines, insulators are widely used in power supply systems to provide tension support for transmission lines and isolate transmission conductors from grounding structures, thereby preventing potential dangers during transmission. They play a vital role in power transmission and transformation. Due to long-term high-load operation and continuous exposure to the natural environment, insulators are one of the components with the highest damage rate in transmission lines. They may suffer from various faults such as breakage and flashover, which seriously threaten the reliable operation of electrical equipment. Therefore, regular inspection and testing of insulator components and early detection of potential insulator faults can effectively avoid major accidents and are of great significance for maintaining the long-term safe operation of power systems.

[0003] The implementation of insulator defect detection includes three parts of work: 1. Insulators usually work in outdoor environments, and the imaging background of their monitoring images is complex. Therefore, insulator defect detection requires first accurately locating the insulator device in the image, which is an important basis for completing insulator defect detection. 2. Determine the state of the insulator device, that is, analyze whether the insulator device is in a normal state or a defective state. Furthermore, it also includes the identification of specific defect types. 3. Locate the damaged area of ​​the defective insulator device; the defective area is the local area that can most directly represent the damage condition, and can usually provide more accurate judgment information and decision-making basis for power supply system operation and maintenance personnel.

[0004] Traditional manual inspection methods usually involve manual analysis of the captured insulator monitoring images. However, long hours of work can cause visual fatigue for analysts, and problems such as missing defects and false fault reports are prone to occur. There are also problems such as low detection efficiency and uneven results. Especially when conducting inspections in complex environments (such as uneven lighting, background interference from buildings, occlusion from other power equipment, and bad weather), its shortcomings are more prominent. In recent years, with the rapid development of high-definition camera technology and computer vision technology, intelligent detection methods have gradually replaced early manual insulator defect detection. Traditional intelligent detection methods mainly rely on threshold segmentation, feature extraction, texture analysis and other technologies to segment the acquired insulator device images and extract features such as insulator contours, edges, and textures, and then combine support vector machines, cluster analysis and other pattern recognition technologies to analyze the insulator device status. However, the above methods require artificial design of features based on specific scenarios, and are easily affected by factors such as noise and background, making it difficult to achieve ideal results. In addition, the algorithms are complex and have poor generalization capabilities.

[0005] With the development of a new generation of artificial intelligence technology represented by deep learning technology, intelligent image recognition technology has been successfully applied in many fields. Some researchers consider using deep learning-based target detection or image segmentation technologies to intelligently identify insulators and their defects in insulator device images. For example, Zheng et al. proposed an infrared insulator device image detection model based on an improved feature fusion algorithm, combining multi-scale features with feature enhancement modules to improve the ability to extract infrared image features of insulators. Ling et al. used the Faster R-CNN network to detect the insulator string and segmented the damaged part of the insulator string through the U-Net network. Tao et al. proposed a cascade detection method, which further detected and located the self-explosion defects of the corresponding insulators based on the detection of insulators in the image using a target detection algorithm. Zhang Ye et al. built a Fine Mask R-CNN model by introducing an information fusion module with parallel operations of multi-scale convolution kernels, and realized the detection of bird-involved defects in insulators. However, this type of insulator defect detection method requires the collection of a large number of insulator defect samples to train the network. In the actual detection process, only a small number of insulator defect categories can collect enough defect samples, or it takes a very high cost to obtain enough samples, especially the need to obtain labeled samples for the damaged area. In addition, in addition to the high inter-class variance between different defect categories in insulator devices (i.e., the feature differences between different damage categories), insulator devices also have a large intra-class variance (i.e., the appearance differences between different devices with the same defect type); the distribution of insulator defect categories (damage types) in the actually constructed insulator device image dataset is often highly unbalanced. Therefore, it is difficult to obtain balanced and large number of defect samples for most categories of insulator defects. Therefore, from a practical point of view, small sample learning technology is usually used for training and detection.

[0006] Small sample learning technology is mainly proposed to solve the problem of insufficient sample size of (labeled) data in training. It solves the problem of not being able to obtain a large number of samples in network training by enhancing small sample category data, metric learning and meta-learning. Existing studies have considered exploring insulator defect detection methods under small sample conditions. For example, Li et al. used a relational network to realize railway insulator classification under small sample conditions. This method can only classify insulator defects and cannot locate the location of insulator defects. Zhai Yongjie et al. proposed a disc insulator defect detection method based on hybrid sample transfer learning, generated artificial defect samples and combined them with real samples, proposed a balanced loss function, and increased the model's attention to positive samples, thereby realizing insulator defect detection under few samples. The human interference in this sample transfer learning method is too strong, and the detection accuracy needs to be improved. In addition, the simultaneous demand for classification features and positioning features further complicates the problem, and the aforementioned small sample learning technologies cannot solve such problems at the same time. Summary of the invention

[0007] The object of the present invention is to provide an intelligent detection method and system for insulators in a power supply system to address all or part of the above-mentioned problems, so as to simultaneously complete defect classification of insulator components and location of damaged areas with only a small number of defect area labeled samples.

