Improved Spectral Interrogation of Tilted Fiber Bragg Grating Refractometer Using Residual Convolutional Neural Networks

Abstract

This paper proposes a new approach of demodulating the spectra of tilted fiber Bragg grating (TFBG) based refractive-index (RI) sensor by employing convolutional neural networks (CNN) with residual blocks. The proposed residual CNN model was trained and tested by over 5500 samples recorded at different concentrations of sodium chloride solution, whose RI ranged from 1.3350 to 1.3790 RIU, and the raw spectral data can be directly utilized to train the residual CNN without any pre-processing. Results show that the coefficient of determination could reach 99.82% and the mean square error (MSE) achieves 2.818 × 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−7</sup> RIU. The well-trained model could be applied to demodulate other similar sensors through transfer learning, which demonstrates its good generalization capability of the proposed method. Additionally, the proposed method could maintain high MSE (i.e., 2.938 × 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−7</sup> RIU) although the interrogator with poorer resolution (20 times less) was utilized, showing high potential to reduce the cost of the interrogation hardware.