Research On Fault Diagnosis Of Rolling Bearings Based On CNN

As an advanced technology,fault diagnosis is utilized to avoid machine failure and enhance its reliability.Nowadays,no matter for the security or the stability of rotating machines,rapid expanding productive forces are requiring more and more strictly.As one of the core components in the machine,the diagnosis on rolling bearings has become a vital exploration field.Along with the booming advancement of computer science recently,fault diagnosis has gradually developed from precision to intelligence.Convolutional Neural Network has been in the spotlight because of remarkable computation efficiency.By extracting the feature of bearing vibration data comprehensively,it can identify their faults.Aimed at various situations,this paper designs corresponding diagnosis method based upon Convolutional Neural Network.In the meantime,an experimental platform for fault diagnosis was established to study the availability of the method.The following paragraphs list the major elements of the article.1.Fault features of bearing vibration signal can be easily interfered by background noise,so this paper combines Singular Spectrum Decomposition with 1D Convolution Neural Network to propose a novel diagnosis approach.We decomposed bearing vibration signal which were collected into dozens of Singular Spectrum Components.Then we followed the correlation coefficient and kurtosis rule to choose available components for reconstruction.The reconstructed signal is input into 1D Convolutional Neural Network and the diagnosis result is obtained.At last,we applied the method to the diagnosis experiment and prove its excellent diagnosis performance.Even if we put Gaussian white noise to the testing data,this method still could diagnose the fault accurately.2.This paper designs the Feature-Integration Convolution Neural Network by combining1 D Convolutional Neural Network with Convolutional Neural Network because traditional deep learning diagnosis model always ignores the frequency domain features of bearing data.This model has two kinds of input data including 1D vibration data and time-frequency maps.We proved the availability of the diagnosis model by open bearing data.4.Transfer Learning can deal with the tough challenge to obtain sufficient labelled bearing vibration data in reality.As a result,the paper connects it with Convolutional Neural Network to design a novel model.The fundamental theory of transfer learning is introduced,including the concept of Maximum Mean Difference and Joint distribution adaptation.Then,a fault diagnosis model named TLCNN is designed.During training,we utilized the dynamic learning rate.This method was put to the fault diagnosis experiment.Even under the transfer tasks of different working conditions,the diagnosis result of proposed diagnosis model is more accurate than other methods.