On August 26, KAIST (President Kwang Hyung Lee) announced that a research team led by Professor Jae-Gil Lee from the School of Computing has introduced a novel “time-series domain adaptation” technique. This technology enables AI models to adapt to shifts in manufacturing processes or equipment without the need for additional defect labeling or model retraining.
In smart factories, AI-powered defect detection systems rely heavily on sensor data. However, changes in the manufacturing environment—such as machine upgrades or fluctuations in temperature, pressure, and speed—often cause AI models to misinterpret the new context, leading to a sharp drop in performance.
Professor Lee’s team tackled this issue by designing a method that allows AI systems to remain accurate, even when the environment changes significantly.
The core of their innovation lies in recognizing that performance degradation isn’t just caused by shifts in raw data, but also by changes in how defects occur—a concept known as label distribution shift. For instance, a change in semiconductor equipment might lead to different types of defects (like an increase in ring-shaped versus scratch defects), which standard models aren’t prepared to handle.
To address this, the team developed a method that breaks down sensor data from new processes into three components: trends, non-trends, and frequencies. Much like how a human operator might assess machine health based on pitch, vibration, or rhythm, the AI analyzes data from multiple angles. This approach forms the basis of their new system, TA4LS (Time-series domain Adaptation for mitigating Label Shifts).
TA4LS compares the model’s predictions with clustering patterns found in the new process data. When it detects inconsistencies, such as predictions skewed toward outdated defect patterns, it automatically adjusts them to better fit the current process.
A key advantage of this technology is its simplicity and flexibility. It can be added to existing AI systems as a lightweight plug-in, without the need for complex redevelopment or retraining. This makes it particularly attractive for industries aiming to scale AI quickly and cost-effectively.
In tests using four public benchmark datasets, each representing sensor data with domain shifts, TA4LS improved defect detection accuracy by as much as 9.42 % compared to existing methods.
The technology proved especially effective when label distribution shifts were large. In such scenarios, TA4LS was able to autonomously adjust to new defect patterns, making it well-suited for smart factory environments that produce small batches of diverse products.
This technology solves the retraining problem, which has been the biggest obstacle to the introduction of artificial intelligence in manufacturing. Once commercialized, it will greatly contribute to the spread of smart factories by reducing maintenance costs and improving defect detection rates.
Jae-Gil Lee, Professor, Korea Advanced Institute of Science and Technology
The research was led by Ph.D. student Jihye Na (first author), with contributions from Ph.D. student Youngeun Nam and Junhyeok Kang of LG AI Research. The findings were presented at KDD 2025 (ACM SIGKDD Conference on Knowledge Discovery and Data Mining), one of the most prestigious conferences in AI and data science.
Journal Reference:
Na, J., et al. (2025) Mitigating Source Label Dependency in Time-Series Domain Adaptation under Label Shifts. KDD '25: Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2. doi.org/10.1145/3711896.3737050