두산에너빌리티 · AKB-2026-001
- 과제 단계
- PoC
- 과제 유형
- Solve
배경 & 목적
In the existing manual inspection method for detecting defects in the welding process, there was significant variance depending on the inspector's skill level. The goal was to automate defect detection using vision AI to reduce quality variance.
접근 방식
Welding defect images were collected and labeled by defect type, then used to train an object detection model. A real-time inference pipeline was constructed by integrating with on-site cameras.
핵심 성과
Achieved 94% defect detection accuracy, reducing inspection time by 60%. Significant reduction in judgment variation between inspectors.
어려웠던 점 & 해결 방식
Initially, there was severe data imbalance due to insufficient defective samples. We supplemented this by employing augmentation techniques and synthetic data generation, and adjusted the threshold based on field feedback to reduce false positives.
참고 관계
이 과제를 참고한 과제
댓글 (3)
- [데모] 지주 AI전략팀
유사 공정에 적용 검토 중입니다. 데이터 라벨링 기준을 공유해 주실 수 있을까요?
- [데모] 계열사 AI전담
정확도 94% 달성 과정에서 가장 효과가 컸던 개선 포인트가 궁금합니다.
- [데모] 팀장/임원
저희 사업장에도 동일 이슈가 있어 큰 도움이 되었습니다. 감사합니다.