Abstract: Earthquake early warning （EEW） is an effective approach to reduce human casualty and economic loss resulted from destructive earthquakes. Quick and accurate magnitude estimation after an earthquake is an important part of EEW, and the traditional approaches of magnitude estimation based on empirical parameters have their limits in accuracy and timeliness. This paper proposed a multi-fully connected convolutional neural network to quickly estimate the magnitude based on the information from a single station. The frequency-domain information from 3 065 earthquakes recorded by Japan’s KiK-net and K-NET networks between 1997 and 2019 （arriving waves corresponding to the selected data ranging from 3 s to 9 s）, together with the corresponding information on hypocentral distance, focal depth, and site conditions （v_S30） are used to train and validate the proposed model. The validation results demonstrate that the magnitude estimate accuracy is 89.92% even using as little as 3 s of arriving waves and the accuracy improves as longer duration of arriving waves is used. When 9 s of arriving waves are used, the accuracy increases to 96.08%. Comparison with the traditional P_d method suggests that the proposed approach in this study has smaller mean and standard deviation of the absolute estimation error, thus the proposed method has better accuracy and timeliness and would greatly enhance the disaster mitigation effects of the EEW systems.