Investigation of rolling bearing weak fault diagnosis based on cnn with two-dimensional image

Abstract

In this paper, we choose convolution neural network (CNN) as the method to diagnosis weak fault of rolling bearings. In order to improve the training effect of CNN, different two-dimensional image conversion algorithms which include Gramian angular sum difference fields, wavelet time-frequency diagram, Markov transition field are introduced in to convert one-dimensional time series of bearing vibration signals into images. To relieve the pressure of hardware calculation and shorten the time of training and validation, we use the piecewise aggregate approximation (PAA) to compress the data as much as possible while preserving the whole signal information. We add the batch normalization layer to avoid the gradient saturation problem of ReLU function and minibatch method is used to overcome the instability of stochastic gradient descent with momentum (SGDM) while designing CNN. Each kind of images are made as the training sample, and the results show that both the wavelet time-frequency diagram and the Gramian sum or difference angle field diagram can better identify the fault state, and the wavelet time-frequency diagram was relatively better. By comparing with different recurrent neural network (RNN) diagnosis models, the validity of the model was proved. At the same time, the model is applied to the performance degradation identification of fault parts, and the results shows that the model can effectively identify the degradation of inner ring, outer ring and rolling body, while the accuracy of inner ring and the outer ring is better. This paper provides a new idea for weak fault diagnosis of rolling bearings.