Battery Image Inspection Using AI Drift Detection for New Defects

The patent segments the defect detection function into two distinct AI models: the first model handles classification of known defect types, while the second model specifically segments the task of detecting distribution differences that indicate new defect types. This segmentation allows each model to specialize in its specific function, improving overall adaptability while maintaining manageable complexity through clear functional separation.

The second AI model serves multiple functions: it acts as an anomaly detector for new defect types, a trigger for retraining operations, and a quality gate that determines when the first model needs updating. This multi-functionality allows the system to maintain relatively simple structure while achieving high adaptability through the versatile second model that handles multiple aspects of new defect detection and model maintenance.

The patent implements preliminary action by continuously monitoring the distribution of incoming images using the second AI model before actual defect classification. When distribution differences are detected, the system proactively triggers data augmentation and retraining processes in advance, rather than waiting for sufficient natural defective samples to accumulate. This preliminary detection and proactive retraining ensures the model remains reliable even when training data for new defect types is scarce.

The patent uses copying by creating augmented training data through data augmentation techniques when new defect types are detected. Instead of requiring large quantities of actual defective product images, the system copies and transforms existing images to generate synthetic training samples, thereby overcoming the limitation of insufficient training data while maintaining model training quality.