DAStatFormer achieves 99.4% accuracy with 24 statistical features for DAS pattern recognition

Distributed Acoustic Sensing (DAS) uses optical fibers for large-scale monitoring, but its high-dimensional spatio-temporal data makes event classification computationally expensive. Existing deep learning approaches—CNNs, recurrent models, and Transformers—either fail to capture long-range dependencies or require processing raw DAS matrices at prohibitive cost. To address this, researchers at IMT Nord Europe developed DAStatFormer, a hybrid multibranch architecture that replaces raw signals with 24 statistical attributes per channel, extracted from temporal, waveform, and spectral domains via ANOVA feature selection. This reduces data size by orders of magnitude while preserving discriminative information.

DAStatFormer processes each domain through dedicated step-wise and channel-wise attention branches, fused by an adaptive gating mechanism. Experiments on the open Φ-OTDR benchmark and a real-scenario DAS dataset show it achieves up to 99.4% accuracy and near-perfect real-world performance, using significantly fewer parameters and lower inference cost than models like DASFormer and DeepViT. The code is publicly released. This approach makes real-time, scalable DAS-based monitoring feasible for applications such as pipeline surveillance, seismic detection, and perimeter security.