CN-121998922-A - Distribution network component defect detection method and device and readable storage medium

CN121998922ACN 121998922 ACN121998922 ACN 121998922ACN-121998922-A

Abstract

The invention belongs to the technical field of distribution network defect detection, and relates to a distribution network component defect detection method, a distribution network component defect detection device and a readable storage medium; the method comprises the steps of inputting YOLOv images of a to-be-detected distribution network component into a programmable gradient information network of a network, extracting auxiliary features, inputting an auxiliary feature map into a first head network, inputting YOLOv images of the to-be-detected distribution network component into a backbone network of the network, outputting three-layer feature maps, inputting SPPELAN the third-layer feature map into a module, outputting a target third-layer feature map, respectively inputting the first-layer feature map, the second-layer feature map and the target third-layer feature map into a first REPNCSPELAN module and a second MHSENet module in a neck network, outputting a fusion feature map, inputting the fusion feature map into a second head network, and obtaining defect identification and classification results of the to-be-detected distribution network component based on prediction results output by the first head network and the second head network.

Inventors

WU NING
JIN PENGFEI
GONG TIANTIAN
WANG MANSHANG
CHEN JIAKAI
SHI YANG
He Tianyao
DING LINGWEI
DING BAIWEN

Assignees

国网江苏省电力有限公司镇江供电分公司

Dates

Publication Date: 20260508
Application Date: 20260113

Claims (10)

