Abstract: In order to improve the accuracy of thermocouple dynamic calibration, based on a new thermocouple dynamic calibration system composed of a semiconductor laser, infrared detector and calibration thermocouples, the static calibration of the infrared detector was analyzed in the dynamic calibration system. According to the principle of back propagation(BP) neural network, the structure and the parameters of neural network were determined. The dynamic calibration experiment with ordinary K type armor-loaded thermocouple was performed and the static calibration data of infrared detector was acquired. Nonlinear curve fitting was performed using the least square method and BP neural network. The fitting results of the both the methods were analyzed and the fitting curves were obtained. The results show that, the fitting effect of BP neural network is better than the traditional least square method when there is less and uneven distributed data. The error caused by data fitting is reduced, dynamic characteristics of the thermocouple are acquired more accurately and thermocouple dynamic compensation is realized. The study has an important reference to the research of thermocouple dynamic characteristics.