Rockburst prediction significantly affects the development and utilization of underground resources. Currently, an increasing number of artificial intelligence algorithms are being applied for rockburst prediction. However, owing to the scarcity of data for certain rockburst grades, machine learning models have struggled to accurately train and learn their characteristics, resulting in bias or overfitting. In this study, 321 worldwide cases of rockbursts were collected. Seven indices considering both rock mechanics and stress conditions were selected as input parameters for the model. To address the issue of limited data for certain rockburst grades, the Synthetic Minority Over-sampling TEchnique (SMOTE) algorithm was used for comprehensive oversampling and synthesis of the rockburst data. The theoretical rationality of this method was corroborated by the Spearman’s correlation coefficient. Additionally, the model hyperparameters were optimized using the Bayesian optimization method, and an improved eXtreme gradient boosting (XGBoost) rockburst prediction model (SM–BO–XGBoost) was established. The constructed SM–BO–XGBoost model was compared with decision tree, random forest, support vector machine, and k-nearest neighbor classification machine learning models. The results showed a significant improvement in the prediction accuracy for the None and Strong rockburst categories, which had limited data in the original rockburst dataset. To address the poor interpretability of the XGBoost model, the SHapley Additive exPlanations (SHAP) method was introduced to explain the constructed model, and to analyze the marginal contributions of different features to the model output across various rockburst grades. The SM-BO-XGBoost model was validated using field rockburst records from the Xincheng and Sanshandao gold mines. As indicated by the results, the model demonstrated favorable performance and applicability, with wide potential for predicting engineering rockbursts.

岩爆预测显着影响地下资源的开发利用。目前，越来越多的人工智能算法被应用于岩爆预测。然而，由于某些岩爆等级的数据稀缺，机器学习模型很难准确地训练和学习其特征，从而导致偏差或过度拟合。在这项研究中，收集了全球 321 个岩爆案例。选择同时考虑岩石力学和应力条件的七个指标作为模型的输入参数。针对某些岩爆等级数据有限的问题，采用合成少数过采样技术（SMOTE）算法对岩爆数据进行全面过采样和合成。 Spearman相关系数验证了该方法的理论合理性。此外，采用贝叶斯优化方法对模型超参数进行优化，建立了改进的极限梯度提升（XGBoost）岩爆预测模型（SM-BO-XGBoost）。将构建的 SM–BO–XGBoost 模型与决策树、随机森林、支持向量机和 k-近邻分类机器学习模型进行比较。结果表明，原始岩爆数据集中数据有限的“无”和“强”岩爆类别的预测精度显着提高。为了解决 XGBoost 模型可解释性差的问题，引入 SHapley Additive exPlanations (SHAP) 方法来解释所构建的模型，并分析不同特征对不同岩爆等级的模型输出的边际贡献。 SM-BO-XGBoost 模型利用新城金矿和三山岛金矿的现场岩爆记录进行了验证。结果表明，该模型表现出良好的性能和适用性，在预测工程岩爆方面具有广泛的潜力。