Abstract: Using the initial P-wave early warning parameters to construct onsite ground motion prediction models, so as to quickly release an alarm message when it reaches the predefined threshold, is a key issue of the onsite earthquake early warning system, which is directly related to the accuracy and timeliness of the early warning information. For micro-electro-mechanical-systems-based seismic intensity instrument with poor data quality, the obtained displacement record after two integrations has a large deviation, which will lead to more false and missed alarms. Therefore, for waveforms recorded by seismic intensity instrument, in this paper, we adopted Butterworth filters of different orders （1−4） to build up several onsite ground motion prediction models based on the P-wave 3 seconds data and the whole P-wave window. These models are the relationships between displacement amplitude PD and peak ground velocity PGV, PD and peak ground acceleration PGA, velocity amplitude PV and PGV, PV and PGA, acceleration amplitude PA and PGV, PA and PGA, respectively. The models are then verified using the collected micro-electro-mechanical-systems-based seismic event records from the Sichuan-Yunnan Demonstration Early Warning Network. The results show that for the seismic intensity instrument records, the two optimal onsite ground motion prediction models are the relationship between PV and PGV and the one between PA and PGA obtained by the first-order Butterworth filter processing and derived from the whole P-wave window. In specific applications, two or more statistical relationships should be simultaneously adopted to predict onsite ground motion, and observed ground motion values should be used as additional judgment conditions to reduce the probability of false and missed alarms.