Unsupervised anomaly detection in composite manufacturing using autoencoders and cluster-specific thresholding

Abstract

Artificial intelligence (AI) offers promise for advancing composite manufacturing by enhancing process monitoring, efficiency, and quality while mitigating defects. Nevertheless, AI application for anomaly detection is constrained by limited real-world data and reliance on labeled datasets, necessitating frequent retraining. We propose a novel three-stage anomaly detection framework for composite curing. First, an autoencoder is trained on normal data to extract features. Next, K-means clustering groups similar patterns. Finally, a model combining Mahalanobis distance with an elliptic envelope quantifies deviations using cluster-specific thresholds. Evaluation on autoclave data with a Digital Image Correlation setup yielded an impressive detection accuracy of 99.69% overall.