Gearbox Diagnosis Based On Domain Adaptation Under Different Working Conditions

Most traditional gearbox fault diagnosis methods are based on the data driven model,which takes the training data and the test data to satisfy the probability and distribution hypothesis.While the complexity of gearbox operation conditions makes the monitoring data distribution change,which leads to the adaptability of the diagnosis model greatly reduced.Domain adaptation learning is essentially a technology that uses the source domain training data to help the target domain to construct a model.It can eliminate the different effects of different data distribution in different fields and realize the modeling of data between different working conditions.Therefore,on the basis of domain adaptation learning theory,the study on the method for gearbox fault diagnosis is of theoretical significance and has practical value.Taking gearbox as object,this paper studies feature mining and fault diagnosis methods under different working conditions based on association rules and marginalized stacked denoising autoencoder.The main works are as follows:(1)Based on association rules,a method for mining gearbox vibration signal feature correlation relation under different working conditions is proposed.The feature data is mapped to different interval grades and symbolized by equal probability interval mapping.Then,Apriori algorithm which is modified by adding restriction conditions is used to mine the feature association rules.At last,the rules are used to draw a relationship map between working conditions and features to intuitively display the distribution of the features with different conditions and fault effects,and to analyze the relationship between features and different working and fault conditions.(2)In view of different distribution rules of data under different working conditions,an auxiliary model based domain adaptation method for fault diagnosis is proposed.By establishing a feature learning model of convolution neural network based marginalized stacked denoising autoencoder,the distribution difference of learned feature between source and target working conditions is reduced,while the fault categories can be distinguished at the same time.Finally,a classifier with stronger adaptability and higher accuracy is trained to solve the problem that the original diagnosis system is poor or even invalid due to the change of working conditions.(3)Study the application of historical data and analyze the stability and adaptability of the model.According to a certain proportion,the normal data and label data of target conditions are added to the training data to observe the data dependency of the model.A criterion is proposed for quantitative analysis of stability and adaptability of the model which is used for gearbox fault diagnosis under different working conditions.