Gas path fault diagnosis of aircraft engine using HELM and transfer learning

Abstract

Gas-path fault diagnosis of an aero-engine is a key challenge for flight safety. However, there are two problems: 1) Large collections of healthy condition samples and few fault samples; 2) The diagnosis accuracy cannot meet the practical demands. Therefore this paper proposes a new integrative diagnostic approach for gas-path fault of an aircraft engine to overcome the limitations. A novel model (HELM-TL) is presented to improve fault diagnosis performance, which ensembles the advantages of HELM and transfer learning by fusing target domain data and source domain data effectively. The attention mechanism is adopted to construct global dependency of gas-path monitoring data. Furthermore, this paper develops a novel algorithm for fault diagnosis which dynamically updates learning rate according to loss function. The proposed approach not only obtains more accurate fault distribution, but also avoids over-fitting problem, which greatly improves diagnosis performance and generalization. Finally, experimental data from China Eastern Airlines is adopted to validate. The experimental results show that proposed method has excellent robust according to sensitivity analysis. Moreover, it achieves a good tradeoff between diagnosis accuracy, transfer performance, and runtime cost. The proposed method also improves training accuracy by 5%–15%, reduces RMSE of fault diagnosis by 11%, and improves prediction accuracy by 12.1% compared with the average value of state-of-the-art methods.