A Turboshaft Aeroengine Fault Detection Method Based on One-Class Support Vector Machine and Transfer Learning

Abstract

The fault detection of turboshaft engines is very important to ensure the flight safety of helicopters. Because there are few fault data in engine historical operation data, engine fault detection is often regarded as an anomaly detection problem, which is solved by the one-class classification (OCC) method. However, previous studies on fault detection usually ignored the difference between engine data caused by different engine states and operating conditions. Therefore, this paper considers the existence of these differences and introduces transfer learning to solve the problem. In this paper, based on one-class support vector machines (OC-SVM) and transfer learning (TL), an algorithm named OC-SVM-TL is proposed to detect turboshaft engine faults. In this algorithm, the hyperplane of the OC-SVM is transferred from the source domain to the target domain as a knowledge structure to help the target domain to establish a fault detection model with high accuracy. The training process is divided into two steps. The first step is to train the OC-SVM model with the data of the source domain, and the second step is to train the OC-SVM model with the data of the target domain, and the hyperplane difference between the source domain and the target domain is considered in the training process. Finally, the fault detection experiment of turboshaft engine was designed, and the fault detection of turboshaft engine was carried out under different working conditions and different engine states. The experimental results showed that the proposed algorithm has good fault detection performance when the target domain data are few or the amount of target domain data changes.