Hybrid convolutional neural network optical fiber vector curvature sensing via spectral pattern recognition

Fiber curvature sensing has attracted increasing interest due to its high sensitivity, structural versatility, and real-time response. In this work, we propose an intelligent vector curvature sensing system based on a composite optical fiber structure that combines a multimode–OHTC Mach–Zehnder interferometer (MZI) with an etched triple-core fiber Bragg grating (TCF-FBG). Theoretical and experimental analyses verify that the hybrid MZI–TCF-FBG configuration enables high-sensitivity intensity-modulated vector curvature sensing, achieving a maximum sensitivity of 23.41 dB/m⁻¹. To further enhance sensing speed and avoid reliance on wavelength tracking, a spectral pattern–recognition framework is developed using a hybrid convolutional neural network (CNN) and multilayer perceptron (MLP). With unified data preprocessing and cross-validation, the optimized 2D CNN + MLP model yields a minimum mean-square error of 0.002 m⁻¹ and an R² value approaching 1, demonstrating excellent accuracy and generalization. The proposed method offers a promising route toward rapid, robust, and high-precision vector curvature sensing.