1. The detection method for the defects of the distribution network component is characterized by comprising the following steps of: inputting the image of the distribution network component to be detected into a programmable gradient information network of YOLOv networks, and inputting an auxiliary feature map extracted by the programmable gradient information network into a first head network of YOLOv networks; Inputting the image of the to-be-detected distribution network component into a YOLOv network backbone network to output three-layer characteristic diagrams, inputting the third-layer characteristic diagram into a SPPELAN module for multi-scale pooling and aggregation, and outputting a target third-layer characteristic diagram; Inputting the first layer feature map, the second layer feature map and the target third layer feature map into a first REPNCSPELAN module, a second REPNCSPELAN module and a MHSENet module in a neck network of YOLOv network respectively to perform feature fusion, outputting a fusion feature map, inputting the fusion feature map into a second head network of YOLOv network, wherein inputting the target third layer feature map into MHSENet module to perform feature fusion comprises: Extracting local features of the target third-layer feature map layer by utilizing a multi-convolution sensing frame; The MHSEBlock double-path structural unit is utilized to carry out attention enhancement on the local features extracted by the multi-convolution sensing frame; And obtaining defect identification and classification results of the distribution network components to be detected based on the prediction results output by the first head network and the second head network.
2. The method for detecting defects of a distribution network component according to claim 1, wherein the multi-convolution sensing framework comprises a first convolution combination unit, a channel cutting unit, a second convolution combination unit, a third convolution combination unit and a convolution unit; MHSEBlock the dual path building block includes a first MHSEBlock unit and a second MHSEBlock unit.
3. The method for detecting defects of a distribution network component according to claim 2, wherein the step of extracting local features of the target third layer feature map layer by using the multi-convolution sensing frame, and the step of enhancing the attention of the local features extracted by using the MHSEBlock double-path structural unit comprises the steps of: inputting the target third layer characteristic diagram into a first convolution combination unit, and outputting a first depth characteristic diagram; inputting the first depth feature map into a channel cutting unit for channel dimension cutting; inputting the cut first depth feature map into a first MHSEBlock unit for attention enhancement, and outputting a target first depth feature map; inputting the first characteristic map of the target depth into a second convolution combination unit, and outputting a second depth characteristic map; inputting the second depth feature map into a second MHSEBlock unit for attention enhancement, and outputting a target second depth feature map; inputting the target second depth feature map into a third convolution combination unit, and outputting a third depth feature map; And adding elements of the first depth feature map, the second depth feature map and the third depth feature map, inputting the added elements into a convolution unit, and taking the output of the convolution unit as the output of a MHSENet module.
4. A method for detecting defects of a distribution network component according to claim 3, wherein the MHSEBlock unit performs attention enhancement on the input depth profile, and includes: inputting the depth feature map into a first convolution combination subunit and a second convolution combination subunit which are parallel, and outputting a first target feature map and a second target feature map; Inputting the first target feature map into a feature extraction subunit, and outputting a third target feature map; And adding elements of the second target feature map and the third target feature map, inputting the second target feature map and the third target feature map into a third convolution combination subunit, and outputting a target depth feature map.
5. The method for detecting defects of a distribution network component according to claim 4, wherein, The feature extraction subunit comprises a first feature extraction network and a second feature extraction network which are parallel, and outputs of the first feature extraction network and the second feature extraction network are subjected to cascade superposition to obtain a third target feature map; Or, the feature extraction subunit includes a first feature extraction network and a second feature extraction network connected in series, and obtains a third target feature map based on an output of the second feature extraction network.
6. The method for detecting defects in a distribution network component according to claim 5, wherein the output of the first feature extraction network is expressed as: , Wherein, the Representing a first feature extraction network An output of (2); A feature map representing an input first feature extraction network; Representing the output of the first 1*1 convolution block after the depth feature map is convolved; representing the output of the second 1*1 convolution block after the depth feature map is convolved; Representing the output of the third 1*1 convolution block after the depth feature map is convolved; representing a transpose; Representing a normalization operation; the output of the second feature extraction network is expressed as: , Wherein, the Representing a second feature extraction network An output of (2); a feature map representing an input second feature extraction network; representing average pooling; representing a fully connected layer; representation ReLu activates a function; Representing the Sigmoid activation function.
7. The method for detecting defects of a distribution network component according to claim 1, further comprising performing iterative training on the YOLOv network by using image samples of the distribution network component before inputting the image of the distribution network component to be detected into the YOLOv network, wherein a construction process of a loss function during the iterative training comprises: calculating the cross-over ratio loss and the weighted cross-over ratio weight coefficient based on a predicted target frame output by YOLOv network and a real target frame of an image sample of the distribution network component; Obtaining a first weighted cross-over ratio loss based on the product of the weighted cross-over ratio weight coefficient and the cross-over ratio loss; Based on the product of the first weighted cross-ratio loss and the gradient gain thereof, obtaining a second weighted cross-ratio loss, so that the convergence rate of YOLOv network is gradually increased along with the increase of iterative training times; based on the product of the first weighted cross-correlation loss and the non-monotonic focusing coefficient thereof, a third weighted cross-correlation loss is obtained, so that the convergence speed of YOLOv network is in direct proportion to the accuracy of the output prediction target frame along with the increase of the iterative training times; and obtaining a loss function based on the weighted sum of the first weighted cross-ratio loss, the second weighted cross-ratio loss and the third weighted cross-ratio loss.
8. The method for detecting defects of distribution network components according to claim 7, wherein the weighted cross-correlation weight coefficients The calculation formula of (2) is as follows: , Wherein, the Representing the width and height of the predicted target frame; representing the width and height of a real target frame; Representing the width and height of the target frame after the predicted target frame and the real target frame are combined; ; First weighted cross-ratio loss Expressed as: , Wherein, the Representing the cross-ratio loss; ; second weighted cross-ratio loss Expressed as: , Wherein, the Representing the gradient gain; a sliding average value representing the cross-ratio loss; Indicating an index super-parameter; third weighted cross-ratio loss Expressed as: , Wherein, the Representing a non-monotonic focusing coefficient; Representing the dynamic outlier factor(s), 、 Representing the adjustable parameter.
9. A device for detecting defects of a distribution network component, comprising: The first prediction module is used for inputting the image of the to-be-detected distribution network component into a programmable gradient information network of YOLOv networks, and inputting an auxiliary characteristic diagram extracted by the programmable gradient information network into a first head network of YOLOv networks; the characteristic diagram extracting module is used for inputting an image of the to-be-detected distribution network component into a backbone network of YOLOv network to output three-layer characteristic diagrams, inputting the third-layer characteristic diagram into the SPPELAN module for multi-scale pooling and aggregation, and outputting a target third-layer characteristic diagram; The second prediction module is configured to input the first layer feature map, the second layer feature map, and the target third layer feature map into a first REPNCSPELAN module, a second REPNCSPELAN module, and a MHSENet module in a neck network of YOLOv network respectively to perform feature fusion, output a fusion feature map, and input the fusion feature map into a second header network of YOLOv network, where inputting the target third layer feature map into the MHSENet module to perform feature fusion includes: The local feature extraction submodule is used for extracting local features of the third-layer feature map of the target layer by utilizing the multi-convolution sensing framework; An attention enhancer module for enhancing attention to local features extracted by the multi-convolution sensing framework by utilizing MHSEBlock dual-path structural units; the detection result acquisition module is used for acquiring defect identification and classification results of the distribution network component to be detected based on the prediction results output by the first head network and the second head network.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements the steps of the method for detecting defects of a distribution network element according to any of claims 1 to 8.

