Abstract: On June 1, 2022, a M6.1 earthquake occurred in Lushan County, Sichuan, and on June 10, another M5.8 earthquake and M6.0 earthquake occurred in Ma erkang City, Sichuan. This paper proposes an earthquake prediction framework based on machine learning, which is used for the continuous magnitude estimation and the prediction of on-site instrumental seismic intensity in earthquake early warning (EEW). Based on the support vector machine (SVM) method, to explore the feasibility of the framework for the three earthquakes, this paper uses the acceleration records obtained from these three earthquakes to perform offline simulation. Meanwhile, this paper sets different instrumental seismic intensity thresholds to analyze the alarm accuracy. The results show that for the three earthquakes, compared with the existing EEW system, the SVM method exhibits more robust performance in magnitude estimation. And within 10 s after the first station is triggered, the magnitude estimation error is within ±0.6 magnitude units. At 3 s after the P-wave arrival, the percentage of the prediction error (within ±1) of on-site instrumental seismic intensity reaches 75.9%, and the average absolute error is 0.7; When the alarm thresholds of instrumental seismic intensity are VI and VII, the percentage of successful alarm reaches 98.4% and 100.0%, respectively. The results indicate that the machine learning method has the potential for application in Chinese EEW system, and provides a reference for the upgrading of EEW system.