Using Neural Network for Diagnostics of an Industrial Gas Turbine

Abstract

Gas Turbines are being utilized in increasing numbers for industrial applications because of their increasing in efficiency and reliability. However, they become degraded during operation and their associated maintenance costs can become extremely high for the owners. Hence, successful maintenance techniques are those which are able to reduce maintenance costs and down-time. In recent decades industry has started to use predictive maintenance techniques because of their benefits in reducing down-time compared to traditional techniques like breakdown maintenance, as a result different predictive maintenance and diagnostics techniques have been developed during the last fifteen years. This study, in particular, will focus on performance diagnostic techniques based on Neural Networks. The network features and training algorithms will be discussed to develop an appropriate model for gas turbine diagnostics. In addition, it will be shown how training data can affect training performance. This study follows on from previous work carried out at Cranfield University to develop engine health monitoring techniques; however it will attempt to investigate the different abilities of neural networks for use in industrial gas turbine diagnostics, especially in non-standard ambient temperatures and its advantages compared to Gas Path Analysis (GPA) techniques.