Hierarchical Consistency Fusion for Cross-Domain Anomaly Detection

Industrial Anomaly Detection plays a pivotal role in aircraft engine blade manufacturing, yet current methods exhibit significant limitations in addressing dynamic domain adaptation challenges. While existing methods predominantly rely on comparative analysis between normal and anomalous samples for defect localization, their discriminative capability deteriorates substantially when training data distribution shifts occur. This paper proposes a novel Hierarchical Consistency Fusion for Cross-Domain Anomaly Detection (HCF-AD), which introduces a dual-phase optimization strategy: 1) The Hierarchical Feature Consistency Learning (HFCL) module disentangles content-style features during training to establish robust representations; 2) The Feature Distribution Fusion (FDF) mechanism aligns cross-domain feature spaces during inference while preserving critical semantic information. Experimental results demonstrate the framework's exceptional defect identification capability under complex operational conditions, accurately detecting critical blade surface anomalies including cracks and corrosion, thereby significantly enhancing detection system reliability.