Intelligent Fault Diagnosis Based on Multi-Resolution and One-Dimension Convolutional Neural Networks

Intelligent fault diagnosis (IFD) plays an important role to increase the safety and reliability of rotating machinery. In recent years, there is a large number of deep-learning-based algorithms applied to IFD. Many studies have shown that the one-dimension convolutional neural network (1D-CNN) performs well in fault diagnosis on original vibration signals. However, the original vibration signals are very complicated and difficult to analyze directly. Some critical information is often hidden in different frequency subbands, such as some fault state is only related to specific subbands. In order to enhance the learning progression well organized for the 1D-CNN model, the proposed approach is based on wavelet packet decomposition (WPD) and 1D-CNN. First, decompose the original vibration signal into different frequency subbands with WPD to make the signals more concise. Next, to be compatible with the 1D-CNN model, we proposed a novel 1D-CNN with multi-resolution (CNN-MR) and the experiment result according to Case Western Reserve University (CWRU) Bearing Data Center. Through CNN-MR, it achieves high accuracy for fault diagnosis.