Traceable Point-Cloud Inspection of Variable-Diameter Steel Pipes Using Multi-Scale Residual Projection and Field-Oriented Joint Calibration

Abstract

Surface inspection in continuous steel-pipe production is increasingly expected to provide timely and traceable quality information for downstream review and disposition, rather than only identifying visible defects. This requirement becomes challenging when changes in pipe specification simultaneously alter point-cloud scale and cylindrical surface geometry, while surface disturbances, processing-time constraints, and asymmetric field costs further affect inspection decisions. To address these coupled issues, this study develops a traceable 3D inspection framework that connects point-cloud organization, defect-sensitive representation, and production-oriented detector calibration. Source-index-preserving regularization first converts variable-size point clouds into a unified representation while retaining access to the original data. A multi-scale residual projection method is then introduced to suppress cylindrical trends and slowly varying surface undulation while preserving local geometric deviations for efficient detection. For deployment, class-specific confidence thresholds, selective 3D candidate review, and overlap-graph consolidation are jointly calibrated according to production priorities related to missed defects, review workload, and computational cost. Experiments using data acquired from an actual steel-pipe production environment show that the proposed method improves defect-sensitive representation while maintaining a P95 front-end latency of 83.62 ms. The calibrated workload-balanced profile further increases defect recall from 87.28% to 92.76% compared with the regular best model. The results demonstrate that source data and production requirements can be incorporated into a unified inspection chain, providing a practical route for adapting general vision detectors to variable-specification, continuously operated industrial inspection.