High-fidelity spatiotemporal feature extraction for heterogeneous DAS signals

Abstract

Disturbance event recognition in Distributed Acoustic Sensing (DAS) systems faces significant multi-scale heterogeneous challenges. Traditional fixed-length sliding windows often lead to morphological truncation or high data redundancy, severely destroying macroscopic topological semantics. This paper proposes a high-fidelity feature extraction framework based on spatiotemporal joint optimization. Universally, multi-channel Pearson cross-correlation spatial fusion suppresses fading dead zones and background noise, while an adaptive hysteresis wrapping mechanism captures the complete lifecycle of heterogeneous events. To ensure fairness, a unified Time-Scale Normalization (TSN) protocol maps features into standard 2D receptive fields. Coupled with the spatial fusion, it effectively decouples the contradiction between scale compression and noise densification. Experimental results demonstrate that the framework overcomes semantic destruction from blind cutting, mitigating the directional misclassification of Clip as Knock from 90.9% down to 62.5%. Compared to traditional 1.0 s windows, the proposed method reduces the system front-end data representation payload by 59% while maintaining an excellent comprehensive recognition rate of 95.62%, offering a high-energy-efficiency edge computing paradigm for computationally constrained DAS terminals.