Research Progress on Distributed Fiber Optic Event Recognition Methods Under Data Constraints
1005-8788
Abstract
Due to the dynamic monitoring capability, Distributed Acoustic Sensing (DAS) has been widely deployed in various industrial scenarios. Deep learning greatly boosts the event detection accuracy of fiber sensing systems. Nevertheless, practical engineering suffers from limited labeled samples and imbalanced class distribution, which severely degrades the generalization of deep neural networks. This paper systematically reviews state-of-the-art distributed fiber event recognition algorithms under data scarcity. Two mainstream technical frameworks, i.e., data augmentation and transfer learning, are elaborated in detail. We analyze the core mechanisms, merits and technical bottlenecks of each method, and prospect future research trends, offering comprehensive references for follow-up investigations.