Application Research On Fault Diagnosis Of Rolling Bearing Based On Deep Learning

Rolling bearing is an essential component of rotation machinery equipment a fault occurs,it will seriously affect the normal operation of the rotation machinery and is closely related to the safety and reliability of the machinery equipment.Moreover,the vibration signal is the main performance indicator to evaluate the health condition of the rolling bearing.Therefore,fault monitoring and diagnosis of the bearing vibration signal can not only achieve real-time monitoring of its health condition,but also provide guidance and reference for troubleshooting and the formulation of maintenance plans.Furthermore,it is of great difficulty to extract deep fault features from vibration signals and easy to fall into dimensional disasters for shallow neural networks.Meanwhile,deep belief networks(DBN)have the ability to process high-dimensional nonlinear data and extract fault effectively.Considering these technical difficulties and the characteristics of DBN,it is of great necessity to establish a fault diagnosis model of rolling bearing based on DBN network.The rolling bearing is taken as the research object,and in-depth research is conducted on key issues in this paper,mainly including vibration signal noise reduction,fault feature selection,and the establishment and optimization of the DBN fault diagnosis models.The main contents of the research are as follows:Firstly,to obtain accurate and reliable experimental data and clarify the failure form and mechanism of rolling bearings,the fault type and failure processing method were selected according to the failure mechanism and laboratory conditions.Then,the experimental device was installed and debugged,and the vibration signal was collected.Due to the interference of the acquisition system and the external environment,the vibration signal was doped with noise to reduce the noise and the noise reduction method of ensemble empirical mode decomposition was adopted.The cross-correlation coefficient and kurtosis criterion were used to evaluate the intrinsic modal function,and the bearing vibration signal was reconstructed in this way,the vibration signal noise reduction was achieved.Furthermore,the characteristic parameters of the time domain and frequency domain were extracted and normalized for the reconstructed vibration signal.Secondly,considering the difficulty of setting DBN network parameters and the problem of low diagnostic accuracy,the network parameters were determined through verification analysis,including the number of hidden layers,the optimal data type,and the activation function of the DBN,Besides,the reconstruction ability of restricted boltzmann machine(RBM)was also verified and analyzed.Moreover,the DBN and other shallow neural networks were compared and analyzed.The results show that the DBN network has high diagnostic accuracy and strong stability,which proves the superiority of the DBN network in the fault diagnosis of rolling bearings.Finally,considering the problem that the accuracy of the DBN network diagnosis is not ideal and it is difficult to determine the number of hidden layer nodes and the learning rate,a modified genetic algorithm(MGA)was used to optimize the DBN.Furthermore,the MGADBN model was applied to the fault diagnosis of rolling bearings,and its fault diagnosis performance was compared with those of BP,DBN,GA-DBN and MGA-BP.The effectiveness of the MGA-DBN diagnosis model was verified.