As a non-contact and highly sensitive monitoring method, optical fiber curvature measurement is widely applied in structural health monitoring, environmental sensing, and other fields. However, traditional optical fiber speckle pattern analysis methods face issues of insufficient information extraction and low measurement accuracy when dealing with large curvatures or complex environments, which limits the widespread application of optical fiber sensing technology. Previous research has explored the use of single-mode to multimode fiber structures, as well as the optimization of image processing algorithms, but these methods have relatively high requirements for camera resolution and light sources. Deep learning algorithms can improve anti-interference capability and prediction accuracy when they are introduced into optical fiber sensing technology. We proposed a high-precision curvature measurement method based on the VggNet 16 model and specialty optical fiber speckle patterns. By designing a single-mode-multimode-specialty optical fiber structure and utilizing deep learning technology, the accuracy and anti-interference capability of optical fiber curvature measurement have been significantly improved.

First, a single-mode-multimode-specialty optical fiber structure is designed, in which the multimode fiber increases the mode field diameter, while the dual-core single-side hole fiber excites asymmetrically distributed modes, enhancing the information content of speckle patterns and improving curvature measurement accuracy. The single-mode fiber serves as the signal input channel, ensuring high fidelity of the signal source and avoiding intermodal dispersion. The multimode fiber supports multiple transmission modes, laying the foundation for complex interactions. The specialty fiber section promotes optical field redistribution and strong coupling between modes, enhancing the complexity of the light spot morphology and increasing the response sensitivity to stress and curvature changes. Then, an automated experimental platform is established. A device is constructed that automatically provides curvature variations, automatically changes the degree of fiber bending, and saves the speckle patterns along with curvature labels. The experimental platform consists of a coherent light source, fiber optic sensing structure, complementary metal-oxide-semiconductor (CMOS) camera, three-dimensional displacement stage, stepper motor, and host computer. A 632.8 nm laser from a He-Ne laser source is coupled into the single-mode fiber through a collimating mirror and then coupled into the multimode fiber through mode mismatch, exciting multiple conduction modes. After passing through the multimode-specialty fiber fusion point, the various modes are coupled into the asymmetric dual-core off-center hole fiber, where coupling and interference occur. The superimposed optical field is focused and magnified by a lens, and finally, the speckle pattern is captured by a CMOS camera. After collecting and processing a large dataset, we conduct tests using the VggNet 16 classification model. Analysis of prediction errors reveal that the model will misidentify speckle patterns with similar curvatures, rather than making chaotic or random misjudgments, indicating that the speckle patterns presented by this fiber structure have a correlation when curvatures are similar. Subsequently, the dimension of the final output layer of the VggNet 16 model is adjusted to train a regression model, and its performance and accuracy are compared with traditional algorithms.

The combination of specialty optical fiber and the VggNet 16 model achieves extremely high prediction accuracy on the test set, with 100.00% of samples having an error less than 0.1 m^-1, 97.03% of samples having an error less than 0.07 m^-1, and 96.04% of samples having an error less than 0.05 m^-1. The mean square error (MSE) of the prediction is 5.877×10^-4 m^-2,^{and the root mean square error (RMSE) is 2.424×10-2 m-1. The prediction results are then compared with typical structures and traditional algorithms. First, Fig. 14 reveals that traditional algorithms will make misjudgments in cases where timestamps are similar. Fig. 15 compares typical optical fiber structures with the structure designed in this paper, showing that the addition of specialty optical fiber improves this phenomenon, specifically by reducing the numerical differences in feature indicators at the same curvature. Fig. 18 compares the VggNet 16 model with the ResNet 50 model, and the results indicate that VggNet 16 performs better under certain error thresholds, especially when the error is less than 0.05 m-1 and 0.1 m-1, where VggNet 16 has higher accuracy. Therefore, the superiority of VggNet 16 in curvature sensing has been verified, and the advantages of the optical fiber structure have been further enhanced on the foundation of deep learning.}

The stress measurement sensing method for specialty optical fiber structures based on the VggNet 16 model proposed in this paper can improve the accuracy and stability of curvature measurement. By comparing the performance between different optical fiber structures, as well as various traditional algorithms and deep learning algorithms, this method has been verified to demonstrate superior measurement performance across different curvature ranges and achieve high-precision regression prediction of curvature values. Through the use of various gradient-weighted class activation mapping techniques for activation area visualization analysis, the key regions focused on by the VggNet 16 model during the prediction process are revealed, enhancing the model’s interpretability and reliability. Simultaneously, the designed single-mode-multimode-specialty optical fiber structure effectively increases the information content of the speckle images, improving the performance of curvature measurement. The automatic curvature setting device designed in the experiment greatly reduces experimental costs and provides support for future research on optical fiber curvature detection methods. Future research can further explore model optimization strategies and the application of these technologies in other related fields.