Vibration State Recognition Method Based On Depth Feature Learning

In order to solve the problem that the fault diagnosis method based on signal processing technology needs to rely on artificial feature extraction and expert knowledge,and the fault diagnosis method based on traditional machine learning theory has insufficient learning depth and difficult feature extraction,a vibration state recognition method based on deep feature learning is proposed.This method is based on Deep Neural Network(DNN)and takes one-dimensional data structure as the processing object.It can extract vibration signal features deeply and adaptively,learn the feature representation of signal and realize the unification of feature extraction and state recognition process.At the same time,this method can be end-to-end fault diagnosis,is an intelligent fault diagnosis method.First of all,is proposed based on a one-dimensional Convolutional Neural Networks(CNNs)of rotor fault diagnosis methods.In this method,one-dimensional vector of vibration signal is input to one-dimensional CNNs for analysis.By utilizing CNN's powerful feature learning and extraction ability,the characteristics of different fault states are deeply learned from the original vibration signal,and finally the signal state is recognized.Secondly,in order to solve the problem that the vibration signal components are complex,nonlinear and non-stationary in actual faults,which affect the deep learning effect,Hilbert Vibration Decomposition(HVD)method and multi-feature information fusion technology are introduced into the deep learning process of vibration signals,and a rotor fault diagnosis method based on Vector Convolutional Neural Network(VCNN)multi-feature fusion learning is proposed.Firstly,the vibration signals collected by multiple sensors are decomposed by HVD to obtain IMF(Intrinsic Mode Function),then fuses them into multi-feature information matrix,and inputs them into VCNN in the form of multi-vector for learning and classification.Since the multi-feature information matrix can better represent the state characteristics of the equipment,it is beneficial to improve the accuracy of the deep learning model in fault identification.Then,a Fully Connected Vector Deep Neural Network(FVDNN)is designed to solve the problem that it is difficult for VCNN to extract the overall characteristics of one-dimensional vibration signal.The model introduces a fully connected neural network to simulate the signal processing process of traditional signal processing technology,so that the deep learning simulation can be adaptive and comprehensively extract the characteristics of the signal.Finally,through the establishment of rotor fault data set for the above diagnosis methods of neural network structure of the experimental study,the most suitable for fault classification network parameters,the establishment of a diagnosis model.Through the experimental comparison with other machine learning diagnosis methods,it is proved that the vibration fault diagnosis method based on deep learning has certain advantages in diagnosis accuracy.