CN114862768A

CN114862768A - Improved YOLOv5-LITE lightweight-based power distribution assembly defect identification method

Info

Publication number: CN114862768A
Application number: CN202210392794.5A
Authority: CN
Inventors: 颜宏文; 万俊杰; 马瑞
Original assignee: Changsha University of Science and Technology
Current assignee: Changsha University of Science and Technology
Priority date: 2022-04-14
Filing date: 2022-04-14
Publication date: 2022-08-05

Abstract

The invention discloses a distribution assembly defect recognition method based on improved YOLOv5-LITE lightweight, which is characterized in that in order to facilitate the model to be deployed to a mobile equipment terminal, the method uses ShuffleNet V2 as a backbone network to extract features, constructs a YOLOv5-LITE lightweight neural network model, removes 1024 volume and 5 x 5 pooling of ShuffleNet V2, adopts global average pooling operation to replace, reduces the network parameter number and improves the model detection speed; by introducing 152 multiplied by 152 characteristic layers beneficial to fine-grained target detection, large, medium and small-scale defect prediction is realized; the adoption of deep separable convolution in the PANet architecture instead of downsampling makes the network more lightweight.

Description

Improved YOLOv5-LITE lightweight-based power distribution assembly defect identification method

Technical Field

The invention relates to the technical field of distribution lines, in particular to a lightweight power distribution assembly defect identification method based on improved YOLOv5-LITE, which is applied to the field of electric power.

Background

Distribution lines are the fundamental part of the operation of electric power systems and are also the links connecting users, distribution stations, power plants. 80% -90% of power system faults occur in a distribution network, a distribution line can be influenced by factors such as lightning, storms, magnetic fields and the like in the operation process, and once faults occur, the faults can have great influence on economy, industrial production and normal life of residents. The transformation from service digitization to digital service of the power transmission and distribution equipment is realized through advanced wireless sensing technology, communication technology, edge calculation and data analysis technology.

In recent years, the deep convolutional neural network has a great superiority in the field of fault location and identification of distribution lines, and current researches mainly include two types: the first is a two-stage target detection model represented by R-CNN, Fast R-CNN and Faster R-CNN; the other is a one-stage target detection model represented by YOLO and SSD, which directly performs position regression.

Therefore, the distribution line fault rapid positioning and identification device has great application value in researching distribution line fault rapid positioning and identification, and accurately repairing the defects and improving the fault identification efficiency for maintainers.

Disclosure of Invention

In order to realize the problem of positioning and identifying the defects of the power distribution assembly, the invention provides the following technical scheme:

an improved YOLOv5-LITE lightweight neural network model is provided, and defect positioning and identification are realized end to end; introducing a 152 x 152 scale feature layer beneficial to small target detection, realizing fine-grained detection, removing 1024 volume and 5 x 5 pooling operations of a ShuffleNet V2 backbone network, and replacing by adopting global average pooling operations to reduce model parameters. Meanwhile, pre-training weights are loaded through transfer learning and a plurality of convolution layers are pre-frozen, so that the weights at the initial training stage are prevented from being damaged, and the model precision is improved. The detection flow chart is shown in fig. 1. The model can realize accurate positioning and identification of three defects of cable separation gasket, cable and insulator separation and no-ring insulator.

In the deep learning model, in order to reduce the over-fitting problem, a large number of data samples are required for model training, and thus the number of samples is expanded by a data augmentation method. The data augmentation strategy adopted by the invention comprises morphological operations such as angle rotation, saturation adjustment, de-noising and defogging, image up-and-down turning, translation and the like, and 9700 images are obtained as a data set after enhancement. The lightweight YOLOv5 model uses a Mosaic data enhancement mode at the input end, four defect images are spliced in the modes of random scaling, random cutting and random arrangement, and the training speed of the model is accelerated.

And manually marking a rectangular frame for three defects, namely separation of a distribution line cable from a gasket, separation of the cable from an insulator and no-ring insulator in the image by using a Labelimg or LabelMe tool, and establishing a corresponding image database and a label database. After the labeling is finished, the data set is divided according to the proportion of 8:2, wherein 80% of the data set is divided into a training set and a verification set according to the proportion of 8:2, the rest 20% of the data set is used as a test set, and the number of samples of the training set, the verification set and the test set is 6208, 1552 and 1940 respectively.

The background of the distribution line image is complex and various, and under different angles and light, the detection of faults by the unmanned aerial vehicle can be interfered, so that a robust network model with high robustness needs to be selected. Relative to a two-stage target detection model based on regions of R-CNN, Fast R-CNN and Fast R-CNN, the YOLOv5 algorithm belongs to a one-stage target detection model, and can directly predict the relative positions of candidate frames to realize object Classification (Classification) and Bounding boxes (Bounding boxes). The YOLOv5 network structure is divided into four parts: input (Input), Backbone (Backbone), Neck (neutral) and Prediction (Prediction).

