Vision-Based Urban Rail Vehicle Fault Detection Using Enhanced YOLO Framework

Urban rail transit systems serve as pivotal city transport modalities. However, conventional vehicle condition monitoring methodologies often fail to provide realtime respon-siveness, potentially giving rise to safety hazards. Predominant vehicle state detection research leans on data from TVDS and TEDS systems, engendering inconsistencies due to diverse vehicle types and data quality variations. This consequently impacts the optimal performance of existing algorithms in urban rail transit contexts, particularly in minor entity detection. To mitigate this, a novel vehicle state detection approach premised on YOLOv5 is presented in this study. A Coordinate Attention Mechanism (CA) is implemented to augment local detail extraction in vehicle status images. Additionally, the conventional Feature Pyramid Network(FPN) is supplanted with a Progressive Feature Pyramid Network(PFPN), facilitating superior multilevel feature fusion devoid of information degradation. On application of this model to an urban rail vehicle state dataset, noteworthy results, including an accuracy of 0.98, recall rate of 0.96, and mean average precision (MAP) of 0.983, are realized. Such outcomes illustrate excellent object detection performance, sans significant augmentation of the model's parameter size.

I. Introduction

Urban rail transit systems have emerged as critical components of modern urban transportation, offering efficient, rapid, and dependable public transit services [1]. In bustling cities like Beijing, the daily average of train operations exceeds 10,700, with 820 trains running concurrently during peak hours. This extensive and high-frequency train operation, covering a total distance of 2.14 million kilometers per day and accommodating approximately 7.78 million passengers, presents significant challenges in maintaining the health status of vehicles during on-track operations.