Modification of U-Net Network Architecture for Core Samples Analysis Automation
| Parent link: | Industrial Engineering, Applications and Manufacturing (ICIEAM): Proceedings 2025 International Conference, Sochi, Russia, May 12-16, 2025. P. 863-868.— .— Piscataway: IEEE, 2025.— 2993-4060.— 979-8-3315-1219-4 |
|---|---|
| Main Author: | |
| Other Authors: | , |
| Summary: | Title screen For oil and gas facilities, digital twins are a promising technology that can improve the quality of forecasts, enhance production safety, and determine the most optimal development strategy for enterprises. To reduce operational costs, increase production volumes, and improve oil processing efficiency, oil companies are now digitizing their assets. The development of digital field concept involves creating a model where not only wells but also the data under study, as well as the results of the studies themselves, are digitized. Furthermore, the direction of modern industrial development also implies the automation of these processes. To achieve these goals, artificial intelligence is being actively implemented. This work focuses on one aspect of automating the core analysis process, specifically the task of modifying the U-Net network architecture to improve the efficiency of working with rock samples. The proposed changes to the original architecture are as follows: adding intermediate connections between decoder layers (for more accurate transfer of features at different levels of abstraction), introducing residual blocks (to address the vanishing gradient problem), replacing 2D convolution with depthwise separable convolution (to make the model lighter and faster) and incorporating attention mechanisms (to increase the network's accuracy). Additionally, the article discusses an “online” augmentation algorithm developed as part of this work. As shown in the study, these modifications enable higher accuracy, better resistance to overfitting, and improved computational efficiency. These achievements can be further utilized in solving classification and segmentation tasks on rock samples Текстовый файл |
| Language: | English |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://doi.org/10.1109/ICIEAM65163.2025.11028233 |
| Format: | Electronic Book Chapter |
| KOHA link: | https://koha.lib.tpu.ru/cgi-bin/koha/opac-detail.pl?biblionumber=683153 |
| Summary: | Title screen For oil and gas facilities, digital twins are a promising technology that can improve the quality of forecasts, enhance production safety, and determine the most optimal development strategy for enterprises. To reduce operational costs, increase production volumes, and improve oil processing efficiency, oil companies are now digitizing their assets. The development of digital field concept involves creating a model where not only wells but also the data under study, as well as the results of the studies themselves, are digitized. Furthermore, the direction of modern industrial development also implies the automation of these processes. To achieve these goals, artificial intelligence is being actively implemented. This work focuses on one aspect of automating the core analysis process, specifically the task of modifying the U-Net network architecture to improve the efficiency of working with rock samples. The proposed changes to the original architecture are as follows: adding intermediate connections between decoder layers (for more accurate transfer of features at different levels of abstraction), introducing residual blocks (to address the vanishing gradient problem), replacing 2D convolution with depthwise separable convolution (to make the model lighter and faster) and incorporating attention mechanisms (to increase the network's accuracy). Additionally, the article discusses an “online” augmentation algorithm developed as part of this work. As shown in the study, these modifications enable higher accuracy, better resistance to overfitting, and improved computational efficiency. These achievements can be further utilized in solving classification and segmentation tasks on rock samples Текстовый файл |
|---|---|
| DOI: | 10.1109/ICIEAM65163.2025.11028233 |