[0008] The technical solution adopted by the present invention is as follows:

[0009] An intelligent detection method for insulators in a power supply system, comprising:

[0010] The process of training the detection model:

[0011] Extracting a first characterization feature of an insulator device image in a sample set, wherein the sample set includes insulator device images with damaged areas marked and insulator device images without damaged areas marked, and all insulator device images are marked with insulator states;

[0012] Performing a first feature enhancement on the first characterization feature to obtain a first feature view;

[0013] classifying the insulator state according to the first characteristic view;

[0014] Locating and extracting a second characterizing feature of a damaged area of ​​the insulator device image;

[0015] Performing a second feature enhancement on the second characterization feature to obtain a second feature view;

[0016] Training the detection model with the goal of maximizing the consistency between the second feature view and the first feature view and minimizing the error in classifying the insulator state;

[0017] Testing process:

[0018] The trained detection model is used to identify the image of the insulator device to be inspected, and the insulator status and damage area are obtained.

[0019] The present invention also provides an intelligent detection system for insulators in a power supply system. The system is configured with a processor and a computer-readable storage medium. The processor runs a computer program in the computer-readable storage medium to execute the above-mentioned intelligent detection method for insulators in the power supply system.

[0020] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0021] 1. Under the condition that there are only a few samples with damaged areas marked, the present invention adopts a multi-task learning method. Aiming at the problem of imbalanced defect categories in the sample set, the supervised classification learning method using cross entropy loss is used to classify defects. Aiming at the situation that there are only a few samples with defective areas marked in the sample set, a weakly supervised deep contrast learning method is designed to locate damaged areas. In addition, the features of the two tasks are mutually enhanced during the learning process, so that both defect classification and damage location pay more attention to damaged areas, thereby improving the accuracy of the model in detecting defect categories and defective areas under small samples.

[0022] 2. Based on the insulator feature analysis subnetwork, the present invention designs an expert knowledge analysis subnetwork to predict the damaged area and extract features of unlabeled samples according to the prior expert knowledge features, and compare and learn with the insulator feature view extracted by the insulator feature analyzer to generate positive samples of image feature views for the same sample, forming a mutually reinforcing feedback loop between the image features of the insulator device and the expert knowledge features, realizing mutual enhancement of features to produce more effective knowledge-enhanced features, and at the same time enhancing the extraction accuracy and training efficiency of the two subnetworks for defect features, thereby improving the speed and accuracy of simultaneous detection of defect areas and defect categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0024] Figure 1 It is a data flow diagram of the intelligent detection method of insulators in the power supply system.

[0025] Figure 2 It is the structural diagram of the insulator feature analysis sub-network IFANet.

[0026] Figure 3 This is the IFEi submodule structure diagram.

[0027] Figure 4 This is the FEi unit structure diagram.

[0028] Figure 5 It is a structural diagram of a preferred embodiment of a feature expression subunit.

[0029] Figure 6 It is a structural diagram of a preferred embodiment of a feature update subunit.

[0030] Figure 7 It is a structural diagram of a preferred embodiment of the residual connection submodule.

[0031] Figure 8 It is a structural diagram of the expert knowledge feature mapping module EKFMM.

[0032] Fig. 9 This is the structural diagram of the KFEr submodule.

[0033] Fig.10 It is a structural diagram of a preferred embodiment of a defect area feature enhancement module DAFEM.

[0034] Fig.11 This is a comparison chart of the average performance of defect classification between the present invention and five typical methods.

[0035] Fig.12 , Fig.13 The figures are performance comparison diagrams of the present invention and five typical methods for defect and flashover classification.

[0036] Fig.14 This is a comparison chart of the average performance of the present invention and five typical methods for damage area positioning.

[0037] Fig.15 , Fig.16 The figures are performance comparison diagrams of the present invention and five typical methods for locating defect areas and flashover areas.

[0038] Fig.17 It is the average performance comparison chart of the defect classification ablation experiment.

[0039] Fig.18 , Fig.19 They are the performance comparison diagrams of ablation experiments for defect and flashover classification. DETAILED DESCRIPTION

[0040] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0041] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other equivalent or alternative features having similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0042] The present invention aims to improve the defect category recognition and damage area location of insulators by introducing weakly supervised deep learning. By using expert knowledge as prior information, expert knowledge features for insulator detection are effectively generated, and by using expert knowledge features as auxiliary information, the comparative learning effect of insulator device image features is enhanced.

