A timely damage state assessment of gantry cranes has a significant impact on the post-earthquake reconstruction and economic recovery in earthquake-stricken areas. This study aims to propose a methodology to rapidly predict the seismic damage states in light of nine classification-based machine learning methods. The 48 earthquake parameters is presented and of which relative importance and influence on the structural responses of the employed simple gantry crane are examined based on the data set matrix of 2760 (ground motions) ×48 (earthquake parameters). Meanwhile the innovative method is proposed to mitigate the class imbalance problem in the training data. Finally, the proposed method is applied to predict the fragility of a gantry crane subjected to ground motions and the efficiency and accuracy of nine machine learning methods are compared herein. The results demonstrate that the parameter of spectral acceleration S a at the first self-vibration period of the examined structure is of great significance to predict the accurate damage states. Random Forest, Neural Networks, Logistic Regression, and Support Vector Machine are preferable in all selected machine learning methods hereon. And the predictive fragility curves and the fragility curves computed in FE model are consistent approximately in spite of maximum error of 7.5%.