This innovation enhances efficiency, reduces costs, and expands the adoption of AI across industries.
The Challenge of Changing Conditions
Recently, AI-based defect detection systems have been adopted in smart factory environments. However, when manufacturing processes change—due to machine replacement, or variations in temperature, pressure, or speed—existing AI models often fail, causing a sharp drop in performance.
The New AI Solution
A KAIST research team led by Professor Jae-Gil Lee introduced a time-series domain adaptation technology that maintains stable AI performance without retraining. This innovation allows AI models to continue defect detection accurately even in new or shifting environments.
How Time-Series Domain Adaptation Works
- The technology decomposes process sensor data into trends, non-trends, and frequencies.
- It enables AI to analyse multiple perspectives, much like how humans detect anomalies through sound and vibration patterns.
- The TA4LS (Time-series domain Adaptation for mitigating Label Shifts) method corrects predictions automatically, even when defect occurrence patterns change.
Easy Integration into Existing AI Systems
The solution works like a plug-in module that can be added to current AI systems without complex redevelopment. This makes it highly practical for manufacturers looking to scale AI across diverse and dynamic environments.
Real-World Performance Gains
In benchmark experiments with four time-series datasets, KAIST’s method improved defect detection accuracy by up to 9.42% compared to existing approaches. It proved especially effective in environments with changing defect patterns, such as semiconductor wafer processes.
Implications for Smart Factories
This breakthrough solves a major challenge - AI retraining costs. Once commercialized, it will allow smart factories to deploy AI defect detection more broadly, reducing maintenance costs while improving accuracy and reliability in production.
Research Recognition
The results were presented at KDD 2025, the top academic conference in AI and data mining. The project was supported by Korea’s Ministry of Science and ICT through the SW StarLab initiative.
