The drawbacks of the traditional Kalman filter arise from its requirement of accurate model and noise statistics which generally can' t be satisfied in engineering applications. An improved Kalman filter including inaccuracy in filter's initial condition is presented in this paper. By reducing the requirement of model accuracy, it can improves the robustness of Kalman filter under the model mismatch. Fault detection with this filter reduces the fault alarm rate owing to its stronger robustness. Because the accuracy of missile attitude control system model is limited, it is difficult to detect fault for the traditional Kalman filter owing to its weak robustness. The improved algorithm can detect faults effectively . A simulation example shows its validity.