A Two-Order Transfer Model for Gearbox Fault Diagnosis

摘要

To enhance gearbox fault diagnosis performance under varying working conditions, a two-order transfer model based on manifold regularization (MR) projection/maximum variance (MV) projection and domain selection machine (DSM) is proposed. In the first order transfer learning (TL), each source domain is mapped with MR and the target domain is mapped with MV from high dimensional spaces to the low dimensional space. Meanwhile, the minimum mean difference (MMD) is used to minimize the difference between the two domains in the low dimensional space. In the second order, the DSM model is utilized to select high-quality source domains by designing the domain selection vector in new space. Experimental results using the SpectraQuest's Drivetrain Dynamics Simulator show that the proposed method has better gearbox diagnosis accuracy under varying working conditions than each single transfer learning model.