Aiming at the problems that MEMS gyroscopes are easily affected and have large random errors, resulting in inaccurate model establishment and low measurement accuracy, an improved adaptive Kalman filter method is proposed in this paper. Firstly, the ARMA model is established, and the attenuation coefficient is introduced into the traditional Kalman algorithm to reduce the influence of the old value of the system, at the same time, the prediction error matrix based on the system new interest mutation is introduced to eliminate the system mutation values. The Allan variance is used to analyze and compare the original gyroscope data and the filtered gyroscope data. The results show that the angle random walk, bias instability and angular rate random walk of the gyroscope used in the experiment are at least an order of magnitude smaller, and the standard deviation is significantly reduced, which indicates that the improved algorithm effectively suppresses the random noise and improves the performance of MEMS gyroscope.