KAIST Develops AI That Detects Defects Even When Smart Factory Manufacturing Processes Change

An artificial intelligence (AI) technology has been developed that can reliably detect defective products even when smart factory manufacturing processes change.

Recently, there has been a surge in the adoption of defect detection systems using AI sensor data in smart factory manufacturing sites. However, existing AI models have shown limitations in performance because they cannot understand new conditions when machines are replaced or when manufacturing processes change in terms of temperature, pressure, or speed.

In contrast, the new AI model can accurately detect defects without retraining, even when manufacturing processes change. Its performance can be improved by up to 9.42%, which is expected to help reduce operational costs and expand applicability in the industry.

(From left) JiHye Na, PhD candidate, Jaegil Lee, Professor. Provided by KAIST

KAIST announced on the 26th that the research team led by Professor Jaegil Lee in the School of Computing has developed a "Time-series Domain Adaptation" technology, which allows existing AI models to be used even when manufacturing processes or equipment change, without the need for additional defect labeling.

The "Time-series Domain Adaptation" technology enables AI models that handle time-varying data-such as temperature changes, machine vibrations, power consumption, and sensor signals-to maintain stable performance without additional training, even when the actual deployment environment differs from the environment in which the model was originally trained.

Before developing the technology, the research team focused on the fact that the confusion experienced by AI models during environmental changes is not limited to differences in data distribution, but also occurs when the pattern of defect occurrences (label distribution) changes.

To address this, the team developed a method to decompose sensor data from new processes into three components-trend, non-trend, and frequency-to identify each characteristic. This approach allows AI to analyze data from multiple perspectives, much like how humans comprehensively judge anomalies by considering pitch, vibration patterns, and periodic changes in machine sounds.

In addition, the team developed the "TA4LS (Time-series domain Adaptation for mitigating Label Shifts)" technology, which automatically adjusts prediction results by comparing the predictions of existing models with cluster information from new process data. This enables the model to recalibrate predictions that may be biased toward defect patterns from previous processes, adapting them to the new process.

Conceptual diagram of TA4LS technology developed by the research team. Provided by KAIST

This technology is highly practical because it can be easily integrated as a plug-in module into existing AI systems. Its main advantage is that it can be applied immediately with only additional procedures, regardless of the AI technology currently in use.

In experiments using four benchmark datasets for time-series domain adaptation (involving four types of sensor data with changes), the research team confirmed that accuracy improved by up to 9.42% compared to previous methods.

Notably, during the experiments, the AI demonstrated remarkable performance in autonomously correcting and distinguishing differences even when the distribution of labels, such as defect occurrence patterns, changed significantly due to process changes. This result proves that products can be produced more effectively and without defects in environments-such as smart factories-where small quantities of various products are manufactured, which is a key advantage of smart factories.

Professor Jaegil Lee stated, "The biggest obstacle to introducing AI in manufacturing is the need for retraining when processes change. Our team has developed a technology to solve this problem, laying the foundation for reducing maintenance costs and improving defect detection rates on the factory floor."

Meanwhile, JiHye Na, a PhD candidate in the KAIST School of Computing, participated as the first author of this research. YoungEun Nam, another PhD candidate, and JunHyuk Kang, a researcher at LG AI Research, contributed as co-authors. The research results were recently presented at the Knowledge Discovery and Data Mining Conference 2025, a leading event in the field of artificial intelligence and data.

This content was produced with the assistance of AI translation services.