Spatial-temporal feature fusion-based Brillouin frequency shift extraction method

Brillouin optical time-domain analysis (BOTDA) enables high-precision distributed sensing by extracting the Brillouin frequency shift (BFS). However, the BFS extraction based on traditional Lorentz curve fitting (LCF) suffers from limited accuracy and low efficiency. This work proposes a spatial-temporal feature fusion method integrating pseudo-video denoising (PVD) network and deep learning to improve the BFS extraction in BOTDA. In our proposal, a PVD network was introduced to improve the signal-to-noise ratio of temporal data by simulating video processing, utilizing multi-frame time redundancy to suppress noise while preserving Lorentz line-shape features. Furthermore, a spatial-temporal feature fusion network combining convolutional neural network (CNN), long short-term memory (LSTM), and multi-head attention (MHA) was constructed. After extracting local spectral features through CNN, LSTM was used to capture the long-term temporal dependence of Brillouin gain spectrum (BGS) signals along the fiber axis, and MHA was introduced to dynamically weight critical frequency points to suppress noise interference. Experiments on a BOTDA system with a sensing fiber of 1.02 km single-mode fibers (SMFs) demonstrated that the proposed method achieved a root mean square error (RMSE) of 0.333 MHz for the BFS extraction, benefiting a higher accuracy by 3.5% as well as 18 times faster processing speed with respect to that with the LCF method. This work provides a high-precision and time-efficient BFS extraction solution for distributed Brillouin sensing, especially for applications under complex noise conditions, exhibiting potentials in structural health monitoring for critical infrastructure such as energy pipeline monitoring, high voltage cable monitoring, and so on.