Unsupervised anomalous sound detection(ASD) is a technique aimed at identifying potential abnormal sound of machines by learning features from normal samples. However, due to the changes in machine operating conditions or background noise, ASD systems trained under conventional conditions may not be effective in new environments, which can be considered a domain generalization problem. In this paper, we propose a new method based on meta-learning to address domain generalization in ASD. By splitting source domains into meta-train and meta-test domains, our system can simulate domain shift during training and exhibits good generalization ability to target domains. Experiment conducted on the DCASE2023 task2 demonstrate that our approach performs well and surpasses existing advanced methods on most machine types.