Rockburst Severity Prediction Under Few-Shot Conditions Using Feature Fusion Learning

Abstract Accurate and efficient prediction of rockburst severity is crucial for the safe construction of deep rock engineering, but only a limited number of recorded rockburst data cannot enable the reliable prediction. This paper presents an innovative method for predicting rockburst severity, termed Feature Fusion Learning (FFL). It combines feature fusion with machine learning models and category voting techniques, enabling square-level amplification of sample data to address the challenge of few-shot in predicting rockburst severity. The performance of this method is evaluated against two typical data augmentation techniques (i.e., noise injection and synthetic minority over-sampling technique) combined with three machine learning models (i.e., extreme gradient boosting, light gradient boosting machine, and random forest). The results indicate that the FFL method exhibits superior performance across all metrics compared to the other models. In addition, feature importance assessments and a comparative study on the effect of amount of prior data on the prediction performance of the FFL method are conducted. The results reveal that the FFL method effectively enriches the feature dimensions while preserving the information and properties of the original feature variables. Moreover, the FFL method shows high robustness when applied to datasets with varying sample sizes.