Physics-guided deep learning for vector vibration sensing with multicore optical fiber Bragg gratings

To improve the accuracy of vector vibration sensing based on multicore fiber Bragg gratings (MCFBGs), we propose a physics-guided deep learning framework for high-precision reconstruction of vibration direction vectors, alleviating the degradation caused by the ideal assumptions in conventional physical models. The framework establishes an end-to-end mapping between multichannel wavelength responses and vibration vectors, incorporating inter-core differential feature augmentation and a staged transfer learning strategy. These innovations significantly reduce the reliance on large-scale experimentally labeled datasets. Experimental results demonstrate that the proposed method achieves a mean absolute error (MAE) of 1.48° and a relative error as low as 0.82%, representing an 83.7% improvement over conventional geometric-model-based approaches. The method maintains robust reconstruction accuracy across varying vibration conditions. Furthermore, our approach demonstrates significant potential for complex applications, including three-dimensional shape sensing and intelligent sensing.