Accurate Brillouin frequency shift determination in dual-pulse BOTDA via deep learning

In this work, we proposed and demonstrated a convolutional neural network (CNN) to overcome the detrimental impact of the distortion peaks in the measured Brillouin gain spectrum (BGS) from a dual-pulse Brillouin optical time-domain analyzer (DP-BOTDA) system, enabling precise extraction of the Brillouin frequency shift (BFS) and achieving a sub-meter spatial resolution (SR) by using two pulse sequences with a time interval shorter than the lifetime of phonon. A simulation model of DP-BOTDA was established to generate numerous training data under various conditions. Then, for proof of concept, the data acquired from a DP-BOTDA experimental setup were processed via the pre-trained CNN, and the results verify that the proposed CNN is able to achieve fast, high- precision and distributed BFS extraction along a 2.3 km sensing fiber, with a much smaller root mean square error (RMSE) of 1.08 MHz along the fiber compared with the 12.47 MHz RMSE by Lorentzian curve fitting. The proposed CNN paves the way for applying DP-BOTDA to structural health monitoring, offering a robust and reliable option for sub-meter-resolution Brillouin sensing.