Efficient Spectrum Demodulation Technique for Data-Sparse Optical Fiber Sensors With Feature Transfer Strategy

Optical fiber sensors are widely used in detection systems due to their high sensitivity and electromagnetic interference immunity. However, the sensor's spectrum demodulation accuracy is often affected by noise and limited training data. In this work, a hybrid framework combining variational mode decomposition (VMD) for spectrum reconstruction and transfer learning for sparse data efficient demodulation has been proposed. A novel evaluation index integrating statistical and physical properties has been proposed to assist VMD in reconstructing the noisy spectrum, for enhanced physical fidelity and feature clarity. Clear features facilitate the identification of similarities between sensors, thus enabling knowledge transfer from sufficient sample sensors to those with limited samples through similarity mapping. A hybrid model that extracts spatiotemporal spectrum features, achieving 99.39% demodulation accuracy even with limited hardware. This work offers a cost-effective and noise-resistant solution for high-precision optical fiber sensor spectrum demodulation under data-sparse constraints.