Comprehensive 3-D Defect Detection of Bonding Wire Based on Point Cloud Segmentation Network and Feature Recognition

In semiconductor packaging, bonding wire transmits signal and power between the chip and the substrate, and its quality directly impacts the device’s lifespan and performance stability. However, due to the extremely small size and complex geometry of bonding wires, conventional image processing methods often fail to achieve high detection accuracy. To address this limitation, this paper proposes a defect detection method based on a point cloud segmentation network and 3D point feature recognition. Based on the point cloud characteristics obtained from the line-spectral confocal sensor, we optimize the data loading method and network architecture. Two feature extraction methods, normal vector angle variation and ellipse fitting are proposed for analyzing Cu-BW ( ball-wedge copper) and Al-WW (wedge-wedge aluminum) bonding wire. Experiments show that the accuracy of defect detection reaches 96.19% for Cu-BW and 97.43% for Al-WW. Both results meet the production requirements for bonding wire defect inspection.