Abnormality Detection and Identification Algorithm for High-Speed Freight Train Body

Abstract

Abnormality detection and identification for high-speed freight train body is an indispensable part of the Train Operation Status Monitoring System. Generally, abnormality detection and data recording are performed manually, which is very prone to cause problems such as false detections, missing detections and recording errors because of lots of freight trains passing through the station simultaneously. In order to tackle these problems, we proposes an Improved-YOLO model, based on the YOLOv4, adding the SE-Block to optimize the feature selection method, switching to Cascade PConv Module and Integrated BN combination instead of PANet for multi-level feature fusion, using random data augmentation to improve the generalization. Meanwhile, model training is assisted by introducing negative sample mechanism. With 4594 positive samples (including freight train body abnormalities) and 5406 negative samples (excluding freight train body abnormalities) collected at the actual station as the train set, and 4705 images of freight train body within 24 h as the test set, the Improved-YOLO has dropped by 18.72% and 7.10% in false detection rate and missing detection rate respectively compared to the original YOLO model and reached 8.95% and 9.38%.