Interpretable Health Indicators Construction Under Complex Scenarios for Comprehensive Multicomponent System Monitoring

Abstract This paper presents an interpretable framework for constructing health indicators (HIs) to support reliable monitoring and prognostics in complex multicomponent 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 artificial intelligence 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 interpretable-by-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 (tep) and the Commercial Modular Aero-Propulsion System Simulation (c-mapss) turbofan datasets, both featuring high-dimensional, multisensor 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.