Research On The Damage Detection Method Of Urban Rail Train Wheel Set Tread Based On Machine Vision

With the rapid development of urban rail transit,the safe and efficient operation of urban rail trains has become an important issue.The tread of the wheel set is the direct action part between the train and the rail,and its working state directly affects the driving stability of the train.In order to realize the intelligent detection of tread damage,a non-contact detection method based on machine vision technology is proposed in this paper,and the improved YOLOv5 target detection algorithm is used to identify and locate tread damage.The main contents are as follows:First of all,to solve the problems of the original tread image with more background information and uneven illumination,an algorithm for tread extraction and illumination compensation is designed.Firstly,by using the nature of uneven illumination of the tread image,a tread extraction method that first enhances the contrast and then performs edge and line detection is adopted.The tread images were subsequently illuminated using a modified Retinex image enhancement algorithm.Secondly,in view of the problems of large scale differences and variable shapes of tread damage,the YOLOv5 model has been improved in three aspects.One is to change the detection method from the anchor frame-based detection method to the anchor-free detection method.The second is to replace the standard convolution in the model with a deformable convolution.The third is to enhance the shallow structure of the model and add a detection layer for small targets.Finally,the effectiveness of the proposed method is verified by ablation experiments and comparative experiments.Thirdly,aiming at the problem that tread damage images are few and difficult to collect,an image sample expansion method based on Info GAN is proposed.The augmentation task of the tread damage dataset was accomplished by training Info GAN with a dataset similar to the tread damage.And aiming at the problem of unstable training of Info GAN,combining the experience of DCGAN and WGAN,the Info GAN has been improved in terms of model structure and loss function.Finally,the experiment proves that the method can effectively expand the data set.Finally,the design of the wheel tread detection system is completed and field deployment verification is carried out.The effectiveness of the image preprocessing method proposed in this paper and the improved YOLOv5 damage detection method are proved by actual tests.