<（左から）Ph.D候補のJihye Na、Jae-Gil Lee教授>
Recently, defect detection systems using artificial intelligence (AI) sensor data have been installed on smart factory manufacturing sites. However, if the manufacturing process changes due to machine replacement or variations in temperature, pressure, or speed, existing AI models fail to properly understand the new situation and significantly reduce performance. KAIST researchers have developed AI technology that can accurately detect defects in such situations without retraining, achieving performance improvements of up to 9.42%. This achievement is expected to contribute to reducing AI operational costs and increasing applicability in a variety of areas, including smart factories, healthcare devices, and smart cities.
On August 26, Kaist (President Kwang Hyung Lee) announced that a research team led by Professor Jae-Gil Lee of the Computing School has developed a new “time series domain adaptation” technology that will allow existing AI models to be utilized without additional defect labels, even when manufacturing processes or equipment changes.
Time-series domain adaptation technology allows AI models that handle time-varying data (e.g. temperature changes, machine vibrations, power usage, sensor signals, etc.) to maintain stable performance without additional training, even when the training environment (domain) and the actual application environment are different.
Professor Lee's team paid attention to the fact that the core problem of AI models being confused by changes in the environment (domain) is not only in differences in data distribution, but also in changes in defect occurrence patterns (label distribution) themselves. For example, in semiconductor wafer processes, the ratio of ring-type defects to scratch defects can vary with equipment changes.
The researchers have developed a method to separate the new process sensor data into three components: trend, non-trend, and frequency, and analyze the properties individually. AI has now been enabled to analyze data from multiple perspectives, just as humans combine periodic changes in pitch, vibration patterns, and mechanical sounds to detect anomalies.
In other words, the team applies a method of applying a method of automatically modifying predictions by comparing results predicted by existing models with clustering information in new process data. This allows predictions biased towards defect-inducing patterns of existing processes to be accurately adjusted for the new process.
In particular, this technology is extremely practical as it can be easily combined like additional plug-in modules inserted into existing AI systems without the need for separate complex development. This means that it can be applied immediately with just a simple additional step, regardless of the AI technology currently in use.
<図1。研究チームが開発した「TA4LS」技術の概念図。新しいプロセスからのセンサーデータは、同様のパターンに従ってコンポーネント（トレンド、非トレンド、および周波数）によってグループ化されます。これらを既存のモデルによって予測される欠陥の傾向と比較し、不一致を自動的に修正することにより、テクノロジーはプロセスが変更された場合でも高性能を維持します。 >
In experiments using four benchmark datasets for time series domain adaptation (i.e., four types of sensor data with changes), the research team has improved accuracy by up to 9.42% compared to existing methods.[TT1]
In particular, when changes in the process caused significant differences in label distribution (e.g., defect occurrence patterns), AI showed significant performance improvements by autonomously correcting and distinguishing such differences. These results proved that the technology can be used more effectively without defects in environments that produce small batches of various products, one of the main advantages of smart factories.
Professor Jegil Lee, who oversees the research, “The technology solves the biggest obstacle to the adoption of artificial intelligence in manufacturing, the retraining problem. When commercialized, it will significantly contribute to the widespread adoption of smart factories by reducing maintenance costs and improving defect detection rates.”
This study was conducted in a PhD, Jihye Na. Kaist students are completing their PhD along with Youngeun Nam. Junhyeok Kang, a student and researcher in LG AI research, is co-author. The results of the study were presented at KDD (ACM SIGKDD Conference on Knowledge Discovery and Data Mining) in August 2025.
*Paper title: “Reducing source label dependencies in time series domain adaptation under label shifts”
doi: https://doi.org/10.1145/3711896.3737050
This technology was developed as part of the development of the SW StarLab project (RS-2020-II200862, DB4DL: DB4DL development), the original technology development program for the SW computing industry.
/Public release. This material of the Organization of Origin/Author is a point-in-time nature and may be edited for clarity, style and length. Mirage.news does not take any institutional position or aspect, and all views, positions and conclusions expressed here are the views of the authors alone.
