Aiming at the problems of insufficient sensitivity and significant nonlinear response of fiber Bragg grating sensors in the low-temperature interval (-20 ℃~15 ℃) and the slightly poor accuracy of the traditional algorithm in the room-temperature to medium-high-temperature interval (15 ℃~85 ℃), this paper takes the Gaussian process regression algorithm as the core algorithm to process the FBG temperature signal data. The final experiment shows that: under the composite sensitization scheme as the premise, the GPR algorithm has an average absolute error of 0.05 ℃~0.14 ℃ under the condition of small samples at low temperatures, which is a significant advantage over the linear regression (0.15 ℃~0.31 ℃) and polynomial regression (0.04 ℃~0.17 ℃), and meanwhile maintains the accuracy advantage of 0.03 ℃~0.37 ℃ in the temperature zone above room temperature. In addition, GPR quantifies the data uncertainty by 95% confidence interval (e.g., ±0.15 ℃ at -20 ℃), which verifies the adaptability of the scheme in complex environments over a wide temperature range (-20 ℃~85 ℃). This study provides a new path of highly sensitive and reliable FBG sensing technology for cryogenic monitoring and industrial temperature control scenarios.