Hybrid Physics-Informed Neural Networks for Digital Twin in Prognosis and Health Management
, University of Central Florida
Highly Rated
Predictive digital-twin models are used to guide decisions in prognosis and health management. We'll challenge the myth that building digital twins using deep learning requires large amounts of labeled data. We introduce a framework that allows for the simultaneous use of physics-informed and machine learning by implementing recurrent neural networks for cumulative damage modeling. The main advantage of this approach is the compensation of limitations in physics-informed kernels as well as labeled data. Our framework handles highly unbalanced datasets formed by few output observations and data lakes containing time series used as inputs. GPU computing enables scaling up computations for fleets of hundreds of assets, keeping training under a few hours and inference under a few seconds. We'll show applications in wind energy (optimizing service intervals at a turbine level), civil aviation (scheduling detailed inspection by tail number), and autonomous vehicles (forecast of battery life).