Research on a Method for Identifying Small Fire Hazards in Enclosed Spaces Based on 1-DDCNN

This study addresses the resolution limitations of traditional methods in UWFBG overlapping spectrum demodulation by proposing a demodulation method based on a one-dimensional dilated convolutional neural network (1DDCNN), which is applicable to two types of overlapping spectra and maintains high accuracy even under signal overlap and noise conditions. Simulation results show that when demodulating four parallel overlapping UWFBG groups, the root mean square error (RMSE) is less than 0.10 pm, with an average demodulation time of 15.2 ms; when demodulating four series-connected overlapping UWFBG groups, the RMSE is less than 1.5 pm, with an average demodulation time of less than 30 ms. Experimental validation demonstrates that this method achieves an RMSE of 0.0983 pm at the central wavelength of $\mathbf{1 5}$ parallel-connected overlapping UWFBGs and can accurately demodulate and locate long-distance signals from 100 UWFBGs.