Transfer condition assessment of gas turbines via double multi-task Gaussian process

Abstract

Predictive maintenance and health management are crucial means to enhance the operational reliability and safety of gas turbines, with time series condition assessment being a pivotal step in this process. However, the complexity of gas turbine and the harsh operating conditions raise the high cost and significant difficulty in acquiring data via sensors, which pose challenges for the conventional data-driven time series modeling methods. These result in the scarcity of available historical data and the deteriorate of condition assessment accuracy. To this end, this paper proposes an end-to-end, probabilistic transfer learning approach based on multi-task Gaussian process (MTGP) for the condition assessment of gas turbines with limited sensor data. For achieving effective transfer condition assessment, the proposed method, named double MTGP (dMTGP), employs (i) a recurrent neural networks (RNN) projector to extract the explicit evolution patterns, (ii) adopts the MTGP module to transfer knowledge from the source domain to the target domain, and (iii) presents a recursive MTGP strategy to have shared knowledge transferred cross outputs. Finally, the superiority and robustness of the proposed method are comprehensively validated against state-of-the-art competitors on four practical gas turbine units.