Anomaly detection is a key task in Prognostics and Health Management (PHM) system. Specially, in most practical applications, the lack of labels often exists which makes the unsupervised anomaly detection very meaningful. Furthermore, unsupervised anomaly detection is also considered as a challenging task due to the diversity and information-lack of data. Variational Auto-Encoder (VAE) is a stochastic generative model which is designed to reconstruct input data as close as possible. In this paper, VAE is applied to extract valuable features for the unsupervised anomaly detection tasks. Comparison experiments are conducted on KDD CUP 99 dataset and MNIST dataset. Results show that features obtained by VAE can make unsupervised anomaly detection approaches perform better. Auto-Encoder (AE) and Kernel Principle Component Analysis (KPCA) were applied as comparisons. The result demonstrates that VAE gets best performance among them.