基于随机树嵌入的随钻测井岩性识别方法

Lithology Identification Method for Logging While Drilling Based on Random Tree Embedding

岩性识别是储层评价中的一项重要工作.随着机器学习方法的不断发展,岩性的智能识别也成为热门研究方向.随钻测井技术目前已经得到了广泛的应用,但是受限于高温高压的钻井作业条件,随钻测井仪器只能测得少量测井参数.由于随钻测井参数较少,直接输入机器学习模型无法充分挖掘其中的信息.对此,本文将随机树嵌入引入随钻测井资料的岩性识别.该方法将低维随钻测井数据通过二叉树编码并转化为高维稀疏特征,利用升维后的数据进行训练从而提升机器学习模型的判别能力.对比实验结果表明,使用随机树嵌入的随机森林方法具有最佳的识别效果,准确率和F_(1)值较直接使用随机森林分别提升了3.16%和3.25%,且优于梯度提升树、极随机树和粒子群优化支持向量机算法.

Lithology identification is an important task in reservoir evaluation.With the development of machine learning methods,intelligent lithology identification has become a popular research direction.Logging while drilling(LWD)technology has been widely used.However,in the actual production process,due to the high-temperature and high-pressure operating conditions,only a few logging parameters can be measured by LWD.Due to the small number of logging parameters,machine learning model is not able to fully tap into the few parameters.To solve this problem,this paper introduced random tree embedding into LWD lithology identification.The low dimensional LWD data was encoded by the binary tree and transformed into high dimentional sparse features,and the upgraded data was used for training to improve the discriminative ability of the machine learning model.The comparative experiment results in this paper show that the random forest method with random tree embedding has the best recognition effect,the accuracy and F_(1) value are improved by 3.16%and 3.25%respectively,compared with the direct use of random forest,and outperforms the gradient boosted tree,extremely random tree and particle swarm optimization support vector machine algorithms.