Qualitative Diagnosis of Bearing Fault Based on Convolutional Neural Network

Abstract

The traditional methods of qualitative diagnosis of bearing require complex domain knowledge. The method based on deep belief network overcomes the shortcomings of traditional methods, but the amount of network parameters are so huge that the network is difficult to train. And Convolutional neural network based on time-frequency image requires wavelet transform to obtain time-frequency image. Based on the strong feature learning ability and generalization ability of convolutional neural network, a qualitative fault diagnosis method of bearing based on convolutional neural network is proposed, which is trained directly on one-dimensional vibration signal. There are several advantages of this method. The number of parameters are much less and training is more effective than the deep belief network. In addition, input doesn't require time-frequency image obtained by wavelet transform. A series of comprehensive tests are carried out by using the data of Case Western Reserve University and our laboratory. The result shows that the network can diagnose the bearing fault accurately and the accuracy is higher than the other methods. And the convolutional neural network trained on Case Western Reserve University's data can also accurately diagnoses the fault type of our laboratory's bearing, which indicates that the method can be used in practical.