Long-Term Prediction of a Gas Turbine Based on Vibration Analysis Using Fully Connected Recurrent Neural Network

Availability of large systems such as gas turbines is a critical issue during their operation. Condition Based Maintenance (CBM) is one of the effective methods to increaseavailability level and inhibit catastrophic failures which lead to break down. Prognostic technique is decision making part of CBM to estimate future condition of the machine. This paper describes an improved prognostic approach to estimate longterm prediction of vibration condition of a gas turbine compressor shaft. In order to estimate the condition of compressor shaft, Fully Connected Recurrent Neural Network (FCRNN) and Jordan Recurrent Neural Network (JRNN) are applied with Levenberg-Marquardt (LM) training method based on historical vibration data of the shaft. Time domain speed and vibration signals of the shaft are collected from an industrial gas turbine when fault is occurred in the healthy shaft. According to the results achieved, the FCRNN method has a better MSE performance than the JRNN model. Simulation results illustrate that the FCRNN method with LM training is capable of providing an accurate long-term prediction to estimate both fault and normal conditions in CBM approach

Keywords:

Condition Based Maintenance , Long Term Prediction , Prognostic , Recurrent Neural Network , Vibration Analysis