MA-SPRNet: A multiple attention mechanisms-based network for self-piercing riveting joint defect detection

Abstract

Efficient detection of defects in riveted joints during the self-piercing riveting (SPR) process will help improve riveting quality. Due to the complexity of SPR defects under actual working conditions, it is difficult for traditional visual technology to detect the forming quality of SPR joints effectively. To detect SPR defects and improve the efficiency of SPR joint forming quality, we proposed a defect detection model based on a multi-attention mechanism, named Multiple Attention Self-Piercing Riveting Network (MA-SPRNet), for the detection of SPR defects. Specifically, to alleviate problems such as unclear object features in complex environments, a multi-level fusion enhancement network (MFEN) is constructed. It fuses features into each level and improves the fusion effect by adding more levels of features. In addition, to alleviate the information redundancy generated during the feature fusion process, the triple attention module (TRAM) and the efficient multi-scale attention module (EMAM) were introduced to enhance the attention of the network to SPR defective. These modules are designed to refine the attention of the network, ensuring a more targeted analysis of riveting features. In addition, the Wise Intersection over Union (WIoU) loss function is introduced, aiming to guide the network to characterize features within the region of interest and to enhance the accurate positioning of riveting defects by the network. Finally, to verify the performance of the MA-SPRNet, an SPR defect dataset was constructed, and a series of experiments based on this dataset were conducted. The detection $mA{P}_{0.5}$ of MA-SPRNet was 82.83%. The results of experiments show that MA-SPRNet effectively realizes the detection of riveted joint defects.