Abstract This paper presents an interpretable framework for constructing health indicators to support reliable monitoring and prognostics in complex multi-component systems. Modern industrial environments generate large volumes of heterogeneous sensor measurements, and extracting meaningful and interpretable indicators from these signals remains a critical step in system health assessment. While data-driven methods have achieved promising predictive performance, they often operate as black-box models, which limits transparency, trust, and practical adoption in safety-critical industrial applications. Explainable AI has emerged as a key solution to address this black-box limitation by providing transparent and understandable model behaviors. In this context, this work develops an interpretableby-design health indicator construction framework that offers clear insight into sensor relevance, component degradation behavior, and system-level health evolution. To this end, the proposed framework introduces three key innovations: (i) a context aware sensor selection mechanism that identifies the most informative sensing channels under different operating conditions, enhancing robustness to noise and variability; (ii) a flexible module that incorporates expert knowledge to guide the shape and behavior of component-level health indicators when domain insights are available; and (iii) an adaptive aggregation strategy that automatically selects appropriate functions to synthesize component-level health indicators into a system-level representation consistent with the system?s structural and operational characteristics. The framework is validated on the Tennessee Eastman Process and the C-MAPSS turbofan datasets, both featuring high dimensional, multi-sensor measurement data with interacting components. Experimental results show that the method produces monotonic, trend-consistent health indicators that capture component and system degradation more accurately than existing approaches, offering improved interpretability, adaptability, and robustness for measurement-informed monitoring in complex industrial settings.
Nguyen et al. (Fri,) studied this question.