Abstract: To improve the applicability of drilling parameter-based methods for surrounding rock classification, an intelligent classification model driven by drilling parameters is developed using a Stacking ensemble learning method based on improved evidence reasoning (ER) rules. First, using numerous on-site samples of drilling parameters and surrounding rock grades, six conventional machine learning algorithms, including support vector machine and random forest, are employed to construct base models. Second, an improved ER rule is developed by introducing two calculation approaches: (1) a base model reliability calculation method that considers classification accuracy and misclassified sample types of each base model on the validation set and (2) a base model weight calculation method that accounts for the diversity of classification probability vectors produced by each base model on the validation set. Finally, based on meta-classifier inference under the improved ER rule, an integrated intelligent surrounding rock classification model is constructed using the classification probability vectors of each base model on the test set as inputs, with the reliability and weight of the base models synthetically considered. The results indicate that compared with individual base models, the overall accuracy of the ensemble model on the test set increases from 90.0%－92.5% to 96.0%, and its reliability increases from 86.0%－89.0% to 94.7%. The ensemble model correctly classifies all samples correctly identified by each base model and partially corrects the misclassified samples identified by most base models. Although the ensemble model′s computation time is approximately 4－10 times that of the individual base models, the overall efficiency remains at the millisecond level, meeting field application requirements.

Key words: tunnel, surrounding rock classification, evidence reasoning rules, Stacking ensemble learning, drilling parameters, diversity measure