Abstract: This paper proposed a key component detection algorithm for metro train roof based on YOLOv5s model to address the current problems of manual inspection, high labor intensity, and high missed detection rate in metro train roof components. Considering the practical situation of insufficient on-site computing power, the paper presented a lightweight design for the YOLOv5s model, replaced the C3 module in the YOLOv5s model Backbone with the Ghost-C3 module, and used Ghost convolution instead of traditional convolution in the YOLOv5s model to reduce model complexity and computational complexity. To compensate for the loss of model performance caused by lightweight design, the paper introduced a CA attention mechanism in the Ghost-C3 module to enhance feature perception of key components and improve model accuracy. The experimental results show that the improved YOLOv5s model has a frame rate of 102.04 f/s, mAP of 97.98%, Pars of 4.47 MB, and FLOPs of 10.2 GB. Compared with the original YOLOv5s model, mAP has increased by 1.36%, Pars has decreased by 33.98%, and FLOPs has decreased by 36.65%. The proposed algorithm can provide technical support for identifying the service status of key components on the roof of subway trains in the future.