Since joining ENIT, UTTOP as an Associate Professor in September 2017, my research has focused on Prognostics and Health Management (PHM), which leverages condition monitoring data and operational insights to assess system health, detect anomalies, diagnose defects, and predict Remaining Useful Life (RUL). With the growing industrial adoption of PHM, research has explored model-based, data-driven, and hybrid approaches to enhance maintenance planning and ensure system reliability, availability, and safety. However, Industry 4.0 necessitates not only improved system reliability but also autonomous health management. Thus, shifting from traditional PHM to AI-driven solutions is essential. Our research develops AI-enabled PHM methodologies that enhance system resilience through real-time diagnostic and prognostic decision-making.

This axis addresses the collection and extraction of critical information from diverse data sources to construct health indicators and predict degradation levels, essential for effective system monitoring. A particular focus is given to the engineering challenges associated with acquiring reliable monitoring data, emphasizing the significance of industry collaboration in developing trustworthy benchmark datasets. Robust methodologies for leveraging heterogeneous data in health indicator construction are examined, with applications spanning various systems, including rotating machinery and turbofan engines. Furthermore, attention-based deep learning models are explored for aligning multimodal data streams and extracting underlying characteristics.

In this axis, we introduce advanced methodologies to navigate through data-related challenges inherent in PHM, including managing sparse, missing, and unlabeled data sets. The first challenge pertains to unlabeled data, particularly in the context of unknown fault types in online monitoring. To address this, we propose a methodology that integrates direct and indirect monitoring information to detect and diagnose unknown robotic drifts. The second challenge involves sparse condition monitoring data for fault diagnostics. In response, we present a multimodal learning approach that combines multiple data modalities while incorporating domain knowledge. Additionally, we introduce physics-informed machine-learning techniques to extract meaningful features from limited datasets. Finally, we tackle the issue of sparse run-to-failure data in prognostics by developing a novel contrastive self-supervised learn- ing framework to improve prognostic accuracy.

This axis proposes new techniques for dealing with the inherent uncertainties in prognostic analyses, thereby supporting more informed and reliable post-prognostics decisions. We first focus on the development of advanced methodologies for managing uncertainties in prognostics, considering the variability in system degradation trends, uncertainties arising from diverse operational conditions, and challenges related to model assumptions and measurement noise. Building on this foundation, we develop robust strategies for post-prognostic decision-making. These strategies include optimizing mission profiles to extend system lifetimes and introducing a dynamic predictive maintenance framework that integrates prognostic uncertainties into maintenance planning. By accounting for these uncertainties, the proposed approach enhances decision-making reliability and improves system availability.

This research advances system-level prognostics by incorporating the dynamics of component interactions and the influence of operational profiles, thereby enhancing predictive accuracy and reliability. We begin with a comprehensive review of existing System-level Prognostics (SLP), analyzing their advantages, challenges, and existing gaps. This review underscores the critical need for robust and advanced SLP techniques to manage the increasing complexity of industrial systems. Building on this foundation, we introduce a methodology for quantifying uncertainty in multi-independent-component systems with diverse structural configurations. This approach utilizes probabilistic deep learning techniques to predict RUL distributions at the component level and subsequently assess system-level reliability. The adaptability of this methodology to various system architectures highlights its practical significance and potential for large-scale industrial applications.

With the vision of building trustworthy and transparent AI for industrial predictive maintenance, we have initiated a comprehensive integration of XAI principles into the field of PHM. Our objective is not only to improve prediction accuracy but also to ensure that critical insights generated by AI models are interpretable, actionable, and aligned with domain expertise.

To this end, we have systematically evaluated leading XAI techniques such as SHAP, LIME, and attention-based attribution methods, particularly in the context of complex multimodal learning models. These techniques have been instrumental in enhancing model transparency and user confidence, especially for high-stakes industrial applications. Furthermore, our approach emphasizes the construction of hierarchical health indicators, both at the component level—to capture localized wear, fault signatures, and degradation patterns—and at the system level, where we model the interdependencies and cascading effects across subsystems. This dual-level health monitoring framework, coupled with XAI, allows for proactive maintenance strategies, supports informed decision-making by human operators, and facilitates the certification of AI systems in safety-critical environments.