Multiscale and Multiorientation Log-Gabor Network for Feature Learning of Vibration Signals in Bearing Fault Diagnosis

Vibration signal analysis is one of the most important methods of machinery fault diagnosis. However, the fault-related components in vibration signals are often buried by the strong background noises, which makes it difficult to perform feature extraction. To address this challenge, a new deep neural network (DNN), multi-scale and multi-orientation Log-Gabor network (MMLGNet) is constructed in this study to extract features from complex vibration signals. Firstly, log-Gabor filters are embedded in MMLGNet to manipulate the learned convolution filters, which allows features to be extracted directly from the vibration signal. Secondly, a multi-scale and multi-orientation log-Gabor (MMLG) layer is proposed in MMLGNet, in which the Gabor filters with different scales and orientations are used to extract the features from vibration signals. The impulses that contain fault information can be highlighted through MMLG to enhance the feature learning performance of the model. Thirdly, an energy entropy-based feature fusion method is proposed to address the issue of redundant feature extraction, which is proposed to enhance channels with strong impulsive features. Finally, the effectiveness of the MMLGNet is validated by the experiments of bearing fault diagnosis with two datasets. The results indicate that MMLGNet has a better feature learning ability than those state-of-the-art DNNs.