General Information           Program          Registration

Chris Mechefske, Queens University: Hybrid (Data Driven, Physics-Based, Digital Twin) Mechanical System Fault Detection/Prognosis
In situations where there is not enough collected data available to allow a useful data-driven model to be developed, knowledge from analytical (physics-based) or computational (digital twin) models may be used to fill in the gaps. As an example, “synthetic” data from an analytical or digital twin model could be used to augment training, or knowledge from these models could be integrated into the machine learning. Physics-based analytical models of various mechanical components and systems (rolling element bearings, gears, and gear systems) already exist. The tools to develop digital twin models for components, as well as small and large mechanical systems also exist (eg. MATLAB toolbox). A larger data set generated by these models would allow exploration of different paradigms (such as reinforcement machine learning) that are out-of-reach with the existing cataloged data and potentially allow for more accurate machine fault detection and remaining useful life estimation.