Title: DAStatFormer: Integrating Statistical Features into a Hybrid Multibranch Transformer for Pattern Recognition in DAS

Abstract: While Distributed Acoustic Sensing (DAS) facilitates extensive monitoring via optical fibers, the classification of events is complicated by the technique's high dimensionality and intricate spatio-temporal structures. Current deep learning solutions, including CNNs, recurrent architectures, and various Transformer models, often struggle to capture long-range dependencies or impose prohibitive computational costs when processing raw DAS matrices. To address these challenges, we introduce DAStatFormer, a hybrid multibranch Transformer that integrates Gated Transformer Networks with compact multidomain statistical features. Rather than relying on raw signals, the model extracts 24 attributes per channel—selected via ANOVA—from temporal, waveform, and spectral domains. This approach significantly reduces data volume by several orders of magnitude without sacrificing discriminative power. The architecture processes each domain through specialized step-wise and channel-wise attention branches, which are then combined using an adaptive gating mechanism. Evaluations on both the open $\Phi$-OTDR benchmark and a real-world DAS dataset reveal that DAStatFormer attains accuracy rates as high as 99.4%, delivering near-perfect performance in practical scenarios. Furthermore, it requires substantially fewer parameters and incurs lower inference costs compared to existing models like DASFormer and DeepViT. These findings highlight the model's potential for scalable, real-time DAS monitoring. The source code is available at https://github.com/MichelD-git/DAStatFormer

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC