A Novel Bi-LSTM-Based Defect Detection Approach for Self-Piercing Riveted Joints: Focal Loss and R-Drop Fusion

Self-piercing riveting (SPR) is widely used in modern manufacturing due to its adaptability to multi-material connections and eco-friendly properties. However, the detection of defects in SPR joints remains a critical challenge, especially given the imbalanced datasets and limited generalization of current methods. This paper proposes an improved attention-based Bi-LSTM model incorporating R-Drop and focal loss techniques to improve SPR defect detection. R-Drop addresses training-inference inconsistencies, improving generalization, while focal loss mitigates the effects of dataset imbalance by emphasizing hard-to-classify defective samples. Experiments demonstrate the superior performance of the proposed model over traditional methods based on deep learning in terms of accuracy, precision, recall, and f1 score, with robust generalization across different datasets. These findings suggest that the method is highly effective for online, non-destructive SPR defect detection.