2024 Volume 89 Issue 2 Pages 125-135
Currently, companies and institutions that have traditionally focused on oil and gas development are trying to innovate their business through digital transformation. Among them, machine learning/deep learning （ML/DL）is one of the main factors of them and can be used in various ways. Here, we have introduced more recent analysis methods for attribute analysis of seismic data, to which ML/DL has been applied for a long time and improved the extraction to create more appropriate reservoir distribution images. In the case of applying the method to a certain oil field data in Vietnam, we tried to analyze the seismic data with some targeted thin and heterogeneous sandstone reservoirs using the frequency response by spectral decomposition as the input feature. In addition to principal component analysis as a conventional approach, we applied kernel principal component analysis and autoencoder that expresses and extracts non-linear features and used such unsupervised learning analyses to image seismic facies distribution that are more suitable for interpretation. Various hyperparameters, such as the number of output principal components in kernel principal component analysis and the parameters of multi-layer perceptrons that composed the autoencoder process, were set appropriately. As a result, the dimensionality reduction of the responses from the target reservoirs is effectively achieved, and the image of its sedimentary environment is improved by clustering. In addition to this, another case of analysis for injectite sandstone in an undisclosed oil field data is also reported here to show an example of using a convolutional neural network approach in the autoencoder, providing complex distribution images of the sandstones successfully. The ML/DL analysis environment and methods used here can be applied not only to seismic data but also to general technical data, and then it is expected that efforts will be made to utilize them for improvement of technical operations.
