TDDA-Net: A transitive distant domain adaptation network for industrial sample enhancement

Abstract

The small sample problem is an open challenge in data-driven industrial fault diagnosis, which leads to low accuracy and weak generalization for modeling. Domain adaptation (DA) attempts to transfer samples of the source domain to the target domain for small sample enhancement. However, existing approaches cannot be directly applied to distant domains, due to insufficient exploration of domains with large discrepancies and negative transfer. To tackle the above issues, this paper proposes a transitive distant domain adaptation network (TDDA-Net), in which the sample features are decomposed into three orthogonalized dimensions to reliably express the sample information. Then, distant domain samples are explored and negative transfer is alleviated in different feature dimensions. In particular, distant domain samples are transitively explored to obtain abundant types of samples, wherein useful information from intermediate domains can be employed. Thus, the modeling generalization is improved. Additionally, the marginal and conditional distributions of the samples are adaptively matched to correct distribution drift, such that the negative transfer is alleviated. Thus, the modeling accuracy is improved. Benchmark simulated experiments and real-world application experiments are conducted to evaluate the proposed network. All the results demonstrate that our TDDA-Net performs favorably against the state-of-the-art methods.