Bending recognition based on learning enhanced fiber specklegram sensor

Abstract

In this paper, we propose and demonstrate a deep learning-enhanced fiber specklegram sensor for bending recognition. A segment of multimode fiber is used to sense bending, and tiny bending changes lead to significant variations in the speckle pattern generated at the distal end of the fiber. Principal component analysis (PCA) is utilized to optimize the collected samples and remove noise and redundant information by mining internal features, which makes the mapping relationship between the speckle image and corresponding curvature clearer and is conducive to reducing computational complexity. Back propagation neural network is employed to learn the mapping relationship between speckle image and curvature on the optimized dataset according to the optimized direction provided by PCA. The testing results show that the prediction error of the trained model for the learned bending state is 5.9 × 10−4 m−1, and the prediction speed is 0.05 ms per frame. The proposed scheme has a strong generalization ability and can be applied to predict bending states that have never been learned or seen with a prediction error of 3.8 × 10−2 m−1, which cannot be realized by the previously reported fiber specklegram sensor based on the classification neural network. The bending recognition scheme enhanced by deep learning proposed in this paper provides an enlightening reference for solving fiber sensing problems with deep learning methods, and it has the potential to be applied in more fields as a general scheme.