Towards The True Hybrid: Physics-Informed Trainable Models for Prognos

One of the areas where machine learning has been applied extensively has been Prognostics and Health Management (PHM) a discipline encompassing fault detection, damage quantification, failure prediction and related predictive maintenance and health-aware decision making.

Although such data-driven tools have been extremely promising, they require large amount of data to refine predictions outside of the training distribution. This requirement hinders their application in domains where data are scarce, e.g., run-to-failure of critical structures and systems, a foundation of PHM. Recently, one of the key area of research in computer science has been physics-informed machine learning, where prior knowledge in the form of partial differential equations, ordinary differential equations, or more general physical constraint is embedded in machine learning architectures, thus trying to guide the training process through physical knowledge, or to constrain the solution through a set of requirements. 

This presentation starts with the foundation of PHM and its benefits, the recent introduction of physics-informed machine learning for PHM and then the challenges of its application. The presentation will concludes with examples of works in the area of hybrid neural networks and physics-constrained Gaussian processes.