Fiber specklegram torsion sensor based on residual network.

• High accuracy and good generalization ability. • Easy saturation and unable to perceive continuous variation of parameters are solved. • The relationship between the specklegram and the accuracy is analyzed. • Analytical process is conductive to the selection of dataset. • Transfer learning is used to tackle the poor portability and reusability. A novel fiber specklegram torsion sensor based on multimode fiber (MMF) is proposed and experimentally demonstrated in this paper. The Residual Network (ResNet) is used to establish the relationship between the changing process of the specklegram and the torsion angle of the MMF. The experimental results indicate that the sensor exhibits the advantages of high accuracy and large measurement range. Within a torsion angle range of 0–360°, the prediction error is within ±2° for 99.1% of the specklegrams in test set. By analyzing the decorrelation process of the specklegrams from different kinds of MMFs, we find that the difference in torsion angle prediction accuracy of different MMFs stems from the decorrelation angle of the specklegram, and the larger decorrelation angle of specklegram holds higher torsion angle sensing accuracy of the trained ResNet. Meanwhile, the difference of the decorrelation angle of specklegram is attributed to the difference of the fiber normalized frequency. This analytical process may be applicable to the analysis of other specklegram based sensing schemes and the selection of dataset for fiber specklegram sensor. The possibility of transfer learning is also investigated to show an improved portability and reusability. In addition, the fiber specklegram torsion sensor owns other merits such as easy fabrication and good robustness, which shows its wide range of potential applications. [ABSTRACT FROM AUTHOR]