Efficient Bearing Fault Diagnosis by Fast-Residual Network With 2-D Representation of Vibration Signals

Effective intelligent bearing fault diagnosis has long been a research focus on the condition monitoring of rotary machinery systems. Traditional diagnosis methods require complex computation of feature vectors and rely on domain knowledge and diagnostic expertise. In recent years, many deep learning-based methods have been applied in this field, such as deep belief network and convolutional neural network. However, these diagnosis methods are hard to train and lack of efficiency. In this paper, we use a new residual network structure to solve the bearing fault diagnosis problem. Firstly, we give a brief description to the original residual network. Then, we redesign the structure and propose a new paradigm: Fast-Residual Network. To fit in the new model, we develop an effective pre-processing method using frequency slice wavelet transform and shape the one-dimensional representation signals to two-dimensional representation image. The final experimental results show that proposed model achieves higher classification and is faster and more accurate than other deep learning-based models.