Cosine-Initialized MAE for Cross-Domain Few-Shot Recognition in Distributed Fiber-Optic Vibration Sensing Systems

Existing deep learning models often underperform in cross-domain few-shot tasks for distributed vibration sensing (DVS), largely due to domain shifts introduced by variations in device types and deployment scenarios. To address this challenge, we introduce the cross-domain few-shot learning (CDFSL) paradigm to the field of distributed fiber-optic vibration signal recognition for the first time, and propose a novel framework, Cosine-Initialized MAE. This approach begins with self-supervised pretraining on large-scale unlabeled source-domain data using a masked autoencoder (MAE), enabling the extraction of transferable, general-purpose feature representations. Subsequently, the model is efficiently adapted to the target domain using only a few labeled samples, via a single-step fine-tuning procedure that incorporates Weight Imprinting and Cosine Similarity Classification. Departing from conventional episodic evaluation, our method employs a fixed evaluation protocol: the model is fine-tuned once on a k-shot support set (k∈{1,5,10}) and then evaluated on the remaining (500−k) samples per class. The proposed framework achieves accuracies of 78.92%, 87.84%, and 91.80% on 5-way 1-shot, 5-shot, and 10-shot tasks, respectively—substantially outperforming established self-supervised and few-shot learning baselines. These results highlight the framework’s strong generalization under limited supervision and underscore its potential for low-resource cross-domain recognition in DVS systems.