Deep-Learning-Assisted Vector Bending Sensor Based on Dual-FPI Structure

In this work, an omnidirectional bending sensor based on dual Fabry-Perot interferometers (DFPIs) is designed and experimentally demonstrated. By eccentrically embedding a thin-diameter fiber (TDF) between two segments of seven-core fiber (SCF), DFPIs are formed and their spectrums are strongly correlated with the bending direction and curvature. A multi-scale convolutional and transformer neural network (MSCTnet) based on a hybrid loss function is employed to model the relationship between the bending state and the spectrum of DFPIs. Compared with other combinations of sensor structures and deep learning methods, the proposed DFPIs integrated with MSCTnet achieves higher prediction accuracy for bending states while requiring fewer training iterations. Experimental results demonstrate that the DFPIs achieve a curvature prediction accuracy of 99.53% with a mean absolute error (MAE) of 0.085 m⁻¹. The azimuth angle prediction accuracy is 99.92%, with an MAE of 1.39°.