Label-efficient event recognition for Φ-OTDR distributed acoustic sensing using self-supervised spatiotemporal...

Area of Science:

• Geophysics and Sensor Technology
• Signal Processing and Machine Learning

Background:

• Distributed Acoustic Sensing (DAS) using phase-sensitive optical time-domain reflectometry (Φ-OTDR) is crucial for long-range vibration monitoring.
• Challenges in DAS event recognition include ultra-long temporal sequences, environmental interference, and scarcity of labeled data.

Purpose of the Study:

• To propose SS-BiMamba, a label-efficient framework for Φ-OTDR event recognition.
• To address limitations in current DAS monitoring systems by improving event recognition accuracy with limited labeled data.

Main Methods:

• Developed a label-efficient framework integrating self-supervised spatiotemporal pre-training, few-shot adaptation, and confidence-guided pseudo-label refinement.
• Employed a shared spatiotemporal encoder with hierarchical spatial convolution, a bidirectional Mamba2 temporal module, and gated fusion.
• The encoder captures cross-channel correlations and long-range temporal dependencies directly from raw Φ-OTDR signals.

Main Results:

• Achieved high accuracies: 94.90% (1-shot), 98.10% (4-shot), 98.63% (8-shot), and 99.37% (16-shot) on a public Φ-OTDR dataset.
• In the 4-shot setting, SS-BiMamba demonstrated competitive performance against fully supervised methods using significantly less labeled data.
• In the 16-shot setting, performance became comparable to established fully supervised methods.

Conclusions:

• SS-BiMamba offers an effective solution for label-efficient event recognition in practical distributed acoustic sensing scenarios.
• The framework demonstrates the potential of integrating self-supervised learning and advanced temporal modules for robust vibration monitoring.