The average lumpiness of ore rock is an important index to measure the blasting quality. The early research mainly relies on empirical formula summary, rock mechanics model calculation, which have shortcomings such as insufficient accuracy and strong subjectivity. Recently,, machine learning algorithm is applied for prediction, but still have problems such as empirical feature selection, insufficient model prediction stability, and poor generalization ability for the prediction of blasting material fragmentation. Aiming at above shortcomings, an extreme Gradient Boosting (xgboost) blasting fragmentation prediction model based on Feature Engineering is proposed. Taking Yuanjiacun Iron Mine in Taiyuan as the research area, engineering data are collected, Random Forest (RF) and Mutual Information (MI) are used for feature selection respectively, and the two feature subsets are integrated to obtain the best feature subset based on the value of MSE. XGBoost is used to predict the block size on the optimal feature subset, and the evaluation system is composed of two indexes: Mean Square Error (MSE) and Mean Absolute Error (MAE). The proposed method is compared with other traditional machine learning algorithms, and the results show that it is better than others. Furthermore, it can provide scientific guidance for the management and control of blasting.