Objective Long-distance pipelines, which face potential hazards such as cracking and corrosion due to illegal branch taps (oil theft) and prolonged service, require online safety monitoring with enhanced automatic defect analysis capabilities. However, existing procedures typically depend on manual interpretation of single-representation magnetic signal images, resulting in low efficiency, high subjectivity, and increased false detection and omission rates, particularly in complex interference scenarios.

Methods To address these bottlenecks and detection requirements in engineering practice, this paper proposes an intelligent pipeline defect detection method based on interactive learning of multi-representation images. First, three complementary representations are created using line graphs, grayscale images, and pseudo-color images, with a random input strategy employed to mitigate single-representation bias during mini-batch training. Furthermore, CutMix—a method that clips and splices images—is introduced to enrich local context by augmenting scarce samples. The augmented data is then fed into YOLOv11 for target detection training, utilizing C3k2 as the backbone to capture axial peaks and narrow-band textures, and C2PSA at the neck to emphasize slender structures and weak contrast boundaries, thereby yielding more robust responses to small targets and subtle textures. In the subsequent reasoning stage, a multi-representation iterative recheck mechanism is proposed to progressively suppress noise and steadily increase recall through candidate screening and sensitivity verification. Finally, to enable rapid deployment for practical detection, a real-time detection and visualization module via video streaming is developed. The established detection model is integrated with existing in-line inspection software to support online analysis, balancing latency and usability in engineering applications.

Results In experimental testing based on data from 12 in-service pipelines inspected with detectors of 508 mm and 813 mm diameters, the proposed method achieved a precision of 89.92％, a recall of 90.64％, and an mAP@0.50 of 91.91％ in typical defect detection tasks such as identifying oil-stealing branch taps. Under more stringent mAP@0.50∶0.95 criteria, it attained 78.57％. Ablation studies demonstrate that both the multi-representation random input approach and CutMix contribute to improved detection performance. Practical engineering applications further show that the method can provide consistent warnings on unannotated data, substantially reducing the cost associated with manual review.

Conclusion The proposed method provides reliable technical support for combating oil theft and ensuring the safe operation of in-service pipelines. Future efforts will focus on expanding deployment across larger pipeline networks by incorporating self-supervised learning to reduce dependence on manual annotation. In addition integration with multi-source sensor data is expected to facilitate the extension from isolated defect detection to comprehensive risk assessment and informed operation and maintenance decision-making.