Novel rockburst evaluation model based on a KPCA-IF-BorderlineSMOTE-ENN preprocessing framework

Rockburst is a major dynamic hazard in deep underground engineering. However, field case-history datasets are often constrained by small sample sizes, multi-source heterogeneity, feature redundancy, anomalous observations, and severe class imbalance, which may reduce the reliability of machine-learning models, particularly for high-risk rockburst identification. To address these challenges, this study proposes a reproducible preprocessing and evaluation framework that integrates interaction-feature construction and selection, radial basis function kernel principal component analysis, Isolation Forest-based outlier screening, and BorderlineSMOTE-ENN boundary-aware rebalancing. Based on 248 published rockburst cases, six ensemble classifiers, namely AdaBoost, CatBoost, LightGBM, GradientBoosting, ExtraTrees, and Random Forest, were evaluated using nested cross-validation combined with Bayesian optimization. In addition to accuracy, balanced accuracy, and macro-F1, safety-oriented indicators for Strong rockburst events, including the missed detection rate and false alarm rate, were reported. The proposed framework improved both overall and class-balanced performance on the present dataset compared with the baseline strategy. ExtraTrees achieved the best overall performance, with an accuracy of 99.00%, balanced accuracy of 99.17%, macro-F1 of 99.00%, Strong-rockburst missed detection rate of 3.33%, and false alarm rate of 0.00%. Ablation analysis showed that BorderlineSMOTE-ENN made the largest contribution to minority-class recognition, while Isolation Forest improved model robustness and KPCA provided a model-dependent nonlinear representation. Overall, the proposed framework provides a reproducible and safety-oriented preprocessing strategy for rockburst intensity classification in small, heterogeneous, and imbalanced datasets.