Insulator Defect Detection Based On MobileNetV3-YOLOv5 Network

In our country,the transmission lines have characteristics of large span and complex grid structure.The insulators in the overhead line play an insulation and support role in the whole power system.Insulators are exposed to the outdoor sun or thunderstorm environment for a long time,so in the regular inspection work of power lines,insulators are one of the focus of attention.With the continuous development and expansion of the scale of transmission lines,the traditional manual inspection of insulators is inefficient and poor security,so in recent years,electric workers often adopt more intelligent electrical inspection.Therefore,it is of great practical significance to study the detection method of insulator defect fault.In this paper,an improved Mobile Net V3-YOLOv5 insulator defect identification and damage location detection algorithm is proposed,so that the recognition accuracy and calculation speed can be taken into account.The insulator detection model of deep learning has many parameters and requires a large amount of calculation,while the Mobile Net V3 network,as a lightweight network,has the characteristics of fewer model parameters and fast training speed.Therefore,aiming at the problem of low accuracy of insulator defect classification by traditional neural network,an insulator image classification algorithm based on improved Mobile Net V3 is proposed.Since Mobile Net V3 network is mainly composed of deep separable convolution,the Inception structure of multi-scale feature fusion network is firstly used to improve the deep separable convolution,and three parallel branches are adopted to extract the multi-scale features of 1×1,3×3 and 5×5.The improved depth separable convolution is introduced into the linear bottleneck structure of Mobile Net V3 network to enhance the ability of multi-scale feature extraction.Secondly,in order to expand the sensitivity field of network convolution and extract more complete insulator feature information,hollow convolution is introduced into the depth separable convolution of Mobile Net V3 network,so as to ensure that more insulator features are covered and classification accuracy is improved,while avoiding the increase of redundant network calculation.Finally,compared with VGG16,Resnet50 and Mobile Net V3 network,the classification accuracy,the number of model parameters,the amount of computation and the recall rate are compared.The classification accuracy of defects and faults of the improved Mobile Net V3 network reaches 92.36%,which is 2.52% higher than that of the original network.Therefore,the improved Mobile Net V3 network can identify damaged insulators more accurately.Aiming at the problems of multiple parameters,large computation amount and long training time of YOLOv5 network model,the YOLOv5 backbone network is first replaced with the Mobile Net V3 model improved in this paper to reduce the number of parameters of the model and improve the calculation speed.Secondly,the CSP module structure in PANet is redesigned,and the parameter quantity of the model is further optimized by introducing deep separable convolution.At the same time,due to the certain accuracy loss of lightweight processing of the network,the CBAM attention mechanism is added to the neck layer of the network to make the model pay more attention to the main information,thereby strengthening the insulator feature extraction ability in complex background and improving the network detection accuracy.Then,DIo U-NMS is used to replace NMS in the post-processing process,which improves the network’s ability to identify shielded insulators and reduces the probability of missed insulators in complex environments.Finally,the improved network is subjected to ablation experiments,and compared with SSD,Faster R-CNN and YOLOv5 algorithms,the number of parameters of the improved algorithm model is reduced by37%,the m AP value reaches 94.5%,and the detection speed is 50 frames per s.Therefore,the improved Mobile Net V3-YOLOv5 detection algorithm proposed in this paper can reduce hardware cost and effectively realize the rapid location of insulator defect,so as to meet the real-time requirements of insulator defect location.