Anomaly detection for composite manufacturing using AI models

The application of artificial intelligence (AI) in composite manufacturing offers new opportunities for automated quality inspection. However, its practical implementation is limited by the lack of in-situ imaging data and the absence of robust anomaly detection models trained specifically for defect detection during composite curing. This study aims to address these challenges by developing an innovative dataset and creating an explainable AI model tailored for anomaly detection in the composite curing process. We used a custom-built autoclave with viewports to develop a novel dataset of the composite curing process. Later, using a unique explainable AI technique, a zero-bias deep neural network (ZBDNN) model was developed by transforming the final dense layer of the standard DNN model into a dimensionality reduction layer and similarity matching layer. ZBDNN model performance was then compared against a one-class support vector machine (OC-SVM) and an autoencoder for abnormality detection. The ZBDNN model demonstrated superior anomaly detection capabilities with an accuracy of 99.71%, outperforming the autoencoder (96.26%) and the OC-SVM (92.38%). The results indicate that the ZBDNN model offers a robust and practical approach for AI-driven defect detection in composite manufacturing process.