Research On Fault Diagnosis Method Of Rolling Bearings Under Noise And Cross Domain Conditions

Rolling bearings are the indispensable critical component of mechanical systems,and their health conditions directly affect the secure and smooth operation of devices.Condition monitoring and fault diagnosis of bearings are significant works to ensure the efficient production of devices and avoid safety accidents.However,the operating environments of bearings in actual engineering are complex,and the existing intelligent diagnosis methods are exposed to problems such as lack of robustness to noise and adaptability to working conditions,making it difficult to meet the requirements of practical applications.Consequently,on the basis of the existing researches,fault diagnosis methods for bearings under complex conditions such as high noise and diverse working conditions are studied,and four different intelligent diagnosis models are proposed.The main details of this thesis are as follow:(1)Aiming at the problem of insufficient feature information of single signal,a multi-scale mean reconstruction method is presented for data expansion and information augmentation.On this basis,a fault diagnosis framework based on multi-scale mean permutation entropy(MMPE)and grey wolf optimization support vector machine(GWO-SVM)is established.MMPE algorithm calculates the permutation entropies of reconstructed sub-signals at multiple scales,which are regarded as the features for describing the health condition of bearings.And,the application of GWO-SVM classifier improves the diagnostic accuracy and intelligence.Fault diagnosis experiments under diverse noise conditons are conducted using CWRU bearing data,and the superiorities and robustness of the proposed method are fully demonstrated.(2)A multi-channel convolutional neural network(MCCNN)is established for end-to-end bearing fault diagnosis under high noise conditions.The model reconstructs the original signal through the proposed multi-scale mean reconstruction method,and rich fault features are extracted out of the sub-signals via multiple parallel 1D-CNNs,which improves the diagnostic accuracy and robustness.Moreover,the applications of residual learning and global average pooling enhance the model training performance.Based on the CWRU bearing data,multiple fault diagnosis experiments under diverse noise conditions are carried out,and the results showed that the MCCNN has great anti-noise capacities.(3)In order to solve the problem that the existing intelligent diagnosis models have insufficient generalied diagnostic capabilities under diverse working conditions due to data distribution discrepancy,a tranfer multi-branch convolutional neural network(MBCNN)is proposed.The average pooling layer is adopted to complement the multi-scale reconstruction towards the input signal,and a deeper network structure is well designed.Then,an appropriate transfer learing method is established,thus transfer fault diagnosis can be achieved with merely a small number of samples on target working condition.A series diagnosis experiments are launched using the CWRU and SELF bearing data,the transfer diagnostic accuracies on different pathes are exceed 98%,thus it has better portability than the existing methods.(4)In order to achieve fault diagnosis in the condition where the target domain merely contains unlabeled data,a multi-branch domain adaptation network(MBDAN)is designed.The labeled data on source domain and the unlabeled data on target domain are simultaneously fed into the model,and the MK-MMD feature distribution alignment and adversarial learning are combined to force the multi-branch feature extractor to capture the domain invariant features,thus the fault classifier is capable of diagnostic tasks on both source and target domain.Comparative cross domain diagnosis experiments are conducted based on CWRU and SELF bearings data,the results indicate that MBDAN not only has great cross-domain diagnosis capabilities under diverse working conditions,but also performs well between different devices.To sum up,this thesis proposes four intelligent diagnostic frameworks for bearing fault diagnosis under noise and cross domain conditions,establishes a complete system of bearing fault diagnostic methods under complex conditions.Moreover,the effectiveness and progressiveness of the proposed methods are fully demonstrated via multiple experiments and results analysis.