This thesis explores the development of data-driven and machine learning methods in application to the health monitoring of rotating plant items being used in the primary and secondary cycles of the Advanced Gas-cooled Reactor (AGR) nuclear power plants in the UK. The methods fall broadly into two categories: the statistical augmentation of a pre-existing knowledge-based system for turbine generator vibration alarm analysis, and the development of a machine learning model for the exploration of long-term predictive measures of asset health for AGR gas circulator units. Both of these topics are unified in their engineering context, and the overall aim of the approaches employed: to provide improved decision support using data to reliability staff tasked with monitoring key nuclear assets. A self-tuning methodology for knowledge-based system parameterisation and data selection in rotomachinery vibration monitoring is introduced, providing a comparative study of numerous methods and case studies for features of interest in both steady-state and step change conditions. These approaches were developed using a historical dataset taken from a turbine generator in use at an AGR, with time series streams from multiple component channels. An event-driven approach to asset health is presented, utilising a support vector machine & logistic regression hybrid model to estimate particular states of interest associated with the gas circulator duty cycle. This approach to health monitoring (examining responses during semi-regular refuelling events) is shown to correlate highly with the remaining useful life of a circulator unit which eventually underwent an unexpected failure, and provides a potential quantitative metric for preventing repeat instances.
|Date of Award||19 Sep 2019|
- University Of Strathclyde
|Sponsors||British Energy Generation (UK) Ltd|
|Supervisor||Stephen McArthur (Supervisor) & Campbell Booth (Supervisor)|