Description

Distribution network component defect detection method and device and readable storage medium Technical Field The invention relates to the technical field of distribution network defect detection, in particular to a distribution network component defect detection method and device and a computer readable storage medium. Background With the continuous expansion of the scale of the power system, the structure of the power distribution network is increasingly complex, and the operation safety and stability of the power distribution network are directly related to the reliability of electric energy supply. The distribution network equipment is exposed in an outdoor environment for a long time, is easily influenced by factors such as wind, rain, corrosion, animal interference and the like to generate a plurality of defects such as binding wire deficiency, transformer oil leakage, self-explosion of a glass insulator, lightning stroke flashover of a porcelain insulator, breakage of the porcelain insulator, loose wire drawing, bird nest and vine winding and the like, and if the defects cannot be found and treated in time, electrical short circuit, equipment damage and even line tripping can be caused, so that the running safety of a power grid is seriously threatened. Therefore, the automatic detection and identification of the defects of the distribution network components have important engineering significance. With the development of deep learning and computer vision technology, a target detection algorithm based on a convolutional neural network is widely applied to distribution network component defect detection. At present, researchers generally adopt main stream detection frames such as fast R-CNN, SSD and the like to carry out feature extraction and defect identification on the distribution network part images. For example, an improved SSD target detection algorithm is provided in the prior art, is used for identifying power abnormal equipment in the power inspection process, is also based on an improved regional convolutional neural network algorithm architecture, is combined with a pre-training VGG16 (Visual Geometry Group-layer model) deep convolutional network to construct a feature extraction module, and achieves breakthrough progress in the field of intelligent diagnosis of electric power fittings, and aiming at the phenomena of surface breakage and pollution flashover of insulators, a YOLOv 8-based surface defect identification algorithm is provided in the prior art, small target defects under a complex background can be effectively detected, and the accuracy and the practicability of power inspection aiming at insulators are improved. However, because the defects of the distribution network component have the characteristic of high similarity among classes, the convolutional neural network in the prior art mainly relies on a local receptive field to perform feature extraction, and is difficult to effectively capture the long-range dependency relationship and the weight difference among channels in the feature map, so that the model has weak distinguishing capability on defects (such as porcelain insulator breakage and glass insulator self-explosion) with similar appearance but different properties, and subtle defect omission and class confusion are easy to occur. In summary, the existing distribution network component defect detection method based on the traditional convolutional neural network has the problem that the detection accuracy is low because the distribution network component image characteristics cannot be effectively captured. Disclosure of Invention Therefore, the invention aims to solve the technical problem that the detection accuracy is low because the distribution network component defect detection method based on the traditional convolutional neural network in the prior art can not effectively capture the image characteristics of the distribution network component. In order to solve the technical problems, the invention provides a detection method for defects of a distribution network component, which comprises the following steps: inputting the image of the distribution network component to be detected into a programmable gradient information network of YOLOv networks, and inputting an auxiliary feature map extracted by the programmable gradient information network into a first head network of YOLOv networks; Inputting the image of the to-be-detected distribution network component into a YOLOv network backbone network to output three-layer characteristic diagrams, inputting the third-layer characteristic diagram into a SPPELAN module for multi-scale pooling and aggregation, and outputting a target third-layer characteristic diagram; Inputting the first layer feature map, the second layer feature map and the target third layer feature map into a first REPNCSPELAN module, a second REPNCSPELAN module and a MHSENet module in a neck network of YOLOv network respectively to perform fe