Supervised machine learning from digital twin data for railway switch fault diagnosis

This study concerns the development of fault diagnosis methods for railway switches based on supervised machine learning. The lack of data led us to build a digital twin that allows us to build the dataset necessary to train the model. The data measured on the switches correspond to time series. We are therefore interested in supervised learning methods based on time series. Our study shows that the STSF model is the best suited for the classification of switch faults. The study then focused on characterizing the robustness of the model with respect to noises that can disturb the measurements. It also showed that the current I and the angular velocity (j) are the most relevant data for learning. The results give the main conditions to ensure an optimal implementation by transfer learning on real systems.