Fault Diagnosis of Aero-engine Bearing Using a Stacked Auto-Encoder Network

Abstract

As an important part of the rotor system of the aero-engine, the bearing has always been the key site of fault. However, it is hard to monitor and diagnose the bearing's health state. In this paper, the authors used a deep learning methods of Stacked Auto-Encoder Network (SAE) to diagnose and classify the bearing faults, which have different varieties and different levels, basing on the bearing failure data obtaining by the bearing failure test bed. In this paper, the accuracy and convergence rate of SAE algorithm are studied by changing the sample length of the collected data and transforming the original time domain signal into frequency domain signal by fast Fourier transform (FFT). In the case of equal input, it is compared with Deep Believe Network (DBN), which is another method of deep learning. The results show that with the increase of the sample length, the diagnostic accuracy is also increased. And the diagnostic accuracy of algorithm which input data is frequency domain parameter is higher than the one which input data is by original time domain parameter. When the input data is frequency domain parameters, the diagnostic accuracy is up to 97%, and the algorithm is also more stable.