An improved YOLOv7 for the state identification of sliding chairs in railway turnout

The sliding chairs are important components that support the switch rail conversion in the railway turnout. Due to the harsh environmental erosion and the attack from the wheel vibration, the failure rate of the sliding chairs accounts for up to 10% of the total failure number in turnout. However, there is little research carried out in the existing literature to diagnose the deterioration states of the sliding chairs. To fill out this gap, by utilizing the images containing the sliding chairs, we propose an improved You Only Look Once version 7 (YOLOv7) to identify the state of the sliding chairs. Specifically, to meet the challenge brought by the small inter-class differences among the sliding chair states, we first integrate the Convolutional Block Attention Module (CBAM) into the YOLOv7 backbone to screen the information conducive to state identification. Then, an extra detector for a small object is customized into the YOLOv7 network in order to detect the small-scale sliding chairs in images. Meanwhile, we revise the localization loss in the objective function as the Efficient Intersection Over Union (EIoU) to optimize the design of the aspect ratio, which helps the localization of the sliding chairs. Next, to address the issue caused by the varying scales of the sliding chairs, we employ K-means++ to optimize the priori selection of the initial anchor boxes. Finally, based on the images collected from real-world turnouts, the proposed method is verified and the results show that our method outperforms the basic YOLOv7 in the state identification of the sliding chairs with 4% improvements in terms of both mean Average Precision@0.5(mAP@0.5) and F1-score.