Contextual Auto-Encoder Anomaly Detection for Micro-Fractures

The patent applies preliminary action by training the autoencoder model in advance using only defect-free images to learn the normal characteristics of products. This pre-trained model then serves as the basis for anomaly detection, where deviations from the learned normal pattern indicate defects. This approach prepares the system beforehand with knowledge of normality, enabling more accurate detection of subtle anomalies without requiring labeled defect data during training.

The patent implements self-service by using the autoencoder model to automatically learn normal product characteristics from defect-free images without requiring manual labeling of defective samples. The model self-trains by reconstructing normal images and identifying deviations as anomalies, eliminating the need for complex data annotation processes and expert labeling of defect types, thereby reducing both complexity and cost.

The patent applies inversion by reversing the conventional supervised learning approach. Instead of training a model to recognize defects directly using labeled defective images, the autoencoder is trained to reconstruct only defect-free images. The model learns what normal products should look like, and any significant deviation from this learned pattern is flagged as an anomaly. This inverted approach enables reliable training with abundant normal images while detecting rare defects without requiring scarce labeled defective data.