Volume 50

Fault diagnosis for industrial gas turbines based on multi-fidelity data Baoyu Zhu, Shaojun Ren, Qihang Weng, Yijia Zhang, Fengqi Si

Abstract

Autoencoders (AEs) are widely used in industrial gas turbines for fault diagnosis. However, traditional AEs often perform poorly with limited training data, causing frequent false alarms. Insufficient data can cause a mismatch between the model and the actual system, especially under operational conditions absent from the training samples. This paper introduces a novel multi-fidelity autoencoder (MFAE) model that integrates limited high-fidelity operational data with abundant low-fidelity simulation data. The MFAE model learns latent features from a pre-trained low-fidelity AE model and establishes correlations between low- and high-fidelity data via linear and nonlinear networks. The effectiveness of the proposed methods is evaluated on an industrial gas turbine unit. MFAE effectively captures essential process characteristics and adapts to various operating conditions with limited operational data, enabling accurate gas turbine monitoring.

Keywords gas turbine, multi-fidelity data, fault diagnosis, process monitoring, autoencoder