Ensuring the quality of 3D-printed polymer parts is crucial, as defects can undermine their functionality and integrity. Warping, stringing and cracking defects can significantly affect the functionality and durability of 3D-printed parts. This study aims to compare the performance of various deep learning (DL) models in detecting these defects individually (warping/no warping, stringing/no stringing and cracking/no cracking) as well as combinedly (warping, stringing, cracking and no defect).
A Raspberry Pi-based data acquisition system was used during the printing of polylactic acid and acrylonitrile butadiene styrene on a Delta 3D printer. The investigation used a Taguchi design of experiments approach with L9 orthogonal array by considering 3 levels of each of the selected process parameters (extruder temperature, bed temperature and print speed), to generate a diverse data set of defect images, which were pre-processed for enhanced computational efficiency. DL models, namely, Dense-Net121, MobileNetV2, ResNet50, VGG16 and XceptionNet were trained using transfer learning approach for both individual and multi-class defect classification.
The models’ performance was assessed based on accuracy, loss, F1-score and receiver operating characteristics metrics. DenseNet121 achieved the highest 98.59% accuracy in warping detection, MobileNetV2 excelled in stringing detection with 99.38% accuracy and XceptionNet performed best in cracking detection at 99.32%. For multiple defect detection, MobileNetV2 outperformed with 98.90% accuracy.
This research presents a novel approach for defect detection in 3D-printed parts by comparing DL models in detecting individual defects as well as multiple defects, highlighting their capabilities for improving accuracy, robustness and real-time monitoring.
