Defect analysis of 3D printed object using transfer learning approaches

Abstract

Additive manufacturing (AM) is rapidly gaining traction across diverse industries, including healthcare, aerospace, and automotive, mainly due to its ability to enhance production efficiency. However, a critical challenge in AM is the early detection of defects, which can significantly reduce production costs and improve productivity. Addressing this need, our study explored the effectiveness of machine learning (ML) approaches, mainly focusing on transfer learning (TL) models, for defect detection in 3D-printed objects. We utilized a range of TL models such as visualgeometrygroup (VGG)-16, inception-residualnetwork (InceptionResNet)-V2, residual network (ResNet)-50, mobilenetwork (MobileNet)-V2, VGG19, ResNet101, extreme inception (Xception), efficientnetwork (EfficientNet)-B0, EfficientNetB7, efficientnetworkV2medium (EfficientNetV2M), and neural architecture search network large (NASNetLarge), analyzing images of 3D-printed objects across three datasets using various statistical measures. An ablation study was also conducted using adaptive moment estimation (Adam), stochastic gradient descent (Sgd), and root mean square propagation (Rmsprop) optimizers. We identified that TL models such as VGG16, MobileNetV2, InceptionResNetV2, and NASNetLarge significantly outperform others, especially when optimized with Adam and Rmsprop. Conversely, we found that models like EfficientNetB0, EfficientNetB7, and EfficientNetV2M exhibited lower performance when paired with the Sgd optimizer. Additionally, Local Interpretable Model-agnostic (LIME)-based approaches are used to provide explanations of the models’ predictions. These findings offer substantial insights into model optimization and integration, aiming to enhance AM product quality through advanced image-based monitoring and inspection.