Prediction of Rockburst Risk in Coal Mines Based on a Locally Weighted C4.5 Algorithm

Abstract

Rockburst is a dynamic phenomenon characterized by the sudden, abrupt, and violent release of deformation energy in coal and rock masses around mine shafts and slopes that can result in considerable destruction. For prediction and evaluation methods are essential for the prevention and control of rockburst disasters, many machine learning methods are developed for this purpose. To accurately predict rockburst risk, the present study addresses this issue by developing a locally weighted C4.5 decision tree algorithm for predicting the risk of rockburst in coal mines. In the proposed processing, the minimum description length principle is first applied to discretize the continuous attribute data in the raw training dataset. Then, the prediction model based on the C4.5 algorithm is trained by 10-fold cross validation using the adjacent samples selected by the k-nearest neighbors method. Finally, the decision tree is completed by applying pessimistic pruning. The rockburst prediction accuracy of the proposed locally weighted C4.5 algorithm is compared with that obtained by the standard C4.5 algorithm based on field data derived from the Yanshitai coal mine, Chongqing, China. The rockburst risk prediction accuracies obtained by the proposed and standard C4.5 algorithms for the samples in the testing dataset were 100% and 71.43%, respectively. Accordingly, the proposed locally weighted C4.5 algorithm greatly outperformed the standard C4.5 algorithm for the prediction of rockburst risk based on the data considered.

Keywords

• C45 decision tree
• k-nearest neighbors
• locally weighted learning
• minimum description length principle
• rockburst prediction