Challenge
In high-speed battery production settings, AI inference must be completed in under two seconds per unit. To maintain consistent model performance despite environmental changes—such as lighting fluctuations or camera angle shifts during operation—a robust retraining and deployment system is required. A machine learning operations (MLOps) system was established to ensure reliable model performance, enabling the filtering of abnormal X-ray images (e.g., missing or partially captured batteries) and the accurate detection of defects such as foreign objects or electrode deformation.
Approach
To improve dataset quality and defect detection accuracy, defect criteria were redefined and applied during model development. Various CNN-based models were evaluated with consideration for the trade-off between speed and performance, and the optimal model was selected for deployment. A pre-trained base model—trained on a large dataset—was used to enable efficient learning from limited data, allowing the system to quickly adapt to evolving production conditions. The model also supports feature exploration to distinguish between normal and abnormal X-ray images in real time.
Value Delivered
AI models were deployed across four battery production lines, improving quality control efficiency through real-time defect detection. The system is now expanding to additional factories and lines, enabling scalable AI-based inspection across manufacturing sites.
