Modelling the vibration response of a gas turbine using machine learning

Abstract

AbstractThis work deals with modelling the vibration response of a gas turbine obtained during the start‐up process until reaching the nominal speed for power generation. Analysing the vibrations of a complex systems like a gas turbine is useful for the diagnostic of faults or damages in the internal mechanical components of the different stages that integrate a turbine. This work focuses on the study of the shaft vibrations of the bearing radial type mounted between the shaft and the bearing compressor associated with the speed of the turbine. This relationship is studied using experimental data collected from a particular gas turbine model. In particular, we propose a methodology to synthesize a computational model following a supervised learning approach implemented through different machine learning techniques, including a multi‐layers perceptron network, support vector machine (SVM), random forest (RF) and genetic programming (GP) with local search. Results show that SVM, RF and GP perform very well in this task, producing accurate predictive models. Moreover, there are some interesting trade‐offs between the methods, regarding generalization error, overfitting and model interpretability that are relevant for future applications and research.