An Improved Graph Attention Network for Semantic Segmentation of Industrial Point Clouds in Automotive Battery Sealing Nail Defect Detection

The problem

Sealing nails plug the electrolyte injection holes of automotive batteries and are laser-welded in place. Welding can leave defects such as bursts, molten beads and pinholes that compromise the battery’s safety and durability. The metal surface is highly reflective and textureless, which makes 2D imaging unreliable: it needs frequent lighting re-tuning and suffers from high false-detection rates. Traditional inspection therefore combines 2D vision to locate a defect with 3D vision to measure it, a multi-stage workflow that is complex, less efficient and accumulates errors from one stage to the next.

A 3D point cloud gives direct access to the shape of the surface, but many state-of-the-art segmentation networks, such as Point Transformer and Swin3D, rely heavily on RGB-D or intensity input. That limits their robustness when only geometry is available, which is common on reflective industrial parts. The defects themselves are also hard to learn. Defects of the same type can take very different shapes, a scan of about 100,000 points contains only about 1,000 defect points, and near a defect boundary the point coordinates barely differ from those of the surrounding surface.

How it works

LGASS (Local Graph Attention for Semantic Segmentation) segments defects directly in the 3D point cloud, so localisation and measurement happen in a single stage. It uses geometry only: the network input is each point’s surface normal, and the point coordinates determine which points are neighbours. The network is a U-Net-style encoder–decoder built around two new modules, one for attention and one for pooling.

1. Scan the weld. A 3D laser profiler (OPT-LPC20, blue 405 nm laser stripe) scans the sealing nail at a lateral resolution of about 0.05 mm. An improved RANSAC algorithm fits the reference planes of the nail and the battery cover, and a circular region around the nail’s centre isolates the weld region for inspection.
2. Watch the normals. Nearly every defect region shows clear changes in height and surface normal. Near a defect boundary, two adjacent points can have almost the same coordinates but clearly different normals, so the difference in normals between neighbours is a key indicator of a defect.
3. Local Graph Attention Filter (LGAF). Each point is linked to its k nearest neighbours. For each neighbour, LGAF correlates the point’s own feature with its feature difference to that neighbour and weights the result by their spatial distance. The resulting attention weights are used to aggregate the neighbours’ features. A large normal difference produces a high attention weight, which focuses the network on defect boundaries.
4. Spatial Attention Graph Pooling (SAG-Pooling). When the encoder downsamples, each point is scored by multiplying a learned spatial attention weight with its normal difference to a neighbour. Only the top-ranked points are kept, so points survive only if they are both spatially important and geometrically distinctive, such as those on defect edges.
5. Label every point. The decoder uses graph-convolution blocks with positional encoding to upsample back to the full point set, with skip connections from the encoder. It assigns each point one of six classes: normal surface or one of the five defect types (burst, pit, warpage, molten bead, pinhole). A class-weighted loss gives the rarer defect classes more influence during training.

Results

The authors built a new dataset for this work from real sealing nails on inspection lines, with naturally occurring defects: 271 normal and 658 defective samples covering five defect types (170 burst, 112 pit, 132 warpage, 102 molten bead and 142 pinhole; Table 1). Code and dataset are released online. LGASS was compared with GACNet, PointNet, PointNet++, Point Transformer (PTv1 and PTv3) and OA-CNN in this geometry-only setting. Over five runs with different random seeds its mIoU stayed between 78.97% and 79.61%, and it achieved the best IoU on four of the five defect classes: pit, molten bead, warpage and pinhole.

The ablations show where the gain comes from. Replacing graph convolution with LGAF adds 2.19 points of mIoU, SAG-Pooling alone adds 0.46, and the two together add 2.59 (Table 5). Normals are the better input: in the input ablation, normals only give 79.23% mIoU, against 76.94% with coordinates only and 78.31% with both (Table 7). The class-weighted loss raises IoU on the minority classes, by 2.3 points for pinholes and 2.13 for molten beads, and lifts mIoU from 78.73% to 79.26% (Table 8). Inference needs about 2.8 GB of GPU memory. The authors note that LGASS may struggle with highly concave geometry, such as recessed nails with overlapping defects, where a fixed k-nearest-neighbour graph limits long-range context, and that rare classes such as pinholes still have lower IoU than the others.