Controllable Transfer Learning and Physics-informed Bayesian Methods for Advanced Predictive Analytics in Dynamic Heterogeneous Environments

Abstract

Predictive analytics has become a vital transformative tool across many industries, enabling proactive action planning, informed decision-making, and improved operational efficiency and safety. At its core, predictive analytics aims at converting raw historical data to actionable insights using various analytical methodologies, including traditional and modern methods. By leveraging recent advancements in Artificial Intelligence (AI) technologies and advanced computing resources, modern predictive methods can analyze and extract insights from high-dimensional big data, thereby revealing hidden nonlinear patterns that traditional methods would miss. Despite the outstanding capabilities of existing modern (AI-driven) methods, such as classical Machine Learning (ML) and Deep Learning (DL), the effectiveness and reliability of their predictive performance are limited by several critical factors. These limiting factors include limited historical data, training and testing data domain discrepancy, poor long-term estimations, training data heterogeneity, data drift due to dynamic operating conditions, and lack of domain or physics-based knowledge integration. Therefore, there is a growing need for advanced predictive methodologies capable of enhancing the effectiveness and performance reliability of AI-driven methods, which is the main objective of this study.

To address the identified literature gaps, this study proposes innovative predictive frameworks by introducing various methodological improvements to conventional AI-driven methods. More specifically, we develop a controllable deep transfer learning method with theoretical guarantees on both the effectiveness and controllability of knowledge transfer to tackle the prevalent issues of limited historical data, domain discrepancy, and poor long-term estimations. The developed method advances the transfer learning paradigm by utilizing a novel adaptive domain adaptation strategy, thereby enabling effective and controllable knowledge transfer from the source domain to the target domain. To tackle training data heterogeneity and data drifting issues, we develop probabilistic predictive frameworks that aim to harness the strengths of statistical mixture modeling, parametric regression modeling, and Bayesian statistics. The developed methods utilize a Bayesian updating scheme to effectively adapt the parameters of the trained data-driven model as more measured data become available in real-time for the machine system of interest, thereby enabling reliable prognostic analysis with uncertainty quantification measures. Furthermore, we propose an integration of a deep learning method, Bayesian statistics, and a physics-informed adaptive sampling algorithm to address the issues of poor long-term estimations, domain discrepancies, and lack of domain knowledge integration of conventional AI-driven methods.

In essence, the advanced predictive methodologies developed in this study are capable of (1) handling training data heterogeneity, (2) extracting highly relevant and physically explainable health indicators from raw signal data, (3) effective and controllable transfer of health degradation knowledge, (4) effective and controllable domain adaptation, (5) accurate and robust long-term estimations, (6) providing uncertainty quantification around the predictions, and (7) leveraging domain or physics-based knowledge. The effectiveness and reliability of the developed methods for predictive analytics are demonstrated using various case studies of battery-powered systems. Overall, this study advances the field of Prognostics and Health Management (PHM) by developing innovative methodological improvements for AI-driven predictive methods. These advancements would facilitate more reliable health condition monitoring and prognostic analysis of engineering systems. Moreover, the developed methods can be applied to wider applications, including non-engineering applications.