Adaptive Manifold Partial Transfer Learning for Cross-Domain Fault Diagnosis
Zhengyi Wang, Yi Qin, Quan Qian · 2023
The domain adaptation method realizes the transfer learning task by reducing the domain discrepancy. However, it is more common that the label space of the two domains is inconsistent in the actual condition. This paper proposes a new adaptive manifold partial transfer learning (AMPTL) method, which realizes partial domain adaptation by reducing the distribution discrepancy between classes in two domains. Moreover, AMPTL proposes a new geometrical structure discovery method to enhance performance in domain alignment, which uses affinity matrix instead of adjacency matrix by manifold learning. Specifically, we use the similarity between samples instead of distance to discover the global geometrical structure of data sets in low dimensional manifold space. AMPTL applies the manifold learning method to the proposed partial domain adaptation method, aiming at adaptively capturing the geometrical structure of data sets. The manifold structure can change adaptively with the spatial mapping to avoid the negative transfer learning caused by the forced maintenance of sample distance. An effective iterative optimization algorithm is designed to optimize the proposed objective function under the condition of theoretical convergence. This method is applied to partial fault diagnosis of bearing datasets with unmarked target domain samples. The experimental results show that AMPTL has higher performance than other typical DA methods.