Research on bending sensing based on optical fiber mode decomposition

A study on bending sensing based on optical fiber mode decomposition is proposed in this paper. First, a novel mode decomposition method for few-mode fibers (FMF) is introduced, which combines the ResNet50 deep learning network with an attention mechanism. ResNet50 is used to extract the rough mode amplitudes and relative phases, while the attention mechanism enhances the important regions during the feature extraction process, thus improving the accuracy of mode decomposition. Compared to the current deep learning-based optical fiber mode decomposition methods, which typically require datasets of over 100,000 near-field intensity images, the method proposed in this paper significantly reduces the training datasets size. With only 20,000 near-field intensity images in the training set, mode decomposition accuracy exceeding 0.99 is achieved for 3-mode, 5-mode, and 6-mode cases.Based on this, the bending characteristics of different modes obtained from the decomposition under different fiber curvatures are analyzed. The results show that as the fiber bending curvature changes, higher-order modes in FMF are more sensitive to bending. As the bending curvature changes from 0.07 to 0.49, the discrepancy of the LP11 mode with the unbent fiber increases more significantly compared to the LP01 mode, rising from 0.05 to 0.72. In contrast, the discrepancy of the LP01 mode increases from 0.12 to 0.61.This indicates that higher-order modes respond more significantly to bending.