基于因子分析的岩性识别智能模型对比

Comparative study of intelligent lithology identification models based on factor analysis

摘要:

岩性识别是钻探中十分重要的环节，传统岩性识别方法耗时较长且准确率较低，基于机器学习的智能岩性识别可以有效加快岩性识别的效率，准确率更高，泛化能力更强。本文以实现基于钻进参数的智能岩性识别预测为目标，使用逻辑回归、支持向量机、K近邻算法、随机森林、神经网络5种算法，先对原始数据进行预处理，然后对地层参数和其他参数进行因子分析、降维处理，建立了以钻进参数矩阵为输入，岩性分类识别和地层参数预测为输出的岩性智能识别模型，实现了钻遇地层的智能预测。实验表明，5种算法下岩性识别在测试集上的准确率和F1分数大部分达到70%以上，个别算法准确率较低但也在60%左右，5个模型地层参数预测的指标MAE和RMSE大部分都在1之下，识别预测准确。

Abstract:

Lithology identification is a very critical process in geological exploration and development. With the development of the geological exploration industry， the traditional lithology identification is limited by time and space， and can no longer meet the increasing data scale and dimensionality. Therefore， there is an urgent need for in-depth research on the intelligent prediction of lithology identification， so as to promote the development of lithology identification in the direction of digitalization， intelligence and timeliness. In this paper， with the goal of realizing intelligent lithology identification and prediction based on drilling parameters， five algorithms are used to establish intelligent lithology identification and prediction models， including Logistic Regression （LR）， Support Vector Machine （SVM）， K-Nearest Neighbor （KNN） algorithm， Random Forest （RF）， and Neural Network （NN）. They are designed to realize the intelligent prediction of drilling encounters. The main research contents are as follows： A regional stratigraphic identification model based on drilling data was designed. Through the data preprocessing of the original data， and then the factor analysis of the formation parameters and other parameters were reduced by factor analysis and modeled by five machine algorithms. An intelligent identification model was established with the drilling parameter matrix as the input， the lithology classification identification and formation parameter prediction as the output. Experiments show that for most of the five algorithms， the lithology identification accuracy and F1-score on the test set exceed 70%， and the individual accuracy is low but also about 60%， and most of the MAE and RMSE predicted by the stratigraphic parameters of the five models are below 1， and the identification and prediction are accurate.