A comparison of OCMPM and OCSVM in motor and sensor fault detection for traction control system

Fault detection is critical to ensure the safe operation of high speed trains. One class support vector machine (OCSVM) and one class minimax probability machine (OCMPM) are two domain-based single class classification methods and commonly used for fault detection. This paper systematically analyzes their training and detecting complexity, principle of optimization and hyperparameter influence of both methods, and compares their performance on motor and sensor fault data from the simulated traction control system of the high speed train. It shows that OCMPM achieves higher fault detection rate than OCSVM given the same false alarm rate. But OCMPM is unfeasible used for real-time fault detection when the training dataset is large.