Artificial Intelligence for Process Monitoring of Automated Fibre Placement - Real-time Defect Detection and Classification
Topic(s) : Manufacturing
Co-authors :
Gabriel BURKE (UNITED KINGDOM), Duc NGUYEN (UNITED KINGDOM), Iryna TRETIAK (UNITED KINGDOM)Abstract :
Automated Fibre Placement (AFP) has seen rising interest within the composites manufacturing industry, in part due to its ability to produce complex parts with high precision and speed. Despite its benefits, like any process, AFP is susceptible to defects. The detection of defects during AFP can be challenging, thereby requiring time-consuming manual inspection to maintain quality control. This process is often conducted post-layup and frequently takes up a large proportion, if not a majority, of the total manufacturing time. In recent years, there has been increased use of profilometry sensors to automate the inspection process. Recognising a shift towards industry 4.0, this research focuses on the use of artificial intelligence, specifically convolutional neural networks (CNNs) trained on laser profilometry data, to detect and classify defects in real-time during AFP.
To collect training data and test the system, a lab-scale AFP setup has been created. Profilometry data has been collected for defect-free samples and three types of defective samples through the manual introduction of defects in the material before layup. The resulting database of sixty unique samples of each classification (defect-free, fold, twist, pucker) forms the basis of the CNN’s training data. CNN training was conducted using data augmentation techniques alongside this database to increase the number of training items, therefore improving the system’s robustness to overfitting. Both network architecture and hyperparameter optimisation studies were conducted on the CNN to achieve high classification accuracy and minimise the computation time for each classification, making real-time capabilities possible.
The CNN system is capable of detecting and classifying various types of defects, including folds, twists, and puckers, very shortly after they occur. This rapid knowledge of the occurrence of defects allows manufacturing processes to be halted as soon as necessary, preventing the completion of defective parts which are, in industry, often discarded and non-recyclable. The proposed system therefore presents potential for reducing material wastage and increasing the sustainability of AFP manufacturing processes. The results of this study demonstrate the potential application of AI for automated inspection during composite manufacturing. Such systems not only address the limitations of traditional quality control methods, but also open avenues for enhancing production efficiency, reducing costs/wastage, and ensuring the consistent delivery of high-quality composite components.
To collect training data and test the system, a lab-scale AFP setup has been created. Profilometry data has been collected for defect-free samples and three types of defective samples through the manual introduction of defects in the material before layup. The resulting database of sixty unique samples of each classification (defect-free, fold, twist, pucker) forms the basis of the CNN’s training data. CNN training was conducted using data augmentation techniques alongside this database to increase the number of training items, therefore improving the system’s robustness to overfitting. Both network architecture and hyperparameter optimisation studies were conducted on the CNN to achieve high classification accuracy and minimise the computation time for each classification, making real-time capabilities possible.
The CNN system is capable of detecting and classifying various types of defects, including folds, twists, and puckers, very shortly after they occur. This rapid knowledge of the occurrence of defects allows manufacturing processes to be halted as soon as necessary, preventing the completion of defective parts which are, in industry, often discarded and non-recyclable. The proposed system therefore presents potential for reducing material wastage and increasing the sustainability of AFP manufacturing processes. The results of this study demonstrate the potential application of AI for automated inspection during composite manufacturing. Such systems not only address the limitations of traditional quality control methods, but also open avenues for enhancing production efficiency, reducing costs/wastage, and ensuring the consistent delivery of high-quality composite components.