Feature-based fault detection of industrial gas turbines using neural networks

Abstract

Gas turbine (GT) fault detection plays a vital role in the minimization of power plant operation costs associated with power plant overhaul time intervals. In other words, it is helpful in generating pre-alarms and paves the way for corrective actions in due time before incurring major equipment failures. Hence, finding an efficient fault detection technique that is applicable in the online operation of power plants involved with minor computations is an urgent need in the power generation industry. Such a method is studied in this paper for the V94.2 class of GTs. As the most leading stage for developing a feature-based fault detection system and moving from a fixed time-scheduled maintenance to a condition-based one, principal component analysis is used for dimension reduction in the sensor data space and dimensionless key features are employed instead. One healthy condition and 6 faulty conditions are used to provide a realistic data set that is used for feature extraction, training, and testing artificial neural networks. In the proposed method, multilayer perceptron (MLP) and learning vector quantization (LVQ) networks are used for the fault classification. The good performance of the LVQ networks is presented by properly selecting the network architecture and respective initial weight vectors. When comparing the results of the MLP and LVQ networks for the fault classification, the LVQ network shows better classification results.