[0043] Example 1

[0044] An intelligent detection method for insulators in a power supply system includes a process of training a detection model and a process of identifying defect categories and damaged areas of an insulator device image using the trained detection model.

[0045] The process of training the detection model:

[0046] Extracting insulator features (i.e., first characterization features) of insulator device images in a sample set, wherein the sample set includes insulator device images (usually only a small portion) with damaged areas marked, and insulator device images without damaged areas marked, and all insulator device images are marked with insulator states;

[0047] Performing a first feature enhancement on the insulator feature to obtain an insulator device image feature view;

[0048] According to the insulator device image feature view, the insulator state (i.e., defect category) is classified, and the insulator state generally includes three states: normal, defective, and flashover;

[0049] Locate and extract expert knowledge features (i.e., second characterization features) of the damaged area of ​​the insulator device image;

[0050] Performing second feature enhancement on the expert knowledge feature to obtain an expert knowledge feature view;

[0051] The detection model is trained with the goal of maximizing the consistency between the expert knowledge feature view and the insulator device image feature view and minimizing the error in insulator state classification.

[0052] The above detection model involves the extraction and enhancement process of insulator features and expert knowledge features, respectively. In some embodiments, Figure 1 As shown, the detection model is designed to include an insulator feature analysis subnetwork IFANet and an expert knowledge analysis subnetwork EKANet to extract and enhance insulator features and expert knowledge features respectively.

[0053] (1) Insulator feature analysis subnetwork IFANet: It is composed of the insulator feature extraction module IFEM (insulator feature extraction module), the insulator feature mapping module IFMM (insulator feature mapping module) (composed of two multi-layer perceptrons MPL with ReLU) and the insulator classification unit ICU (insulator classification unit,).

[0054] (2) Expert knowledge analysis subnetwork EKANet: It consists of a defect area feature enhancement module DAFEM (defect area feature enhancement module) and an expert knowledge feature mapping module EKFMM (expert knowledge feature mapping module) (composed of two KFEMs modules and a multi-layer perceptron MLP with ReLU).

[0055] The insulator feature extraction module IFEM is used to extract the insulator features of the insulator device image; the insulator feature mapping module IFMM performs the first feature enhancement on the insulator features, and the insulator classification unit ICU classifies the insulator status as normal, damaged or flashover according to the insulator device image feature view. The defect area feature enhancement module DAFEM is responsible for locating and extracting the expert knowledge features of the damaged area of ​​the insulator device image, and the expert knowledge feature mapping module EKFMM performs the second feature enhancement on the expert knowledge features.

[0056] In the process of training the detection model, for a group of insulator device images (only a small number of insulator device images are marked with damaged areas), the detection model uses the above two sub-networks to construct two feature views: the insulator device image feature view IFView (insulator featureview) constructed by the insulator feature analysis sub-network IFANet, and the expert knowledge feature view EKFView (expert knowledge feature view) constructed by the expert knowledge analysis sub-network EKANet.

[0057] Insulator device image feature view IFView, first, the insulator feature extraction module IFEM outputs the image representation IR (image representation) of the insulator device image image, that is, the insulator feature: IR = IFE (image), IFE (·) represents the internal processing operation of the insulator feature extraction module IFEM. Then, the insulator feature mapping module IFMM performs the first feature enhancement on the image representation IR and outputs the image mapping IP (image projection), that is, the first feature view: IP = IFMM (IR).

[0058] Expert knowledge feature view EKFView: First, the expert knowledge feature EKR (expert knowledge representation) is obtained by using the defect area feature enhancement module DAFEM. Then the expert knowledge feature mapping module EKFMM outputs the expert knowledge feature view RIP: RIP = EKFMM (EKR), EKFMM (·) represents the internal processing operation of the expert knowledge feature mapping module EKFMM.

[0059] When two feature views are obtained, the consistency between IP and RIP is maximized, and the classification error from the image mapping IP is minimized. In this embodiment, for the task of locating the damaged area in the insulator device image, the consistency loss between the expert knowledge feature view and the insulator device image feature view is characterized by a contrast loss function, and in view of the imbalance of defect categories in the insulator device image sample set, the error loss of the insulator state classification is characterized by a weighted cross entropy classification loss function. Therefore, the detection model performs a multi-task learning process, one task is responsible for supervised defect classification, and the other task is responsible for damage area location based on weakly supervised contrast learning.

[0060] The total loss of the detection model is expressed as L, which consists of two parts: weakly supervised contrast loss L1 and supervised classification loss L2.

