Optimization of random forest through the use of MVO, GWO and MFO in evaluating the stability of underground entry-type excavations
Jian Zhou,Shuai Huang,Yingui Qiu
DOI: https://doi.org/10.1016/j.tust.2022.104494
IF: 6.9
2022-06-01
Tunnelling and Underground Space Technology
Abstract:The stability evaluation of underground entry-type excavations is a prerequisite of the entry-type mining method, which directly affects whether workers can be provided with a safe and reliable working environment and whether subsequent mining operations can be carried out normally. The design and stability assessment of entry-type excavations in current mining engineering largely relies on an empirical design method called the critical span graph, which has been widely applied in the initial span design of various cut and fill stopes. In recent years, with the wide application of various intelligent algorithms in the field of mine engineering, models based on intelligent algorithms provide new research methods and ideas for the assessment of rock stability in entry-type excavations. This study aims to introduce several hybrid models based on the random forest (RF) algorithm into the stability evaluation work to find new data-driven methods with higher accuracy to update the critical span graph. To pursue better classification performance, this paper selects three optimization strategies, namely multi-verse optimizer (MVO), grey wolf optimizer (GWO) and moth-flame optimization (MFO) algorithm, to optimize two core parameters of RF, and establishes three corresponding hybrid models, abbreviated as MVO-RF, GWO-RF and MFO-RF, based on the database containing 399 samples from seven Canada mines. There are two input parameters in the database, i.e., opening span and rock mass condition (expressed as RMR), and the output parameter is rock mass stability, which is specifically divided into three categories: stable, potentially unstable and unstable. In addition, five commonly used measurement indexes applicable to multiclassification problems were adopted to verify the classification ability of the models, i.e., the accuracy (ACC), precision calculated using macro-average (PREM), recall calculated using macro-average (RECM), F1 score calculated using macro-average (F1M) and Kappa index (Kappa). The results indicate that the three hybrid models performed well based on the test set accounting for 25 % of the original database, in which the accuracy of the MFO-RF model was the highest: ACC = 0.9300; PREM = 0.9288; RECM = 0.8983; F1M = 0.9116; Kappa = 0.8666. To evaluate whether the three optimization strategies can effectively improve the performance of RF and judge the degree of improvement, the performance of an unoptimized RF model was discussed in this study. In addition, two support vector machine (SVM) models with different kernel functions were selected as references for performance evaluation. The results indicated that compared with the RF and two SVM models, the classification accuracy of the three hybrid models was obviously more satisfactory. The classification accuracy of the three hybrid models reached 0.91, which was sufficient to explain the excellent classification ability of these models. After tuning the RF hyperparameters of each hybrid model, the critical span graph was further updated according to the optimized classification models, which was the focus of this research. By comparing the critical span graphs obtained by the three hybrid models with the single RF model and two kinds of SVM models, it is certain that the three hybrid models proposed in this paper, MVO-RF, GWO-RF and MFO-RF, are promising in the study of evaluating the stability of entry-type excavations and may be deemed auxiliary decision tools to define the stability region of the critical span graph.
construction & building technology,engineering, civil