Abstract
Signal averaging imposes a significant trade-off between measurement time and measurand resolution in a sub-meter differential cross-spectrum Brillouin optical time domain reflectometry (DCS-BOTDR). This article introduces an optimized post-processing strategy that integrates wavelet denoising (WD) function with a traditional Lorentzian curve fitting (LCF) under relatively low signal averages to mitigate this problem. Symlet, Daubechies, Coiflet, and Biorthogonal Spline WD functions with a 4-level decomposition were executed on a six-core CPU utilizing a single-program-multiple-data (SPMD) paradigm and compared. Experimental validation was executed over distances of 350 meters and 1.21 kilometers of single-mode fiber. The experimental results demonstrate a 2.7 times reduction in required signal averages, in which the proposed LCF + WD method achieved a 2.7 MHz Brillouin frequency shift (BFS) resolution with only 21000 averages, a performance that required 56000 averages using the LCF method alone. To manage the computational load on the large experimental datasets, the algorithm was implemented on a parallel six-core architecture, accelerating the data processing speed by up to 4.8 times compared to serial computation. The method also successfully preserved a 0.4 m spatial resolution and improved temperature resolution to 3°C across a 1.21 km fiber at just 14000 signal averages. In comparison to other methods such as machine learning-based enhancements, the proposed strategy presents a more straightforward, training-free execution that attains comparable BFS and temperature resolutions without the necessity of extensive datasets or rigorous model training. Together with the multicore architecture, the proposed strategy is particularly beneficial for real-time distributed sensing applications where computational resources may be constrained.