Attention-LightNet: A Lightweight Deep Learning Real-Time Defect Detection for Laser Sintering

Part defects in additive manufacturing (AM) operations like laser sintering (LS) can negatively affect the quality and integrity of the manufactured parts. Therefore, it is important to understand and mitigate these part defects to improve the performance and safety of the manufactured parts. Integrating machine learning to detect part defects in AM can enable efficient, fast, and automated real-time monitoring, reducing the need for labor-intensive manual inspections. In this work, a novel approach incorporating a lightweight Visual Geometry Group (VGG) structure with soft attention is presented to detect powder bed defects (such as cracks, powder bed ditches, etc.) in laser sintering processes. The model was evaluated on a publicly accessible dataset (called LS Powder bed defects) containing 8514 images of powder bed images pre-split into training, validation, and testing sets. The proposed methodology achieved an accuracy of 98.40%, a precision of 97.45%, a recall of 99.40%, and an f1-score of 98.42% with a computation complexity of 0.797 GMACs. Furthermore, the proposed method achieved better performance than the state-of-the-art in terms of accuracy, precision, recall, and f1-score on LS powder bed images, while requiring lower computational power for real-time application.