Intelligent detection of grouting fullness defects and 3D porosity quantification in prefabricated building sleeves via multiscale Swin feature fusion and approximate uncertainty quantification

Grouted splice sleeve connections are critical to precast-concrete construction, yet concealed under-grouting defects are difficult to assess from external inspection. This study presents Swin-MFF-BUQ, a dual-pathway framework for uncertainty-aware sleeve assessment. The 2D pathway processes impact-echo and ultrasonic-array time–frequency maps and predicts an analysis-domain anomaly-support mask together with a five-class fullness grade; the mask represents signal-domain anomaly support rather than a physical-space void map. The separately parameterized 3D pathway processes industrial computed-tomography (CT) volumes and estimates physical void geometry and porosity descriptors. Both pathways use deformable-window Swin encoders, bidirectional cross-scale weighted fusion, and uncertainty-aware heads. Experiments used 240 sleeve specimens at five fullness levels, yielding 12,000 records per sensing modality, and a 36-volume CT development collection. On a specimen-held-out signal test split, the 2D pathway achieved 91.4±0.3% mIoU and 95.1±0.2% image-wise Dice, a 2.8-point mIoU improvement over the strongest of eight 2D-compatible baselines. Fullness grading reached 96.8±0.4% accuracy and 96.5% macro-F1. On held-out volumes within the CT development collection, porosity regression yielded an MAE of 0.62±0.05 percentage points, an RMSE of 0.84 percentage points, and R2=0.976. Temperature scaling reduced ECE from 0.034 to 0.021; uncertainty correlated with segmentation error (Spearman ρ=0.81), and deferring the 10% most uncertain cases increased retained-set mIoU to 93.6%. Standard 2D inference required 38 ms per 512×512 map. The findings support uncertainty-aware signal screening and CT-based volumetric characterization; independent sleeve-level validation remains necessary for the developmental CT results.