Enhanced Neural Network Implementation for Temperature Profile Extraction in Distributed Brillouin Scattering-based Sensors

In this work two Neural Network (NN) based solutions are proposed to recover the distributed temperature profile of a sensing fiber, measured using a commercial Brillouin Optical Time-Domain Analysis (BOTDA) interrogator. A detailed analysis in terms of temperature accuracy and processing speed is carried out for both the proposed methods, comparing the results with the ones obtained from the application of classical fitting techniques, namely cross-correlation (CORR), Lorentzian fitting (LF) and Pseudo-Voigt fitting (PV), through both simulations and real measurements carried out in laboratory environment. The results show that the first NN implementation, which aims to maximize the accuracy of the temperature profile and the processing speed, can handle different width of frequency acquisition window but needs to be optimized for a specific frequency acquisition scanning step. The second NN implementation, however, can also handle different values of the acquisition scanning step with a minor performance drop. Simulations and experimental data show a massive advantage of NN implementations in terms of processing speed with respect to classical fitting techniques, with a slightly better accuracy of the estimated temperature profiles.