Machine learning for high-accuracy signal processing in ultra cryogenic temperature optical fiber sensors

The conventional optical fiber interference signal processing method (the peak tracking approach) is fundamentally limited by the free spectral range (FSR). Ultra-wide experiment measurement ranges and ultra-high sensitivity have always been difficult trade-offs. In this paper, a Bayesian-optimized Support Vector Machine (SVM) model is proposed to address this challenge by mapping the relationship between the entire spectrum and temperature. The mean absolute error (MAE) of the never-before-seen spectrum is as low as 0.034145 K, and the R² is as high as 0.99999. This method avoids the cumbersome data processing process of the peak tracking approach, and can not only improve the accuracy and efficiency of signal demodulation, but also has wide application potential in various types of optical fiber sensors. We discuss the difference in sensitivity of different interference dips in the same temperature measurement range in the Sagnac interference sensors. We also analyze the factors affecting the sensitivity of Sagnac interference sensors and the disadvantages of the traditional peak tracking method. This theoretical principle is helpful for more accurately evaluating the performance of the Sagnac interference sensor or for conducting an optimized design. Cryogenic temperature optical fiber sensors with both ultra-high sensitivity and ultra-wide measurement range provide precise and reliable temperature monitoring capabilities, which are crucial for the stability and safety of certain systems.