Interpretable real-time monitoring of short-term rockbursts in underground spaces based on microseismic activities.

In this study, two novel hybrid intelligent models were developed to evaluate the short-term rockburst using the random forest (RF) method and two meta-heuristic algorithms, whale optimization algorithm (WOA) and coati optimization algorithm (COA), for hyperparameter tuning. Real-time predictive models of this phenomenon were created using a database comprising 93 case histories, taking into account various microseismic parameters. The results indicated that the WOA achieved the highest overall performance in hyperparameter tuning for the RF model, outperforming the COA. RF-WOA model accurately predicted the occurrence of this phenomenon with an accuracy of 0.944. Additionally, for this model, precision, recall and F1-score were obtained as 0.950, 0.944 and 0.943, respectively, indicating that the proposed model is robust in predicting damage severity of rockburst in deep underground projects. Subsequently, the Shapley additive explanations (SHAP) method was employed to interpret and explain the prediction process and assess the influence of input features based on RF-WOA model. The results showed that three parameters including cumulative seismic energy, cumulative microseismic events, and cumulative apparent volume have the greatest impact on the occurrence of rockburst events. This study provides an interpretable and transparent resource for accurately predicting rockburst events in real time. It can facilitate estimating project costs, selecting a suitable support system, and identifying essential ways to limit the danger of rockburst.