Enhancing PHM with XAI: Review and Dataset for System Health Indicator Construction

The integration of Explainable Artificial Intelligence (XAI) into Prognostics and Health Management (PHM) holds the promise of transforming maintenance strategies across various industrial sectors by enhancing the interpretability and accuracy of predictive models. Despite progress, integrating XAI into PHM encounters hurdles, such as the absence of consistent methodologies and a unified framework, complicating the integration of these AI techniques into PHM systems. Additionally, the development of interpretable and robust health indicators (HIs) for systems with multiple components is still largely unexplored. It is hindered by the lack of benchmark data that fully represents system complexities, including component interdependencies and the variability of operational conditions. Therefore, this paper aims at addressing these challenges by offering two main contributions: a focused review of XAI in PHM and the introduction of a novel run-to-fail (RTF) dataset from the Tennessee Eastman Process (TEP). This dataset is a leap forward, enabling the detailed monitoring of component degradation and the exploration of component interactions under various environmental conditions. It allows enhancing the development of explainable HIs and prognostics algorithms at the system level to better reflect the challenges faced in industrial settings.