Vibration data is used to detect anomalies in rotating equipment such as bearings, but the conventional RMS value and crest factor are not always the good features for anomaly detection because the detection threshold must be determined depending on the usage conditions. In this paper, we adopt a method to learn the optimum features and thresholds automatically by using an artificial neural network. Specifically, a VAE (Variational Auto-Encoder) that compresses measurement data into a latent space with a small number of dimensions is adopted, and the degree of deviation from the normal distribution in the latent space determined by VAE is used as the anomaly score. Unlike the ordinal VAE, linear processing is performed, and the number of dimensions is corrected according to the degree of correlation of the latent space axes. As the result, it is shown that anomaly detection can be performed more effectively than existing methods for synthetic data and actual measurement data.