数据领域选择与空间迁移在齿轮箱故障诊断中的应用

To improve the performance of gearbox fault diagnosis, a space transfer strategy is proposed. Here the source domains are composed by multiple auxiliary channels and the target domain is composed by single target channel. With transfer learning (TL), the fault diagnosis models in the former can be applied in the latter to overcome the problem of lacking target data. Firstly, the domains are selected according to the band selective independent component analysis (BS-ICA) rule and the original five-dimension spaces are constructed by extracting their time-domain features. Secondly, the source and target domains are mapped to a public two-dimensional space using the equilibrium density projection (EDP). Meanwhile, the minimum mean difference strategy is used to minimize the difference between two projection spaces. Finally, the logistic regression (LR) and support vector machine (SVM) classifiers are both carried out for sample classification. Also, the diagnostic model can be updated by removing low-quality samples while adding high-quality samples in source domains. Based on the Spectra Quest's gear drive system, the performance between proposed method and classical transfer strategies including transfer composition analysis (TCA) and domain selection machine (DSM) are compared, which indicates that the former has higher diagnostic accuracy as well as faster running speed when facing with rapid change of working conditions, thus possessing high value in application of engineering.