Bearing Fault Diagnosis Research Based On Improved Convolutional Neural Network

Rolling bearings are the key components in mechanical equipment.The quality of the bearings is inseparable from the normal operation of the equipment.With the high efficiency,intelligence and diversification of production,a large number of fault data appear diversified,the requirements for intelligent fault diagnosis of equipment are becoming higher and higher,especially the fault diagnosis of rolling bearings has become the top priority.However,the existing methods still have some shortcomings in the fault diagnosis of rolling bearings:traditional bearing fault diagnosis is to process some fault signals,which requires a deep mathematical foundation and strong professional knowledge to extract fault features,and the fault diagnosis and recognition accuracy is low;The convolutional layer of the convolutional neural network is to extract the spatial position of the image.The pure one-dimensional signal spatial features are not obvious,and the one-dimensional signal needs to be converted into a two-dimensional image;the preprocessing of the input data and the network appear during training Overfitting,the application of the activation function will bring about a mean shift;the diagnosis result of single fault cannot accurately reflect the real working condition of the equipment.Because the spatial characteristics of the one-dimensional vibration signal are not obvious,the convolutional neural network has a good feature learning ability for the two-dimensional image.The texture information of the two-dimensional image indicates the damaged feature of the fault,and the efficiency of the fault feature extraction by the convolutional neural network model is improved.In view of the shortcomings of traditional bearing fault diagnosis methods,this first built a simple convolutional neural network model to diagnose two-dimensional vibration images of bearing faults,of which there are only two convolutional layers and pooling layers.Experiments were conducted on the data set,and 86.52% accuracy of bearing fault diagnosis was obtained.The feature learning ability of the convolutional neural network is verified,and the problemthat the traditional bearing fault diagnosis method brings difficulty in extracting feature information and complicated process when extracting fault features is solved.Aiming at the use of sigmoid and modified linear unit(Re LU)activation functions,gradient disappearance and mean shift are brought about.In order to effectively extract fault feature information and reduce the amount of calculation,this introduces a residual layer and proposes an improved Re LU function t Re LU,The improved Re LU solves the neuron necrosis and mean shift of the x<0 part.It is verified by the two data sets of the Western Reserve University bearing data set and the German Paderborn University.The accuracy of bearing fault diagnosis is 99.18% and 100%,respectively.Aiming at the problem of single and multiple fault identification in rolling bearing pitting faults,the vibration time series image contains a lot of redundant information,and a convolutional neural network model of wavelet decomposition image is proposed.The model first performs wavelet decomposition on the vibration image,extracts contours and detailed features,and inputs them into the convolutional neural network model.Experimental verification of 15 kinds of bearing failures using the hybrid data set of bearings from Case Western Reserve University and Padelborn University in Germany,The accuracy rate reached 95.33%.