Wavelet transform based convolutional neural network for gearbox fault classification

Abstract

As a valid method of time-frequency analysis, Wavelet transform (WT) can offer great help for gearbox fault diagnosis. However, it requires much human expertise and prior knowledge to diagnose the faulty conditions of gearbox according to the time-frequency distribution. In addition, the coupling of different failures and noise makes it hard to accurately diagnose the running conditions of the gearbox. In this paper, the convolutional neural network (CNN) is applied for the classification of gearbox health conditions with the time-frequency image generated by WT. As a typical model of deep learning, CNN has distinguished capacity in image recognition. It can automatically extract faulty features from time-frequency images, which can depress the uncertainty of artificial feature extraction. For comparison, S-transform (ST) and short time Fourier transform (STFT) are combined with CNN for the same classification task. Experimental result indicates that the combination of WT and CNN is superior to other methods.