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

Area of Science:

• Sensor Technology
• Machine Learning
• Data Science

Background:

• Distributed optical fiber sensing using phase-sensitive optical time-domain reflectometry (Φ-OTDR) is vital for infrastructure monitoring.
• Challenges include limited data, high computational costs, and slow training for event classification.

Purpose of the Study:

• To develop an efficient event classification method for Φ-OTDR systems using limited data.
• To integrate tabular machine learning with few-shot learning (FSL) for improved performance and speed.

Main Methods:

• Extracted time-domain and frequency-domain features from Φ-OTDR data, converting them into structured tabular data.
• Employed a pre-trained tabular prior-data fitted network (TabPFN) model for few-shot learning.
• Evaluated performance using five-fold cross-validation and independent testing with varying sample sizes.

Main Results:

• Achieved 90.0% cross-validation accuracy and 89.0% test accuracy with only 10 samples per class.
• Performance improved to 97.78% (CV) and 100% (test) with more samples (50-60 per class).
• Model fitting completed within seconds, demonstrating low computational overhead.

Conclusions:

• The proposed FSL method significantly enhances event classification accuracy in Φ-OTDR systems, especially with limited data.
• The approach offers superior performance compared to traditional ML and other FSL baselines in few-shot scenarios.
• The method's efficiency and accuracy support rapid industrial deployment for applications like pipeline monitoring.