In view of the problems that there are many factors affecting the sintering drum and it is difficult to predict accurately, data analysis and field experience are combined to propose an adaptive prediction model of sinter tumbler strength based on sintering state and ensemble learning. Firstly, a prediction data collection of high-quality sinter tumbler strength is obtained through such data preprocessing processes as data collection and integration, abnormal data identification and missing data processing. Then, the Boruta algorithm combined with Pearson linear and XGBoost nonlinear correlation analysis is used to select the input features of the tumbler strength prediction model to avoid from the incompleteness of the input features of the model. Finally, five different machine learning algorithms are used to construct the ensemble learning prediction model of the tumbler strength, and the weight of the prediction results of each sub-model could be updated adaptively according to the actual situation of sintering production. Combined with the on-site production practice and the analysis of the historical data of sintering, the sintering state is distinguished by the temperature of the sintering end point and the negative pressure of the main pipe, and the predicted hit rate of the tumbler strength is more than 90% when the error range is 0.5%.