Institute of Transportation, MOTC

       The rail fastener is the most important component of the rail fastener system. It can fasten the rail tightly on the track, and has great influence on the train’s stability and safety. At present, it is hard to avoid the problem of omission due to the reliance on human visual inspection. In this study, we not only collect the related research results of track inspection abroad, but also set up the image collecting equipment of track fastener, including image recording equipment and lighting equipment. YOLOv4 was adopted for deep learning model training, and the recall rate of defective fasteners was verified from the test data. In the experiments, the image was captured by GoPro camera, and there are 70 km images of track fastener were recorded. The images were verified by Yolo V3 in the previous phase, the Precision rate and the recall rate were 82.5% and 74.5%, respectively. In this second phase, the recall rate was increased to 87% and the mean average accuracy rate (mAP) was increased to 94.8%. Furthermore, the identification system runs over 55 frames per second (Table 4-15 Notebook C). The SSD recognition model also worked well, and was tuned to match the YOLOv4. In order to make it easier for track inspectors to find the defective fastenings, cloud storage, identification, mobile phone APP, query, GPS/gyro positioning and other functions have been developed. GPS position of the defective fastener is used to find the nearest hundred meter of track which in turn is used to calculate the offset to search for the defective fastener. Through this way, human efforts could be greatly reduced. BENEFITS AND APPLICATIONS: The Ministry of Transportation and Communications or the Taiwan Rail Company can refer to the results for policymaking