Novel anomaly detection algorithms for major and critical items of plant

For an electrical utility, the unplanned outage of a major plant item can lead to a substantial loss of revenue. In addition, outage costs must take into account the cost of any required repairs, or even replacements. It is therefore best to avoid long outages through a combination of thorough maintenance and comprehensive condition monitoring. The condition monitoring techniques available to a utility are related to the quality and quantity of data present. In order to capture the dynamics of complex systems, high frequency data is preferred. Lack of a sufficiently high sampling rate can render traditional methods of modelling the dynamics inaccurate or impractical. This thesis proposes an adaptation of machine learning algorithms to model major and critical plant items. Artificial neural networks (ANNs) are shown to allow anomaly detection in pulverising mills and gas turbines (GTs). Hidden Markov models (HMMs) are shown to allow the modelling of GT dynamics despite the absence of high frequency data, and provide tracking of machine health degradation. In addition, a multi-agent system (MAS) is proposed that incorporates several models based on subcomponents of the turbine. The use of MAS technology also allows other algorithms to be easily added to the system, to provide additional condition monitoring capability. Finally, the use of evidence combination allows the agents of different sub-components and differing algorithms to work together to improve condition monitoring performance.

Name dc:type.qualificationname

phd

Level dc:type.qualificationlevel

doctoral-pg

Grantor dc:publisher.institution

University of Strathclyde

Year dc:date.issued

2013

Author dc:creator

• Kenyon, Andrew D.

DOI

10.48730/xf1b-y729

Identifier

T13343

OAI identifier oai:identifier

oai:strathclyde:6q182k12h