Scholarlink

<p>The rapid adoption of additive manufacturing has intensified the demand for real-time, in-situ quality inspection to reduce material waste and improve production efficiency. Conventional post-process inspection methods often result in significant losses of material, energy, and time, while existing deep learning approaches primarily emphasize single-layer detection accuracy without considering resource utilization. To address these limitations, this study proposes an attention-enhanced multi-task learning framework for real-time defect detection during the 3D printing process. The proposed model simultaneously identifies defects in the current layer and predicts potential failures in subsequent layers, enabling early termination of prints likely to fail and thereby reducing unnecessary layer consumption. Unlike prior studies that rely on simple or single-geometry datasets, this work employs a comprehensive dataset containing diverse and irregular shapes to better reflect practical manufacturing scenarios. Experimental results demonstrate that the proposed Multi-Task SE-CNN consistently outperforms existing methods, achieving performance improvements of 4.44% in defect-dominant scenarios and 12.56% under balanced conditions, while significantly enhancing resource efficiency. The results confirm the effectiveness of the proposed framework in achieving a practical balance between inspection accuracy and sustainable manufacturing objectives.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Early failure detection in additive manufacturing using attention-based multi-task learning on complex multi-geometry datasets

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Keng-Pei Lin,
  • Yu-Sheng Yang,
  • Cheng-Jung Yang,
  • Tzu-Lin Chang

摘要

The rapid adoption of additive manufacturing has intensified the demand for real-time, in-situ quality inspection to reduce material waste and improve production efficiency. Conventional post-process inspection methods often result in significant losses of material, energy, and time, while existing deep learning approaches primarily emphasize single-layer detection accuracy without considering resource utilization. To address these limitations, this study proposes an attention-enhanced multi-task learning framework for real-time defect detection during the 3D printing process. The proposed model simultaneously identifies defects in the current layer and predicts potential failures in subsequent layers, enabling early termination of prints likely to fail and thereby reducing unnecessary layer consumption. Unlike prior studies that rely on simple or single-geometry datasets, this work employs a comprehensive dataset containing diverse and irregular shapes to better reflect practical manufacturing scenarios. Experimental results demonstrate that the proposed Multi-Task SE-CNN consistently outperforms existing methods, achieving performance improvements of 4.44% in defect-dominant scenarios and 12.56% under balanced conditions, while significantly enhancing resource efficiency. The results confirm the effectiveness of the proposed framework in achieving a practical balance between inspection accuracy and sustainable manufacturing objectives.