Comparison Between Model-Based and Data-Based Methods for Fault Diagnosis of Railway Turnouts

Railway switch machines are important elements in a railway network, allowing trains to change directions. Any failure in such safety critical equipment can induce economic losses caused by traffic interruptions and even sometimes catastrophic accidents. Hence, the diagnosis of such failures is recognized to be a primordial task. In the consulted literature, the methods considered are those based on an analytical model or those based on data (artificial intelligence) or hybrid with suitable justifications. The present paper deals with synthesis and discussion of two approaches applied to railway switch machine: Linear Fractional Transformation (LFT) Bond Graph as multiphysic model based robust diagnosis and a data-driven method based on k nearest neighbor (kNN) algorithm. A simulation platform is developed to test those two algorithms under various scenarios of faults that may affect the switch machine. Results show that the LFT Bond Graph model based diagnosis whose performances depend mainly on model accuracy is robust with respect to parameters uncertainties and able to detect and isolate the majority of faults, including sensor, plant and actuator faults, while for classification-based diagnosis, even if a good accuracy is achieved with any need of complex equations, a great amount of computational time is devoted to parameters tuning.