Optimizing low-resolution spectral demodulation for long-period fiber gratings using residual convolutional neural networks.

Extracting valid information from low-resolution spectra is challenging and often compromised in many application scenarios due to the high costs associated with acquiring high-resolution spectra, which typically necessitates more sophisticated equipment. To address this challenge, this paper proposes a well-designed novel residual convolutional neural network (ResNet) to enhance spectral resolution for long-period fiber gratings (LPFG) under large stress measurement conditions. This method captures transmission spectra using a low-resolution interrogator (0.15 nm) and employs a CNN with a residual block to augment sample points. The model rapidly acquires high-resolution spectra without requiring preprocessing, thereby improving resonant wavelength detection accuracy. Additionally, an arrayed waveguide grating (AWG) extracts features from the transmission spectrum shifts, which are then processed by a multilayer perceptron (MLP) to model the nonlinear relationship between channel intensity and resonant wavelength. Using this trained model, minor shifts in the resonant wavelength can be accurately predicted. Experimental validation of this demodulation method under various LPFG stress conditions demonstrated a coefficient of determination () of 99.87% and a mean squared error (MSE) of 0.065, significantly surpassing the capabilities of traditional low-resolution interrogators and enabling high-precision resonant wavelength demodulation.