269-278 2026 [춘계학술대회]
| 제목 | 딥러닝 기반 자동차 산업 품질 평가를 위한 SCM420 미세 조직 등급 분류 연구 |
|---|---|
| 분야 | |
| 언어 | English |
| 저자 | 선우재호(에이치엘만도) |
| Key Words | |
| 초록 |
This study proposes a deep learning-based automated quality assessment system for normalized steel- chromemolybdenum420 low-carbon steel microstructure. Traditional quality control has been limited by the subjective visual inspection of operators. To address this, we developed a binary classification model using both high-resolution images captured in a laboratory environment and industrial images collected from partners in actual industrial environments to evaluate quality based on banded structures. To overcome the limitations of the small dataset, a 4-fold static data augmentation strategy was applied. For model optimization, we conducted a comparative analysis between a custom convolutional neural network and various ResNet architectures (18, 34, and 50). Experimental results showed that ResNet18 achieved a validation accuracy of 96.47% on the internal dataset; however, when predicting real industrial data collected from a different environment, the accuracy significantly decreased due to discrepancies in equipment and specimen preparation. To mitigate this, the study observed that errors occurring in different environments could be effectively resolved with a small amount of data through transfer learning. This suggests the technical feasibility of rapid industrial optimization even during the introduction of new equipment or sudden changes in data collection environments. Furthermore, training the model on a unified dataset showed high classification performance, suggesting the potential for a scalable inspection system that maintains reliability across diverse production sites. This research provides a practical foundation for implementing a robust, automated quality management process that is resilient to variations in industrial environments. |
| 원문(PDF) | 다운로드 |