Differentiable Neural Architecture Search for Domain Adaptation in Fault Diagnosis

Abstract

In recent years, deep transfer learning has been widely applied in fault diagnosis. In particular, domain adaptation method to solve unsupervised transfer learning problem has become a research hotspot. However, these researches mainly focus on the construction of loss function, and the design of network structure still depends on manual work. A differentiable neural architecture search method is proposed to automatically search for network structures suitable for domain adaptation in fault diagnosis to solve this problem. The proposed method search for two components in the Inception search space: structures of Inception blocks and feature mask. Inception blocks are stacked to form a complete network, and the feature mask selects the features for domain adaptation. Besides, to improve the search efficiency, we search and adapt multiple target domains at the same time, expanding the application scenarios of domain adaptation. The proposed method is evaluated on two datasets, and experiments show that this method can generate networks with excellent performance.

Keywords: Fault diagnosis, domain adaptation, convolution neural network, differentiable neural architecture search