Research On Fault Diagnosis Of Rolling Bearing Based On Time-frequency Decomposition And Deep Learning

Rolling bearing is one of the most important components in rotating machinery,the good working state of which is the premise to ensure the normal operation of machines.Due to frequently operating in high temperature,high speed,heavy load and other harsh conditions,rolling bearings will inevitably occur various failures,affecting the working performance or even resulting in huge economical losses and occasional casualties.To maximize the work benefit and minimize the maintenance cost,it is of great practical significance and theoretical value to study the fault diagnosis techniques of rolling bearing.Vibration monitoring signals contain abundant information of rolling bearings,which are one of the most commonly used analysis signals in fault diagnosis.Due to the frequent change of conditions,strong ambient noise and disturbance of moving parts,vibration signals contain extremely complex frequency components,which makes it difficult to achieve high precision and high reliability fault diagnosis by traditional methods.In this paper,the rolling bearing fault identification problem is studied,vibration signals are mapped to the timefrequency domain to better characterize the key information,and bearing status identification is carried out combining the self-learning advantages of deep learning algorithms.Through continuous optimization of deep learning network structure,several improved models are proposed in depth step by step.The main work contents of this paper are as follows:(1)Aiming at the problem that the unsteady characteristics of original signals will reduce the identification accuracy and reliability of diagnosis models,a bearing diagnosis model based on a structure optimized deep convolutional neural network(SO-DCNN)combining with wavelet packet transform(WPT)was proposed.This model takes the wavelet packet coefficient matrix obtained by the WPT of vibration signals as input,increases the kernel size at the beginning of network to improve the capability of mining key input features,introduces the identity shortcut structure into the main body of network to alleviate training gradient diminishing problem.adds a global average pooling layer at the end of network to improve the model generalization performance.The experimental results prove the necessity of using wavelet packet transform in the diagnostic process and the superiority of SO-DCNN model compared with classical machine-learning based fault diagnosis algorithms.(2)In order to improve the quality and efficiency of model mining features from wavelet packet coefficient matrix and further enhance diagnosis performance,a multilevel correlation information deep convolutional neural network(MCI-DCNN)based on SO-DCNN was proposed.This model considers the difference of information contents between wavelet packet coefficient matrix and traditional gray image,optimizing the information extraction efficiencies of the front-end convolutional kernels greatly by adjusting kernel sizes pointedly.This model is capable of extracting features from multiple information correlation levels,exponentially expanding the dimensions of fault feature space and highlighting the difference between different diagnostic categories.The experimental results prove MCI-DCNN can improve the feature learning ability and fault recognition performance freely without extra adding model parameters.(3)Aiming at the problem that the quality of WPT forms of signals will affect the final fault diagnosis performance,an integrated WPT deep convolutional neural network(IWPTDCNN)was proposed.This model constructs a special recursive convolutional structure on the basis of MCI-DCNN,making original signals automatically realizing the WPT process inside the diagnostic model.By optimizing all the training model parameters synchronously,the adaptability of WPT network module to the diagnostic task is improved,enhancing final diagnosis performance.The experimental results show that the accuracy of IWPT-DCNN is better than that of the MCI-DCNN model.(4)Aiming at the problem that a single sensor is easy to lose part of bearing information in the signal acquisition process,and multiple sensors are necessary for more reliable fault diagnosis,a multi-source sensor fusion deep attention convolutional neural network(MSFDACNN)was proposed.This model is developed on the basis of IWPT-DCNN,receiving multiple signal sources as input and utilizing attention mechanism to subtly handle the multisource sensor information redundancy problem.By adaptively amplifying the salient features and filtering the redundant features,all the information can be fused effectively,achieving the higher accuracy and reliability bearing fault diagnosis.The experimental results prove that attentional mechanism is effective and the diagnostic reliability of MSF-DACNN is better than that of IWPT-DCNN which only uses a single sensor source.