摘要：The identification of wireline log shapes is of great significance for research in paleogeography and sedimentology, and further for natural resources characterization. Generally, identification of log shapes is manually achieved by geologists through careful observation of the morphological features of log curves. This process is subjective and tedious, and it may introduce artefacts depending on the bias of the interpreters. Therefore, we developed a model which exceeds average performance of experienced geologists in identifying basic log shapes. The excellence of the model mainly results from a large annotated dataset and a specially designed Recurrent Neural Network (RNN) architecture. The dataset contains 2676 Spontaneous Potential (SP) log segments regarding four basic log shapes (bell, funnel, egg and cylinder shapes). There are six layers in the model's architecture, and the core is a layer of Long Short Term Memory (LSTM) units which are designed to extract the morphological features of the log curve. A committee of four experienced geologists annotated a gold standard testing dataset, based on which different approaches can be compared in regard to their accuracies. We also tried standard artificial neural network (ANN), Support Vector Machine (SVM) and an analytical method (Trendline Matching) for this problem. It turns out the proposed RNN model is more accurate than geologists' manual labelling and also the above three methods. The research for this problem is still at the initial stage and expecting more work and ideas to further improve the robustness of the model. Therefore, we also put forward some ideas for improving the model and enlarging the dataset.

关键字：Deep learning RNN model LSTM Log dataset Sedimentary facies

ISSN号：0098-3004

卷、期、页：卷: 134

发表日期：2020-01-01

收录情况：SCIE（科学引文索引网络版）

发表期刊名称：COMPUTERS & GEOSCIENCES

通讯作者：宋随宏,窦鲁星,孙爽

第一作者：侯加根,宋泽章

论文类型：期刊论文

论文概要：宋随宏,侯加根,窦鲁星,宋泽章,孙爽，Geologist-level wireline log shape identification with recurrent neural networks，COMPUTERS & GEOSCIENCES，2020，卷: 134

论文题目：Geologist-level wireline log shape identification with recurrent neural networks