Probabilistic Machine Learning Approach Towards Diagnostic and Prognostic of Degrading Engineering Systems

Abstract

Damage detection and prognosis are critical to managing degrading systems in the engineering infrastructure. The damage detection relying on sensor data and prognosis on historical failure data, usually both are treated separately. However, a holistic approach is needed to capture degradation at initial stages and prognosis. When implementing a machine learning framework for remaining useful life (RUL) estimation, two primary scenarios emerge: (1) utilizing a complete run-to-failure dataset for training, and (2) relying on a partial dataset. The latter presents significant challenges due to the necessity for extrapolation. Furthermore, limited research has explored extending damage detection results to RUL estimation by modeling the underlying degradation process using a surrogate measure. Moreover, few studies extend damage detection results to remaining useful life (RUL) estimation by modeling the underlying degradation process using a surrogate measure. An integrated probabilistic machine learning framework that systematically undertakes diagnosis, prognosis, and RUL estimation is lacking in the literature. Additionally, parameters for the surrogate degradation model used for prognosis are often estimated using conjugate gradient methods, which often fall into local minima. These surrogate models also ignore general population trends.