High-accuracy neural network model for multi-FBG quantity classification and wavelength addressing in AWG-based demodulation system

• A cascaded Res-SE/LNN-BiLSTM model demodulates up to six FBGs via eight AWG channels, achieving 99.21% peak counting and 3.85 pm wavelength RMSE—surpassing prior AWG limits without extra hardware. • Hybrid training fuses simulated and measured spectra (r = 0.987), leveraging physics-aware loss and adaptive hyperparameters to suppress aliasing, yielding > 81% accuracy even at 0 nm spacing. • Experimental validation with 0.1 pm reference confirms real-time six-sensor interrogation over 15 nm, expanding multiplexing capacity by 50% and enabling compact, low-cost distributed FBG networks. We report a deep-learning-enabled interrogation scheme that unlocks high-precision wavelength demodulation of densely multiplexed fiber Bragg gratings (FBG) using only eight channels of an arrayed waveguide grating (AWG). A two-stage cascade is devised: (i) a Res-SE network classifies the quantity of FBGs (1–4) with 99.21% peak accuracy under 2 nm spacing and the number of gratings (1–6) with 81.69% at 0 nm spacing; (ii) dimension-adaptive LNN-BiLSTM regressors then retrieve individual Bragg wavelengths, attaining RMSE of 3.85 pm for a single sensor and 13.78 pm for four-sensor arrays. Experiments on five- and six-grating cascades yield 25.80 pm and 42.76 pm accuracy, respectively-surpassing prior AWG-based limits without extra hardware. The proposed framework redefines the accuracy-capacity frontier of sparse-sampling AWG interrogators and paves the way for compact, low-cost distributed FBG sensing networks.