Research On Rolling Bearing Diagnosis Based On Full Vector Convolutional Recurrent Neural Network

The rolling bearing guarantees the normal and stable working state of the rotating shaft.With the development of industrial big data,enterprises choose multiple sensors to obtain multi-dimensional information of units for monitoring.For the traditional fault diagnosis methods,the fault characteristics of the signal are usually selected manually or the general signal analysis method is adopted.Compared with using convolutional neural network and recurrent neural network to extract signal features adaptively,manual feature selection is more random and less generalized.The feature enhancement of the data set was carried out by using the full-vector spectrum.The one-dimensional convolutional neural network(ID-CNN)was constructed to extract the signal classification information and the LSTM was constructed to extract the signal timing information.The integrated neural network was constructed and the Full-vector convolutional recurrent neural network(FV-CRNN)was proposed for the fault diagnosis of rolling bearings.The main research work is as follows:(1)The theoretical knowledge and derivation of convolutional neural network and recurrent neural network are introduced.Experiments on real fault signals verify the effectiveness of the vector spectrum theory compared with the single channel signal analysis method.(2)A fault diagnosis method of rolling bearing based on one-dimensional convolutional recurrent neural network is proposed and its effectiveness is verified by experiments.Firstly,the feature fusion of homologous dual-channel fault signals was carried out using the vector spectrum technology to obtain the data set of the principal vibration vector of the vector spectrum.Secondly,the data set was manually labeled and divided according to the ratio of 8:2 between the training set and the test set.The one-dimensional convolutional neural network(1D-CNN)and the short-short time memory neural network(LSTM)in the cyclic neural network were respectively constructed.The data set noise was added by denoising coding,and the cross-entropy was used as the loss function.After fine-tuning by Adam optimizer,the full-vector convolutional recurrent neural network(FV-CRNN)model was obtained.The experimental results show that the accuracy and generalization of vector CRNN in the test dataset are higher than that of single-channel CRNN.At the same time,the advantages of this model in fault classification task are verified intuitively by t-sne feature dimensionality reduction and visualization of principal components.(3)Manual selection of CEEMD energy entropy characteristics,combined with Softmax multiple classifier for bearing fault classification.The hidden layer features extracted from single-channel CRNN and full-vector CRNN were input into Softmax multi-classifier respectively,and compared with the energy entropy feature classification results of CEEMD.Experimental results show that the features extracted by full-vector CRNN adaptive algorithm are more effective for fault classification than those extracted by manual selection.