Planetary gearbox fault feature learning using conditional variational neural networks under noise environment

Abstract

The features signals of early fault collected from planetary gearbox are usually weak. It is difficult to extract effective fault features from the collected vibration signals under noise environment. In this paper, a new feature learning method for fault diagnosis of planetary gearbox based on deep conditional variational neural networks (CVNN) is proposed. First, the new method utilizes multi-layer perceptron (MLP) to model the normal distribution features of frequency spectra from noisy vibration signals. Second, the new features are obtained by resampling normal distribution features in order to eliminate the effect of noise. Then the denoised features are compressed and reduced dimensionally by MLP. Third, the effective denoised features are input to classifier. Finally, the trained CVNN is applied for intelligent fault diagnosis of planetary gearbox. The experimental results confirm that CVNN method can extract effective fault features from noisy vibration signals, and it has higher accuracy of fault diagnosis than other methods in the case of low signal to noise ratio (SNR) values.