Domain Adaptation Method And Its Application In Fault Diagnosis Of Rolling Bearings Under Variable Conditions

As one of the most important parts of rotating machinery,rolling bearings are widely used in modern industry,such as metallurgy,chemical,mining and transportation.When they are damaged,the stability of machinery will be affected directly,which will then cause unnecessary economic losses and even casualties.So the research on rolling bearings fault diagnosis is of great practical significance.As rolling bearings often work under variable conditions,the target data can not directly available and the distribution of test data and training data is different.Traditional fault diagnosis methods of rolling bearings based on machine learning apply the classification models which trained under one working conditions to another working conditions with unlabeled data directly,this has led to low diagnostic accuracy.The transfer learning which based on domain adaptation train the model by using source domain data with known labels to assist the target domain data for diagnosis.In order to improve the feature mining ability and domain adaptation effect of the transfer learning model,Several new methods of domain adaptation are proposed and applied to fault diagnosis of rolling bearings.The main research of the paper are as follows.(1)Vibration signals of rolling bearings with labels are difficult to obtained under variable working conditions,that leads to low accuracy of fault diagnosis.Aiming at this problem,a new shallow domain adaptation fault diagnosis method based on complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)and manifold embedding distribution alignment is proposed.Firstly,CEEMDAN is used to decompose the vibration signals of rolling bearings under different working conditions,and some intrinsic mode components(IMF)are obtained.Secondly,the time-domain and frequency-domain features of IMF components with larger kurtosis are extracted to construct multi-features sample set.The extracted features are embedded into the manifold space for manifold feature transformation and the transformed manifold features are aligned dynamically.Finally,the classification model is trained with source data and target data to obtain the fault diagnosis results of rolling bearings with unknown labels.The experimental results show that the proposed method can minimize the difference of feature distribution between domains,and improve the accuracy of rolling bearings state recognition effectively.(2)To solve the problem that shallow domain adaptation rely heavily on a priori knowledge and cost lots of time when selecting features,an unsupervised deep domain adaptation method is proposed to achieve end-to-end fault diagnosis.Firstly,one-dimensional dual residual squeeze-and-excitation module is designed to reinforce or suppress different channel features according to the task,so that the proposed features of network can maximize the fault information.Then,the joint maximum mean discrepancy is used to adaptive minimizing joint distribution discrepancy between source domains and target domains.The experimental results show that the proposed method can not only focus fault-sensitive features adaptively,but also has higher accuracy of classification.(3)Aiming at the problem of convolutional neural network(CNN)does not take into account the time sequence of signals when deal with vibration signal,a domain adaptation method based on fused temporal and spatial features is proposed to enhance the domain adaptation performance of the time sequence signal.Firstly,CNN is used to extract the spatial features of rolling bearings vibration signals,and the extracted spatial features are put into the bidirectional gated recurrent unit(Bi GRU)network for further extract temporal features,so that we can obtain fused spatio-temporal features.Secondly,the joint maximum mean difference domain adaptation and domain adversarial adaptation are used to reduce the differences in feature distribution simultaneously.The experimental results show that the proposed method has higher classification accuracy,robustness and better generalization ability.