Transfer-learning based gas path analysis method for gas turbines

Abstract

Data-driven gas path analysis is a state-of-health diagnostic method. The method utilizes input-output information to solve the health assessment problem of gas turbine engines. In recent years, the rapid development and intense competition of the gas turbine industry bring new challenges to the method. A novel method is required to achieve life-cycle monitoring and further increase diagnostic accuracy. This study proposes a transfer-learning based gas path analysis method. The method combines transfer learning with data-driven gas path analysis for the first time. The proposed method can not only learn fault patterns from engine operating data but also recognize the different diagnostic value of data. This ability enables the developed GPA model to automatically adjust its training dataset to maintain a steady diagnostic accuracy in life-cycle monitoring. Experiments were presented and discussed on a two-spool split flow turbofan engine platform. The transfer-learning based gas path analysis method achieves averaged 94% diagnostic accuracy in a life-cycle simulation. The method improves diagnostic accuracy by 11–20% as compared with traditional data-driven methods.