Aiming at the inaccurate data modeling of MEMS gyroscope or the inability to give a model, a method to reduce the noise of gyroscope data by BP neural network (BPNN) assisted Kalman filtering is proposed in this paper. Analysis of the systematic noise variance Q of the Kalman filter shows that when the model is not accurate, it can be compensated by Q. Based on the principle of BP neural network to optimize Q value, the acquired MEMS gyroscope data were input into the Kalman filter to obtain Q firstly. Then the innovation, filter gain and measurement noise variance are input into the neural network, and Q is used as the output of the neural network. The system noise covariance matrix is optimized by the neural network to obtain Q*. And finally Q* is used as the noise variance matrix of the Kalman filter system. The experimental results show that the method can effectively improve the accuracy of the gyroscope in the case of inaccurate modeling.