A support tensor machine-based fault diagnosis method for railway turnout

Turnouts play a crucial role in the safety and efficiency of trains. Traditionally, the railway turnout fault diagnosis method relied on vectorized data from time series monitoring. However, such data format fails to fully capture the signals' spatial structure and profile information, which are crucial for inspectors to analyze and make judgments. In this study, a novel fault diagnosis method for the railway is developed with the hyperdisk-based supervised tensor machine (HDSTM) and monitoring signal images, which solves the limitations of the existing method. Besides, a novel tensor-form multi-class classifier called HDSTM is proposed to address the limitation of the convex -hull-based support tensor machine (CHSTM) in the underestimation problem. First, the time series monitoring signals are preprocessed and transformed into two-dimensional images. Next, CANDECOMP/PARAFAC decomposition is used to calculate the feature tensor. Then, the HDSTM model is built with the extracted feature tensor to implement the fault diagnosis. The proposed method's performance is evaluated using real-world operational current and power datasets. Experiment results indicate that the proposed method achieved higher average accuracy and precision than existing methods.