A Subspace Projection Approach to Autoencoder-based Anomaly Detection