Intelligent Demodulation of Overlapping Spectra in FBG-PMFBG Sensing Based on a Spiking Neural Network

This paper presents a leaky integrate-and-fire spiking neural network (LIF-SNN) for direct demodulation of spectrally overlapped fiber Bragg grating–polarization-maintaining fiber Bragg grating (FBG–PMFBG) responses. The hybrid spectrum contains one conventional FBG peak and two PMFBG peaks, producing partial-to-severe overlap as the FBG wavelength shifts toward either PMFBG peak. Because controlled overlap measurements are costly, data acquisition is concentrated in informative overlap regions. The proposed LIF-SNN uses surrogate-gradient learning and temporal spike integration to directly estimate all three peak wavelengths without explicit peak separation. Multi-seed validation, additive white Gaussian noise (AWGN) augmentation, and few-shot fine-tuning further improve robustness under limited measured data. On the held-out test set, the model achieves a mean absolute error (MAE) of 0.0786 nm and a mean squared error (MSE) of 0.0130 nm2, outperforming mixture density network (MDN), long short-term memory (LSTM), and gated recurrent unit (GRU) baselines. Few-shot adaptation reduces the MSE to 0.0085 nm2, while 30–50 dB signal-to-noise ratio (SNR) tests show consistently lower MAE than the compared models. A matched non-spiking multilayer perceptron (MLP) ablation further shows that performance depends on both architecture and training strategy: the MLP remains competitive on clean experimental spectra, whereas the LIF-SNN shows stronger robustness under noise and benefits from AWGN augmentation and adaptation. These results demonstrate accurate, adaptive, and noise-robust three-peak demodulation for overlap-affected FBG–PMFBG strain sensing.