Unsupervised Learning for Improving Fault Detection in Complex Systems

Complex mechatronic systems are becoming more widespread such as in industrial machinery and robotic systems. Detecting faults in such systems proves to be a challenging task due to the multitude of components that are interacting. In this paper, we demonstrate how we use an unsupervised learning technique to detect accelerated wear patterns in complex systems, where wear interactions between components are present. We use Gaussian Mixture Models (GMM), to uncover the intricate wear process that takes place when old worn out components are coupled with new healthy components. Then through a numerical simulation of a complex system, and experimental data gathered from a gearbox accelerated life testing platform, we demonstrate that this new-old component coupling leads to an accelerated rate of wear of the new components, and so they would have a lowered life expectancy that would jeopardize the reliability of the system. This approach shows that more fault prevention related information is gained if we take all interacting components into account when monitoring and modelling wear processes of complex systems. Such gained information could lead to more accurate Remaining Useful Lifetime (RUL) estimations and more robust fault prevention.