The physical and mechanical parameters of rock and soil are indispensable for engineering investigation， design， construction， and other operations， but conventional laboratory testing or in-situ tests have obvious accuracy errors. Accordingly， a real-drilling machine learning model was proposed in this paper which is used to predict the soil physical and mechanical parameters from drilling parameters. By collecting the actual data from several holes with the depth of 20 meters located in the National High-tech Industrial Development Zone of Zhuhai， the drilling pressure， torque， and triaxial vibration collected by the EP-200G drilling rig in real-time were used as input data， and the test data of soil cohesion， internal friction angle， water content and elastic modulus were used as output. Based on the established model， it is proved that the prediction accuracy of the machine learning models using single algorithms （including support vector machine， artificial neural networks and decision tree） can only reach 0.78 at most， while the integrated model based on the stacking concept can increase the prediction accuracy to a maximum of 0.98. Combined with this model， a sensitivity analysis between the drilling parameters and soil parameters was carried out， which confirmed that the drilling parameters would change significantly with the change of soil parameters， proving the reliability and applicability of using drilling parameters to predict soil parameters.