The input end adopts the Mosaic data enhancement and the self-adaptive anchor frame calculation. The Mosaic data enhancement mode splices images in a mode of random scaling, random cutting and random arrangement, and increases the detection effect on small targets. Compared with an anchor frame with the length and the width set initially, in network training, the optimal anchor frame values of different training sets are calculated in a self-adaptive mode, so that the difference between the optimal anchor frame values and a real frame (Ground channel) can be reduced, reverse updating is performed, iterative network parameters are reduced, and the target detection speed is increased.

A Backbone network (Backbone) adopts CSPDarknet to extract abundant information characteristics from an input image. CSPNet solves the problem of gradient information repetition of network optimization in a deep convolutional neural network framework backbone network, and integrates gradient change into a characteristic diagram uniformly, so that the parameter number of a model and the floating point operation times (FLOPS) of a computer per second are reduced, the speed and the accuracy of reasoning are ensured, and the size of the model is reduced.

The neck network of YOLOv5 employs an FPN + PANet structure, as shown in fig. 2. FPN mainly promotes the effect of target detection through fusing high-bottom characteristic, especially can improve the detection to little target. The PANet introduces a Bottom-up path enhancement (Bottom-up path augmentation) structure on the basis of the FPN, can fully utilize shallow features of a network to carry out segmentation, enables a top-level feature map (feature map) to enjoy rich position information brought by a Bottom layer, improves the detection effect on a large object, and enables a model to identify the same object with different sizes and scales.

Drawings

FIG. 1 is a schematic flow diagram of the present invention.

Fig. 2 is a diagram of the FPN + PANet network architecture.

FIG. 3 is a diagram of a modified YOLOv5-LITE network architecture.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention are described in further detail below with reference to specific embodiments.

The research provides a lightweight network based on YOLOv5, and on the basis of a traditional YOLOv5 mode, a ShuffleNet V2 architecture replaces CSPDarknet53 to serve as a feature extraction network, so that a YOLOv5-LITE lightweight network model is constructed. ShuffleNet V2 both inherits the characteristics of ShuffleNet packet convolution and channel rearrangement and follows 4 criteria for designing lightweight networks. Under the same conditions, ShuffleNet V2 is faster and more accurate than other models. The shuffle netv2 changes the short circuit connection (shortcut) structure from Add to Concat for feature reuse by referencing to a DenseNet network, but unlike DenseNet, the shuffle netv2 is not dense Concat, but has channel rearrangements (channel shreds) after splicing to mix features.

The YOLOv5 Loss function Loss includes boundary regression Loss (coord), confidence prediction Loss (Loss) (conf), and class prediction Loss (cls). YOLOv5 calculates the boundary regression (bounding box) Loss using GIoU Loss, the beligs Loss function calculates the confidence (object score) Loss, and the cross entropy Loss function (bceclloss) calculates the class prediction (class probability score) Loss. The loss formula is as follows:

Loss＝Loss(coord)+Loss(conf)+Loss(cls) (1)

Loss(coord)＝1-GIoU

distribution lines are used as identification targets, and defects of the distribution lines are of three types: cable separation gasket, cable and insulator separation and no-ring insulator. Therefore, a 152 x 152 scale feature layer beneficial to small target detection is introduced, and fine-grained detection is realized. From a backbone network ShuffleNet V2 of a YOLOv5-LITE lightweight neural network model, four scale features of 19 × 19 (feature layer P5 '), 38 × 38 (feature layer P4'), 76 × 76 (feature layer P3 '), and 152 × 152 (feature layer P2') are extracted, as shown in FIG. 3. The large receptive field of P5 'is suitable for large target detection, the large receptive field of P4' is suitable for medium target detection, upsampling is carried out on the basis of P3 ', and fault detection on a small target is realized by fusing a P2' feature layer.

In the feature transmission process, the 19 × 19 scale feature layer is subjected to multi-scale fusion through the SPP structure in the maximum pooling mode of 1 × 1, 5 × 5, 9 × 9 and 13 × 13 to obtain a feature layer P5'. When the feature fusion obtains a richer feature map, in order to prevent the network from being excessively redundant, pruning operation is performed on the 152 × 152 feature layer output after the Feature Pyramid (FPN) fusion, that is, the 152 × 152 feature layer output by the feature pyramid is not subjected to the prediction output of YOLO Head, and the upsampling is directly performed on the PANet structure. The improved algorithm thus still retains the 19 × 19, 38 × 38, 76 × 76 three-scale feature-layer prediction output of YOLO Head.

When a 152 × 152 feature layer is added to a backbone network of YOLOv5-LITE to be responsible for small target detection, the number of networks is inevitably increased, and in order to reduce network parameters, deep separable convolution (Depthwise separable convolution) is used in a PANET structure to replace down-sampling (Down sampling) of an original model, so that feature information interaction of each layer is realized, and network calculation amount and parameters of the model are effectively reduced. Because the acquired data has 3 defect types, 1024 volume and 5 multiplied by 5 pooling operations of a ShuffleNet V2 backbone network are removed, and global average pooling operations are adopted for replacement, so that the speed and the precision of the model are improved while the memory consumption is saved.