[0061] L=β1*L1+β2*L2 (1)

[0062] β1 and β2 allow different weights to be set for the weakly supervised contrast loss L1 and the supervised classification loss L2. Their value ranges are both 0 to 1, and the default values ​​are both 0.5.

[0063] 1) Weakly supervised contrast loss L1.

[0064] The weakly supervised contrast loss function utilizes two modal information (insulator device image feature view IFView and expert knowledge feature view EKFView). The construction process is as follows: a batch of samples (N) are randomly selected, and a contrast prediction task is defined on the paired augmented samples derived from the batch of samples. Assume that in a batch of samples (N), IFView dp It is an image of the defect area defect and the position pos in the insulator device. Defect sample (IFView dp ,EKFView dp ) Defined as:

[0065]

[0066] Where L2||·|| represents the L2 distance. η dpis a small constant used to ensure Not 0, IFView(v dp ) represents the image feature view corresponding to the damaged area p and the insulator part d, EKFView(v' dp ) refers to the expert knowledge feature view corresponding to the damaged area p and the insulator position d, EKFView(v' kp ) refers to the expert knowledge feature view corresponding to the damage area p and the insulator part k, and M represents the total number of insulator parts in an insulator device image. The weakly supervised contrast loss L1 is calculated on all the damage area-insulator device images in a batch of samples.

[0067]

[0068] 2) Supervised classification loss L2. We send the output of the insulator feature mapping module IFMM to the insulator classification unit ICU (simple linear classifier). The supervised classification loss is defined as:

[0069]

[0070] state 1 and state 0 They represent the weights configured for defective samples and normal samples, y' represents the predicted insulator state label, y represents the insulator state label annotated for the sample, y = 1 represents the insulator state of the sample is defective, y = 0 represents the insulator state of the sample is normal, and hard represents the difficulty learning weight configured for samples of different insulator states to drive the model to learn more information from difficult samples. η2 is a small constant used to ensure that L2 is not 0.

[0071] This embodiment uses contrastive learning for the following two reasons: (1) In actual scenes, only a small number of insulator device images have damaged area positioning annotations (bounding boxes), and contrastive learning can effectively utilize insulator device images without labeled damaged areas; (2) Contrastive learning is robust in classification tasks of category-imbalanced datasets. In practice, most insulator device image datasets have an imbalance in the number of defect categories. In classical contrastive learning, the two feature perspectives often come from the same field. The difference is that the feature perspectives of the positive samples (defective insulators, including defects and flashovers) of the method of this embodiment come from knowledge in different fields (existing research has shown that the feature perspectives from multi-field knowledge should also be consistent), while the feature perspectives of the negative samples (normal insulators) come from knowledge in the same field.

[0072] 1. Positive sampling: In order to obtain a pair of positive sampling feature views, this embodiment randomly selects a labeled image with a specific defect and generates two feature views for it. The first view is its insulator device image feature view, and the other view is its expert knowledge feature view (obtained and enhanced by the defect area feature enhancement module DAFEM).

[0073] 2. Negative sampling: In order to identify negative difficult samples, unlike randomly selected images, this embodiment further uses "similar difficult sample" images. Therefore, normal insulator components or images of insulator components of the same type but with different defects are used as negative samples and are called "similar difficult sample" images.

[0074] For the test process after training, only the defect area feature enhancement module DAFEM in the insulator feature analysis subnetwork IFANet and the expert knowledge analysis subnetwork EKANet are retained to perform defect classification and damage area location.

[0075] like Figure 2 As shown, in some embodiments, the insulator feature analysis sub-network IFANet includes three IFEMs (insulator feature extraction modules) modules connected in sequence; each feature extraction module IFEMs is composed of two insulator feature extraction sub-modules IFEi connected together.

[0076] like Figure 3 As shown, the IFEi submodule includes three FEi units, and the first FEi unit is connected to the second FEi unit through a residual connection and then connected to the third FEi unit.

[0077] like Figure 4 As shown, the FEi unit includes two groups of feature expression subunits FESU and two groups of feature update subunits FUSU. Specifically, the FEi unit includes the first convolution module x, the first feature expression subunit FESU, the first feature update subunit FUSU, the second feature update subunit FUSU, the second feature expression subunit FESU and the second convolution module x in sequence. The first convolution module x is also connected to the first feature update subunit FUSU, the second feature update subunit FUSU and the second feature expression subunit FESU respectively, the first feature expression subunit FESU is connected to the second feature update subunit FUSU, and the second feature update subunit FUSU is connected to the second convolution module x.

