PHYSICS-INFORMED DEEP LEARNING FRAMEWORK FOR PROBABILISTIC MODELING OF ENVIRONMENTALLY INDUCED DEGRADATION

Abstract

Evaluating the degradation behavior and estimating the lifetime of engineering systems and structures is crucial to ensure their safe and reliable operation. Deep learning (DL) models, which are in the form of multi-layer neural networks (NN), have been widely used for the prognostics of such systems and structures, primarily by estimating their degradation intensity and remaining useful life (RUL). Although DL prognostic models have shown promising performance, there are limitations with such models that need to be considered. Firstly, they only learn the data patterns without consideration of the governing physics of degradation. Excluding physics, accompanied by the lack of interpretability in DL models, makes them prone to violating physical laws while showing a good fit to the training data. This issue may lead to weak generalization, mainly for predicting situations outside the training dataset. Secondly, they require significant data for sufficient training, which may not always be available. To estimate degradation and lifetime, NNs are typically trained in a supervised setting using labeled data that ideally have been collected at different levels of degradation up to the failure points. However, collecting that data is usually expensive and time-consuming, particularly for durable systems with long lifetimes, as material degradation (e.g., corrosion, fatigue, wear, or creep) is often slow. Therefore, there is a need for a model that possesses interpretability and follows the underlying physics of degradation that occurs in real-world conditions. Additionally, this model should be trainable with limited data.This dissertation proposes a novel data-driven framework to address the abovementioned limitations, including disregarding physics, lack of interpretability, and the need for big data in DL prognostics models. The framework comprises two NNs: a physics discovery NN and a predictive NN. The former models the underlying physics of degradation, while the latter makes probabilistic predictions for degradation intensity. The physics discovery NN guides the predictive NN and forces it to follow the underlying physics of degradation, which results in better life estimations. In this way, less data is required for sufficient training as the physics discovery model acts as a constraint and limits the search space for the parameters in the training of the predictive model. Additionally, integrating the state-of-the-art feature importance measurement methods into the physics discovery model makes it possible to identify the primary environmental factors that significantly impact the degradation process. This work enhances the interpretability by shedding light on the dominant factors influencing the system's degradation. The application of the proposed approach is demonstrated through two case studies based on publicly available datasets for degradation phenomena. The outcome of this research study can be used to develop a prognostics and health management system that can facilitate a low-cost and high-performance predictive maintenance strategy for systems experiencing environmentally induced degradation. Also, the proposed method can guide data collection from the field by revealing the influential factors that play crucial roles in the degradation of systems. Moreover, the proposed approach offers valuable benefits to designers, enabling them to incorporate appropriate preventive and mitigation strategies into their designs.