DAStatFormer: A Hybrid Multibranch Transformer with Statistical Feature Integration for DAS-Based Pattern Recognitions

Researchers introduce DAStatFormer, a hybrid Transformer model that dramatically improves Distributed Acoustic Sensing (DAS) event classification by extracting 24 statistical features per channel instead of processing raw signals, achieving 99.4% accuracy on benchmark datasets while reducing computational requirements significantly compared to existing deep learning approaches.

DAStatFormer addresses a critical bottleneck in geophysical and structural monitoring applications where Distributed Acoustic Sensing generates massive high-dimensional datasets through fiber-optic networks. The innovation lies in feature engineering rather than brute-force model scaling—the researchers strategically extract temporal, waveform, and spectral domain features using ANOVA selection before feeding them into a gated Transformer architecture. This approach reduces data dimensionality by orders of magnitude while maintaining discriminative power for pattern recognition tasks.

The hybrid architecture combines step-wise and channel-wise attention branches with adaptive gating, enabling the model to capture both local and global spatio-temporal dependencies without the prohibitive computational costs associated with processing raw DAS matrices. Existing solutions like DASFormer and DeepViT struggle with this tradeoff, requiring substantial memory and compute resources during inference. DAStatFormer's reported 99.4% accuracy on the open Φ-OTDR benchmark and near-perfect performance on real-world datasets validates the feature-engineering strategy.

This work has implications for industrial monitoring infrastructure, seismic monitoring networks, and structural health surveillance—domains where real-time processing at scale remains computationally challenging. The efficiency gains make DAS-based monitoring economically viable for larger deployments and enable edge deployment scenarios previously impossible with resource-intensive deep learning baselines. The open-source release democratizes access to this methodology, potentially accelerating adoption across monitoring applications.

• →DAStatFormer achieves 99.4% accuracy on DAS pattern recognition benchmarks using statistical feature extraction instead of raw signal processing
• →The model requires significantly fewer parameters and lower inference costs than competing architectures like DASFormer and DeepViT
• →Strategic feature engineering using ANOVA-selected attributes reduces data dimensionality by orders of magnitude while preserving discriminative information
• →Hybrid multibranch Transformer design with adaptive gating enables efficient capture of spatio-temporal dependencies in distributed acoustic sensing
• →Real-world deployment feasibility improves substantially through computational efficiency gains, enabling scalable monitoring infrastructure applications