The application provides a method for training the defect rate prediction model of a first semiconductor chip production device in a semiconductor chip production process. The method comprises the following steps that: a logistic regression model is adopted to select one or a plurality of most suitable prediction parameters from the parameter set of sample data in a historical semiconductor chip sample data set of a first semiconductor chip production device; and the parameter values of the selected one or plurality of most suitable prediction parameters are adopted to train the defect rate prediction model, wherein the defect rate prediction model is used for predicting the defect rate of semiconductor chips produced by one or more second semiconductor chip production devises located at the downstream side the first semiconductor chip production device. With the method adopted, the defect rate prediction model applied to the first semiconductor chip production device can be trained, and therefore, defect maintenance processing can be performed on the first semiconductor chip production device in time when a predicted defect rate exceeds a predetermined threshold value.