Research On Visual Anomaly Detection Method Of Pipe Gallery Based On Deep Learning

A public pipe gallery is an important public supporting facility to maintain the sustainable development of the chemical industrial park.It is an important carrier to ensure the material transmission of enterprises in the park.It transmits important means of production for a large number of enterprises in the park and maintains a lot of high temperatures,high pressure,flammable,and explosive material transmission tasks.At present,the abnormal detection of the pipe gallery is still carried out by manual 24-hour remote monitoring or personnel climbing the corridor detection,most of which rely on experience and visual inspection,which may face major security risks.In recent years,with the rapid development of machine vision technology,especially deep learning technology,it has surpassed the human detection level in some detection fields.This project takes this as the breakthrough point,mainly studies the methods of defect detection(supervised learning)and anomaly detection(semi-supervised learning)of public pipe gallery based on deep learning,which is used to assist the intelligent inspection robot of pipe gallery instead of manual inspection,It has important research value.Firstly,aiming at the problems of "machine replacing human" and intelligent detection in abnormal detection of pipe gallery defects,and comprehensively considering the factors of real-time,easy deployment,and portability,YOLOv5 s is selected as the defect detection model of pipe gallery.To overcome the shortcomings of insensitivity to small targets such as damage and corrosion and low recall rate in the detection process,an improved YOLOv5 s detection model based on the transformer is proposed,Finally,the experiment has shown that transformer can improve the speed of network training and the ability of feature extraction,to improve the ability of small target perception and detection,as well as the generalization ability of the network.The overall average accuracy is improved by 24%.Secondly,aiming at the problem of unbalanced data in abnormal detection of pipe gallery(the normal samples are far more than the abnormal samples and defects)and the "open problem" in abnormal detection can not be solved in abnormal detection,an abnormal detection method of pipe gallery based on generation and reconstruction model is proposed.In this paper,two existing anomaly detection methods of the generative model,the convolutional autoencoder(CAE)and generative adversarial networks(GAN)are studied.Finally,the experiment shows that the improved Gan model is superior to the improved CAE method in anomaly detection of pipe gallery,and the overall improvement is 35.3%,which can meet the requirements of anomaly detection of pipe gallery.Finally,according to the above research results,relying on the guide rail intelligent inspection robot developed in our laboratory,the visual anomaly detection software of pipe gallery is designed with Python.The software includes anomaly detection methods and defects anomaly detection methods so that they can be used in the local control center or robot body.The test results have shown that the system can meet the basic requirements of anomaly detection in the pipe gallery,and have achieved the expected detection accuracy.