Institute of Transportation, MOTC

       This study attempted at automatic classification of railway track components with AI technology, so as to reduce the manpower and time of manual inspection, also to put this technology to on-site test and promote its application. We began by gathering worldwide and domestic cases of railway track component inspection and analyzing the deployed equipment/method for local applicability. We then built the image capturing equipment for samples of track components, including image recording and lighting. The gathered samples of track components in top and side views were divided into 10 types and 4 types of defect classes defined for the top and side view of track, respectively, besides 2 types of normal class of image mark. Yolov4-Tiny model was employed in deep learning training, before verifying recall rate of defective components from the data under test. A positioning system was further employed to increase the efficacy of go-around inspection, which will help track maintenance and management in Taiwan, in turn, increasing rail traffic safety. In the experiment, the track is captured by using GoPro motion camera for both top and side view. More than 70 km of track images were recorded. Previously, the defective component inspection using Yolov4 resulted in a 87% recall rate, with mAP at 94.8%, and a speed of 50 frames per second (fps). In this year’s study, Yolov4-Tiny was used instead, in the training with data set of top and side views of defective components. The results were 91% and 94% recall rates, and 91.7% and 99.2% mAP for top and side views, respectively, suggesting an apparently improved indicator of classification, with executive efficacy improved to 150 fps. at the beginning of experiment, a comparison of the AI model and human visual inspection showed the former was outperformed, which was found to be because some positive samples were labelled mistakenly or the defective samples on site were very different from the training set. when the wrong labels were corrected and new defective samples for retraining were added, the overall performance became better than human inspections.

BENEFITS AND APPLICATIONS: The findings of this study can be provided as reference for the Ministry of Transportation and Communications or the Taiwan Rail Administration in routine track inspections to effectively manage railway safety and for subsequent track maintenance and reinforcement