A deep learning algorithm ADPNet for strain and temperature decoupling of fiber bragg gratings

Abstract

Fiber grating sensor signals can be affected by both strain and temperature, and decoupling strain and temperature for fiber optic sensing is a challenging task. We propose a deep learning-based spectral segmentation model ADPNet (Adaptive Dilated Pyramid Net) to distinguish strain and temperature variations by extracting fiber Bragg wavelength features. To train and validate our model, we built an FBG demodulation system and collected spectral data under different strain and temperature conditions. The experimental results show that the root mean square error (RMSE) of ADPNet is 141.755 με and the decoupling time is 0.54 ms. ADPNet has a significant advantage in RMSE over compensation and machine learning regression methods on the same data set, demonstrating the great potential of deep learning algorithms for the spectral decoupling problem.