Real-time layer-by-layer defects detection in 3D-printed multi-geometry polymer components using deep learning

• Novel defect detection for 3D-printed parts using a Raspberry Pi-based system • Custom CNN model detects warping, stringing, and cracking with high accuracy. • Achieves 99.3285% accuracy in multi-defect classification, surpassing pre-trained models. • Total cycle time of 288.97 ms exhibits real-time performance of the developed system. • System improves real-time defect detection and reduces material wastage. Minimizing material wastage remains a critical challenge in extrusion-based additive manufacturing (AM). This research introduces a real-time defect detection and monitoring framework utilizing a Raspberry Pi (RPi)-based image acquisition system to identify surface defects during 3D printing and terminate defective builds at an early stage. A design of experiments (DoE) approach was employed to generate a comprehensive image dataset capturing three common defects, namely warping, stringing, and cracking, in parts fabricated using polylactic acid (PLA) and acrylonitrile butadiene styrene (ABS). The collected images were pre-processed and subsequently used to train and evaluate multiple deep learning (DL) models including DenseNet121, MobileNetV2, ResNet50, VGG16, XceptionNet, and a custom convolutional neural network (CNN). Both binary classification tasks (defect versus no defect) and multiclass classification tasks (warping, stringing, cracking, and no defect) were investigated. The custom CNN demonstrated the highest performance, achieving accuracies of 99.0071% for warping, 99.7701% for stringing, 97.840% for cracking, and 99.3285% for multiclass classification. To evaluate its feasibility for real-time deployment, the operation cycle time was determined by aggregating average latency across all processing stages, yielding 288.97 ms. This low cycle time confirms the system’s capability to detect defects almost instantaneously and to abort printing promptly, thereby reducing material wastage. Furthermore, the lightweight architecture and limited parameter count of the CNN enhance its suitability for embedded hardware deployment. The developed system provides an effective low-cost, vision-based strategy for in situ quality monitoring in AM, enabling resource-efficient and reliable production.