Feature-Preserving Learnable Downsampling of DAS Signals Using Residual Neural Networks

Distributed acoustic sensing (DAS) based on phase-sensitive optical time-domain reflectometry (Φ-OTDR) continuously generates large volumes of high-dimensional time-series data, which poses challenges to data storage, transmission, and real-time processing. To address this issue, a residual neural network downsampling scheme is proposed. The network automatically learns the intrinsic features of the beat signals and preserves diagnostically relevant vibration information under aggressive dimensionality reduction. Experiments on a heterodyne Φ-OTDR system show that, compared with the traditional downsampling scheme, the proposed scheme achieves a higher signal-to-noise ratio (SNR) and better waveform fidelity at the same compression ratio. When combined with μ-law quantization, an overall compression ratio of up to 32 is achieved while maintaining acceptable signal quality, which significantly improves the efficiency of data storage and transmission.