Spatial Shape Sensing of Multi-core Optical Fiber Based on Distributed Strain Measurement and Deep Learning

Shape sensing using optical fiber has attracted more attention in many fields with the increased demand for spatial shape perception. Frenet-Serret equation based on solving differential equations is commonly associated with complex numerical computations and error accumulation. These issues have become the obstacle in high precision optical fiber shape sensing. Therefore, a deep learning-based shape sensing method using distributed strain measurement information of multi-core optical fiber is proposed in this paper, which enables obtainment of fiber spatial positions directly from distributed strain measurement of three symmetrical cores in multi-core optical fiber. The proposed deep learning neural network model, which combines a multi-head self-attention temporal convolutional network (MSTCN) with a bidirectional long short-term memory (BiLSTM) network, focuses on extracting sequence dependent features of distributed strain information and mapping them to spatial shapes, thereby achieves shape reconstruction without the need of numerical computations. The shape sensing experiment utilizing optical frequency domain reflectometer is implemented for verification. The two-dimensional (2D) and three-dimensional (3D) shape sensing results in the experiment have shown that the proposed method is able to achieve shape sensing and accuracy enhancement for shape sensing compared with the usual Frenet-Serret equation. For 2D and 3D shape sensing, the best accuracies have been improved by 55.78% and 66.54%, respectively. This proposed method, which achieves shape sensing directly from distributed strain measurement of multi-core optical fiber without complex computation and cumulative errors, will have potential for high precision optical fiber shape sensing in practical applications.