Interpretability analysis of rockburst risk prediction in coal mines based on machine learning-SHAP

The traditional prediction model has significant limitations in the risk assessment of rockburst in deep coal mines, which is mainly reflected in its insufficient ability to analyze complex geological conditions and mining disturbance factors. These models are usually difficult to accurately capture the impact of various environmental factors on rockburst risk, lack the dynamic assessment ability of the importance of characteristics, and the traditional models are difficult to formulate targeted prevention measures, which can not meet the refined needs of deep coal mine safety production. This limitation not only reduces the reliability of prediction results, but also limits the effectiveness of control measures. Therefore, in view of the problems that the traditional prediction model lacks the corresponding interpretation and analysis of rockburst in deep coal mines, cannot distinguish the characteristic importance of various environmental factors, and is difficult to carry out targeted prevention and control measures, In view of the problems that the traditional prediction model lacks the corresponding interpretation and analysis of the rockburst in deep coal mines, cannot distinguish the characteristic importance of various environmental factors, and is difficult to carry out targeted prevention and control measures, this paper uses three algorithms in machine learning, namely random forest (Random Forest), support vector regression (Support Vector Regression), and extreme gradient lifting (Extreme Gradient Boosting), and takes the daily footage, the number of coal seam pressure relief holes, the number of drill cuttings detection holes, the number of bottom coal pressure relief holes, the number of bottom coal blasting holes, the total energy of microseismic, the frequency of microseismic, the maximum microseismic energy, the maximum stress, the amount of drill cuttings, and the depth of drill cuttings as 12 eigenvalues. After data processing, through cross validation and superparameter adjustment,the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and The algorithm with the best regression performance is comprehensively screened out under the conditions of R² and STD deviation. The corresponding interpreter is used to analyze the interpretability of the algorithm and SHAP. Finally, the model is applied in the field to verify its feasibility. The research results indicate that the XGBoost algorithm performs the best on this dataset, with a sample MSE of 0.0001, RMSE of 0.0024, MAE of 0.0011, MAPE of 1.3164 %, R² of 0.9930, and Std Deviation of 0.0021 the XGBoost-SHAP model, the feature importance sequence was obtained using global interpretation, and the feature value with the highest weight in this environment was selected; By combining local explanations to make the inference and prediction processes within the model transparent, the black box problem of machine learning has been solved; The model can better capture the real-time impact of environmental factors on impact risk, and can provide a scientific basis for dynamic assessment of impact risk in different environments. The dynamic assessment and targeted prevention and control decision support framework for rockburst risk can provide important guidance for the targeted prevention and control of rockbursts at the site.