Condition Based Maintenance of Gas Turbines Using Simulation Data and Artificial Neural Network: A Demonstration of Feasibility

Abstract

Gas turbine maintenance is crucial due to high cost for the replacement of its components and associated loss of power during shutdown period. Conventional scheduled maintenance, based on equivalent operating hours, is not the best alternative as it can require unnecessary shut downs. Condition based maintenance is an attractive alternative as it decreases unnecessary shut downs and has other advantages for both the manufacturers and the plant owners. However, this has shown to be a complex/difficult task. A number of methods and approaches have been presented to develop condition monitoring tools during the past decade. Condition monitoring tools can e.g. be developed by means of training artificial neural networks (ANN) with historical operational data. Such tools can be used for online gas turbine performance prediction where input data from the plant is fed directly to the trained ANN models. The predicted outputs from the models are compared with corresponding measurements and possible deviations are evaluated. With this method both recoverable degradation, caused by fouling, and irrecoverable degradation, caused by wear, can be detected and hence both compressor wash and overhaul periods optimized. However, non-availability of operational data at the beginning of the gas turbine operation may cause problems for the development of ANN based condition monitoring tools. Simulation data, on the other hand, may be generated by using a manufacturer’s engine design program. This data can be used for training artificial neural networks to overcome the problem of non-availability of operational data. ANN models trained with simulation data could be used to monitor the engine from the very beginning of its operation. A demonstration case using a Siemens gas turbine has been shown for this proposed method by comparing two ANN models, one trained with operational data and the other with simulation data. For the comparison an arbitrary section of operational data was used to produce predictions from both models, whereupon these were plotted with corresponding measured data. The comparison shows that the trends are very similar but the parameter values for the measured and the simulated data are shifted by a constant. Using this knowledge, one can provide an ANN based engine monitoring tool that could be adjusted to a certain engine using engine performance test data. The study shows promising results and motivates further investigations in this field.