Abstract: In torque verification, the fluctuation of torque value over time is not a known common function, and it is vital to explore the trend of its change and fluctuation for torque verification and dissemination of value of quantity. To explore the fluctuation characteristic of the torque sensor, the influence of medium- and long-term stability on the change of torque value in torque measurement is investigated, and a neural network is used to train the detection data under given conditions to predict the torque value at other time points. The torque value data obtained under strictly controlled experimental conditions for the last 12 years were used for training and analysis, the accuracy and reliability of the model were demonstrated by cross-validation, the torque value of the next calibration cycle was predicted, and the probability that the torque value of the next cycle meets the long-term stability of 0.03 level was finally calculated to be greater than 71.8%.