Vis enkel innførsel

dc.contributor.advisorDuffaut, Kenneth
dc.contributor.advisorWestad, Frank
dc.contributor.authorCaceres, Veronica Alejandra Torres
dc.date.accessioned2022-05-24T12:00:47Z
dc.date.available2022-05-24T12:00:47Z
dc.date.issued2022
dc.identifier.isbn978-82-326-6585-3
dc.identifier.issn2703-8084
dc.identifier.urihttps://hdl.handle.net/11250/2995893
dc.description.abstractWell log data formats remain rigid, outdated, and difficult to understand and to work with. The amount of available documentation is limited, thus much of the valuable information held in these data formats is left behind inside the files. In recent years, more and more data have been acquired, and the capacity to process and use them in a wiser manner is rapidly increasing. This exponential growth has led many organizations and industries to replace or improve existing procedures and best practices about storage, access, and use of data. A clear example of this is the DLIS (Digital Log Interchange Standard) well log data format. DLIS is a complex binary file that can be read by a limited amount of software. Also, the extraction of the data can be complicated and most of them can easily be bypassed. Another critical problem when it comes to well log data is associated with preprocessing steps. In general, the preprocessing steps such as depth matching, splicing, and general data quality checking are time-consuming and very tedious tasks. However, they are also considered critical procedures that must be properly performed to avoid misleading analysis and wrong interpretations. The most crucial preprocessing step is the well log depth matching because no proper alignment of log measurements severely hinders the identification of well log correlations. Well log correlations are the basis of any petrophysical analysis and interpretation. Therefore, before any log data analysis, a proper depth matching process must be performed. This thesis investigates innovative ways to generate as much automated as possible a well log database by using Open-Source applications and well-known programming language as HDF5 (Hierarchical Data Format version 5) and Python, respectively. In this way, our proposal can also be implemented and taken further by others. The second aspect of this research is to develop fully automated well log depth matching workflows aided by machine learning algorithms. This thesis is mostly focused on finding new ways of integrating and implementing current and well-known algorithms to provide new alternatives to solve the well log depth matching problem. We developed a prototype well log database using two wells from the Ivar Aasen field in the North Sea. These two wells differ in their drilling plans, companies involved in their drilling and logging process, as well as the amount of data acquired. Therefore, they are good examples to show the potential of our proposed database. First, having a single HDF5 file that gathers all the available information held in several DLIS files for each well and presents the information in a hierarchical structure considerably simplifies the access, organization, and use of data. Second, we also investigated the feasibility of integrating crucial preprocessing steps into the database by developing Python functions. These functions extract relevant information from the database, perform well log depth matching and retrieve the updated information allocating it in the database. This interaction between the database and a depth matching workflow alleviates the downside of having multiple independent files of processed log curves, which are highly prone to get lost. Third, our proposed database takes care of any metadata associated with the principal data. Additionally, during this research, we focused most of our efforts on developing well log depth matching workflows to align raw logging while drilling (LWD) and electrical wireline logging (EWL) logs. To do this, we first created and integrated several Python functions that perform a semi-automatic flow using cross-correlation and dynamic time warping (DTW). We qualitatively and quantitatively evaluated the performance of our implementation. We visually compared well log profiles before and after depth matching, and we also compared Pearson correlation, Euclidean distance, and Predictability of the trace values to validate our results against a manual depth matching performed by a petrophysicist. The cross-correlation method showed the best results even though it is limited to a constant bulk shift. DTW is a competitive alternative and more automated than cross-correlation. However, DTW is prone to overfitting, introduces artifacts in the log measurements when excessive stretch/squeeze effects are present, and it has longer execution times that can rapidly increase with the number of logs. Also, we expanded these studies toward full automation of the workflow. Therefore, we investigated and tested deep learning techniques like one-dimensional convolutional neural network (1D CNN) to carry out well log depth matching. We trained, validated, and tested seven different CNN models, which correspond to the seven different log measurements pairs gamma-ray, resistivity, P wave, and S sonic logs, density, neutron, and photoelectric factor (PEF). At this stage, we used four wells from the Ivar Aasen field. During depth inference, we used the same two wells that were used to test our previous workflow. Similarly, we evaluated the results using log profiles and Pearson correlation and Euclidean distance values before and after depth matching using CNN and cross-correlation. We showed that CNN can easily be implemented, substantially reducing the user intervention, and providing competitive results compared to cross-correlation. After showing the potential of CNN, we investigated further how multimodal machine learning techniques can be integrated into our CNN approach to aggregate the estimated depth shifts from each CNN model. This test attempts to replace the user-assisted weighted average performed in the cross-correlation workflow to determine a single common depth shift for all the log measurements simultaneously. We deployed several late fusion strategies using simple average, weighted average, linear and no linear learners, and model-level fusion. By comparing depth matching results between all the fusion strategies tested, the late fusion weighted average showed the highest Pearson correlation values between all log measurements in the same two wells in the Ivar Aasen field. Therefore, we further compared these results to cross-correlation and the manual depth matching results. We found that the Pearson correlation values for the cross-correlation and manual depth matching are higher than the CNN fused approach. However, the differences in Pearson correlation values are insignificant between the three methods. Our two topics in this research, such as the well log database and the depth matching approaches show promising results. Hence, there is plenty of room for improvements and future research based on these topics. For example, the CNN approach is now based on a simple training set, which does not include several real-data effects. Therefore, our proposed method could have a great potential to be further improved to achieve better results than the current more time-consuming depth matching methods while providing the advantage of being fully automated. The proposals in this work could substantially reduce the processing time needed to perform depth matching compared to the current state of the art on a large scale.en_US
dc.language.isoengen_US
dc.publisherNTNUen_US
dc.relation.ispartofseriesDoctoral theses at NTNU;2022:104
dc.titleWell-Log Database and Automation of Well Log Depth Matching by Using Analytical Methods and Deep Learningen_US
dc.typeDoctoral thesisen_US
dc.subject.nsiVDP::Technology: 500::Rock and petroleum disciplines: 510en_US
dc.description.localcodeDigital fulltext is not availableen_US


Tilhørende fil(er)

Thumbnail

Denne innførselen finnes i følgende samling(er)

Vis enkel innførsel