Research On Detection Methods Of Key Parts In Metro Based On Computer Vision

In recent years,the subway train has ushered in a new golden period of development.Due to the dark and humid underground environment,the heavy load,the frequent braking and other factors,the parts of subway trains are easy to break down,which seriously endanger the safety of people taking public transportation.It puts forward higher requirements for the safety inspection of subway trains.The traditional method for metro train inspection relies on manual labor,which has some shortcomings such as low efficiency,high cost and unstable detection results.In this paper,the defect detection algorithm of key parts in metro is studied based on computer vision.The pictures are obtained by the high-precision cameras which are deployed at the bottom and side of the train.By adopting technology of deep learning and anomaly detection,the problems of rare negative samples and complex environment are overcome.A new feasible technical scheme is provided for the safety inspection of trains in this paper.The main contents include the following points:1)Based on the analysis of the current research status of train inspection and anomaly detection at home and abroad,a technical solution for the detection of important and difficult problems in key parts is proposed in this paper.The basic concept,principle and development process of convolutional neural network are further introduced to provide theoretical basis and technical groundwork for the following paper.2)This paper improves Faster RCNN by Mobile Net V2,light region proposal network and attention mechanism.In this way,the size of network model is greatly reduced and the detection speed is improved without loss of accuracy.Under the condition of the detection accuracy satisfied,the model size is reduced to 21.2MB and the positioning speed of target in every image is 45 ms.Its comprehensive performance is better than that of other compared algorithms.3)Aiming at the industrial problem of insufficient negative samples,the one-class algorithm based on convolutional neural network is adopted to ensure the good ability of feature extraction.At the same time,the network structure is improved so that it can couple the characteristics of real negative samples and simulated negative samples during training,which makes the network model is more capable of detecting subtle faults.In addition,the novel Efficient Net is used to replace the basic convolutional network,which improves the accuracy and reduces the model size significantly.The experimental results show that the value of AUC reaches above 0.99 in abnormal detection of various types of bolts.Combined with the algorithm of object detection in the previous section,the fault-detection speed of each input image reached 0.4s and the false alarm rate of all kinds of faults was less than 10% under the condition that the false negative rate is 0.4)The effect of generative adversarial networks on defect detection of air spring is analyzed and studied.In response to the fact that the adversary network in the original GANomaly network was likely to divergent and fail,a new structure of adversary network was designed to better guide the coding and reconstruction operation of image,resulting in an overall AUC of 0.9985.The experimental results show that this method is superior to other methods of comparison,and it has a good detection performance on the picture of air spring from the actual scene.