Detection of Fastener Abnormal State Based on Deep Style Migration Synthetic Data
The service status of the fastener system,a key basic component in the track structure,directly affects the normal operation of railroad trains.Existing machine vision-based inspection schemes are prone to missed or false detec-tion due to uneven samples.In this paper,an algorithm was proposed to detect the abnormal state of the fastener system of slab ballastless track based on image style migration learning.By performing forward modeling of railroad track struc-tures based on actual scenes,the algorithm generated BIM models with identical geometric parameters according to the constraint relations in the design phase,rendering details through a lightweight physics engine,and randomly deploying fastener systems with different types of status to output virtual inspection images.On this basis,the virtual inspection im-ages were realistically migrated using recurrent adversarial generation network to obtain a highly simulated synthetic data-set with balanced positive and negative samples.Finally,this dataset was used to fully train the deep target detection net-work to achieve accurate detection of the abnormal state of the fasteners of slab ballastless track.The experimental results show that the Faster R-CNN network fully trained from the synthetic dataset obtained by style migration has the best ac-curacy in the detection of the abnormal state of the fasteners of the slab ballastless track,with the MAP value of 94.91％for fastener detection for four types of normal,fractured,missing and displaced states,5.39％higher than the real data-set and 2.37％higher than the virtual BIM dataset.Among them,the detection accuracy of missing fasteners is improved the most,with a 10.13％increase compared to the real dataset,achieving an effective improvement of fastener state rec-ognition accuracy.
万方数据
维普
