Data-driven approaches to prognostics and health management for intelligent maintenance

Modern Cyber-Physical Systems (CPSs) tightly integrate physical processes with digital control and communication layers, operating in increasingly complex, interconnected, and safety-critical environments. Ensuring their reliability and availability requires advanced monitoring and maintenance strategies capable of predicting faults and managing degradation before failures occur. Within this context, Prognostics and Health Management (PHM) has emerged as a comprehensive framework that combines sensing, data processing, diagnosis, and life prediction into intelligent decision support for condition-based and predictive maintenance planning. This dissertation explores data-driven methodologies to enhance the PHM pipeline across both prognostic and diagnostic dimensions. The first part focuses on Remaining Useful Life (RUL) estimation, introducing autoencoder-based feature extraction techniques for health indicator generation, along with multi-class, similarity-based, and degradation-based models for RUL estimation. These approaches are further extended to edge-level implementations to evaluate feasibility in resource-constrained environments. The second part addresses anomaly detection as a core diagnostic task, proposing a general framework for systematic benchmarking of detectors through synthetic yet realistic anomaly generation on time-series data. Combining theoretical modeling, numerical validation, and implementation studies, this work advances the development of robust PHM methodologies. The proposed frameworks offer adaptive and computationally efficient tools to support intelligent, data-driven maintenance in modern monitoring applications.