A PHYSICS-INFORMED NEURAL NETWORK FRAMEWORK FOR BIG MACHINERY DATA IN PROGNOSTICS AND HEALTH MANAGEMENT FOR COMPLEX ENGINEERING SYSTEMS

Abstract

Big data analysis and data-driven models (DDMs) have become essential tools in prognostics and health management (PHM). Despite this, several challenges remain to successfully apply these techniques to complex engineering systems (CESs). Indeed, current state-of-the-art applications are treated as black-box algorithms, where research efforts have focused on developing complex DDMs, overlooking or neglecting the importance of the data preprocessing stages prior to training these models. Guidelines to adequately prepare data sets collected from CESs to train DDMs in PHM are frequently unclear or inexistent. Furthermore, these DDMs do not consider prior knowledge on the system’s physics of degradation, which gives little-to-no control over the data interpretation in reliability applications such as maintenance planning.In this context, this dissertation presents a physics-informed neural network (PINN) architecture for remaining useful life (RUL) estimation based on big machinery data (BMD) collected from sensor monitoring networks (SMNs) in CESs. The main outcomes of this work are twofold. First, a systematic guide to preprocess BMD for diagnostics and prognostics tasks is developed based on expert knowledge and data science techniques. Second, a PINN-inspired PHM framework is proposed for RUL estimation through an open-box approach by exploring the system’s physics of degradation through partial differential equations (PDEs). The PINN-RUL framework aims to discover the system’s underlying physics-related behaviors, which could provide valuable information to create more trustworthy PHM models. The data preprocessing and RUL estimation frameworks are validated through three case studies, including the C-MAPSS benchmark data set and two data sets corresponding to real CESs. Results show that the proposed preprocessing methodology can effectively generate data sets for supervised PHM models for CESs. Furthermore, the proposed PINN-RUL framework provides an interpretable latent variable that can capture the system’s degradation dynamics. This is a step forward to increase interpretability of prognostic models by mapping the RUL estimation to the latent space and its implementation as a state of health classifier. The PINN-RUL framework is flexible as it allows incorporating available physics-based models to its architecture. As such, this framework takes a step forward in bridging the gap between statistic-based PHM and physics-based PHM methods.