Torsional sensing dataset based on multimode fiber scatterplot and its deep learning demodulation approach

Abstract The dataset comprises speckle pattern images collected from a multimode fiber (MMF) subjected to varying torsional angles. These experimental data were specifically designed for training deep neural networks aimed at developing fiber-optic torsion sensors, enabling precise recognition of speckle pattern variations corresponding to applied angular deformations. The dataset includes a total of 7110 images acquired over a torsion range from 0° to 180°, with an angular increment of 2°. Key experimental parameters are as follows: a laser source with a wavelength of 632 nm, a 1 m-long MMF (core diameter of 105 μ m and numerical aperture of 0.22), and a CCD imaging system with a spatial resolution of 1280 × 1024 pixels. This dataset establishes a benchmark for deep learning-based fiber-optic torsion sensing research, supporting both neural network training and validation, and facilitating comparative studies with other fiber-optic sensing datasets.