Motor Bearing Fault Diagnosis Based On Deep Learning Convolutional Neural Network

In the field of electrical equipment,all kinds of large electric units and other major equipment are developing in the direction of high speed,high precision and high efficiency.It is necessary to establish a reliable fault diagnosis system in order to protect personal safety and prevent major economic losses caused by accidents.Rolling bearing is a workpiece which extremely prone to damage,and because it is also an indispensable component in the rotating equipment of electric machinery,there possess very important economic and social value in the research of it’s diagnosis.With the development of industrialization,a large amount of industrial monitoring data has been accumulated,which also catalyzes the rapid development of machine learning in modern times.In the context of the era of big data,traditional fault diagnosis methods have been unable to keep up with the pace of development now due to the need for complex manual preprocessing or the support of specific expert experience and knowledge,And as a new show,data-driven fault diagnosis method has been more and more favored by scholars and experts at home and abroad.As the most cutting-edge research achievement in the field of pattern classification recognition and machine learning,deep learning theory has equally achieved good performance in speech recognition,machine vision,emotion analysis,semantic translation and other personalized fields.As a typical network in deep learning,convolutional neural network is widely used in image processing,target tracking,target detection,time-frequency analysis and other scenarios due to its extremely powerful feature extraction capability for all kinds of complex information.And the research purpose of this paper is to explore a new convolutional structure or method in the field of motor bearing fault diagnosis under deep learning.The main research contents of this paper as follow:Firstly,this paper introduces two of the most common neural network structures in the field of fault diagnosis based on deep learning,and focuses on the research status of motor bearing diagnosis,points out some problems that still exist in the field now.The whole process of this method is simple,and it has a good diagnostic effect as well.Secondly,contrapose to the situation that diagnosis accuracy of traditional one-dimensional convolutional structure is not high enough to give full play to convolutional network which possess powerful advantages in the field of image recognition,the diagnosis of bearing faults under the two-dimensional convolutional structure is mainly explored.In this method,the onedimensional original vibration data is simply normalized and expanded into a two-dimensional matrix structure,and the data value in the matrix corresponds to the chromaticity value of each pixel,which is finally converted into the corresponding grayscale image.Thirdly,aiming at the problems of gradient dispersion/explosion and network degradation that may exist in a pure two-dimensional convolution structure,a double-flow method combining low-latitude information and high-latitude information is proposed for diagnosis,that is,one-dimensional and two-dimensional information of fault signals are prepared simultaneously in the input layer in advance,the optimization of hyper-parameters in the algorithm model is discussed,and several groups of comparative experiments are carried out in different sample sets.The double-flow structure not only takes into account the original onedimensional signal information,but also improves the receptive field of the network in higher dimensions.Experimental results show that the proposed structure has a better classification performance than the one-dimensional convolutional structure.Finally,in view of the problem that the two-dimensional information in the original doubleflow method is only a simple arrangement of one-dimensional information,and it can’t dig deeper information,a time-frequency double-flow method is proposed,and the twodimensional information is replaced by a wavelet time-frequency graph which contains deeper information.The optimization of each super parameter in the algorithm model is also analyzed.Multiple groups of comparison tests are carried out in different sample sets,and the classification performance of different algorithm models for motor bearing fault data sets is compared and analyzed in multiple angles.It can be seen from the experimental results that compared with the original double-flow algorithm,the performance of the algorithm under the double-flow structure is better.