Learning transfer feature representations for gas path fault diagnosis across gas turbine fleet

Abstract

Intelligent data-driven fault diagnosis based on conventional machine learning techniques has been extensively studied in recent years. However, these methods often assumed that the data used for training and testing are drawn from the identical distribution, which is impractical in real application. Such idealized hypothesis may confine these promising data-driven techniques to well-designed experimental environments rather than actually putting them into real-world applications. In practice, the distribution discrepancies between source domain and target domain will degrade the diagnostic performance. To this end, this work introduces a transfer learning based extreme learning machine to align the distribution discrepancies of the data collected from a turbofan engine, which is rarely studied in the fault diagnosis for aero-engine. The proposed method is capable of learning the transferable cross domain features while preserving the properties and structures of source domain as much as possible. Meanwhile, the marginal distribution and conditional distribution discrepancies are matched. Through these transfer data representations, a relatively high diagnostic accuracy is guaranteed. Finally, extensive experiments have been performed on gas path fault diagnosis of turbofan engine, including hybrid transfer cases and complete transfer cases, to verify the effectiveness and feasibility of the proposed method.