Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper puts forward a novel time series anomaly detection framework, denominated as Channel Mixing and Temporal Convolution Network (CHaMTeC). This framework incorporates an inverted embedding strategy, multi-layer temporal encoding, an MSE-based feedback mechanism with dynamic thresholding.

Nevertheless, several notable concerns exist. To begin with, the overall procedure of the proposed framework is quite simplistic. Secondly, within the comparative experiments, the methods utilized lack appropriate references and appear rather outdated. Third, the introduction of this paper is not logical enough. Additionally, the proposed framework fails to exhibit a high degree of innovation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper focuses on time series anomaly detection in an industrial setting using a novel approach called ChaMTeC (CHAnnel Mixing and TEmporal Convolution Network). This method integrates an inverted embedding strategy, multi-layer temporal encoding, and a Mean Squared Error (MSE)-based feedback mechanism with dynamic thresholding to enhance anomaly detection performance, which is descried in Figure 1. The introduction gives the motivation for the topic, and the literature review section provides valuable insights into different topics essential to understanding the work. The efficacy of the suggested approach is shown on 5 data examples. The paper also shares another new data set called waterlog for time series anomaly detection, which is designed to reflect real-world industrial control system scenarios with reduced anomaly rates. Overall, this method is suitable for industrial settings where anomalies/defects are rare and challenging.

Below are my comments:

• What does ITAD in line 80 stand for? A lot of the abbreviations are never defined.

• “Feature-wise Transformation” should be “Feature-wise Transformation:”?

• What is Q in line 224?

• Also, why are the equations not numbered after Equation 13?

• In Table 1, it should be 180,000 instead of 180.000. Similarly for other values. • Can the authors add a LSTM autoencoder comparison as a baseline.

• The papers explains the different metric (F1) values used and why but can the authors comment on why the F1 score is lower compared to other time series anomaly papers using the same datasets, like example: https://www.semanticscholar.org/reader/a46b06a4b8b4deecf96a4e42cd19b4696f999e66 • For some of these common data set, usually the window size is 100, is the same in the paper? If not, why?

• How is thFactor of 2.0 derived?

• Can you comment on why the F1_CPA value is MSL data but similar to F1_PA in waterlog data? • The literature for deep learning reconstruction and autoencoder method can be enriched by adding the following paper: https://ieeexplore.ieee.org/abstract/document/10020482

• Please proofread the paper.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear Authors,

The manuscript presents a compelling approach to time series anomaly detection, particularly in the context of industrial control systems, where the identification of rare and subtle anomalies remains a significant challenge. The integration of inverted embedding, multi-level temporal encoding, an MSE-based feedback mechanism, and dynamic thresholding within the ChaMTeC framework appears well-aligned with the complexities of real-world ICS environments.

I appreciate the clarity of your methodology and the use of multiple datasets, including the carefully designed WaterLog∗ dataset, which helps ground the evaluation in realistic scenarios.

That said, I would like to offer a few suggestions that may help strengthen the paper further:

It would be valuable to consider incorporating statistical significance testing (e.g., t-tests, ANOVA) to support the performance comparisons between models. For instance, the F1CPA difference between ChaMTeC and TimeMixer on the WaterLog∗ dataset is relatively modest, and readers may benefit from understanding whether such differences are statistically meaningful. Including significance indicators in Tables 2–5 could enhance the robustness of the evaluation.

A brief discussion of which types of anomalies are most challenging for the model, and why, would enrich the analysis and offer useful insights into the framework’s limitations and areas for future improvement.

Finally, elaborating on the individual contributions of each model component (e.g., the role of inverted embedding or the MSE-based feedback loop) through ablation results or discussion would help clarify the design choices and their impact on performance.

Overall, I believe this is a well-executed study with practical relevance, and with a few refinements, it could be a valuable contribution to the literature on time series anomaly detection in industrial contexts.

Respectfully, 

Comments on the Quality of English Language

The manuscript is generally clear and understandable. A few minor grammatical improvements and editorial refinements could help improve the overall flow and readability, but the language does not hinder comprehension.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Decision: Reject

 

Summary:

I think this study challenges new time series analysis. However, I think the explanation in Section 3 “Methodology” is not specific enough. I did not understand the structure and advantages of “ChaMTeC” used in this study. In addition, there are some parts that are explained too simply, such as “Encoder (z)” in Equation 6 and “RNN(x,r,h)” in Equation 9. Please provide specific explanations.

Section 4 involves a comparison experiment, but please provide a separate section on setting up the experiment. Please also provide appropriate explanations for “LMSAutoTSF,” “iTransfotmer,” “TimeMixter,” and “PatchTST” used in the comparison. You use “F1,” “F1PA,” and “F1CPA” as metrics. Please provide a clear explanation of these indexes.

The text as the whole, I think the description of the research you have performed is not enough for the reader to understand. I would recommend adding much more to the overall explanation that it is not misleading to the reader.

Comments on the Quality of English Language

I don't think there are any major problems with English grammar. I think the expression of the sentence is not appropriate because this sentence is not enough explanation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The author has carefully revised all my concerns and I agree to publish this version.

Author Response

Thank you very much for taking the time to review this manuscript and for your valuable opinions and guidance.

Reviewer 4 Report

Comments and Suggestions for Authors

Decision: Reject

 

Summary:

I think the “Introduction” and “Literature review” are very comprehensive. However, the explanation of the methodology is not enough.

The problem is that the definition of “anomaly” is not clear, even though this is a research of anomaly detection. The problem is that it is not clear what kind of data “x” in the equation is. For the mathematical equations, expressions such as “DataEmbeddingInverted”, “RNN”, and “Concat” are used without definitions, expression which is difficult for the readers to understand is used without a definition. I think this kind of description is not possible for RNNs in Equation (11), because RNNs is commonly required setting the hyperparameters for learning.

The pictures shown on the paper have been replaced by Figure.4. In the end, I am not understanding why Figure.4 is so important in this study. I think you have not adequately explained your methodology, including calculation methods, various estimation methods to be compared, and definitions of the using dataset.

Currently, the description is very difficult for others to understand what you have performed. Please revise the sentence structure, separate the results from the methods, and improve the description of the methods.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf