A rapid demodulation method for the interference spectrum of optical fiber sensors based on machine learning

Abstract Interferometric optical fiber sensors based on wavelength demodulation play an important role in the field of high-precision measurements. Generally used wavelength tracking method is based on the locations of feature wavelengths (peaks or valleys), and a long time of spectrum scanning and demodulation is required. A rapid demodulation method based on machine learning is proposed in this paper, and a more efficient demodulation of the interference spectra of optical fiber sensors is achieved. It is demonstrated by numerical simulations and experiments that the demodulation performances of the machine learning method are far better than those of the wavelength tracking method when the sampling intervals are sparse. Under a certain accuracy requirement (coefficient of determination larger than 0.9900 and mean square error less than 0.10), the machine learning based method can demodulate the interference spectra with a maximum sampling interval of 8 nm or a minimum wavelength range of 8 nm without feature wavelengths. The demodulation speed can be therefore improved to 400 times that of the wavelength tracking method.