Novel convolutional neural network architecture with attention mechanisms for real-time in-situ defect detection in metal additive manufacturing

Real-time in-situ defect recognition is a big challenge in metal additive manufacturing (AM). This research proposes a novel convolutional neural network (CNN) framework integrated with attention mechanisms to tackle it. The chosen CNN framework is ResNet-18, and the three integrated attention mechanisms are Spatial Attention (SA), Convolutional Block Attention Module (CBAM), and Squeeze-and-Excitation (SE). These mechanisms improve the framework’s capacity to highlight serious flaws and filter out unnecessary background noise. 1800 high-resolution images (six defect categories) from the NEU Steel Surface Defect Dataset were used to analyse the model. The data was split into 60/20/20, and a 5-fold cross-validation was executed. Hyperparameter optimisation and data augmentation strategies were employed to improve the framework’s generalisation. Experimental evaluation revealed notable performance improvements, with the model achieving 1.0, a perfect macro F1-score, indicating that both precision and recall are perfect. Nevertheless, rational performance expectations indicate a 10–15% target enhancement above the baseline. Error analysis revealed demanding categories, including Inclusion and Patches, that require more attention. This research showcases the promising nature of attention-enhanced CNNs in industrial defect detection and directs future research in optimising real-time defect detection in AM systems.