Scholarlink

<p>Additive manufacturing (AM) is widely applied in various industries, with laser powder bed fusion (L-PBF) being a leading technology for producing high-precision and complex metal components. However, the L-PBF exhibits various surface defects with complex morphologies, markedly influencing component fatigue life and mechanical characteristics. Therefore, accurate defect identification and localization in the L-PBF process are critical to ensure product quality and reliability. This paper proposes an improved faster region-based convolutional neural network (Faster R-CNN) model to enhance defect detection performance in the L-PBF process. First, this paper uses region of interest (ROI) Align to replace the traditional ROI Pooling, eliminating positional deviations of candidate boxes and improving defect localization accuracy. Next, a feature pyramid network (FPN) module is introduced to acquire more in-depth and diverse information by multi-scale features. Additionally, a spatial pyramid pooling (SPP) module is incorporated to work with FPN to enhance the model's ability to detect defects of different sizes and shapes, especially small defects that are often challenging to identify. Finally, the effectiveness of the proposed model is evaluated against various object detection frameworks, including SSD, RetinaNet, YOLOv4, YOLOv7, CenterNet, Cascade R-CNN, YOLOv8, and YOLOv10. Experimental results demonstrate that the model achieves a mean average precision (mAP) of 80.68% and a total loss of 0.4362. Compared to the traditional Faster R-CNN model, it achieves an 11.01% improvement in mAP and a reduction in loss by 0.8751. Moreover, compared to alternative algorithms, the proposed model performs better and exhibits more balanced performance between the two defect types.</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

An Improved Faster Region-based Convolutional Neural Network for Detecting Small Object Defects in Laser Powder Bed Fusion

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Yanhui Ma,
  • Shuo Li,
  • Xueqing Wang,
  • Bin Fu,
  • Lusheng Li,
  • Zhiqiong Wang

摘要

Additive manufacturing (AM) is widely applied in various industries, with laser powder bed fusion (L-PBF) being a leading technology for producing high-precision and complex metal components. However, the L-PBF exhibits various surface defects with complex morphologies, markedly influencing component fatigue life and mechanical characteristics. Therefore, accurate defect identification and localization in the L-PBF process are critical to ensure product quality and reliability. This paper proposes an improved faster region-based convolutional neural network (Faster R-CNN) model to enhance defect detection performance in the L-PBF process. First, this paper uses region of interest (ROI) Align to replace the traditional ROI Pooling, eliminating positional deviations of candidate boxes and improving defect localization accuracy. Next, a feature pyramid network (FPN) module is introduced to acquire more in-depth and diverse information by multi-scale features. Additionally, a spatial pyramid pooling (SPP) module is incorporated to work with FPN to enhance the model's ability to detect defects of different sizes and shapes, especially small defects that are often challenging to identify. Finally, the effectiveness of the proposed model is evaluated against various object detection frameworks, including SSD, RetinaNet, YOLOv4, YOLOv7, CenterNet, Cascade R-CNN, YOLOv8, and YOLOv10. Experimental results demonstrate that the model achieves a mean average precision (mAP) of 80.68% and a total loss of 0.4362. Compared to the traditional Faster R-CNN model, it achieves an 11.01% improvement in mAP and a reduction in loss by 0.8751. Moreover, compared to alternative algorithms, the proposed model performs better and exhibits more balanced performance between the two defect types.