Toward Reliable High-Speed Railway Pantograph-Catenary System State Detection: Multitask Deep Neural Networks With Runtime Reliability Monitoring

The pantograph-catenary system (PACs) is the only channel for high-speed trains to obtain electric energy, and its state is crucial to the safety of train operation. There are two key indicators to measure the PACs state: the tension state of the catenary dropper and the zig-zag value of the PACs contact point (CPT). With the development of deep learning, on- board vision inspection equipment has gradually become an effective means of PACs state inspection. However, since more and more deep-learning black boxes are used in the railway industry, the reliability estimation of their decisions has become an unavoidable challenge. As for PACs state detection, there are two other challenges: the lack of defective dropper sample and the difficulty in distinguishing between true and false contact points. To overcome the challenge of defective sample scarcity, this article proposes to detect dropper defect by the semantic features extracted by a multitask pantograph-catenary system state detection network (MPCN), which is trained using only normal droppers. Aiming at the contact point tracking, a multiple hypothesis tracking tree (MHTT) algorithm is proposed to track the contact points using the change rule of the contact point sequence. Meanwhile, in the semantic space, domain knowledge is combined with Bayesian hypothesis testing to monitor the runtime reliability of the MPCN. In this way, the PACs state detection and the MPCN reliability monitoring can be implemented in a unified framework. The effectiveness of our method has been verified in various scenarios of various railway lines.