Detecting the Onset of Machine Failure Using Anomaly Detection Methods

Mohammad Riazi, Osmar R. Zaïane, Tomoharu Takeuchi, Anthony Maltais, Johannes Günther, Micheal Lipsett

Abstract: During the lifetime of any machine, components will at some point break down and fail due to wear and tear. In this paper we propose a data-driven approach to anomaly detection for early detection of faults for a condition-based maintenance. For the purpose of this study, a belt-driven single degree of freedom robot arm is designed. The robot arm is conditioned on the torque required to move the arm forward and backward, simulating a door opening and closing operation. Typical failures for this system are identified and simulated. Several semi-supervised algorithms are evaluated and compared in terms of their classification performance. We furthermore compare the needed time to train and test each model and their required memory usage. Our results show that the majority of the tested algorithms can achieve a F1-score of more than 0.9. Successfully detecting failures as they begin to occur promises to address key issues in maintenance like safety and cost effectiveness.