Investigation on fiber torsion sensing mechanism based on multimode fiber speckle and deep learning

Abstract This study proposes a high-precision fiber speckle-based torsion sensor utilizing a multimode fiber structure and the ResNet34 with self-attention and multi-branch model. The sensor leverages deep learning techniques to establish a mapping between fiber speckle patterns and torsion angles while integrating residual networks, self-attention mechanisms, and multi-branch structures to enhance feature extraction and prediction accuracy. Experimental results demonstrate that the proposed sensor achieves 99.8% accuracy within ±1° and ±0.5° error ranges in the known-angle test set (test set I). In the unknown-angle test set (test set II), it attains an accuracy of 89.67% within a ±1° error range, exhibiting strong generalization capability. Compared to conventional fiber torsion sensing schemes, this approach offers advantages such as structural simplicity, low cost, and a wide measurement range. It is well-suited for structural health monitoring, robotic arm control, and other engineering applications, providing a novel solution for high-precision and cost-effective torsion measurements.