以深度學習進行全方位軌道缺失構件檢測

A COMPREHENSIVE DETECTION SYSTEM OF DEFECTIVE TRACK COMPONENTS BY DEEP LEARNING

軌道扣件負責固定鋼軌於枕木上，避免軌道鬆脫變形。傳統鐵路軌道扣件巡檢方式係採人工，以目視方式進行巡檢。近年來，人工智慧深度學習蓬勃發展，在許多領域獲得突破性進展，因此本研究以人工智慧(Artificial Intelligence，AI)方式進行鐵路軌道缺失構件檢測，得以降低人工巡檢的負擔及時間，並將此技術實地測試與應用推廣。首先蒐集國內外鐵路軌道構件檢測案例並分析其設備及方法，接著建立軌道構件樣本的擷取設備，包含影像記錄及照明，再依收集之上視、側視軌道構件樣本，分為上視10種缺失類別、側視4種缺失類別與2種正常類別。採用YOLOv4-Tiny模型進行深度學習訓練。在實驗中錄製超過70公里軌道構件影像，分別以上視缺失構件與側視構件資料集進行訓練，所得上視缺失與側視缺失檢出率分別為91%與94%，平均精確率各別為91.7%與99.2%，顯示辨識指標良好，執行效能也可達150張/秒。期間比較YOLOv4-Tiny模型與人工徒步巡檢，成本與效能均較佳。

The rail fasteners are responsible for fixing the tracks on the sleepers to prevent the tracks from loosening and deforming. The traditional inspection method of railway track fasteners adopts human visual inspection. Artificial intelligence (AI) deep learning has developed rapidly and made a breakthrough in many fields in recent years. Therefore, this study uses artificial intelligence to detect missing components of railway tracks and reduce the burden and time of manual inspection. The goal is to promote this technology to field applications. Firstly, we have surveyed some commercial track fastener inspection products from both domestic and foreign; analyze the deployed equipment and method for domestic applicability. Then, the image collecting equipment for the track components, including image recording and lighting, is established. There are ten and six types of classes defined for the top and side view of the track, respectively. YOLOv4-tiny, for its superior performance, is selected for our application. More than 70 km of track images were recorded in experiments. Moreover, the performance metrics (recall rate, mAP) are (91%, 91.7%) and (94%, 99.2%) for the top and side-view, respectively. The execution speed of YOLOv4-tiny also reached 150 fps. Meanwhile, the AI model was compared with the human eye-sight inspection, and the overall performance was better than human inspections.