Fiber Bragg Grating Spectral Denoising Based on Multiscale Residual Convolutional Neural Network

Prolonged operation of fiber Bragg grating (FBG) sensing systems inevitably leads to performance degradation in optical fibers, gratings, and demodulation instruments due to component aging. This degradation ultimately induces optical transmission loss and results in significant noise contamination in FBG reflection spectra. To address this critical issue, this study proposes a novel Multiscale Residual Convolutional Neural Network (MSRCNN) model specifically designed for FBG spectral denoising. The proposed methodology integrates a multiscale feature extraction module into a one-dimensional convolutional neural network framework to comprehensively capture spectral characteristics across different scales. Furthermore, residual learning blocks are strategically incorporated through additive skip connections to enhance gradient flow and improve model convergence. Extensive validation using simulated datasets demonstrates that the MSRCNN model achieves superior noise suppression performance, attaining the minimum mean square error (MSE) convergence value and outperforming comparative models by up to 82.87 % in denoising efficacy. The experimental results further illustrate the model's outstanding denoising performance, with processed spectra achieving a remarkable improvement of up to 43.13 % in linearity metrics, conclusively validating the superior performance of the proposed model in practical applications.