Research On Application Of Multi-scale Convolutional Neural Network In Insulator Defect Detection

Power transmission needs to be carried out through the transmission line,and the insulator,as a significant part of connecting the tower and the wire in the line,plays an important role in ensuring the safe and stable operation of the line.Thus,how to efficiently detect power inspection insulators and their defects has become an urgent problem in the development of smart grids.At present,such problems as small image targets,complex backgrounds and uneven illumination can be found in the detection of the insulators and their defects.In the current mainstream target detection algorithm,Convolutional Neural Network(CNN)is the most used feature extraction network and the multi-scale CNN can realize feature analysis at different scales,which is conducive to conducting the defect detection of those small targets.To achieve the detection of the insulators and their defects,this paper studies the insulator detection model based on the multi-scale CNN,combines the multi-scale feature extraction and fusion algorithm and visualizes the multi-scale CNN to interpret the model and classify the defects of the insulators.In this paper,the following aspects are mainly studied and the corresponding research results are obtained as follows:(1)To address the problems of insufficient insulator types and too few defective samples in the initial insulator dataset,a deep adversarial network data enhancement model is used to generate insulator samples with multiple style domains through a cyclic consistent generation adversarial network,and finally a sample library containing three types of insulators with different materials and four different defective types is constructed.Apart from that,a pretraining model is trained with these samples,which provides a database for the subsequent training and testing of the multi-scale model.(2)In terms of the study on the defect detection of the insulators based on the multi-scale CNN,a multi-scale feature extraction network structure is proposed in this paper to realize the extraction of insulator image features at different scales and fusion of features at different scales.Moreover,the design of the target prior frame is incorporated to detect the insulators and their defects,combined with the end-to-end training and optimization method.Finally,the training and testing results of these methods are compared and analyzed in the insulator dataset constructed in this paper.It is demonstrated that the detection result of the multi-scale CNN in the insulator dataset is 91.37%,while the design of the target prior frame makes the detection accuracy improve by 1.3%.(3)The visualization technique is applied to the multi-scale CNN model proposed in this paper,and an interpretable approach after modeling is used to first visualize features of the network layer separately,then analyze the model in this paper by combining the class activation graph mapping method.Ultimately,through visualization,how the multi-scale model plays a key role in identifying the insulators is explained and the decision focus of the insulator defect detection task is figured out.In the experiment,the insulator image features of different layers of the multi-scale CNN are visualized and the image characterization of the insulators and their defects is implemented by the class activation map visualization method.