Deep learning to overcome secondary interaction artifacts in Brillouin optical frequency-domain analysis

Secondary interaction artifacts (SIA) in Brillouin optical frequency-domain analysis (BOFDA) sensors introduce severe measurement errors and significantly degrade the effective spatial resolution. SIA originates from the stimulated Brillouin scattering (SBS) between the pump and the excited acoustic sidebands, leading to pronounced distortions in the Brillouin gain spectrum (BGS). Conventional methods to overcome SIA rely either on complex signal processing algorithms or additional hardware components, which inevitably result in increased computational complexity or system cost. Here, we propose a deep learning method based on deep autoencoder (DAE), to adaptively correct BGS distortion without any system modification. We construct a diversified dataset for DAE training based on the BGS analysis model influenced under SIA. Both numerical and experimental results demonstrate that the DAE effectively learns the nonlinear mapping between distorted BGS and the ideal Lorentzian BGS, enabling highly accurate demodulation of the Brillouin frequency shift (BFS). Under strong SIA conditions, the root mean square error (RMSE) and standard deviation (SD) of the BFS based on DAE demodulation improved by approximately 77.2% and 82.3%, respectively, compared with the conventional Lorentzian curve fitting (LCF) method. Moreover, the DAE effectively suppresses the broadening of the BFS step response induced by SIA, allowing the system to maintain the theoretical spatial resolution under varying strain conditions, thereby significantly enhancing the measurement accuracy and stability of BOFDA sensors.