An Automatic Fault Detection Method of Freight Train Images Based on BD-YOLO

• Published in: IEEE access Vol. 10; pp. 39613 - 39626
• Main Authors: Zhang, Longxin, Wang, Miao, Liu, Ke, Xiao, Mansheng, Wen, Zhihua, Man, Junfeng
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Cluster analysis; Clustering; Convolutional neural network; Convolutional neural networks; Fault detection; Feature extraction; Freight trains; Locking; Object detection; Object recognition; Plugs; Proposals; Rail transportation; Vector quantization
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2022.3165835

The automatic fault detection of freight train image is difficult to locate, and its average detection accuracy is low. In this paper, an object detection model Bidirectional You Only Look Once (BD-YOLO), is proposed. The process of BD-YOLO is divided into four steps. First, the feature extraction network extracts and separates image features. Second, multiscale features are fused to aggregate features of different scales. Third, prediction across scale modules is used to forecast the feature layers obtained after aggregation. Fourth, the final prediction result is obtained by decoding the prediction feature layer. The BD-YOLO model is trained using the mosaic data enhancement method and K-means clustering algorithm to improve detection accuracy and speed. Experimental results show that the mean average precision of BD-YOLO model is improved by 17.57%, on average, compared with the state-of-the-art object detection models on four types of data sets. The BD-YOLO model can immediately detect three typical faults, namely, movement of upper lever, offset of locking plate, and closing of truncated plug door handle. It has high detection rate, low false detection rate, and good robustness.