[0078] The feature expression subunit FESU consists of 4 residual connection submodules, 7 convolutional layers and 1 ReLU activation layer. Figure 5As shown, in some embodiments, the feature expression subunit FESU includes a first residual connection submodule (Residual Block), a first convolutional layer (Conv Layer), a second convolutional layer, a second residual connection submodule and a third convolutional layer connected in sequence, and between the first convolutional layer and the second convolutional layer, there are also a third residual connection submodule, a fourth convolutional layer, a fifth convolutional layer, a first Relu activation layer (Relu Layer), a sixth convolutional layer, a seventh convolutional layer and a fourth residual connection submodule connected in sequence, and the fourth convolutional layer is also connected to the seventh convolutional layer.

[0079] The feature update subunit FUSU consists of 2 residual connection submodules, 4 convolutional layers and 1 ReLU activation layer. Figure 6 As shown, in some embodiments, the feature update subunit FUSU includes a fifth residual connection submodule, an eighth convolutional layer, a ninth convolutional layer, a second Relu activation layer, a sixth residual connection submodule, a tenth convolutional layer and an eleventh convolutional layer connected in sequence, the fifth residual connection submodule is also connected to the tenth convolutional layer, and the sixth residual connection submodule is also connected to the eleventh convolutional layer.

[0080] like Figure 7 The figure shows the hierarchical design of a preferred embodiment of the residual connection submodule, which consists of 3 convolutional layers and two Relu activation layers, and also includes two cross-layer shortcut connections. The twelfth convolutional layer, the third Relu activation layer, and the thirteenth convolutional layer are connected in sequence, the fourteenth convolutional layer, the fourth Relu activation layer, and the fifteenth convolutional layer are connected in sequence, and the output of the thirteenth convolutional layer is connected to the input of the residual connection submodule across layers, and then output to the fourteenth convolutional layer in one way, and the output of the fifteenth convolutional layer is connected to the output of the convolutional layer across layers in another way.

[0081] like Figure 8 The figure shows the structure of the expert knowledge feature mapping module EKFMM. The expert knowledge feature mapping module EKFMM includes two KFEMs (knowledge feature extration modules) connected in sequence and a multilayer perceptron with a ReLu activation layer. The structure of the KFEMs module is the same as that of the IFEMs module mentioned above. Specifically, Figure 8 As shown, the KFEMs module is composed of two KFEr submodules connected together.

[0082] like Fig. 9 As shown in the figure, the KFEr submodule consists of three knowledge feature extraction units FE k The first FEk unit is connected to the second FEk unit by residual connection and then connected to the third FEk unit. The structure of the FEk unit is as follows Figure 4 shown.

[0083] The core component of the expert knowledge analysis subnetwork EKANet is the defect area feature enhancement module DAFEM, which uses a feedback loop to learn the damaged area positioning features from the obtained expert knowledge feature view and obtain the image features of the positive sample. This module is responsible for locating and extracting the expert knowledge features of the damaged area of ​​the insulator device image.

[0084] like Fig.10 The figure shows a structural diagram of a defect area feature enhancement module DAFEM in some embodiments, which includes a bounding box generation unit and a defect information acquisition unit.

[0085] The bounding box generation unit is used to mark the damaged area. Specifically, the marking process includes:

[0086] Step 1: Identify whether there is a damaged area marked in the insulator device image. That is, according to whether there is damaged area marking information in the insulator device image, determine whether it is an insulator device image without a damaged area marked or an insulator device image with a damaged area marked.

[0087] Step 2: For the insulator device image without marked damaged areas, perform gradient weighted class activation mapping on the insulator features to generate a heat map, use an adaptive threshold to predict the damaged area on the heat map, and generate standard information on the location of the damaged area, such as generating a bounding box marking the location of the damaged area;

[0088] For the insulator device images with marked damaged areas, the marked information is directly used.

[0089] The defect information acquisition unit is used to extract the characteristics of the damaged area:

[0090] For insulator device images without marked damage areas, expert knowledge features are extracted from the predicted damage areas; for insulator device images with marked damage areas, expert knowledge features are extracted from the marked damage areas.

[0091] The expert knowledge features used are composed of the following two categories of features:

[0092] 1) First-order statistical characteristics: used to measure the distribution of pixel intensity in the damaged area, including energy (measurement of pixel value intensity), entropy (measurement of pixel value uncertainty), and the maximum, average, and median values ​​of pixel grayscale in the area.

[0093] 2) Grayscale features. These include grayscale co-occurrence matrix features, grayscale size region matrix features, grayscale run matrix features, and neighborhood grayscale difference matrix features.

[0094] Example 2

[0095] This embodiment constructs an insulator device image sample set to conduct a verification experiment on the method of Embodiment 1, and verifies the performance of the method in defect identification and damaged area location.

