Automatic detection of geometrical anomalies in composites manufacturing : a deep learning-based computer vision approach

This thesis focuses on the development of a machine learning-based vision system for quality control of composite manufacturing processes. Deep Convolutional Neural Networks (DCNNs) are used to build a real-time end-to-end solution for the complex process of draping of the fiber-reinforced cut-pieces by conducting online visual inspection. The visual inspection will ultimately help with manufacturing of double-curved composite parts such as aircraft’s rear pressure bulkhead. The developed solution provides accurate and robust measurement without the need for expensive coordinate measuring machines (CMM) in the shop floor. The development of inspection software is completed in the following two stages. In stage I, after creating a hand-labeled visual dataset acquired from a fabric layup robotic system in the German Aerospace Center (DLR), a DCNN was designed, trained and tested for image classification. Then, the idea of combining images from multiple cameras for generalization of the designed model to different wrinkle properties and environments was evaluated. The proposed method employs computer vision techniques and Dempster-Shafer Theory (DST) to enhance wrinkle detection accuracy without the need for any additional hand-labeling or re-training of the model. By the application of the DST rule of combination, the overall wrinkle detection accuracy was greatly improved. In stage II, four state-of-the-art image segmentation DCNN models (DeepLab V3+, U-Net, Mask-RCNN, IC-Net) were evaluated to accurately identify the gripper, fabric, and any probable wrinkle on a dry fiber product. The results show using a DCNN model and transfer learning can lead to acceptable results while training on a small and inaccurately annotated dataset. Also, the impact of human annotation quality on the performance of DCNN models was evaluated by comparing two human-annotated datasets. Then, an approach for detection of wrinkles at the early stages of formation was developed and evaluated. Finally, the challenges of using synthetically generated data for training the models were assessed by conducting complementary experiments. The developed solution can be practically used for visual inspection of the draping process in composite manufacturing facilities. The presented method can be readily adopted to train DCNN models using other datasets and perform visual inspection tasks in different automated manufacturing processes.Applied Science, Faculty ofEngineering, School of (Okanagan)Graduat