Research on Multi-Type Rivet Head Defect Extraction and Classification Based on PointGhost Lightweight Network

Riveting quality inspection is critical for ensuring structural integrity and safety in aerospace, automotive, and civil engineering, as rivet defects during the riveting process may cause catastrophic failures in structural connections. This study focuses on the detection method for multi-type rivet head defects and aims to improve the performance of feature extraction and classification for various head defects. The research is carried out to develop a lightweight classification network with a Dynamic Screening Self-Attention (DSSA) mechanism for 3D point clouds. To achieve the rivet head dataset, we employ Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering to extract each target head data from the dataset of riveted plates. The head dataset can be further simplified using the Non-Maximum Eigenvalue Curvature Method (NMECM). In this way, redundant information can be reduced. The PointGhost network is then designed for the classification of head defects. It contains a sampling module with a Virtual Block Sampling (VBS) mechanism that reduces the computational complexity. In addition, there exists a feature extraction module with a Grouped Pointwise Convolution Ghost (GPC-Ghost) lightweight model that performs local and global feature learning, together with the DSSA mechanism to enhance the riveted head defects. Lastly, the severity levels of rivet protrusion and indentation are quantified using Principal Component Analysis (PCA) and the Total Least Squares (TLS) fitting algorithm. In terms of the experiment, six popular lightweight models are compared, wherein GPC-Ghost shows more significant performance, achieving a 4.31% higher mean accuracy than PointNet++, with less computational cost of 0.66 GFLOPs. Based on the comparative analysis of six attention mechanisms and seven classification networks, the PointGhost model possesses the highest mean accuracy of 99.49%, with an average misclassification rate of 1.19%. The method can balance the accuracy and efficiency effectively, demonstrating its potential for engineering inspection.