[0096] The data set used in this experiment contains 10,000 images of insulator components, among which the number of images corresponding to different types of insulator components is also extremely unbalanced: the number of normal insulator component image samples (7,000 images) is far more than the number of defective samples (3,000 images); further, the number of corresponding defect categories in the above defective samples is also quite different, including 2,250 images of insulator defects and 750 images of insulator flashovers. There are three categories of insulator status labels in the data set, one of which is "Normal", and two defect labels, namely Damage and Flashover. The hardware configuration for the insulator defect intelligent detection experiment is shown in Table 1.

[0097] Table 1 Hardware configuration of the insulator device defect intelligent detection experiment

[0098]

[0099] The sample set also includes 3000 damage area bounding boxes (i.e., damage area annotation information) of two defect types (damage and flashover) annotated by insulator inspection engineers for 3000 defective insulator device images. This embodiment separates some insulator device images with damage bounding boxes from the entire data set. Therefore, there are two groups of images, namely, 600 images (450 insulator defect images and 150 insulator flashover images) with "damage area annotation" and 2400 images (1800 insulator defect images and 600 insulator flashover images) without "damage area annotation". During the experiment, the image data set "without damage area annotation" was divided into three subsets: 70% for training, 10% for verification, and 20% for testing. For the image data set "with damage area annotation", the data set was randomly divided into two subsets: 20% for training and 80% for testing. Note that there is no overlap of insulator device images between all subsets.

[0100] The batch size is set to 8, and the detection model is trained for 150 epochs. The model is optimized by the Adam method, and the initial learning rate is decayed by 10% (starting from 0.001%) every 15 epochs. The weights β1 and β2 of the loss function are set to 0.5. In addition, the relevant models in the proposed method are all implemented based on the PyTorch framework, and four TITAN RTX GPUs are used to train the model.

[0101] Intelligent detection of insulator component defects includes three subtasks:

[0102] (1) Accurately locate the insulator component from the insulator component image;

[0103] (2) Implement status classification for insulator components;

[0104] (3) Locate the damaged area of ​​abnormal insulator components.

[0105] For subtask 1, this embodiment uses the YOLO v8 network to achieve accurate positioning of insulator devices, and its positioning accuracy reaches more than 99%. Therefore, the focus of the test is to evaluate the performance of subtasks 2 and 3, and compare them with existing typical methods. For the insulator device classification task, the area under curve (AUC) of the receiver operating characteristic curve (ROC) is used to measure the model performance. For the damage area positioning task of abnormal insulator devices, the intersection over union (IoU) is used, that is, the positioning performance of the model is measured based on the annotated true bounding box and the predicted bounding box of the model. The positioning results are only calculated on the test set using the damage area annotation dataset. The damage area positioning is defined as correct only when IOU>Thr(IOU), where Thr(·) is the set intersection over union threshold.

[0106] 1 shows the ROC AUC score of the method proposed in Example 1 (denoted by OURS) for specific defect classification and the average ROC AUC score of two defect classifications, and compares them with five typical existing insulator defect classification methods (denoted by METHOD-1 to METHOD-5, respectively). Fig.11 is the average performance (ROC AUC score) comparison of the proposed method and five typical methods for two defect classifications (average value for two defects), Fig.12 , Fig.13 The performance (ROC AUC score) of insulator defect and insulator flashover classification experiments are compared respectively.

[0107] Table 2 AUC and average AUC of defect classification compared with comparison methods

[0108]

[0109] Comparison of existing insulator defect classification algorithms:

[0110] [METHOD-1]: "Insulator defect identification method based on SSIM-Sobel and multi-feature fusion" proposed by Zhao Yuanxi, Qiu Zhibin, Zhu Xuan, etc.

[0111] [METHOD-2]: "Research on intelligent recognition algorithm for defects in aerial insulator images" proposed by Dai Yongdong, Jin Yang, and Dai Yufan;

[0112] [METHOD-3]: "Insulator defect detection method and system based on local rotation feature learning" proposed in Chinese patent document CN114418964A;

[0113] [METHOD-4]: "Aerial insulator multi-defect detection network integrating CAT-BiFPN and attention mechanism" proposed by Kang Jie, Wang Qing, Liu Wenbo, etc.

[0114] [METHOD-5]: “Research on Insulator Fault Detection Based on Deep Learning” proposed by Zhang Yansheng, Wang Chenglong, and Liu Yuanhong.

[0115] The insulator device classification experiment shows that the proposed method obtains better ROC AUC scores in the insulator device classification experiment, and its overall performance of defect classification is significantly improved, including "insulator defect" and "insulator flashover". The above experimental results show that by introducing a cross-modal contrast learning branch and a network structure carefully designed for the image features and expert knowledge features of insulator devices, the performance of insulator device image analysis can be effectively improved, and it has good robustness for different types of defect classification tasks. In summary, the defect classification experimental results reflect the superiority of the proposed method over the five existing typical methods in the task of insulator device classification.

