Gearbox fault diagnosis method based on multi-channel one-dimensional convolutional neural network feature learning

The input of multi-channel signal can collect fault feature information more comprehensively, and CNN has achieved great success in the field of RGB image recognition. Inspired by its RGB image processing, this paper introduces multi-channel CNN into gearbox fault diagnosis. EMD method can decompose the original signal into multiple intrinsic mode function (IMF) signals to obtain more fault features. Therefore, EMD method is used as the preprocessor of the original signal in this paper. After EMD processing, the IMF with obvious fault information is selected by using kurtosis maximization principle to form multi-channel signal. After that, multi-channel 1dcnn (mc-1dcnn) is constructed to fuse and extract multi-channel signal features, and the effective integration of EMD and mc-1dcnn is realized.

A new deep neural network model mc-1dcnn is proposed, and a gearbox fault diagnosis method based on mc-1dcnn is proposed

(1) Based on the effective decomposition of vibration signal by EMD, mc-1dcnn uses multi-channel one-dimensional signal input and directly takes the original signal as input for fault feature extraction and diagnosis, which has stronger generalization ability and higher recognition accuracy;

(2) The fusion features of multi-channel input are more comprehensive than those of single channel input, and the features learned by one-dimensional convolution kernel are more abstract than those learned by full join layer, which can fully mine the hidden features of data;

(3) Sdae further embedded mc-1dcnn can effectively filter the noise and reduce the dimension of feature set, and improve the performance of fault diagnosis.