SEMI-AUTOMATED WELL LOGS QUALITY CONTROL AND PREDICTION WITH MACHINE LEARNING IMPLEMENTATIONS

• Main Author: Trimandiri Putra, Prajaka
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/40050
• Institution: Institut Teknologi Bandung

Well logs are considered as a critical data during the exploration of hydrocarbon resources because they are implicitly expressing the formation characteristics along the well depth, which are able to be interpreted by the experts to characterize the hydrocarbon reservoirs. Unfortunately, there are several factors during well logs acquisition that could lead to poor-quality data such as bad borehole conditions and tools failures. In such cases, tedious tasks such as quality control and assurance towards the data are necessary to validate the measurement results thus may strengthen the interpretation results. To overcome those problems and minimize the tasks intensity, this study aims to develop a semi-automated workflow in controlling and assuring well logs data quality which can be used for prediction improvement by the aid of machine learning. There are two main ideas to achieve such workflow which are outlier detection and removal and log prediction. Consideration of outlier detection necessity could be based on either the indications of bad borehole conditions or reports of tool failures from well site during the measurement. The outlier detection will be supported by a particular classification algorithm called One-Class Support Vector Machine (SVM). The prediction capability of machine learning will be highlighted, before and after the outlier removal. The prediction will be supported by a particular regression algorithm called Multilayer Perceptron (MLP) Regressor. After these algorithms are integrated, a case study is conducted and simple statistical metrics are used to evaluate the algorithms performance. Based on the case study, the performance of One-Class SVM in detecting outlier is outstanding. The algorithm clearly illustrates the separation between inlier and outlier using a rigid boundary where its shape can be adjusted by the users based on their needs. The outlier removal also improves the prediction accuracy but the degree of improvement still varies, depends on the chosen predicted log. The performance of MLP Regressor in prediction is relatively good because it is simple and its accuracy is most likely higher compared to the other simple regression algorithms, also its adjustable parameters have many options to choose. It also practical and effective to predict a parameter by using many predictors. However, the amount of data that is introduced and involved to the algorithms is highly impactful to the final results of those algorithms, just like the nature of other machine learning algorithms. In the end, this study represents a simple, concise, and practical approach in working with well logs especially in quality control and logs prediction by involving the machine learning advantages