[0116] For the insulator damage area location experiment, Table 3 shows the experimental results of the proposed method (referred to as OURS) for locating the damage area of ​​abnormal insulator devices at different IoU thresholds (including 0.3 and 0.5), as well as the comparison results with the above five existing typical methods (still referred to as METHOD-1 to METHOD-5 respectively).

[0117] Table 3 Comparison of the accuracy of damage area positioning under different IoU thresholds

[0118]

[0119]

[0120] Fig.14 Comparison of the average performance (IoU value) of the proposed method and five typical methods for locating the damaged area of ​​two types of insulator defects (average value for two defects). Fig.15 , Fig.16 The performance (IoU value) of the damage area location experiment for insulator defect and insulator flashover is compared respectively. According to the results of the insulator damage area location experiment shown in Table 3, it can be seen that:

[0121] (1) The proposed method has significantly better damage area localization ability than the comparison method at different IoU thresholds (0.3 and 0.5), and the average value of its damage area localization accuracy is significantly higher than that of the comparison method;

[0122] (2) The proposed method has better ability to locate the damaged area of ​​defective insulator components than the comparison method under different IoU thresholds (0.3 and 0.5), and its IoU value is higher than the comparison method;

[0123] (4) The proposed method has better ability to locate the damaged area of ​​flashover insulator components than the comparison method under different IoU thresholds (0.3 and 0.5), and its IoU value is higher than that of the comparison method.

[0124] (4) Compared with the comparison method, the proposed method maintains superior performance in terms of damage area location accuracy as the IoU threshold increases. For example, when the threshold increases, the decrease in the IoU values ​​corresponding to "insulator defect" and "insulator flashover" is significantly smaller than that of the comparison method.

[0125] According to the experimental results, the proposed method outperforms five typical comparison methods. Specifically, the proposed method outperforms the comparison methods in terms of damage area localization performance for the above two defects and robustness when the IoU threshold increases.

[0126] In addition, this embodiment also conducts an ablation experiment on the defect classification task to verify that the designed classification loss function L2 and the defect area feature enhancement module DAFEM improve the analysis performance of the insulator device image sample set with unbalanced defect categories, so as to prove the rationality and effectiveness of the designed detection system as a whole.

[0127] Table 4. Ablation experiments of loss function and LFEM module for defect classification task

[0128]

[0129] The performance of the basic model (using the standard cross entropy loss function for defect classification and not using the defect area feature enhancement module DAFEM, marked as “Ours (BASE)”) is compared with the cases where only the proposed classification loss function L2 is used (marked as “Ours (LOSS)”) or only the defect area feature enhancement module DAFEM is used (marked as “Ours (DAFEM)”). The experimental results are shown in Table 4. Fig.17is the average performance (ROCAUC score) comparison between the proposed method and the ablation experiment method for two defect classifications (average value for two defects), Fig.18 , Fig.19 The performance (ROC AUC score) of insulator defect and insulator flashover classification experiments are compared respectively. The ablation experiment results show that the use of the proposed classification loss L2 and the introduction of the defect area feature enhancement module DAFEM significantly improve the performance of the defect classification task, and the contribution of the DAFEM module is greater because the DAFEM module tends to generate more robust expert knowledge features, which further enhances the insulator feature extraction module IFEM's attention to the defect area.

[0130] Example 3

[0131] This embodiment proposes an intelligent detection system for insulators in a power supply system, which is configured with a processor and a computer-readable storage medium, wherein a computer program is stored in the computer-readable storage medium, and the method of embodiment 1 can be executed by running the program. The processor runs the computer program in the computer-readable storage medium to execute the intelligent detection method for insulators in a power supply system described in embodiment 1.

[0132] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. An intelligent detection method for insulators in a power supply system, comprising: The process of training the detection model: Extracting a first characterization feature of an insulator device image in a sample set, wherein the sample set includes insulator device images with damaged areas marked and insulator device images without damaged areas marked, and all insulator device images are marked with insulator states; Performing a first feature enhancement on the first characterization feature to obtain a first feature view; classifying the insulator state according to the first characteristic view; It is characterized by further comprising: Locating and extracting a second characterizing feature of a damaged area of ​​the insulator device image; Performing a second feature enhancement on the second characterization feature to obtain a second feature view; Training the detection model with the goal of maximizing the consistency between the second feature view and the first feature view and minimizing the error in classifying the insulator state; Testing process: The trained detection model is used to identify the image of the insulator device to be inspected, and the insulator status and damage area are obtained.

