Data-gedreven gezondheidsbeoordeling en prognostiek van technische systemen bij beperkte databeschikbaarheid

Prognostics and Health Management (PHM) is a multidisciplinary framework for assessing the condition of engineering assets and forecasting their remaining useful lifetime to enable predictive maintenance. This thesis addresses the challenge of developing PHM models under limited data availability, a common constraint in real-world industrial systems and across diverse engineering domains. The research explores how hybrid and domain-informed learning frameworks can enhance model performance, interpretability, and reliability in such data-constrained environments. The study is organized around three key objectives: (1) a comparative evaluation of conventional machine learning (ML) and deep learning (DL) methods for constructing health indicators (HIs) from limited time-series data; (2) the integration of domain knowledge into DL models through constraint-guided optimization to improve physical consistency and reduce dependence on annotated datasets; and (3) the quantification of predictive uncertainty to support more reliable decision-making. These objectives are demonstrated across three case studies: (i) anomaly detection and fault diagnosis in a small household coffee vending machine using multi-sensor data; (ii) bearing degradation modeling using constraint-guided gradient descent (CGGD) to enforce monotonic degradation, bounded health states, and energy-aligned signal behavior; and (iii) hybrid modeling for remaining useful time estimation and uncertainty quantification in telecom base transceiver station backup batteries under forecasted variable load conditions. The results demonstrate that embedding domain-informed constraints significantly improves the robustness and interpretability of DL models under limited data conditions. Hybrid frameworks that combine physics-based and data-driven approaches further enhance reliability and uncertainty estimation. Collectively, this work introduces new methods for constraint-guided learning and hybrid PHM modeling, advancing the development of data-efficient, explainable, and trustworthy predictive maintenance systems.