BiLSTM-Based Fault Anticipation for Predictive Activation of FRER in Time-Sensitive Industrial Networks
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsOverall, the text is well-written and organized, the contribution is clear, and the results are discussed in depth.
I have some doubts about the basic idea of the paper. The proposed approach indeed reduces redundancy overhead, but it comes at the cost of weakening the deterministic reliability guarantees that are the very motivation behind FRER. This raises a fundamental question: is the trade-off acceptable in safety-critical industrial deployments, where worst-case reliability is often mandatory? The authors should clarify under what conditions predictive activation can be justified, and how the residual risk of unanticipated faults is handled.
Also, as a future work, I suggest the authors to quantify the benefits of reduced redundancy overhead, because I guess in many use-cases, it's not worth it.
Comments for author File: 
Comments.pdf
Author Response
Reply is available in details in the attached file
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis study introduces a predictive fault-tolerant networking framework that duplicates critical traffic only when needed. Using a machine learning model trained on indicators such as delay, variation, and retransmissions, the system anticipates faults and activates redundancy. Once stability returns, redundancy is removed. Experiments show improved link efficiency while maintaining reliability, demonstrating effective anticipatory fault protection.
The paper is well structured and well written. However, this reviewer wants to see the consideration before the works can be recommended for publication:
+ In Section 2.2 the authors describe several temporal machine learning algorithms that have been used in the literature which includes LSTM, GRU, Transformer, Informer, etc. However, what is the motivation for the authors to use LSTM for that work, in other words, what specific characteristics of this problem make LSTM suitable over others? No justification is given. The authors need to articulate why they think LSTM will bring better results.
+ Since the paper employs LSTM, it should consider citing more works that use LSTM and describe the logic behind using LSTM (e.g., https://doi.org/10.3390/s23115348, https://doi.org/10.1109/JSEN.2021.3105226, etc)
+ In the simulation, the paper uses four cases: No Faults, Rare Faults, Base Faults & Complex Faults. What was the number of samples in each category? This information can be included in Table 1.
+ The paper listed a few metrics, but omitted TN, overall accuracy, F1-score, etc? Table 2 can be expanded.
+ In machine learning, for model evaluation, it is customary to use 5-fold or 10-fold cross-validation to get a more acceptable assessment of model performance. It is not clear from the writing whether they have used cross cross-validation technique. Otherwise, model performance evaluation remains questionable.
+ Whatever system is designed, it would be as good as the fault data is. Therefore, the reliability and trust of data from sensors that collect fault-related parameters in practical implementation in a real industry environment is a vital issue, which is not discussed anywhere in the article. Please consult https://doi.org/10.1109/TR.2024.3416967 to see how the reliability of sensors matters, and this point should be discussed with citations.
+ Adding a ROC with and without the proposed method will show further performance differentiation.
Author Response
Reply is available in details in the attached file
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsJournal IoT (ISSN 2624-831X)
Manuscript ID IoT-3841705
Type Article
Title BiLSTM-Based Fault Anticipation for Predictive Activation of FRER in Time-Sensitive Industrial Networks
Authors Mohamed Seliem * , Utz Roedig , Cormac Sreenan , Dirk Pesch
Special Issue AIoT-Enabled Sustainable Smart Manufacturing
This paper addressed the problem of redundancy overhead in Frame Replication and Elimination for Reliability (FRER) in Time-Sensitive Networking (TSN). To overcome the problem of persistent redundancy even under nominal conditions, the authors proposed a predictive FRER activation framework using a Key Performance Indicator (KPI)-driven bidirectional Long Short-Term Memory (BiLSTM) model. By analyzing multivariate KPIs such as latency, jitter, and retransmission rates, the model anticipates potential faults and proactively activates FRER, while redundancy is deactivated once the KPIs recover. Generally, the paper is well organized. I have the following comments for the authors’ consideration:
- The Introduction could be refined by first introducing the broader background of TSN and reliability challenges, then narrowing to the redundancy issue and the BiLSTM-based predictive solution.
- The selection of the LSTM model should be explained in more details since this model has been well studied before. The improvement from LSTM to BiLSTM should be explained further.
- More justification is needed for the selection of softmax prediction in (3). How are the weights in (4) determined and what if other values are selected?
- Cases studies are studied and analyzed to verify the effectiveness of the proposed method. However, some quantitative index or results should be calculated to verify the effectiveness of the proposed method.
- The discussion part in Section 8 should be discussed together with the experiment results.
Author Response
Reply is in details in the attached file.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have addressed all my concerns, and the quality of the paper is now high and acceptable for publication.
One reference does not appear as a number when referred to within the body of the paper. Please fix that LaTeX error in formatting.
