Anomaly Recognition Algorithm Of Segmented Insulator Based On Deep Learning

In the subway line system,catenary is a special form of high-voltage transmission line used to supply power to electric locomotives.The section insulator belongs to the key equipment of OCS,and once it fails,it may lead to train interruption.At present,the monitoring of OCS equipment in Metro Depot is mainly carried out by manual inspection,but the daily inspection workload is large and the efficiency is low,which can not ensure the normal operation of trains in extreme weather.Based on the location of key equipment in the main line of Jiahe depot in Guangzhou,Guangdong Province,the subject carries out the research on video monitoring algorithm.The purpose is to develop an anomaly identification algorithm that can monitor the operation status of key equipment of catenary in real time,make up for the gap in relevant fields,and improve the efficiency of hidden danger troubleshooting and fault handling in the depot.In this paper,a segmented insulator anomaly recognition algorithm based on deep learning is proposed.The model learns the target and makes the corresponding anomaly early warning through target detection and image segmentation.To sum up,the main research contents of this paper are as follows:(1)Firstly,deep learning is innovatively introduced into the anomaly recognition of segmented insulator in industry.Different from traditional image processing,deep learning method has stronger robustness,better portability and strong generalization,and has little manual dependence.The abnormality of segmented insulator mainly includes two parts:bolt looseness detection and deflector wear detection.The overall monitoring is completed by obtaining the image video analysis of color camera.For each sub problem,this paper adopts the process of target detection,coarse positioning and image segmentation to extract fine information,and finally integrates it into an end-to-end anomaly recognition algorithm.(2)The target detection module is used for the rough positioning of the key positions of bolts and deflectors.The position information of the target can be extracted from the monitoring imaging,and then input into the image segmentation and subsequent processing steps after cutting.In this paper,YOLOv5 model is adopted,which has small volume,fast speed and good accuracy,and ensures a high detection accuracy under the condition of real-time monitoring.(3)After obtaining the target image,input it into the image segmentation network to extract the fine information of the target to make an abnormal early warning.After the comparative test,the UNET + + network can better meet the requirements of real-time monitoring.By analyzing the shortcomings of the original model,this paper makes three improvements to the model used in segmentation.Firstly,hole convolution is used to replace the original pooling layer in the network encoder stage,so that the characteristics of the coding stage have stronger spatial information.In the structure,the simple version of ASPP module is used to replace the basic module,and a small multi-scale fusion module is used to enrich the feature semantic information at this stage.Finally,Se attention mechanism module is added to improve the feature capturing ability of the model to key positions.Through comparative experiments,it is proved that the accuracy of the improved UNet++model is greatly improved,and the performance of the subject experiment is excellent.