When the initial alignment of ring laser gyroscope SINS is performed on a rocking base

the noise of ring laser gyroscopes and accelerometers will increase. Then

it will need more time to accomplish the initial alignment. Sometimes

the large noise will make the alignment fail. Only with a lowpass filter

the output of sensors will still suffer from large noise. Wavelet can suppress the noise brought by the dither motion of base. However

it has a high computational complexity and block-processing structure

which makes it hard to be implemented online. To solve this problem

a new prefilter method is proposed in this paper. After the output of sensors is filtered by the lowpass filter

a steady-state Kalman filter which uses a hidden Markov model will be used to filter it. Then

the large sensor noise brought by the dither motion of base can be lowed down to a very low level. Experiment results show that this prefilter method can work efficiently with a low computational complexity and help SINS accomplish the initial alignment on a rocking base.