Demodulation of Different Quantities of Overlapping Spectra in FBG Sensors Based on Combined Conv-TasNet and LSTM

We propose a demodulation method that combines a convolutional time-domain audio separation network (Conv-TasNet) method with a long short-term memory (LSTM) model to address overlapping spectra in quasi-distributed fiber Bragg grating (FBG) sensing networks. The Conv-TasNet is used for separating and denoising the overlapping FBG spectra, followed by the LSTM model for demodulating the separated spectra. In addition, the different numbers of FBG spectral overlaps are demodulated by the method. Results show that the Conv-TasNet method achieves a root mean square error (RMSE) of 0.00064 and a signal-to-noise ratio (SNR) of 20 dB for separating overlapping spectra. The LSTM model, on the other hand, achieves an RMSE of 0.93 pm when demodulating the separated spectra. Furthermore, when using different wavelength multiplexing and varying degrees of overlapping spectra as inputs, the average demodulation RMSE for two overlapped FBGs is 0.769 pm, and for three overlapped FBGs it is 0.763 pm.