Accurate detection of blast furnace tuyere leaks is critical for operational safety and energy efficiency in the steel industry. However, significant challenges arise from the scarcity of real-world datasets and the subtle, ambiguous nature of leak-related features. Here, we propose a cross-domain detection framework guided by a sparse set of target samples to effectively bridge the sim-to-real gap. The framework incorporates a multi-dynamic attention network designed as a feature enhancement module within the detector’s backbone. By employing a progressive serial fusion strategy, this module amplifies the discriminative representation of faint, multi-scale leak patterns. Furthermore, we introduce a region-refined domain adaptation strategy that utilizes a spatially selective adversarial focal mechanism. Unlike conventional global alignment approaches, this method leverages specific positive and negative samples through RoI-based alignment to achieve precise, region-focused domain adaptation. Extensive experiments conducted on both synthetic and real-world industrial datasets demonstrate that the proposed method significantly outperforms state-of-the-art approaches in terms of detection accuracy and cross-domain robustness.
Huang et al. (Tue,) studied this question.