Rapid few-shot tabular machine learning for Φ-OTDR event classification

Distributed optical fiber sensing systems based on phase-sensitive optical time-domain reflectometry (Φ-OTDR) play a crucial role in infrastructure monitoring. However, event classification in these systems is hindered by limited data availability, high computational complexity, and slow model training. In this paper, we present an event classification method that integrates tabular machine learning with few-shot learning (FSL) through a pre-trained tabular prior-data fitted network (TabPFN) model. This method transforms the comprehensive set of time-domain statistical features (e.g., mean, peak-to-peak) and frequency-domain energy distribution features extracted from a conventional Φ-OTDR into structured tabular data, enabling subsequent machine learning processing. Specifically, the proposed scheme achieves a five-fold cross-validation accuracy of 90.0% ± 6.24% and an independent test accuracy of 89.0% under extremely small sample conditions and with only 10 samples per class available. Furthermore, with an increasing sample size, the performance further improves, reaching 97.78% ± 1.88% (at 60 samples per class) in cross-validation and 100% (at 50 samples per class) in independent testing, while fitting consistently completes within seconds. Comparative analysis against both strong traditional machine learning models and representative meta-learning FSL baselines further validates the superiority of our approach in few-shot settings. These impressive results and low computational overhead during application pave the way for rapid industrial deployment of Φ-OTDR systems in applications such as oil and gas pipeline monitoring.