Frequency-driven Convolutional Neural Network for Enhancing Noise-Robustness of Bearing Fault Detection

Bearing fault detection has been a critical issue, in particular under noisy environments considering that a robust fault detection model should be designed to perform well under harsh and diverse operational circumstances. Recent existing works using deep learning-based methods for noise-robust modeling have contributed to the capability of coping with a variety of environments. However, there still exist challenging tasks to enhance the performance as the influence of the noise increases. Herein, a convolutional neural network (CNN)-based model is proposed to achieve anti-noise bearing fault diagnostics by taking into account spatial and frequency-wise characteristics. The proposed method exploits Hilbert-Huang transform (HHT) in order to extract mode-wise instantaneous and spectral features, as well as the CNN-based modeling via frequency-driven grouping strategies. By utilizing mode-wise frequency characteristics in feature extraction and feature learning of deep learning framework, it is shown that our proposed method yields promising performance improvement for noise-robust fault diagnosis.