Software defect prediction is an effective way for improving the quality of software, and the effect of software defect prediction is closely related to data sets’own characteristics. In regard of feature information redundancy and large dimension of data sets, combining with the powerful learning feature ability of deep learning, we propose a software defect prediction method based on deep autoencoder networks. This method firstly uses an unsupervised learning sampling method to do  sampling for 6 open source projects data sets to solve class imbalance problem of datasets. We then build a deep autoencoder network model through training, which can reduce the dimension of data sets. The model uses three classifiers for connection and employs the training sets with reduced dimension to train the classifiers. Finally, we use the test sets to do prediction. Experimental results show that the proposed method outperforms the basic software defect prediction model and the software defect prediction model based on existing feature extraction methods under the circumstance of the data sets with large dimension and redundant feature information. Besides, it is adaptive to different classifiers.