A Deep-learning approach to predicting availability of Industrial Gas Turbines

McGinty, Jason Kenneth (2019) A Deep-learning approach to predicting availability of Industrial Gas Turbines. Masters thesis, University of Lincoln.

A Deep-learning approach to predicting availability of Industrial Gas Turbines
thesis jason mcginty.pdf - Whole Document

Item Type:Thesis (Masters)
Item Status:Live Archive


In this thesis a new system is proposed which applies the techniques developed in the domain of Deep Learning to predict the ongoing availability of Industrial Gas Turbines (IGT’s) in the range of 5 to 15MW. Deep learning has been targeted due to the recent rapid advances made due to the availability of increasingly mature software platforms, development methodologies and high-performance hardware.

The complexity of problem is explored in the literature review (such as the difficulty of modelling accurately the many different physical processes, noisy or missing data and the complex interdependence of many of the systems), and it is summarised that the focus of this work is one which does not appear to have been comprehensively tackled within the published literature (whose main focus has been on individual component reliability rather than the IGT in its entirety).

This thesis also demonstrates some of the progress that have taken place in recent years with regards the core algorithms and frameworks of deep learning by revisiting previous work and clearly showing the improvements advances such as Batch Normalisation and Adaptive Momentum (Adam) has made. Further investigation into the most appropriate such algorithms for this particular problem is also covered.

The proposed deep learning architecture uses a number of traditional network components (such as Multi-Layer Perceptron’s, Convolutional Neural Network’s etc.) in a way to solve a novel problem of overall IGT ( and which is hoped is applicable in other complicated industrial environments) availability prediction problem and to account for the large gaps in visibility of issues that current systems encounter when attempting to predict availability for IGT’s as single holistic system.

A further unique fact of this work is that it has also leveraged large amounts of data made available by a major OEM in order to realise the full benefit of a deep learning approach to this problem. This is in contrast to the majority of papers published to date which focus solely on a handful of units. This has made possible for the first time a digital twin wrought out of operation of an entire fleet of a diverse range of IGT’s and allow for a comprehensive coverage of the factors pertaining to the maintenance of a high availability of the industrial assets targeted.

Divisions:College of Science > School of Computer Science
ID Code:47505
Deposited On:08 Dec 2021 12:58

Repository Staff Only: item control page