Classification and Prediction of Miners' Emergency Response Competence Under Sudden Events Based on Seed k-Means and Stacking Learning Algorithm
Abstract
This study aims to predict and grade miners’ emergency behavioral competence based on physiological signals to support safety management in high-risk mining environments. A virtual reality platform simulating four mine emergency scenarios—fire, water inrush, roof collapse, and gas explosion—was developed. Fourteen physiological indicators were collected within two minutes of incident onset, alongside behavioral data from reaction, memory, and discrimination tasks. A two-stage approach was employed. First, six unsupervised clustering methods were applied to behavioral data and evaluated using Silhouette, Calinski-Harabasz, Davies-Bouldin, and FPC indices. Then, supervised clustering techniques—Constrained K-means, COP K-means, and Seeded K-means—were used to refine capability grading based on questionnaire-guided labels. Seeded K-means yielded the highest clustering performance and was used to generate final behavioral labels. Next, five classifiers were trained on physiological features. Hyperparameters were optimized via Grid Search, Random Search, and Optuna. A Stacking ensemble further improved prediction performance. Finally, SHAP (SHapley Additive exPlanations) was used for model interpretability. Results showed Seeded K-means achieved the best clustering (ARI = 0.8306, NMI = 0.8123), and the Stacking model reached 86.55% classification accuracy. SHAP analysis revealed that heart rate variability, skin temperature, and electrodermal activity were key predictors. In conclusion, this study establishes a physiological-signal-based framework for predicting and grading miners’ emergency behavior capacity, providing a foundation for individualized training and intelligent risk control.