Data-driven design for fault prognosis: Application to industrial components, subsystems, and systems

Abstract

Technical processes in diverse industries, like manufacturing, chemicals, and power generation, involve intricate operations that aim to achieve specific outcomes. These operations often entail complex interactions among components and systems. However, they can also carry substantial risks to health, the environment, and industry sustainability. Hence, the implementation of a robust fault prognosis system is paramount for safeguarding the safety and dependability of these intricate technical processes.

Conversely, these technical processes frequently accumulate vast quantities of historical data through routine sensor measurements, event logs, and records. This observation fuels a compelling interest in crafting fault prognosis methods solely reliant on this abundant process data. Consequently, the main objective of this thesis was to design effective data-driven fault prognosis strategies tailored to diverse operational scenarios.

This thesis explores fault prognosis across various technical process levels, including key components, subsystems, and systems. It begins with basic component-level issues and progresses to intricate system-level challenges, aiming to gain a holistic grasp of data-driven fault prognosis complexities in industrial contexts.

The first objective aimed to create a reliable fault prognosis system for critical technical process components, specifically roller bearings. A novel data-driven prediction framework was proposed, involving two phases: feature extraction via Empirical Mode Decomposition and Remaining Useful Life (RUL) prediction using an RFs-based model with hyperparameters fine-tuned through Bayesian optimization. Notably, this approach demonstrated substantial enhancements in RUL prediction accuracy compared to conventional data-driven and stochastic methods during an actual run-to-failure experiment involving roller bearings.

The second objective aimed to create an effective fault prognosis system for technical process subsystems, with a focus on preventing operational failures. The research selected an automated fuse test bench, a manufacturing line subsystem, as the subject of study. Initially, an integrated fault diagnosis system based on extreme gradient boosting was introduced, showcasing superior performance in detection and classification accuracy while achieving quicker diagnosis times than standard approaches. Subsequently, the scope of extreme gradient boosting was broadened to encompass fault prognosis through methodological enhancements and the incorporation of supplementary data streams like images.

The third objective focused on designing an accurate predictive model for anticipating future operating conditions in industrial systems to avert catastrophic accidents. The study honed in on a liquid hydrogen storage system as its research subject. A novel application of the random forests algorithm was introduced to enable early detection of hazardous incidents like liquid hydrogen spills, thereby averting catastrophic outcomes like detonation. The model demonstrated remarkable accuracy, surpassing other machine learning methods previously employed for similar experiments. This model, forged through the study, offers valuable insights for comprehensive risk analysis and the identification of prevention and mitigation measures, especially in the context of emerging liquid hydrogen technology applications.

This Ph.D. study offers the potential for enhanced industrial fault prognosis methods, fostering improved safety and sustainability within the industry. Furthermore, it may serve as a valuable reference and launching point for future academic research, offering insights into the merits and complexities of employing data-driven techniques in real-world industrial applications.