School of Energy, Xi'an University of Science and Technology
Key Laboratory of Western Mine Exploitation and Hazard Prevention of Ministry of Education

通过传统的监测手段获取矿压数据并采用统计学或机器学习算法对矿压进行预测已不能满足矿山智能化发展要求，需要寻求新的方法提升矿压数据监测及矿压预测的准确性和实时性。基于三维相似物理模型试验，搭建分布式光纤监测系统，沿模型走向和高度2个方向预埋分布式光纤，在模拟工作面开采过程中采集来压数据，并引入光纤布里渊频移平均变化度作为判断是否来压的指标；通过对光纤监测数据进行噪声去除、归一化及相空间重构等预处理，将一维初始监测数据转换为三维数据；使用贝叶斯算法对CatBoost算法进行迭代参数寻优，在达到最大迭代次数后将最优参数组合装载到CatBoost算法中，通过训练得到矿压显现预测模型。结果表明：贝叶斯算法比传统网格搜索法的迭代次数更少、误差更小；与随机森林（RF）、梯度提升决策树（GBDT）和极值梯度提升树（XGBoost）算法相比，CatBoost算法的预测精度更高、泛化能力更强；基于贝叶斯算法优化的CatBoost矿压显现预测模型能准确预测出测试集中的3次来压，且整体预测趋势与实测值较为吻合，平均绝对误差为0.009 1，均方根误差为0.007 7，决定系数为0.933 9。

Obtaining mine pressure data through traditional monitoring methods and using statistical or machine learning algorithms to predict mine pressure can no longer meet the requirements of intelligent development in mines. It is necessary to seek new methods to improve the accuracy and real-time performance of mine pressure data monitoring and prediction. Based on three-dimensional similar physical model experiments, a distributed fiber optic monitoring system is constructed. The distributed fiber optic cables are pre-embedded along the model's direction and height. Pressure data is collected during the simulated mining process of the working face, and the optical fiber Brillouin frequency shift mean variation degree is introduced as an indicator to determine whether the pressure is coming. By preprocessing the optical fiber monitoring data such as noise removal, normalization and phase space reconstruction, the one-dimensional initial monitoring data is converted into three-dimensional data. The method uses Bayesian algorithm to iteratively optimize the parameters of the CatBoost algorithm. After reaching the maximum number of iterations, the optimal parameter combination is loaded into the CatBoost algorithm. The prediction model for mine pressure appearance is obtained by training. The results show that the Bayesian algorithm has fewer iterations and smaller errors than traditional grid search methods. Compared with random forest (RF), gradient boosting decision tree (GBDT) and extreme gradient boosting (XGBoost), the CatBoost algorithm has higher prediction accuracy and stronger generalization capability. The CatBoost mine pressure appearance prediction model optimized by the Bayesian algorithm can accurately predict the three weighting in the test set. The overall prediction trend is in line with the measured value, with mean absolute error of 0.0091, root-mean-square error of 0.0077, and determination coefficient of 0.933 9.

国家自然科学基金资助项目（41027002,51804244）；

柴敬,张锐新,欧阳一博等.基于贝叶斯算法优化的CatBoost矿压显现预测[J].工矿自动化,2023,49(07):83-91.DOI:10.13272/j.issn.1671-251x.2022110065.

CHAI Jing, ZHANG Ruixin, OUYANG Yibo, et al. CatBoost mine pressure appearance prediction based on Bayesian algorithm optimization[J]. Journal of Mine Automation，2023，49(7)：83-91.