Analysis of Vibration Data and Investigation of Models to Support Maintenance Decision of Rotating Components

Abstract

This PhD project focuses on analysis of vibration data and investigation of mathematical models to support maintenance decision of rotating components and more specifically roller bearings. The PhD work generally consists of experimental work and mathematical modelling. There are 6 papers included in the thesis. The first 4 papers work on developing a laboratory vibration setup to run accelerated life tests, data-processing, and investigation of prognostic models, while Paper 5 uses the insights gained from previous papers to propose a maintenance optimization model to be able to support maintenance-decisions for systems subject to competing failures. Paper 6 by (Alfarizi et al. 2022) studies a random forest technique as a machine learning algorithm for bearings and I contributed in data collection and data processing.

Analysis of vibration data and prognostic models for remaining useful life (RUL) prediction of bearings have been widely studied in recent literature. Bearings are critical components which are widely employed in industrial machines, power generation, and automotive systems. They are often operating in harsh and demanding environment conditions such as high temperature, heavy loads, and contaminated surroundings. Such factors can accelerate the bearings wear and tear, making RUL prediction crucial for an effective maintenance planning, safe operation, and proactive resource management.

RUL prediction of bearings is challenging due to noisy vibration signals, presence of low and high frequency components, nonlinearity of the signals, as well as diversity of fault types and their degradation patterns. In Papers 1-4, we collected the laboratory data, investigated time-domain and frequency-domain statistical features obtained from various signal-processing techniques, and developed deterministic and stochastic RUL prediction models to predict RUL of bearings. Additionally, a comparative analysis was carried out to highlight the benefits of frequency aspect of the vibration signals over time aspect, while the most appropriate health indicators (HI) were proposed for condition-monitoring. Different types of uncertainties were incorporated in the model to address the stochasticity of failure threshold and model parameters. In Paper 5, the findings from previous papers have been used and a numerical maintenance optimization model was proposed for continuously monitored systems that follow a Wiener process. The model is suitable for systems that experience random shocks in addition to internal degradation in their lifetime. A Monte-Carlo simulation-based approach is also developed to validate the results of the numerical model. The “run-to-failure” datasets were collected in Reliability, Availability, Maintainability, and Safety (RAMS) laboratory at NTNU and used in the papers to demonstrate the applicability of the proposed frameworks.

The thesis can serve as a basis for further research and engineering applications from different perspectives. The physical characteristics of bearings (i.e., fault signatures) can be integrated with stochastic models to enhance the RUL prediction performance. Machine learning algorithms such as pattern recognition can be implemented on faulty processed signals for both fault diagnosis and failure prognosis purposes. Moreover, it would be interesting to further improve the maintenance optimization model to consider different maintenance strategies, consider various sources of shocks with varying deterioration rate affected by a change of operating condition, as well as developing the model to take a system-level perspective.