Imbalanced Partial Transfer Network for Intelligent Machine Fault Diagnosis

Transfer learning (TL) has garnered significant interest in mechanical fault diagnosis. Many current TL approaches typically assume that ample data are available and that both the source and target domains possess identical label spaces. However, these TL methods often fail to address real-world issues, particularly when the number of samples in different conditions is unequal (i.e., imbalance) and the target label space is a subset of the source label space [i.e., partial transfer learning (PTL)]. To address these issues, this study proposes the imbalanced partial transfer network (IPTN). The IPTN introduces a weighted maximum density divergence (MDD) loss and a discriminative sample generator (DSG). The DSG identifies distinctive samples in the target domain and expands the dataset by augmenting these distinctive samples to solve the sample imbalance problem. Meanwhile, the new loss function termed weighted MDD promotes the ability of PTL by increasing interclass distance and intraclass density. Experiments on two datasets demonstrate the superior diagnostic performance of the IPTN compared to several comparison methods, highlighting its powerful transfer capability in situations involving sample imbalance and PTL.