Cross-project defect prediction technology can use the existing labeled defect data to predict new unlabeled data, but it needs to have the same metric features for two projects, which is difficult to be applied in actual development. Heterogeneous defect prediction can perform prediction without requiring the source and target project to have the same set of metrics and thus has attracted great interest. Existing heterogeneous defect prediction models use naive or traditional machine learning methods to learn feature representations between source and target projects, and perform prediction based on it. The feature representation learned by previous studies is weak, causing poor performance in predicting defect-prone instances. In view of the powerful feature extraction and representation capabilities of deep neural networks, this study proposes a feature representation method for heterogeneous defect prediction based on variational autoencoders. By combining the variational autoencoder and maximum mean discrepancy, this method can effectively learn the common feature representation of the source and target projects. Then, an effective defect prediction model can be trained based on it. The validity of the proposed method is verified by comparing it with traditional cross-project defect prediction methods and heterogeneous defect prediction methods on various datasets.