Revolutionizing NDT 4.0 with Deep Attention Learning for Anomaly Detection (DAL-AD) in Mg-based L-PBF components

Detection of anomalies in 3D-printed magnesium alloy while printing is difficult because of the reactive nature of the material. In alignment with the principles of Non-Destructive Testing (NDT) 4.0, which emphasizes the inspection of advanced manufacturing processes and fully automated systems, this work presents a novel approach for anomaly detection in additively manufactured parts. Three Mg-based alloy cubes were printed through Selective Laser Melting (SLM) at different scan rates, and X-ray Computed Tomography (XCT) scan was employed to generate the image slices of all three samples. The novel data from all three samples has been selected to segment the anomaly from the printed part. The work has incorporated an innovative approach of adding a saliency map to the model for segmenting the different 3D printed volumes. Incorporating attention layers into the U-net algorithm enhances the learning characteristics of the model by emphasizing the specific region concerning the saliency map. It was found that by using an attention layer in the model, the accuracy in the segmentation of anomalies has been increased compared to simple U-net and other transfer learning approaches as a backbone. The proposed methodology with salient connection has achieved the Dice similarity coefficient (DSC) and Intersection over union (IOU) of 98.29% and 96.67% respectively, demonstrating its effectiveness in the context of NDT 4.0 for the inspection of additively manufactured components. Further aligning the proposed DAL-AD (Deep Attention Learning for Anomaly Detection) framework with broader industrial segments such as Industry 5.0 and ISO 9000, this work enables AI-assisted, sustainable, and in-situ quality control in additive manufacturing.