工业制造领域大量产品由多种结构件构成，在装配过程中伴随的异物混入、零部件错接、缺失、错位等异常情况会显著影响产品性能与安全。不同于表面质量控制中的结构性异常检测问题，装配异常需要考虑柔性组件形变、遮挡、不同装配约束强度等带来的不确定性，本质上属于逻辑异常。因此，传统基于局部一致性的视觉异常检测方法较难适用。为此，本文提出一种面向装配场景的异常检测方法，能够同时应对逻辑或结构异常检测。该方法首先构建基于多代表性特征替换的记忆索引直方图描述符，通过统计重构局部特征的索引分布，从而增强模型对组件位置关系、连接关系及数量约束的刻画能力；其次，引入基于图注意力增强的全局特征记忆库，利用近邻图结构与图注意力机制对跨区域的拓扑依赖关系进行建模，实现对大尺度空间结构的表征。测试阶段，通过直方图记忆库与全局特征记忆库的两阶段匹配，实现对零部件缺失、错接以及异物混入等逻辑异常的准确检测。同时，模型在结构性异常检测方面仍保持竞争力。多个工业数据集与公开数据集的实验验证了该方法的有效性，在具有代表性的电池包数据集Fuse-Cable中取得了99.86%的图像级AU-ROC及93.78%的像素级AU-sPRO。实验结果表明该方法的异常识别与定位性能优于现有局部嵌入建模方法，展现出良好的泛化性、鲁棒性与实际应用价值。

Industrial products consist of multiple structural components. Assembly anomalies—such as foreign objects, incorrect connections, missing parts, and misplacements—can severely affect performance and safety. Unlike structural defects in surface inspection, assembly anomalies are often logical, caused by component deformation, occlusion, or varying constraints, making traditional local-consistency-based methods less suitable. This paper proposes an anomaly detection method for assembly scenarios that handles both logical and structural anomalies. The method first constructs a memory-indexed histogram descriptor based on multi-representative feature replacement, which statistically encodes the index distribution of reconstructed local features to enhance the modeling of positional, connective, and quantitative constraints. Second, it introduces a graph-attention-enhanced global feature memory bank, which employs a neighbor-graph structure and an attention mechanism to model cross-region topological dependencies and represent large-scale spatial structures. During testing, a two-stage matching process between the histogram memory and the global memory enables accurate detection of logical anomalies including missing, misconnected, and foreign objects, while maintaining competitive performance on structural anomalies. Experiments on several industrial and public datasets demonstrate the method’s effectiveness. On the battery-pack dataset Fuse-Cable, it achieves 99.86% image‑level AU‑ROC and 93.78% pixel‑level AU‑sPRO. The results show that the proposed method outperforms existing local‑embedding approaches in both recognition and localization, exhibiting good generalization, robustness, and practical value.