Label-efficient event recognition for Φ-OTDR distributed acoustic sensing using self-supervised spatiotemporal pre-training and bidirectional Mamba2

This is the authors' abstract. We don't add key points for this paper.

Distributed acoustic sensing (DAS) based on phase-sensitive optical time-domain reflectometry (Φ-OTDR) enables long-range vibration monitoring, yet reliable event recognition remains challenging due to ultra-long temporal sequences, environmental interference, and limited labeled data. To address these issues, we propose SS-BiMamba, a label-efficient Φ-OTDR event recognition framework that integrates self-supervised spatiotemporal pre-training, few-shot adaptation, and confidence-guided pseudo-label refinement. The framework employs a shared spatiotemporal encoder consisting of a hierarchical spatial convolution branch, a bidirectional Mamba2 temporal module, and a gated fusion mechanism to capture cross-channel correlations and long-range temporal dependencies directly from raw signals. Experiments on a public Φ-OTDR dataset show that SS-BiMamba achieves accuracies of 94.90%, 98.10%, 98.63%, and 99.37% under 1-shot, 4-shot, 8-shot, and 16-shot settings, respectively. Notably, in the practical 4-shot setting, the proposed method achieves competitive performance compared with several fully supervised baselines while using only a small fraction of labeled data, and in the 16-shot setting, its performance becomes comparable to several fully supervised methods. These results demonstrate that SS-BiMamba provides an effective solution for label-efficient event recognition in practical distributed acoustic sensing monitoring scenarios.