矿井突水是煤矿安全生产面临的主要威胁之一，快速分析突水成因和准确判别突水水源是矿井突水灾害治理的关键步骤。为有效防治矿井突水灾害，准确快速地判识矿井突水水源，提出一种基于粒子群优化算法（PSO）结合极限梯度提升回归树（XGBoost）的矿井突水水源识别模型（PSO-XGBoost），通过高效的参数全局搜索模式进一步提高突水水源识别效率与精度，并将该模型成功应用于辽宁抚顺煤田老虎台矿区以验证模型的实用性。基于老虎台矿40组水样光谱数据，首先利用多元散射校正、平滑去噪、标准化及主成分分析对原始光谱数据预处理，依据分层随机抽样按照7∶3比例进行训练集和测试集划分。其次，初始化粒子个体最优值和全局最优值，利用PSO对XGBoost算法的learning_rate、n_estimatiors、max_depth等7项参数进行迭代寻优，构建最优参数组合下的分类识别模型。为进一步研究该模型的优越性，选取平均判识准确率和对数损失值作为评价指标，对比PSO-XGBoost模型与PSO-SVM、PSO-RF模型的分类识别结果，同时通过100次重复交叉验证评价各模型的泛化能力。对比结果表明，XGBoost、PSO-SVM、PSO-RF和PSO-XGBoost模型对测试集数据的平均判识准确率分别为87.76%、87.56%、91.67%和91.67%。对于重复交叉验证，XGBoost、PSO-SVM、PSO-RF和PSO-XGBoost模型的平均准确度分别为87.76%、87.56%、90.63%和93.18%，相应的对数损失平均值分别为0.5453、0.5460、0.5623和0.4534。综合分析评价指标结果得出，PSO-XGBoost模型在矿井突水水源识别方面具有更高的判别精度和更好的泛化能力。

Mine water inrush is one of the main threats to mine safety production. Rapid analysis of the cause of water inrush and accurate identification of water inrush source are the key steps of mine water inrush disaster control. In order to effectively prevent and control mine water inrush disaster and identify mine water inrush source accurately and quickly, a mine water inrush source identification model (PSO-XGBoost) based on particle swarm optimization algorithm (PSO) and limit gradient lifting regression tree (XGBoost) was proposed. The efficiency and accuracy of water inrush source identification were further improved by the efficient parameter global search model, and the model was successfully applied to the Laohutai mine in Fushun coal field, Liaoning Province to verify the practicability of the model. Based on the spectral data of 40 groups of water samples from Laohutai mine, the original spectral data were preprocessed by multiple scattering correction, smoothing denoising, standardization and principal component analysis, and the training set and test set were divided according to the ratio of 7∶3 according to stratified random sampling. Secondly, the individual optimal value and the global optimal value of particles are initialized, and PSO is used to iteratively optimize seven parameters of XGBoost algorithm, such as learning_rate, n_estimatiors, max_depth, etc., to construct the classification and recognition model under the optimal parameter combination. To further investigate the superiority of the model, the average discrimination accuracy and log loss value were selected as evaluation indexes to compare the classification recognition results of PSO-XGBoost model with PSO-SVM and PSO-RF models, while the generalization ability of each model was evaluated by 100 repetitions of cross-validation. The comparison results showed that the average discrimination accuracies of XGBoost, PSO-SVM, PSO-RF and PSO-XGBoost models for the test set data were 87.76%, 87.56%, 91.67% and 91.67%, respectively. For repeated cross-validation, the average accuracy of XGBoost, PSO-SVM, PSO-RF, and PSO-XGBoost models were 87.76%, 87.56%, 90.63%, and 93.18%, respectively, with corresponding log-loss averages of 0.5453, 0.5460, 0.5623, and 0.4534, respectively. Comprehensive analysis of evaluation indexes shows that PSO-XGBoost model has higher discrimination accuracy and better generalization ability in mine water inrush source identification.