An Extremely-Sparse Data based Demodulation Method of Microfiber Temperature and Salinity Sensors near the Dispersion Turning Point

With the development of optical fiber seawater temperature and salinity sensors, a high accurate and efficient demodulation method is in demand. In this paper, an artificial neural network assisted demodulation method, called as the WGAN-BRBP network, is proposed to interrogate the extremely-sparse sampled spectrum of micro-fiber salinity and temperature sensors operating near dispersion turning point. Experimental results show that, when the sampling interval is 13 nm, the accuracies of the demodulation results obtained by the Bayesian regularized back-propagation (BRBP) network are much higher than those obtained by the traditional cross-sensitive matrix method (CSMM) with a dense sampling spectrum. They are higher than those obtained by other machine learning methods, including the ordinary BP neural network. When the sampling interval reaches to 30 nm, the demodulation results obtained by the WGAN-BRBP network can maintain high accuracies, with a mean relative error of 0.766% for temperature and 0.883% for salinity. Furthermore, the acquisition time for each spectrum can be greatly reduced from 6780 ms to 5.1 ms. The accurate demodulation of the extremely sparse data can improve the demodulation efficiency of micro-fiber sensors, and provide a feasible method for their real-time applications.