在芯片制造过程中，为了确保产品质量，在晶圆加工的最后阶段，需要进行晶圆缺陷检测。晶圆中的缺陷分布往往反映了生产流程中的特定故障，能够为故障溯源提供重要信息，因此，晶圆缺陷分类对芯片制造的故障监控与良率提高发挥重要作用，也是芯片制造领域的研究热点。在处理晶圆缺陷分类问题时通常会面临数据类间不平衡和有标签样本少等问题。本文针对不同的数据场景，研究基于深度学习的晶圆缺陷分类方法，主要创新性工作如下:(1)针对一般场景，提出了基于极分块缺陷密度和残差网络的晶圆缺陷分类方法。根据晶圆数据的特点，设计极分块缺陷密度特征，并提出基于极分块缺陷 密度与残差网络的混合方法。实验表明，该方法能够提高一般场景下的晶圆缺陷分类效果。(2)针对数据类间不平衡场景，提出了焦点辅助分类生成对抗网络。该模型是辅助分类生成对抗网络的改进，在其中结合训练样本的特殊性，采用focal loss损失函数，实现自适应代价敏感学习，提高少数类样本的生成效果。基于焦点辅助生成对抗网络提出数据过采样方法，提高了数据类间不平衡场景下的分类效果。(3)针对有标签样本少的场景，提出了基于旋转滤波对比学习的晶圆缺陷分类方法。设计了基于旋转和滤波的对比学习算法，并提出了三阶段训练框架，第一阶段采用无标签样本进行对比学习，第二阶段固定编码器并采用有标签样本更 新全连接层，第三阶段利用有标签样本更新整个网络。在此框架下充分利用所有样本进行训练，提高了有标签样本少场景下的晶圆缺陷分类效果。

In the chip manufacturing process, in order to ensure product quality, wafer defect inspection is required at the final stage of wafer processing. Defect distribution in wafers often reflects specific failures in the production process and can provide important information for fault tracing. Therefore, wafer defect classification plays an important role in fault monitoring and yield enhancement in chip manufacturing process, and is also a research hotspot in the area of chip manufacturing. When dealing with wafer defect classification, there are usually problems such as imbalanced data and few labeled samples. In this paper, for different data scenarios, deep learning-based wafer defect classification methods are studied. The main innovative works are as follows:(1) For the general scenario, wafer defect classification method based on polar block defect density and residual network is proposed. According to the characteristics of wafer map, the polar block defect density feature is designed, and a hybrid method based on polar block defect density and residual network is proposed. Experiments show that the proposed method can improve the performance of wafer defect classification in the general scenario.(2) For the scenario with imbalanced data, focal auxiliary classifier generative adversarial network is proposed. The focal auxiliary classifier generative adversarial network is modified from auxiliary classifier generative adversarial network, and considering the particularity of the training samples, focal loss is used to realize adaptive cost-sensitive learning, which improves the quality of generated samples of minority class. Based on focal auxiliary classifier generative adversarial network, an oversample method is proposed, which improves the classification performance in the scenario with imbalanced data.(3) For the scenario with few labeled samples, wafer defect classification method based on rotation-filter contrastive learning is proposed. A contrastive learning algorithm based on rotation and filtering is designed, and a three-stage training framework is proposed. In the first stage unlabeled samples are used for contrastive learning to get a encoder; in the second stage, the encoder is fixed and labeled samples are used to update fully connected network; and in the third stage, labeled samples are used to updates the entire network. Under this framework, all samples are fully utilized for training, which can improve the wafer defect classification performance in the scenario with few labeled samples.