A unified out-of-distribution detection framework for trustworthy prognostics and health management in renewable energy systems

Abstract

With the advances in artificial intelligence, there is a growing expectation of more automatic and intelligent prognostics and health management (PHM) systems for the real-time monitoring of renewable energy systems. Although the deep learning significantly promotes the development of PHM, it generally works in a close-world assumption that the real-time monitoring data are in-distribution (ID). These methods may lack the ability to alert the system when encountering the out-of-distribution (OOD) data that are previously unseen/unknown. In this study, a unified OOD detection framework is proposed for the intelligent PHM, so as to enhance its reliability and trustworthiness. Specifically, two types of OOD data from unseen working conditions and unseen fault types are comprehensively considered in the unified framework. A class-wise outlier detection strategy is presented to detect the OOD inputs during decision-making. To suppress the unexpected distribution shift caused by variable working conditions, a novel generalization representation of learning towards unseen working conditions is developed by using supervised contrastive learning. The proposed OOD detection framework can not only flag the unreliable diagnostic output of deep learning models, but also reduce the interference of variable working conditions, showing its applicability in real application scenarios. Extensive experiments demonstrate the advantages and the significance of the proposed unified OOD detection framework to establish highly reliable and trustworthy PHM models.