High-fidelity acoustic signal enhancement for phase-OTDR using supervised learning

Phase-measuring phase-sensitive optical time-domain refiectometry (OTDR) has been widely used for the distributed acoustic sensing. However, the demodulated phase signals are generally noisy due to the laser frequency drift, laser phase noise, and interference fading. These issues are usually addressed individually. In this paper, we propose to address them simultaneously using supervised learning. We first use numerical simulations to generate the corresponding noisy differential phase signals for the given acoustic signals. Then we use the generated acoustic signals and noises together with some real noise data to train an end-to-end convolutional neutral network (CNN) for the acoustic signal enhancement. Three experiments are conduct to evaluate the performance of the proposed signal enhancement method. After enhancement, the average signal-to-noise ratio (SNR) of the recovered PZT vibration signals is improved from 13.4 dB to 42.8 dB, while the average scale-invariant signal-to-distortion ratio (SI-SDR) of the recovered speech signals is improved by 7.7 dB. The results show that, the proposed method can well suppress the noise and signal distortion caused by the laser frequency drift, laser phase noise, and interference fading, while recover the acoustic signals with high fidelity. (C) 2021 Optical Society of America under the terms of the OSA Open Access Publishing Agreement

Automated summary

This paper proposes using supervised learning to simultaneously address noise issues caused by laser frequency drift, laser phase noise, and interference fading, by training an end-to-end convolutional neural network to enhance acoustic signals. After enhancement, the average signal-to-noise ratio (SNR) of the recovered PZT vibration signals increased from 13.4 dB to 42.8 dB, and the average scale-invariant signal-to-distortion ratio (SI-SDR) of the recovered speech signals improved by 7.7 dB.