In this paper, we deal with a model that classifies mixed-type defect patterns that occur during the wafer production process according to the causes and detects new previously unseen defect patterns. The wafer bin map (WBM), which can be obtained through electrical die sorting (EDS) test after wafer manufacturing, has various defect patterns depending on the cause. However, in practice, it is impossible to obtain a sufficient quantity to allow the user to apply a wafer bin map with the desired defect pattern to the DNN. In order to classify mixed-type defect patterns and detect new defect patterns in a situation where only a few samples of the single-label defect pattern exist, we suggest the denoising autoencoder that optimizes the wafer bin map to remove randomly generated noise, a model agnostic meta learning (MAML) that is suitable for few-shot learning, and segmentation method to separate complex defect patterns into regions in the wafer bin maps. The presented model uses a multi-label model to classify mixed-type defect patterns, and (N+1) classes are set to detect unseen out-of-distribution (OOD) patterns. In addition, in order to train the OOD class that is not given for training, the fake parameter of OOD-MAML is modified and improved to be suitable for use in the multi-label model. The proposed model shows high accuracy even when only a small number of single pattern samples exist as training data.