计算机科学
人工智能
语音识别
模式识别(心理学)
生成语法
计算机视觉
特征提取
生成模型
数据建模
光学成像
分布式声传感
光纤
特征学习
图像分割
信号处理
人工神经网络
训练集
分割
作者
Yikai Wu,Rui Cheng,Xu Zhou,Yuhao Qian,Liang Wang,Ming Tang
标识
DOI:10.1109/jiot.2026.3730549
摘要
Distributed optical fiber acoustic sensing (DAS) is a promising technique for intrusion monitoring, but its real-world deployment faces challenges of domain shift and scarcity of both labeled and unlabeled data. To address these challenges, we propose a contrastive learning with generative data augmentation (CLGA) framework for cross-domain few-shot event recognition. The framework performs supervised pre-training of a ResNet-18 on the source domain to provide effective initialization. A multiscale self-supervised generative adversarial network (MSS-GAN) is employed to synthesize high-quality samples, reducing the demand for unlabeled data. And a self-supervised contrastive learning module is then introduced to achieve domain adaptation from the source to target domains. Finally, a few-shot fine-tuning strategy with only a few labeled samples enables rapid adjustment to specific tasks. This integrated framework enables effective transfer from the source domain ImageNet to different DAS target domains when using limited unlabeled and labeled samples. Experimental results show that the framework achieves 98.56% and 98.44% accuracy on laboratory and field datasets with 100% unlabeled data and 50 labeled samples per class. Even with only 10% unlabeled data and 10 labeled samples per class, it also achieves over 90% accuracy in both scenarios. Our model also supports open‑set recognition by integrating temperature‑scaled and uncertainty quantification, and it achieves 93.79% and 91.17% overall accuracy on laboratory and field datasets, showing high robustness in recognizing mixed known and unknown events. The CLGA framework shows excellent generalization capability and cross-domain recognition capability, enabling cross‑domain few‑shot recognition in DAS.
科研通智能强力驱动
Strongly Powered by AbleSci AI