Fast and robust demodulation of temperature from sparse sapphire fiber Bragg grating spectra with machine learning.

Sapphire fiber Bragg gratings (FBGs) have demonstrated their efficacy in sensing at high-temperature harsh environments owing to their elevated melting point and outstanding stability. However, due to the extremely high volume of modes supported by the clad-less sapphire fiber, the demodulation capability of the reflected spectra is hindered due to their irregular and somewhat complicated shapes. Hence, a mode-stripping or scrambling step is typically employed beforehand, albeit at the expense of sensor robustness. Additionally, conventional interrogation of sapphire FBG sensors relies on an optical spectrum analyzer due to the high sensitivity provided by the spectrum analyzer, where the long data acquisition time restricts the system from detecting instantaneous temperature variations. In this study, we present a simple sensor configuration by directly butt-coupling the sapphire FBG multi-mode lead-out fiber to a single-mode lead-in fiber, and detect its reflected spectra via a low-cost, fast, and coarsely resolved (166 pm) spectrometer. We leverage machine learning to compensate for the under-sampling of the measured FBG spectra and achieve a temperature accuracy of 0.23 °C at a high data acquisition rate of 5 kHz (limited by the spectrometer). This represents a tenfold improvement in accuracy compared to conventional peak-searching and curve-fitting methods, as well as a significant enhancement in measurement speed that enables dynamic sensing. We further assess the robustness of our sensor by attaching one side of the sensor to a vibrator and still observe good performance (0.43 °C) even under strong shaking conditions. The introduced demodulation technology opens up opportunities for the broader use of sapphire FBG sensors in noisy and high-temperature harsh environments.