Denoising of Brillouin Gain Spectrum Images for Improved Dynamic Measurements of BOTDR

Abstract

Brillouin optical time-domain reflectometry (BOTDR) is widely used for strain and temperature measurements in various fields. However, the accuracy and reliability of the measurements are often limited by the noise in the sensor signals. Dynamic measurement of BOTDR requires small averaging number and fast measurement, and hence noise reduction is more significant in dynamic measurement. Small gain stimulated Brillouin scattering (SBS) can enhance the Brillouin signal power in BOTDR to realize dynamic measurement, but noise reduction is still important in system. In this work, we investigate the denoising of Brillouin gain spectrum (BGS) images using convolutional neural networks (DnCNN) to improve the accuracy of the small gain SBS STFT-BOTDR measurement of strain vibration. It is shown that the denoising of BGS images along the time axis can result in better detection of the strain vibration compared with denoising of BGS images along the fiber length. The denoising performance was evaluated using frequency uncertainties and R-squared values. The best denoising performance was achieved with a DnCNN network with 8 layers and 200 epochs, leading to a frequency uncertainty of 2.32 MHz and an R-squared value of 0.907. The frequency uncertainty is improved to about 45% of the original value.

Keywords

• Brillouin optical time-domain reflectometry
• deep learning
• distributed fiber optic sensors
• image denoising