Research On Fault Diagnosis Of Rotating Machinery Based On Multi-feature Fusion CNN Network

With the continuous promotion and development of several industrial revolutions,various machinery and equipment are playing a more and more decisive role.Bearings,as one of the most basic and important parts of mechanical equipment,which problems of them in any part may give rise to deterioration and failure of the mechanical equipment,affecting the normal operation of the entire industrial production activities.Also,if the related faults are not found in time,it may lead to staff casualties.Therefore,it is very essential to accurately and efficiently identify the existence of bearing faults.Traditional diagnostic methods are mostly based on feature extraction.However,feature extraction is a difficult process.In addition,it also relies on expert experience,and the extracted features greatly affect the final diagnostic accuracy.Therefore,deep learning provides an automatic and effective method to extract the features of raw data.This article takes the rolling bearing of rotating machinery as the research object,combined with the convolutional neural network in deep learning,the research topic of this paper is introduced,and a series of research work is carried out to solve the problem of bearing fault diagnosis.The main research contents are as follows:(1)The network model is trained by starting from the One Dimensional Convolutional Neural Network(1D CNN)model.Based on the CWRU bearing data set,the original one-dimensional vibration signal is taken as the input of the network model.Further,to improve the diagnosis accuracy of the network model,the methods of batch standardization +Dropout and batch standardization + Support Vector Machine(SVM)are respectively used to optimize the network model.By comparing the fault diagnosis accuracy obtained under different methods,it can be concluded that batch standardization +SVM can better optimize the network model.(2)The 2D time-frequency graph obtained by converting is taken as the input of the network model to train the network model.Short-time Fourier transforms(STFT),wavelet transform(WT),generalized S-transform(GS-T)and Wigner-Ville distribution were used to transform one-dimensional vibration signals,and then the twodimensional time-frequency graphs obtained by their conversion were input into the network model for training.By comparing the fault diagnosis rates obtained by four time-frequency analysis methods,it can be concluded that the effect of generalized S transformation is better.(3)Aiming at the problems of low recognition accuracy of bearing fault diagnosis model and poor generalization ability caused by insufficient fault data features,a multifeature fusion CNN fault diagnosis method is designed.On the one-dimensional dimension,it takes a one-dimensional array of one-dimensional vibration signals encoded by one-hot codes into the built 1D CNN model for training.On the twodimensional dimension,it first transforms one-dimensional vibration signal into twodimensional time-frequency diagrams through the generalized S transformation,and takes them as the inputs of the 2D CNN model,and then trains the network model.Different from the CNN model above,it has an additional convergence layer,which combines the features of the two dimensions,to achieve the effect of multi-feature fusion and enrich the feature information of different dimensions.The experimental verification of the CWRU bearing data set and MFPT bearing data set shows that the multi-fusion CNN model has higher diagnostic accuracy than other similar methods.