S2-DyGNN: A Spectro-Spatial Dynamic Graph Neural Network for Acoustic Event Classification in Distributed Acoustic...

Area of Science:

• Geophysics
• Signal Processing
• Machine Learning

Background:

• Distributed acoustic sensing (DAS) captures complex wave propagation.
• Conventional models struggle with dynamic spatiotemporal interactions in DAS data.
• Existing architectures are limited by static topologies and isolated spatial grids.

Purpose of the Study:

• To introduce a novel architecture for enhanced event classification in DAS systems.
• To address limitations of conventional models in adapting to dynamic feature relationships.
• To improve the modeling of complex spatiotemporal interactions across sensor arrays.

Main Methods:

• Developed the Spectro-Spatial Dynamic Graph Neural Network (S2-DyGNN).
• Employed a 2D frequency-time convolutional front-end for spectro-temporal feature extraction.
• Utilized a dual-matrix graph neural network (GNN) with dynamic spatial topology recalculation.

Main Results:

• S2-DyGNN achieved a peak macro-averaged F1-score of 86.6% and 94.0% overall accuracy on a skewed nine-class DAS dataset.
• The dual-matrix topology improved the F1-score for the 'openclose' class to 55.7% from 48.0%.
• Outperformed conventional models by effectively handling sparse transient events against background noise.

Conclusions:

• Explicitly coupling localized spectro-temporal features with dynamic spatial topologies enhances DAS event classification.
• S2-DyGNN offers a robust and scalable solution for real-world continuous monitoring.
• The dynamic graph approach overcomes limitations of static models for complex acoustic event detection.