为了提高铀矿钻孔地层岩性识别的准确性，解决传统集成学习模型识别地层岩性效果不佳的问题，提出了一种基于主成分分析优化的Stacking 集成学习模型。首先，基于皮尔逊相关系数量化测井参数与目标岩性的线性关联强度，结合测井地球物理机理，筛选出与岩性关系较为密切的6 个测井参数作为输入特征。同时，计算基学习器预测误差的皮尔逊相关系数，并使用Q统计量矩阵评估预测结果的相关性，从中筛选出误差互补性强（即低相关性）且预测模式差异显著（即低Q值）的基模型组合。通过主成分分析算法对基模型的预测结果进行加权融合，并将这些融合后的特征作为输入，构建第二层元模型的训练数据，从而实现一个高精度的多层次集成学习模型。实验结果表明，基于主成分分析优化的Stacking 模型的识别精度达97.19%，明显优于传统Stacking 模型以及所有个体模型的性能，这一结果验证了所提出方法的有效性，为砂岩型铀矿钻孔地层岩性识别研究提供了新的思路和方法。

Abstract：To improve the accuracy of lithology identification in uranium mine borehole strata and addressthe limitations of traditional ensemble learning models in achieving satisfactory lithology recognitionperformance, this study proposes a PCA-optimized Stacking ensemble learning model. The methodology beginsby quantitatively evaluating the linear correlation strength between various logging parameters and the targetlithology using Pearson correlation coefficients. This statistical analysis is further refined by incorporatinggeophysical mechanisms of well logging, leading to the selection of six logging parameters that demonstratestrong lithological relevance as input features. These parameters effectively capture essential petrophysicalcharacteristics while minimizing redundancy in the feature set. Simultaneously, the model employs a dual-facetedapproach to base learner selection. The Pearson correlation coefficients of prediction errors among base learnersare systematically calculated to assess their error interdependence. Additionally, the Q-statistic matrix is utilizedto evaluate the correlation and diversity of prediction outcomes. From this, a combination of base models withstrong error complementarity (i.e., low correlation) and significant prediction mode differences (i.e., low Q-value)is selected. To optimally consolidate the predictive information, the PCA algorithm is utilized to perform aweighted fusion of the prediction results derived from the diverse base models. The resulting principalcomponents, which represent the most discriminative fused features, then form the refined input feature space fortraining the second-layer meta-model, thereby achieving a sophisticated and high-accuracy multi-level ensemblelearning framework. Experimental results show that the PCA-optimized Stacking model achieves a recognitionaccuracy of 97.19%, significantly outperforming traditional Stacking models and all individual models. Theexperimental results indicate that the recognition accuracy of the PCA-optimized Stacking model reaches97.19%, significantly outperforming both the traditional Stacking model and all individual models. Furthermore,the study provides valuable insights and a novel technical framework for lithology identification in sandstonetypeuranium mine borehole strata, contributing to more reliable and efficient geological interpretation in uraniumexploration and mining operations.