2. The intelligent detection method for insulators in a power supply system according to claim 1, characterized in that: The locating and extracting the second characterizing feature of the damaged area of ​​the insulator device image includes: Identify whether there are damaged areas marked in the insulator device image; For the insulator device image without marked damage area, a gradient weighted class activation mapping is performed on the first characterization feature to generate a heat map, and the damage area is predicted on the heat map using an adaptive threshold, and a second characterization feature is extracted from the predicted damage area; For the insulator device image with the damaged area marked, a second characterization feature is extracted from the marked damaged area.

3. The intelligent detection method for insulators in a power supply system according to claim 1, characterized in that: The second characterization features include: first-order statistical features and grayscale features.

4. The intelligent detection method for insulators in a power supply system according to claim 1, characterized in that: The step of training the detection model with the goal of maximizing the consistency between the second feature view and the first feature view and minimizing the error in classifying the insulator state comprises: The detection model is trained by minimizing the loss of the detection model, where the loss of the detection model is calculated as: L=β1*L1+β2*L2, Wherein, L represents the loss of the detection model, L1 represents the consistency loss between the second feature view and the first feature view, L2 represents the error loss of the insulator state classification, β1 and β2 are weights set for the loss L1 and the loss L2 respectively; In the formula, L2||·|| represents the L2 distance, η dp is a non-zero constant, N is the sample batch size, IFView(v dp ) represents the image feature view corresponding to the damaged area p and the insulator part d, EKFView(v' dp ) refers to the expert knowledge feature view corresponding to the damaged area p and the insulator position d, EKFView(v' kp ) refers to the expert knowledge feature view corresponding to the damaged area p and the insulator part k, and M represents the total number of insulator parts in an insulator device image; state 1 and state 0 They represent the weights configured for defective samples and normal samples respectively, y' represents the predicted insulator state label, y represents the insulator state label annotated for the sample, y = 1 indicates that the insulator state of the sample is defective, y = 0 indicates that the insulator state of the sample is normal, hard is the difficulty learning weight configured for samples with different insulator states, and η2 is a set non-zero constant.

5. The intelligent detection method for insulators in a power supply system according to claim 1, characterized in that: An insulator feature extraction module is used to extract the first characterization feature of the insulator device image, and the insulator feature extraction module includes three IFEMs modules connected in sequence; each of the IFEMs modules is composed of two IFEi sub-modules connected; each IFEi sub-module includes three FEi units, and the first FEi unit is connected to the third FEi unit after the first FEi unit is residually connected to the second FEi unit; each FEi unit includes a first convolution module, a first feature expression subunit, a first feature update subunit, a second feature update subunit, a second feature expression subunit and a second convolution module in sequence, wherein the first convolution module is also connected to the first feature update subunit, the second feature update subunit and the second feature expression subunit respectively, the first feature expression subunit is connected to the second feature update subunit, and the second feature update subunit is connected to the second convolution module.

6. The intelligent detection method for insulators in a power supply system according to claim 5, characterized in that: The feature expression subunit includes a first residual connection submodule, a first convolutional layer, a second convolutional layer, a second residual connection submodule and a third convolutional layer connected in sequence; between the first convolutional layer and the second convolutional layer, the third residual connection submodule, the fourth convolutional layer, the fifth convolutional layer, the first Relu activation layer, the sixth convolutional layer, the seventh convolutional layer and the fourth residual connection submodule are connected in sequence; the fourth convolutional layer is also connected to the seventh convolutional layer.

7. The intelligent detection method for insulators in a power supply system according to claim 5, characterized in that: The feature update subunit includes a fifth residual connection submodule, an eighth convolutional layer, a ninth convolutional layer, a second Relu activation layer, a sixth residual connection submodule, a tenth convolutional layer and an eleventh convolutional layer connected in sequence, the fifth residual connection submodule is also connected to the tenth convolutional layer, and the sixth residual connection submodule is also connected to the eleventh convolutional layer.

8. The intelligent detection method for insulators in a power supply system according to any one of claims 5 to 7, characterized in that: The second characterization feature is enhanced by a second feature mapping module using expert knowledge. The expert knowledge feature mapping module includes two KFEMs modules and a multi-layer perceptron connected in sequence. The structure of the KFEMs module is the same as that of the IFEMs module.

9. The intelligent detection method for insulators in a power supply system according to claim 1, characterized in that: The first characterization feature is enhanced using an insulator feature mapping module, wherein the insulator feature mapping module is formed by connecting two multi-layer perceptrons with Relu activation layers.

10. An intelligent insulator detection system in a power supply system, the system being configured with a processor and a computer-readable storage medium, characterized in that: The processor runs the computer program in the computer-readable storage medium to execute the intelligent detection method for insulators in a power supply system according to any one of claims 1 to 9.

Images