Towards efficient deep autoencoders for multivariate time series anomaly detection