Developing an explainable rockburst risk prediction method using monitored microseismicity based on interpretable machine learning approach

Abstract

The short-term rockburst prediction in underground engineering plays a significant role in the safety of the workers and equipment. Due to the complex link between microseismicity and the rockburst occurrence, prediction of short-term rockburst severity is always challenging. It is, therefore, necessary to develop an intelligent model that can predict rockbursts with high accuracy. Besides the predicting capability, it is essential to understand the model’s interpretability regarding the decisions to ensure reliability, trust and accountability. Accordingly, this paper employs the knowledge of explainable artificial intelligences (XAI) by proposing a novel glass-box machine learning model: explainable boosting machine (EBM) to predict the short-term rockburst. Microseismic (MS) data obtained from the underground engineering projects are utilized to build the model, which is also compared with the black-box random forest (RF) model. The result shows that EBM can accurately predict the rockburst severity with high accuracy, while providing with the underlined reasoning behind the prediction from the global and local perspectives. The EBM global explanation reveals that MS energy followed by MS apparent volume and the MS events is the most contributing factor to determining the Rockburst severity. It also gives insights into the relationship between MS factors and rockburst risks, delivering how various MS parameters impact the model predictions. The local explanation extracts the understanding of wrongly predicted samples. The interpretability and transparency of the proposed method will facilitate understanding the model’s decision which adds effective guidance evaluating the short-term rockburst risks.