Machine learning approaches to support life extension of nuclear power reaction

Condition monitoring is the process of observing a parameter, or multiple parameters, extracted from industrial assets which provide an indication of their overall health or presence of faults. In many instances the data produced is numerous, difficult to understand and time consuming to analyse. In the nuclear power industry large volumes of data are captured through inspections and monitoring but only a relatively small amount of the data has associated health information. Machine learning has been identified as a solution to these issues allowing data to be analysed quicker and in greater volumes, alleviating some of the demands on the relevant engineers. Within this thesis two key contributions are made. The first is the development of a model of fuel channel bore estimation and subsequent deployment to make estimations of operational data. Fuel Grab Load Trace (FGLT) data is operational data which was never intended to be used in condition monitoring. Greater understanding of FGLT data was obtained by targeted analysis of specific regions of the data based on physical understanding of nuclear assets. The model was produced by combining techniques from machine learning supplemented with formalised engineering knowledge to provide decision support in a safety-related environment. The model was found to produce estimations of fuel bore within 1mm root mean square error on real historical FGLT data. The second is the exploration and implementation of semi-supervised machine learning techniques applied to the detection of cracks within nuclear reactor monitoring data. Semi-supervised machine learning is attractive as it provides comparable results to supervised machine learning techniques without the costly burden of labelling large volumes of data, a commonly occurring challenge in nuclear reactor core condition monitoring. Three different semi-supervised machine learning approaches have been evaluated and for the detection of cracks using FGLT data it has been found that Transductive Support Vector Machines have the best performance. In this thesis I demonstrate the benefits of machine learning and semi-supervised machine learning to decision support of nuclear reactors through two health monitoring related problems of nuclear reactor cores. Challenges still remain with the application of machine learning in the nuclear condition monitoring domain including the high cost of obtaining labelled data, lack of historical operate to failure data and incorporating future data in the training of algorithms as it is made available.Condition monitoring is the process of observing a parameter, or multiple parameters, extracted from industrial assets which provide an indication of their overall health or presence of faults. In many instances the data produced is numerous, difficult to understand and time consuming to analyse. In the nuclear power industry large volumes of data are captured through inspections and monitoring but only a relatively small amount of the data has associated health information. Machine learning has been identified as a solution to these issues allowing data to be analysed quicker and in greater volumes, alleviating some of the demands on the relevant engineers. Within this thesis two key contributions are made. The first is the development of a model of fuel channel bore estimation and subsequent deployment to make estimations of operational data. Fuel Grab Load Trace (FGLT) data is operational data which was never intended to be used in condition monitoring. Greater understanding of FGLT data was obtained by targeted analysis of specific regions of the data based on physical understanding of nuclear assets. The model was produced by combining techniques from machine learning supplemented with formalised engineering knowledge to provide decision support in a safety-related environment. The model was found to produce estimations of fuel bore within 1mm root mean square error on real historical FGLT data. The second is the exploration and implementation of semi-supervised machine learning techniques applied to the detection of cracks within nuclear reactor monitoring data. Semi-supervised machine learning is attractive as it provides comparable results to supervised machine learning techniques without the costly burden of labelling large volumes of data, a commonly occurring challenge in nuclear reactor core condition monitoring. Three different semi-supervised machine learning approaches have been evaluated and for the detection of cracks using FGLT data it has been found that Transductive Support Vector Machines have the best performance. In this thesis I demonstrate the benefits of machine learning and semi-supervised machine learning to decision support of nuclear reactors through two health monitoring related problems of nuclear reactor cores. Challenges still remain with the application of machine learning in the nuclear condition monitoring domain including the high cost of obtaining labelled data, lack of historical operate to failure data and incorporating future data in the training of algorithms as it is made available