A Novel Few-Shot Deep Transfer Learning Method for Anomaly Detection: Deep Domain-Adversarial Contrastive Network With Time-Frequency Transferability Analytics

This paper tries to solve the challenges in few-shot transfer learning for anomaly detection: how to guarantee the transfer effect on insufficient even limited source domain data, and how to make the transfer process interpretable for getting trustworthy results. This paper proposes a deep domain-adversarial contrastive model with time-frequency transferability analytics. The essential idea is extracting fine-grained information from different frequency bands for reliable transfer. First, a time-frequency domain feature pool is constructed by applying wavelet scattering network (WSN) under different decomposition scales and rotation orientations. An orientation-first selection strategy is further designed to determine the optimal features that can cover the low, medium and high frequency bands. A new transferability metric, named frequency importance metric (FIM), is then built through frequency hypersphere matching to quantify the significance of each frequency band from source domain data. Second, a deep domain-adversarial contrastive network (DDCN) is constructed to realize selective information transfer according to frequency band's significance. In DDCN, a purposeful feature representation can be extracted through the contrastive learning between the deep features and wavelet features that are weighted by FIM, thus leading to valid transfer in few-shot scenario via domain-adversarial training. Experiments are conducted on two typical anomaly detection problems, i.e., image recognition detection on the MNIST USPS and Office-Home datasets, and early fault detection on the IEEE PHM Challenge 2012 bearing dataset. The results not only verify the superior performance of the proposed method to the few-shot transfer learning, but also reveal the frequency saliency in the transfer process.