Abstract: Wafer defect detection is a crucial step in semiconductor manufacturing,directly affecting product yield and production efficiency.Existing wafer defect detection algorithms generally have difficulties in balancing detection accuracy,lightweight design,and inference speed.Moreover,they often suffer from missed detections and misclassifications in mixed defect scenarios.Based on this,to meet the requirements of wafer mixed-type defect detection,using the YOLOv11n model as the core,performance verification experiments are conducted on single and mixed-type defect detection of wafers.Firstly,for issues such as imbalance in defect categories and missing labels in the open-source dataset,the MixedWM38 dataset is processed through data filtering,data augmentation,and data annotation to construct a wafer defect detection dataset containing 15 700samples.Secondly,the average detection accuracy of single defects is quantified using the mAP@0.5 and mAP@0.5:0.95 indicators,and the detection accuracy of mixed defects is quantified using indicators such as hamming loss and exact match.The model is evaluated for lightweighting using parameters and giga floating point operations(GFLOPs),and the inference speed was characterized by the frame per second(FPS).Finally,in the same experimental conditions,the YOLOv11n model was compared with typical object detection models such as SSD,Faster R-CNN,RT-DETR-L,and YOLOv8n.The experimental results show that the YOLOv11n model has significant advantages in all evaluation indicators,particularly achieving the optimal balance of detection accuracy,lightweight design,and inference speed.The mAP@0.5,mAP@0.5:0.95,Parameters,GFLOPs,and FPS reached 99.4%,92.4%,2.6M,6.3,and 99.3,respectively.The research achievement can provide a reliable technical solution for high-precision real-time detection of semiconductor wafer mixed-type defect.