MST-VAE: Multi-Scale Temporal Variational Autoencoder for Anomaly Detection in Multivariate Time Series

Round 1

Reviewer 1 Report

This paper explores multi-scale temporal Conv1D in combination with variational autoencoder based on a previous architecture InterFusion where inter-metric and temporal dependencies are considered. The paper is well written and organized with a proper evaluation comparing several methods using different datasets. Next, only few comments that I would consider to improve the manuscript.

Since the proposed approach is based on Interfusion, the main differences between both works should be more emphatized in order to show clearly the contributions.

It is claimed (in line 209) that dimension reduction can remove abnormal points which might be indistinguishable from normal ones, is it forced in some way? abnormal points would be also abnormal ones in a latent space? Are they noise or a anomalies? Later it is explained the use of the last observation and MCMC imputation, if it is the reason, it should be noticed in that point.


One typo found in line 272 "it it ..."

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes a practical unsupervised learning approach using Multi-Scale Temporal convolutional kernels with Variational AutoEncoder (MST-VAE) for anomaly detection in multivariate time series data. Combining short-scale and long-scale convolutional kernels to extract various temporal information of the time series can enhance the model performance. Extensive empirical studies on five real-world datasets demonstrate that MST-VAE can outperform baseline methods in effectiveness and efficiency.

1、 The contribution of the Deep Learning framework is not clear. It is described that the previous dimension reduction method can not reduce noise. How can the method in this paper reduce noise?

2、 The motivation of long-scale and short-scale is not described clearly enough.

3、 There are 12 entities in ASD and SMD respectively. It is not stated in the paper that the experimental results of ASD and SMD are the average of all 12 entities or the entity with the best effect.

4、 In Table 2,the proportion of hyper-parameter “Train test split” should be replaced by “Train validation split”

5、 Please complement the effect of window lengths of the remaining three datasets in Figure 9.

6、 Please consider the following related reference.

Intelligent Detection for Key Performance Indicators in Industrial-Based Cyber-Physical Systems, IEEE Transactions on Industrial Informatics, 2021, 17(8):5799-5809

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Paper has novel and interesting results. It can be accepted after minor revision. 

Paper need to proof read for possible spelling mistake and to improve grammatical errors.

Conclusion can be written in well organised way. 

Use uniform formatting for references. Only cited references should remain. Remove extra references.  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The author have solved the problems.