Answers are grounded in this paper and cite their sources.
Distributed acoustic sensing (DAS) vibration event recognition often suffers from limited labeled data and severe class imbalance, which significantly degrades model generalization. To address these challenges, this paper proposes an adaptive semi-supervised learning framework based on dynamic pseudo-labeling. Built upon a strong/weak augmentation consistency paradigm, a confidence-aware adaptive thresholding mechanism is introduced to dynamically adjust pseudo-label selection according to the model training state, thereby balancing sample utilization and label reliability. Furthermore, a prediction-consistency-based sample selection strategy is developed, where a Count-Gap indicator is employed to identify low-confidence but stable samples, enabling the mining of potentially informative data. A generalized cross-entropy loss is adopted to improve robustness against noisy pseudo labels, while focal loss is incorporated in the supervised branch to mitigate class imbalance. Experimental results on the self-collected XRDAS dataset demonstrate that the proposed method improves recognition performance under low-label conditions, achieving an accuracy increase from 20.12±0.11% to 39.63±0.11% with only 10% labeled data, and reaching 70.30±0.12% at a 30% labeled ratio. Experiments on a large-scale public DAS event-classification dataset further evaluate the method on a second sensing system with pronounced class imbalance.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: