Wheel Tread Damage Detection Based On Deep Learning

The damage detection of train wheel set is of great significance for the safe operation of the train.The system of tread damage detection based on photoelectric technology and digital image processing technology has broad application prospects because it can perform non-contact on-line detection and can achieve both high accuracy and high efficiency.This paper mainly studies the tread damage image generation technology based on conditional generative adversarial nets,the tread damage detection technology based on YOLOv2 model,and designs a set of wheel tread damage detection system.The main work includes the following aspects:(1)The tread damage image generation technology based on conditional generative adversarial nets.Based on the residual strategy,the identity mapping block is used to replace the downsampling layer in the generative network,which makes improvement to generative network of conditional generative adversarial nets.At the same time,under the constraint of generative adversarial nets losses,the gradient descent algorithm is used to alternatelt optimize the parameters of generative network and discriminative network,and the convergence generative model can be achieved by training.Experimental results show that the generative model can generate wheel tread damage images with damaged texture features.(2)The tread damage detection technology based on YOLOv2.A tread damage detection network based on the YOLOv2 model is built by using 19 convolution layers and 5 max pooling layers.Firstly,the K-Means++ clustering algorithm is used to classify the manually labeled tread damage and determine the aspect ratio of the tread damage as the initial size of Anchor boxes.Then,under the constraint of mean square error loss function,the pre-training model trained by public data set is finetuned to obtain the final convergence detection model.The direct regression of the model is used to detect the location and category of the tread damage target.The model achieved a correct rate of 96.6% when testing the test set,while meeting the requrements of real-time detection.(3)Design and implementation of the train wheel damage detection system.A train wheel damage detection system based on photoelectronic technology and image processing technology was developed.The system consists of a hardware acquisition module and a software processing module.The software processing module includes twp parts: image preprocessing and tread damage detection.The system is trial-run in the maintenance section of the Locomotive Depot in Qingdao,Jinan,Xuzhou,etc.It can completely collect wheel tread images,and the detection performance is good,meeting the system index requirements.