Fault diagnosis of Rolling Bearing Based on One-dimensional Residual Convolution Recurrent Neural Network

Abstract

To resolve the issue that conventional rolling bearing fault diagnosis technology are incapable of extracting features adaptively, a one-dimensional residual convolutional recurrent neural network (1DRCRNN-LSTM) is proposed to obtain signal characteristics directly from the original signal. Firstly, a train-valid-test paradigm dataset with sample overlap is created by data augmentation and one-hot coding. Secondly, a convolutional neural network (CNN) and a long short-term memory neural network (LSTM) are fused and a residual learning mechanism is introduced to build a network model for the extraction of signal characteristics. Finally, a softmax classification layer is implemented to classify ten types of faults and output diagnostic results. The proposed network model is verified using the Case Western Reserve University (CWRU) Rolling Bearing Dataset and the results show that the 1DRCRNN-LSTM model has a fault diagnosis accuracy of 0.987, which is better than that of multilayer perceptron (MLP) and convolutional neural network and other methods. Meanwhile, through t-sne visualization techniques, it is proved that this method has superior feature extraction and fault diagnosis capabilities.