A subtle defect recognition method for catenary fastener in high-speed railroad using destruction and reconstruction learning

Split pins (SPs), as essential fasteners in the pantograph -catenary system, play a critical role in maintaining the structural stability of the catenary support device (CSD) in high-speed railroad. Unfortunately, impact and vibration generated by the interaction between catenary system and high-speed running vehicles will progressively damage SPs, including but not limited to the loose, breakage, and missing of SPs. Current image processing technologies have limited accuracy. Traditional convolutional neural network (CNN) -based systems that directly learn deep features are inefficient in detecting subtle defects (such as fractured and loose SPs). This study presents a novel SP inspection system based on an innovative destruction and reconstruction learning (DRL). In the DRL, a novel cross stage partial network (CSPNet)-based classifier is firstly developed to alleviate the impact of redundant semantic information and optimize gradient regression process. Then, "destruction" in the DRL stream means these partitioned sub -regions in an input image are shuffled by a region shuffling mechanism (RSM); this guides the classifier to pay more attention to the distinguishable regions in the destroyed and original images for a purpose of correct recognition. For "reconstruction", a newly developed position reconstruction network (PRN) is utilized to restore the original spatial distribution of the sub -regions by modeling the semantic correlation among the sub -regions. The proposed DRL can injects more discriminative details to the classifier by jointly training the classification network and PRN. Finally, comprehensive experiments are conducted on CSD datasets captured from real railroad lines, and the results demonstrate that the proposed approach outperforms existing state-of-the-art models yet keeps a reasonable inference speed.