Fast Brillouin Frequency Shift Retrieval by Sparse Frequency Enhanced Neural Network

We propose a novel artificial neural network (ANN) based fitting method to extract Brillouin frequency shift (BFS) from frequency-sparsely sampled Brillouin gain spectrum (BGS) in Brillouin Optical Time Domain Sensor (BOTDS). In the proposed method, only a few time-traces located at different scattering frequencies need to be collected. The sparse frequency sampling strategy, which means using large frequency scanning step, is able to improve the measurement speed. Then an advanced Sparse Frequency Brillouin Spectrum Enhanced Fitting Network (SF-BSEFN) is designed to directly retrieve BFS from low-resolution BGS (BGS $_{\mathrm {LR}}$ ). The proposed network structure possesses a data enhancement layer, which contains the information of high-resolution BGS (BGS $_{\mathrm {HR}}$ ), thus compensating for the degradation of the fitting accuracy. In a proof-of-concept experiment, we use a Brillouin optical time domain reflectometry (BOTDR) sensor to measure the BGSLR of a 3 km sensing fiber. and the acquisition time for BGSLR is 1.7 s. By using SF-BSEFN, the temperature extraction time for the entire BGSLR matrix with $7\times 6001$ data points is 0.028 s. When the BFS is extracted from BGSLR, the fitting accuracy obtained by using the proposed SF-BSEFN is 1.6 times that by using the conventional ANN method. When the averaging times of BGS exceeds 4000, the temperature measurement uncertainty obtained by using the proposed SF-BSEFN is 0.16 °C, which is almost the same as that obtained by using conventional ANN for BGSHR with $151\times 6001$ data points. However, the measurement speed is 20 times faster than that of the conventional frequency scanning-based BOTDR sensors.