Due to the dark and humid underground environment, large subway loads and frequent braking, the bogies of subway trains are prone to failures that seriously jeopardize the safety of public transportation. In this paper, the shortcomings of RCNN’s time-consuming training are solved by introducing Fast, which uses a shared convolutional layer and extracts the region where the target may exist for learning. It is proposed to improve the feature extraction process of Faster-RCNN algorithm by lightweight network by changing the traditional convolutional network structure, using MobileNet, attention mechanism model, and Unet network structure, and experimentally analyze the subway bogie defect detection of the improved Faster-RCNN. The processed Faster-RCNN image is mainly distributed between gray value 50 and 125, the proportion of which is 0.153 and 0.075, respectively, and the visual effect of local details of the image is more obvious, and the defects are easier to identify. The improved Faster-RCNN network model has improved the leakage detection rate of large-scale defects such as A4, A5, C4, C5, etc. The leakage detection rate of A4 and A5 is 0%, and the leakage detection rate of C4 and C5 is 2.63% and 3.57%, respectively, which has a better effect of detecting defects of subway bogies.
