A Comparative Study of Rolling Bearing Fault Classification Using CWT-CNN and STFT-CNN Methods

Early fault detection (EFD) of rolling bearings is critical for the healthy and long operation of rotating machinery. In addition to EFD, fault classification is also important in the overall diagnosis of rolling bearing conditions. Many researchers have focused on time- and frequency-domain-based approaches for EFD. Recently, machine learning-based methods have gained priority in automatic fault feature extraction and classification. This work is focused on the effectiveness of two different methods for EFD and fault classification. The first method consists of combining Continuous Wavelet Transform (CWT) and Convolutional Neural Network (CNN) for fault diagnosis and classification, whereas the second method uses a combination of Short-Term Fourier Transform (STFT) and CNN. Alexnet is a widely used CNN for image classification that provides superior results. In this work, Alexnet is used for the classification of images of vibration signals. Initially, vibration data is acquired from healthy bearings as well as bearings with various seeded defects on the races and balls. The CWT and STFT images are generated from the vibration data, and the image data is used for training and testing in CNN. An accuracy of 95.8% is obtained for the CWT and CNN combination and 92.2% for the STFT and CNN combination.