Variational AutoEncoder to Identify Anomalous Data in Robots

Comment 

Reply 

Lack of the code which would allow reproducing the result 

We agree with the reviewer and in fact we are planning to publish paper’s code on public company’s github account.

can you compare your results with similar works 

We found no current work that would serve as fair comparison. 

better justification of using VAE instead of AE should be provided 

In subparagraph 3.2 "Monte Carlo Reconstruction" modified sentences into: 

“One of the main advantages of using a VAE instead of AutoEncoder is that having a stochastic process as part of the latent space permits to generate multiple reconstructions of the predicted signal starting from a single point in the latent space in a Monte Carlo fashion” 

And 

“Another advantage of using VAE is its robustness against noise” 

it would be interesting to visualize how different data is presented in the bottleneck and this could be used as a form of explanation of the result? This is where VAE could be helpful 

Added Figure 4 showing representation of latent space in case of training and testing data. 

Comment 

Reply 

both sections 6. Discussion and 7. Conclusion are short - maybe to expand them or join together 

Agreed, sections merged together 

acronym VAE should be avoided in title 

Agreed, title changed into “Variational AutoEncoder to identify anomalous data in robots” 

Monte-carlo -> Monte Carlo 

Agreed and fixed 

Spelling should be checked, e.g. row 124: Figures 1a, 1b illustrates; 235: Figure 4. ... reconstruction of same same  ... 

Agreed and fixed 

Comment 

Reply 

The legibility of figures 4 to 6 should be improved. 

Agreed. Figure 4b has been completely modified. 

Colors of Figures 5 and 6 have been improved using standard color palette. Moreover, legend has been modified to improve readability. 

Literature should be checked if there are no newer items. Especially from the last 18 months. 

Add reference to paper from 2020 using similar technique.  

The total number of references published within 18 months is already 3 out of 7 scientific publications. 

At least some references from recent years should be added and the literature review should be updated. 

See comment above 

Line 203-204 - with encoder’s layers dimensioned respectively [512, 256, 128, 64, 32] – one would have to formally distinguish the spelling of the number of layers from the style of citing literature. 

Agreed and modified 

As for a scientific article, it lacks equations describing dependencies, which are only described in the text descriptively. 

It is unclear what equations could be added to give value. 

Comment 

Reply 

As I understood in your research the residual signal is calculated by the difference between normal signal and measured signal. If its correct, how you can solve the challenge of uncertainties? 

No, it is not correct. In our research we make use of trained Variational AutoEncoder to reproduce known signals. The VAE robustness allows us to cope with signal noise. 

A value scoring the difference between signals reconstructed by the VAE and measured signals is calculated. 

Please add a block diagram regarding the proposed algorithm and explain step-by-step about your method there. 

Agreed. Authors added a schematic representation of how our proposed method works. For cosmetic reasons it is presented in form of pseudo-code rather than block diagram. 

n the Figures, these are based on samples or time? how many samples do you select and what is the sampling rate frequencies? 

As stated in all the figures, time is represented as sample. Sampling time added in the text. 

Please highlighted the contributions. 

It is unclear the meaning of this comment. 

The results need to be improved 

Authors already stated that one of the weaknesses of this study is the lack of real anomaly data. This is due to the high reliability of the Kinova robots. 

Based on your response in the first question, the VAE is used to improve the robustness against to uncertainty and external disturbance. How this technique improve the robustness? explain about it in the manuscript and highlighted please.

I think using the VAE for improved robustness is a distraction from the actual reasoning for the VAE use. 

This line from the manuscript explains the purpose behind using the VAE: "One of the main advantages of using a VAE instead of AutoEncoder is that having a stochastic process as part of the latent space permits to generate multiple reconstructions of the predicted signal starting from a single point in the latent space in a Monte Carlo fashion." 

However, we have added:  

"In future works we will investigate more the possibility of using statistics from multiple reconstructions of the same sample to identify anomalies or using a standard auto-encoder to simplify the process." 

To cover the sensible suggestion from the reviewer to consider using an AutoEncoder over a VAE. 

Please highlighted your manuscript's contribution(s) in the introduction," clear ".