Probabilistic and Physics-Informed Machine Learning for Predictive Maintenance with Time Series Data

Physics-informed neural networks are capable of learning from both observation data and the underlying physical laws. Meanwhile, their implementation in real application settings requires additional considerations related to multi-objective optimization of variables with vastly different scales. Besides, many applications benefit from having well-calibrated uncertainty estimate along with the prediction. In this study, we examine physics-informed neural network for a predictive maintenance application with times series data, using a physical fatigue crack propagation model from mechanical engineering. Our goal is to attain good predictive performance, while at the same time producing correct uncertainty intervals and limiting computation cost. Moreover, we also consider as baselines some established uncertainty quantification techniques in deep learning, and we provide a detailed quantitative assessment of their calibration.