Oscar Serradilla Casado obtained the qualification 'CUM LAUDE'.

Oscar Serradilla Casado obtained the qualification 'CUM LAUDE'.

---

Thesis title: Design and validation of a methodology to implement data-driven predictive maintenance in industrial environments

Court:

• Chairmanship: Basilio Sierra Araujo (UPV/EHU)
• Vocal:José María Alonso Moral (Universidad de Santiago de Compostela)
• Vocal:Gian Antonio Susto (University of Padova)
• Vocal: Urko Leturiondo Zubizarreta (Ikerlan S.Coop.)
• Secretary:Jose Ignacio Aizpurua Unanue (Mondragon Unibertsitatea)

Abstract:

• New trends in manufacturing and industry leads to digitalise all processes, machines and communicate them forming cyber-physical systems (CPS), facilitating process monitoring and data acquisition. The analysis of that amount of data provides new insight in product quality, process optimisation and predictive maintenance (PdM). PdM analyses industrial assets to perform maintenance actions that extend their life and anticipate their failures to prevent them, optimising maintenance costs with respect to time-based and corrective maintenance strategies. PdM systems aim to monitor industrial assets to detect anomalies, diagnose their root cause, predict their degradation and propose mitigation actions.
• Research on data-driven PdM systems has increased in the last six years due to their capability to model complex industrial systems by learning from the large amount of data collected from industrial assets. However, they are rarely transferred to industrial production scenarios due to they fail incorporating domain expert knowledge to the system. In addition, most data-driven works do not address industrial requirements such as interpretability, real time execution, novelty detection or uncertainty modelling. No methodology to guide the life-cycle of data-driven PdM models in industrial environments exist, which could facilitate the implementation of PdM systems in real use-cases to reduce maintenance costs and avoid production breakdowns.
• The main contribution of this thesis is the design and validation of the methodology for data-driven techniques and expert knowledge combination for predictive maintenance (MEDADEK-PdM). It defines the stages, steps and tasks to guide the design, development and implementation of data-driven PdM systems according to business and process characteristics. It defines the required working profiles to facilitate their collaboration, and includes a list of deliverables. The methodology is designed in a flexible and iterative way, combining standards, state-of-the-art methodologies and related works of the field.
• The methodology has been validated empirically by its application in three industrial use-cases, where industrial requirements are addressed. The first use-case consists of modelling correct working engine data to detect anomalies in run-to-failure aviation engine data, addressing novelty detection with data-driven PdM systems in a simulated environment. The second use-case consists of estimating and explaining the remaining useful life (RUL) of experiments in a bushing testbed, by combining data-driven PdM systems with explainable artificial intelligence (XAI) techniques and domain knowledge. The third use-case implements an adaptable data-driven PdM system for semi-supervised anomaly detection and diagnosis in press machine process data. The system detects novel anomalies and performs their diagnosis combining XAI, clustering and projection techniques. The adaptability of the system to changing environmental and operational conditions (EOC) is addressed with transfer learning.
• The application of the proposed methodology guides the life-cycle of data-driven PdM systems, integrating human in the loop (HITL) to include domain knowledge. As a result, the obtained PdM systems tackle the specific industrial requirements of the addressed use-cases, obtaining a trade-off between accuracy and explainability.