Defect detection method of weld defect decoupling detection model based on multi-order cross-attention

Abstract Ultrasonic phased array testing is widely used in weld quality evaluation, but practical detection faces challenges such as diverse defect types, susceptibility to weld structure noise and material scattered waves, low detection rate of small targets and low-contrast defects, and limited multi-scale feature fusion capability. To address these issues, this study proposes a Multi-Order Cross-Attention Decoupled Detector (MO-CA DDOD), which achieves accurate detection through a three-stage architecture: (1) A Dynamic Cross-Attention (DCA) module is introduced in the neck to perform cross-scale and cross-channel-spatial dimension weighted adjustment of features, highlighting key information and suppressing background interference; (2) A Multi-Order Gated Aggregation (MOGA) module is integrated in the head, which dynamically selects features through multi-order depthwise separable convolution and a gating mechanism, generating high-discriminative representations for IoU, category, and bounding box prediction. Experiments show that the model achieves 0.845 in mAP@0.5 on the PAUT-Welds dataset, outperforming 11 mainstream models; ablation experiments verify that the DCA and MOGA modules increase mAP@0.5 by 3.8 percentage points, and Precision and Recall by 5.2 and 7.3 percentage points, respectively. In addition, validation on the Steel Pipe x-ray and x-ray datasets confirms that the proposed algorithm has significant advantages in detection accuracy and efficiency, providing an effective solution for weld defect detection in complex scenarios.