Abstract: Since existing methods for wafer map unknown defect detection require new methods to retrain models, making it difficult to reuse existing classification models, a post-hoc unknown defect detection method based on Hyperdimensional Computing (HDC) was proposed. Without modifying the existing classification model, a Hyperdimensional Positional Binding-Rotate Fusion (HPBRF) module and a class hypervector retraining module based on HDC Cosine Similarity Loss (HCosineLoss) were introduced. First, position encoding vectors were generated for multiple layers of features and embedded into feature vectors in a hyperdimensional space through a binding-rotate operation. Secondly, the hypervectors of each layer were fused via a bundling operation to obtain a more informative feature hypervector. Finally, class hypervectors were retrained using HCosineLoss with the fused feature hypervectors to align each class hypervector with its corresponding feature hypervector and separate different class hypervectors. Experiments on the WM811K dataset show that the proposed method improves the average AUROC (Area Under the Receiver Operating Characteristic Curve) by 12.52, 11.70, 6.71, and 6.58 percentage points compared to ASH (Activation SHaping), LTS (LogiT Scaling), KNN (K-Nearest Neighbors), and Gram methods, respectively. Efficiency analysis shows that the average processing time is 0.27 ms, the peak memory usage is 461.82 MB, and the additional storage overhead is only 36 KB. The proposed method improves the detection capability of unknown defects through posterior ensemble with low additional computational and storage overhead.

Key words: HyperDimensional Computing (HDC), unknown defect detection, out-of-distribution detection, feature fusion, loss function, position encoding, wafer map defects

摘要： 针对晶圆图未知缺陷检测中现有方法需引入新方法重训练模型，难以复用已有分类模型的的问题，提出一种基于超维计算（HDC）的后验式未知缺陷检测方法，在不修改已有分类模型的前提下，引入超维位置绑定旋转融合（HPBRF）模块与基于超维余弦相似度损失（HCosineLoss）的类别超维向量重训练模块。首先，HPBRF模块通过为多层特征生成位置编码向量，并通过绑定旋转操作将其嵌入映射至超维空间的特征向量中。其次通过捆绑操作融合各层超维向量得到信息更丰富的特征超维向量。最后使用HCosineLoss与融合后的特征超维向量重训练类别超维向量，促使类别超维向量与对应的特征超维向量对齐，不同类别超维向量相分离。在WM811K数据集上实验结果表明，相较于ASH(Activation SHaping)、LTS(LogiT Scaling)、KNN(K-Nearest Neighbors)和Gram等方法，所提方法平均接受者操作特征曲线下面积（AUROC）分别提升12.52、11.70、6.71和6.58个百分点。效率分析显示，样本平均处理时间为0.27 ms，峰值内存占用461.82 MB，额外存储开销仅36 KB。所提方法通过后验式集成，提升了对未知缺陷的检测能力且额外计算与存储开销较低。

关键词: 超维计算, 未知缺陷检测, 分布外检测, 特征融合, 损失函数, 位置编码, 晶圆缺陷