Data-Driven Performance Prediction Using Gas Turbine Sensory Signals

Abstract

The performance of a gas turbine engine (GTE) deteriorates with degradation and aging. The availability of the operating data from the GTE with the capability to perform data analysis provides an opportunity to identify short-term and longterm performance deterioration and relate to more difficult to detect components degradation. In this work, a data-driven and machine learning-based predictive modeling framework has been developed for performing combined input and model selection towards generating easily interpretable, parsimonious and accurate regression models intended for gas turbine engine performance analysis. The proposed multistage predictive modeling framework incorporates the orthogonal least squares (OLS) learning and multi-criteria decision-making approach for selecting inputs and model structures in a computationally efficient manner while optimizing multiple objectives. The regression models obtained from this framework for predicting power and exhaust gas temperature (EGT) outputs using GTE operational data collected over a period of three years have demonstrated short-term and long-term performance deterioration patterns for the GTE.