Ensemble of DAEs for Fault Detection of Gas Turbine Engines

Aiming at the problem of limited fault-type samples, an ensemble of deep autoencoders (DAE) based fault detection approach for gas turbine engines was proposed. The proposed structure first transferred the measurement features into residual features through the physical model of gas turbine engines. Then a set of one-class classifiers based on DAEs and one-class support vector machines (OCSVM) were constructed as the base classifiers of the ensemble. The final prediction result was determined by majority voting. This structure only uses the flight data of the engine under health conditions and limited degraded conditions. The fault detection system was simulated using a dynamic model of a twin-shaft turbofan engine. The results show that the proposed approach can obtain more accurate and reliable fault detection results than the compared fault detection algorithms.