Kalman filter is a recursive algorithm based on minimum variance estimation, filtering performance and the estimated accuracy depend on the priori knowledge of system model and noise statistical properties, and imprecise priori knowledge can cause significant degradation even disperse in the filtering performance. BP neural network is used for system identification to acquire the precise system equation. The process noise and measurement noise covariance matrix in adaptive estimated Kalman filter algorithm is used to propose a new algorithm of innovation-based neural network adaptive Kalman filter. Matlab simulation results show: compared with the traditional Kalman filter algorithm, the output signal obtained through the improved Kalman filter algorithm is almost identical with the original signal, the noise is significantly suppressed, meanwhile the improved algorithm does not need accurate system mathematical model, which is effective and available in practical application.