Enhancing AFP manufacturing with AI: Defects forecasting and classification

In this work, a novel AI-driven framework for real-time defect prediction and classification for proactive quality control is introduced. By integrating autoencoders, Long Short-Term Memory (LSTM) networks, and Convolutional Neural Networks (CNNs) with the laser profilometry data acquisition into a joint pipeline, the proposed system is able to forecast defects in automated fibre placement tapes before they fully develop, enabling early corrective actions to reduce material waste and rework time. Experimental validation demonstrated the framework’s ability to predict twist defects up to 5 mm before the defect appears under the sensor, and pucker defects 2 mm with an overall 94% accuracy, offering a substantial advantage over conventional AFP defect sensors. The proposed system represents a step towards predictive defect management in AFP, enhancing efficiency of manufacturing and final product reliability. • The predictive framework detects AFP defects before they fully manifest • CNN-based encoder extracts spatial tape features from profilometry in real time • Temporal patterns analysed using LSTM to forecast defect evolution • Defect types classified from predicted profile using a classifier neural network • Pucker and twist defects predicted 2 mm and 5 mm ahead respectively with 94% accuracy