The above embodiments are merely illustrative, and not restrictive, and various changes and modifications may be made by those skilled in the art without departing from the spirit and scope of the invention, and therefore all equivalent technical solutions are intended to be included within the scope of the invention.

Claims

1. A method for identifying defects of a power distribution assembly based on improved YOLOv5-LITE lightweight class is characterized by comprising the following steps: can carry out accurate quick location and discernment to the distribution subassembly defect, be convenient for dispose to the mobile device terminal, have the particularity of real-time, rate of accuracy height, lightweight:

(1) the method uses ShuffleNet V2 as a backbone network to extract features, constructs a YOLOv5-LITE lightweight neural network model, removes 1024 convolution and 5 multiplied by 5 pooling of ShuffleNet V2, replaces by global average pooling operation, reduces the network parameter number and improves the model detection speed;

(2) in the stage of extracting the characteristics of a backbone network, the defect detection of large, medium and small scales is realized by introducing a 152 multiplied by 152 characteristic layer which is beneficial to the detection of fine-grained targets;

(3) the method has the advantages that deep separable convolution is adopted in the PANET framework to replace down-sampling, so that the network is lighter, and common defects of distribution line components such as cable separation gaskets, cable and insulator separation, and loop-free insulators can be identified.

2. The method for identifying the defects of the power distribution assembly based on the improved YOLOv5-LITE lightweight class according to claim 1, wherein a backbone network of the YOLOv5 network adopts the CSPDarknet53 to extract features, so that although the detection accuracy is improved, the model calculation is complex, more memory space is consumed, and the method is not favorable for being deployed to an unmanned aerial vehicle for monitoring the faults of the power distribution line in real time; the invention provides a YOLOv 5-based lightweight network, which is characterized in that on the basis of a traditional YOLOv5 mode, a ShuffleNet V2 architecture replaces CSPDarknet53 to serve as a feature extraction network, and a YOLOv5-LITE lightweight network model is constructed.

3, the ShuffleNet V2 not only inherits the characteristics of ShuffleNet grouping volume and channel rearrangement, but also follows 4 criteria for designing lightweight networks; under the same condition, the ShuffleNet V2 has higher speed and better accuracy compared with other models; the shuffle netv2 changes the short circuit connection (shortcut) structure from Add to Concat for feature reuse by referencing to a DenseNet network, but unlike DenseNet, the shuffle netv2 is not dense Concat, but has channel rearrangements (channel shreds) after splicing to mix features.

4. From a backbone network ShuffleNet V2 of a YOLOv5-LITE lightweight neural network model, four scale features of 19 × 19 (feature layer P5 '), 38 × 38 (feature layer P4'), 76 × 76 (feature layer P3 '), and 152 × 152 (feature layer P2') are extracted, as shown in FIG. 3; the large receptive field of P5 'is suitable for large target detection, P4' is suitable for medium target detection, up-sampling is carried out on the basis of P3 ', and fault detection of small targets is realized by fusing a P2' feature layer.

5. When the feature fusion obtains a richer feature map, in order to prevent the network from being excessively redundant, pruning operation is performed on the 152 × 152 feature layer output after the Feature Pyramid (FPN) fusion, that is, the 152 × 152 feature layer output by the feature pyramid is not subjected to the prediction output of YOLO Head, and the upsampling is directly performed on the PANet structure; the improved algorithm thus still retains the 19 × 19, 38 × 38, 76 × 76 three-scale feature-layer prediction output of YOLO Head.

When a 152 × 152 feature layer is added to a backbone network of yollov 5-LITE to be responsible for small target detection, the number of networks is inevitably increased, and in order to reduce network parameters, deep separable convolution (Depthwise separable convolution) is used in a PANet structure to replace down-sampling (Down sampling) of an original model, so that feature information interaction of each layer is realized, and network computation amount and parameters of the model are effectively reduced.

7. Because the acquired data has 3 defect types, 1024 volume and 5 multiplied by 5 pooling operations of a ShuffleNet V2 backbone network are removed, and global average pooling operations are adopted for replacement, so that the speed and the precision of the model are improved while the memory consumption is saved.

8. A YOLOv5-LITE model is taken as an object network, and two training modes are respectively adopted for the object network: i.e., a training mode in which pre-training weights are loaded and partial layers are frozen using transfer learning and a training mode in which pre-training weights are not used.

9. The method is characterized in that a training mode of loading pre-training weights by transfer learning is used for loading the weights of a ShuffleNet V2 structure in a COCO data set, a ShuffleNet V2 network layer in YOLOv5-LITE is frozen during initial training, at the moment, a model trunk is frozen, a feature extraction network is not changed, and only the network is finely adjusted; when the model is not frozen, the feature extraction network changes, and all parameters of the network change.