The filter state components are not completely observable in a Kalman filter on an inertial frame. Aiming at this problem, we propose a method of twoposition alignment. Based on the theory of piecewise constant system (PWCS) and the method of singular value decomposition (SVD), we can get the highest degree of observability as rotating the heading axis among the three axes. In addition, the simulation is carried out with the heading error specification, and the conclusion rotation with 180° along the heading axis is the optimal.