TYUST-PAUT and PAUT-MSFDet: a real-world dataset and detection framework for industrial PAUT weld defects

Phased array ultrasonic testing (PAUT) is widely used for internal weld inspection in pressure vessels, yet automated defect recognition remains challenging because of scale and morphology variations, weak echoes, noise artefacts, inter-class similarity, and class imbalance. This study constructs the real-world TYUST-PAUT weld defect detection dataset from pressure-vessel manufacturing sites, covering five representative defect categories: air-hole, incomplete-penetration, unfused, non-critical indication, and crack. Based on this dataset, we propose PAUT-MSFDet, a multi-scale spatial-frequency enhanced detector with category-aware soft candidate filtering. Built on Task-aligned One-stage Object Detection (TOOD), PAUT-MSFDet integrates Cross-Scale Defect-Aware Refinement (CSDR), Hybrid Spatial-Frequency Feature Fusion (HSFF), and Category-Aware Candidate Quality Filtering (CACQF) to improve weak-defect representation, spatial-frequency discrimination, and positive-sample supervision. PAUT-MSFDet achieves 82.5% mAP@0.5 and 44.5% mAP on TYUST-PAUT, improving mAP@0.5 by 8.0 percentage points over TOOD. It also reaches 158.7 FPS with 6.301 ms per-image latency. Ablation studies, category- and scale-level analyses, visualisations, supplementary weld defect experiments, and prototype-level software integration further demonstrate its effectiveness and practical feasibility for offline PAUT weld defect inspection.