Online Pipeline Weld Defect Detection for Magnetic Flux Leakage Inspection System via Lightweight Rotated Network

Developing an online pipeline magnetic flux leakage (MFL) weld defect detection method deployed in industrial sites is essential to expedite the localization and repair of defects. However, most existing methods suffer from large computational burdens, which hinder their deployment on resource-limited industrial personal computers (IPCs) in sites. Additionally, weld defects in MFL signals exhibit irregular contours, and there exist strong couplings between welds and defects. To tackle these limitations, this study proposes an online weld defect detection method for MFL inspection systems via a lightweight rotated network (LRNet). Initially, an efficient feature encoder is devised to mine distinguishable representations of weld defects to clarify the coupled relations between welds and defects. Subsequently, a dual-path feature aggregation is established to enrich the semantic and localization information of the extracted features through top-down and bottom-up subnetworks. Finally, a refined rotated boundary decision is designed to accurately characterize and predict the irregular contours of weld defects. The three tailored strategies follow the lightweight design principle aimed at alleviating the model's computational burden. Experimental results on MFL data from experimental sites and the real world demonstrate that LRNet improves accuracy by approximately 10% under comparable computational costs, which facilitates its deployment in IPCs.