Insulator Defect Detection Based On Digital Image Processing

With the rapid development of economy and society,the demand for civil and commercial power is increasing while the demand for the stability of power transmission is also increasing.As an important component of transmission lines,insulators are usually exposed to the natural environment continuously,which will cause various defects such as self-explosion,damage and deformation.These defects seriously threaten the reliable operation of electrical equipment.Therefore,it is necessary to detect the state of insulators efficiently and accurately.The huge difference in sample size of different types of insulator defects increases the difficulty of defect detection.To solve this problem,this research studies from the perspective of detection methods based on sufficient samples and few samples.This paper proposes a novel few-shot insulator defect detection method based on local feature depth information.The main contributions and contents of this paper are as follows:1.For defects with sufficient sample size,such as self-explosion of insulators,this paper conducts comparative experiments of nine object detection algorithms training with different number of samples,including YOLOv7.The experimental results show that the performance of YOLOv7 outperforms other networks under most conditions.When the training samples of YOLOv7 above 100,the AP50,AP75 and mAP all exceed 0.97,0.86 and 0.62 respectively.The average inference time of the network is 11.4ms,which can satisfy the requirements of accuracy and speed in insulator defect detection.In addition,this paper carries out multi-class detection by YOLOv7 which adopts a dataset with sufficient samples of selfexplosion defects and few samples of broken defects.The results show that YOLOv7 can only identify defects with a large number of training samples,and cannot realize detection of defects with few training samples.2.In order to solve the problem of insufficient defect samples in insulator defect detection,this paper proposes a novel few-shot insulator defect detection method combining oriented object detection and deep EMD(Earth Mover’s Distance)network.The proposed method combines large sample insulator string identification with few-shot defect detection.Firstly,the insulator strings are extracted using the Oriented R-CNN.Next,the insulator string features are divided into sub-blocks.Based on the local features of the insulator sub-blocks,a small amount of defect samples and normal samples are used to construct prototype features.The insulator defect condition is determined by the deep EMD distance between the local features and prototype features,and finally the precise detection of insulator defects is realized.The training strategy of large sample pre-training and small sample fine-tuning is adopted to train the network.The proposed fewshot defect detection method provides a new solution and an implementation method for intelligent defect detection of the power equipment with few defect samples.3.Several experiments are conducted to verify the effectiveness of the proposed few-shot insulator defect detection method.Firstly,comparison experiments are conducted on the dataset of glass insulator self-explosion with sufficient samples.The detection accuracy of the proposed few-shot detection algorithm and the conventional large sample algorithms are compared to verify the effectiveness of the proposed method.In order to further verify the detection effect of the proposed method on few-shot datasets,experiments on composite and porcelain damage insulator dataset are also carried out.The experimental results show that mAP scores of the proposed method reach 0.51,0.55 and 0.65 respectively for detecting glass self-explosion insulators under the condition of 2,5 and 10 defect samples,and 0.71,0.73 and 0.91 respectively for detecting damage insulators.Our method uses a few samples can achieve similar results as hundreds of samples of common object detection methods.