| Aug 26, 2025 |
New AI system detects factory defects accurately without retrainingResearchers built an AI system that adapts to process changes, maintaining defect detection accuracy and lowering retraining costs in smart factories. |
| (Nanowerk News) Artificial intelligence is transforming modern manufacturing, but one persistent obstacle remains: models trained on factory data often fail when conditions shift. A change in machine type, production speed, or even temperature can throw off AI systems that monitor sensor data for defects, leading to sharp drops in accuracy. |
| Researchers at KAIST have unveiled a new approach that overcomes this weakness (Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, "Mitigating Source Label Dependency in Time-Series Domain Adaptation under Label Shifts"). Their method, called time-series domain adaptation, allows existing AI models to adapt to new manufacturing environments without retraining or costly relabeling of defects. In tests across multiple benchmark datasets, the system improved defect detection accuracy by as much as 9.42 percent. |
| The key lies in how the AI interprets time-varying data such as vibrations, power use, or pressure fluctuations. Instead of treating sensor inputs as a single stream, the team developed a way to break down the data into three components—trends, non-trends, and frequencies—then analyze each separately. By comparing these patterns with predictions made by existing AI models, the system can automatically correct mismatches and adjust to new defect patterns. |
| This is particularly important in industries where the types of defects change with equipment updates. For instance, in semiconductor manufacturing, the ratio of scratch defects to ring-shaped defects can shift when machinery is replaced. The KAIST technology, named TA4LS (Time-series domain Adaptation for mitigating Label Shifts), recognizes these shifts and recalibrates predictions in real time. |
| A practical advantage is its plug-and-play design. The technology can be added as a lightweight module to existing AI systems without reengineering, lowering barriers to adoption. That makes it well-suited not only for smart factories but also for applications in healthcare devices and smart cities, where sensor-driven AI must adapt to dynamic environments. |
| Professor Jae-Gil Lee of KAIST, who led the research, emphasized the significance of this advancement. “This technology solves the retraining problem, which has been the biggest obstacle to the introduction of artificial intelligence in manufacturing,” he said. “Once commercialized, it will greatly contribute to the spread of smart factories by reducing maintenance costs and improving defect detection rates.” |
| By enabling AI to stay accurate even when conditions change, the KAIST team’s work marks a step toward more resilient and cost-effective automation across multiple sectors. |
| Source: Provided by Korea Advanced Institute of Science and Technology (Note: Content may be edited for style and length) |
