Steelmaking Predictive Analytics Based on Random Forest and Semantic Reasoning
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis manuscript present a combination study employs both semantic reasoning for inference tasks and ML for predictive tasks. In this paper, the RF classification is re-created using rule-based reasoning.
The topic is very important to steel manufacturing, metallurgists and ML research community. The target fits the scope of the journal Applied Science. However, the authors described a lot on the methodology of the present models. The application part and results is rare. I cannot agree the publication of the present form in Applied Science since the application part is somehow missing.
The detailed comments:
1. The motivation were the method development for combination of semantic reasoning and ML. From the application background, the motivation of this study is optimization of the refurbishment process of work rolls, and optimized the removed surface stock, and finally maximize the yield of the work rolls without overworking the work rolls. This motivation should be emphasized in the introduction part.
2. The research status on this topic may be reviewed. The work rolls are under constant pressure during operation and get worn after heavy usage, requiring regular refurbishment. More detailed data on the surface stock quantity, etc. may be given for details.
3. The authors engage with domain experts and stakeholders to obtain valuable domain knowledge of cold rolling processes. The data are translated into an expert rules. As I understood, this method is somehow like a simplified version of expert system. What is the difference between the present method and expert system?
4. There are many impacting factors that affect the rolls during operation, ranging from dynamic sensor data, such as tons rolled or the chemical composition of the steel, to static data, such as the roll production information. All this data is collected. Which data are collected in details? Could you give an example of it in a cold rolling plant?
5. Besides, the authors conducted interviews and questionnaires with domain experts and plant operators to construct our domain expert knowledge rule set. How about the accuracy of those knowledge rule?
6. In results part, the application of this method, results in the cold rolling mills may be given for better application purpose.
Author Response
Thank you for taking the time to read and review our manuscript. Please view the attachment document for our full response.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThank you for contributing to industrial maintenance, specifically predictive maintenance practice. The study is exciting and valuable for the research field. However, although the title states predictive maintenance, I see little actual description and deliverables that will leave the impression that the study is dealing with predictive maintenance, more like decision-making contribution and rule-based assistance. The study would be more fruitful if it contained more descriptions and conceptual propositions within the actual framework of research and an explanation of what this study brings to the table. My suggestions, recommendations, and questions for improvement are in the following.
2. Literature review
I suggest to the authors that the literature review should be further expanded to cover predictive analytics (as a precursor to predictive maintenance) and the contributions of similar work within the ontological-based domain and semantics for predictive maintenance. The authors should also emphasize the gaps detected and the contribution to the literature in terms of their approach. For instance, following the search strings ("semantic reasoning" AND "machine learning" AND "predictive analytics"), there are at least 150+ papers dealing with the proposed topics. Hence, elaborating only on five articles sounds disingenuous and lacks depth.
3. Methods-4.Framework
As for the methodology, I would argue to add more explanation on the RF. As a specific ensemble algorithm, why is he better than other ones for this particular example? Why not use different ensemble algorithms? Also, please add more explanation if you already want to describe RF, how bagging here works, how node splitting works, and the selection of features (and subfeatures). Why not use Bagging (Bootstrap Aggregating); Boosting (e.g., AdaBoost, Gradient Boost); LightGBM; XGBoost; Stacking, Voting, Bayesian Model Averaging or Model Combination, a mixture of models, etc.?
Also, how did domain experts contribute to the model? Can you provide an explanation and role? What specific expert and in what part of semantic reasoning?
Also, I would advise the authors to switch from an 80/20 train test to 60/20/20 or 50/30/20 for the train/test/validate procedure and provide results, and if needed, conduct bootstrapping of parameters. It would also be good to provide parameter values for the model obtained to check for the validity of results alongside the dataset used for the procedure.
I would suggest major improvements.
Kind regards,
the reviewer.
Comments on the Quality of English LanguageSlight improvement to the language is required.
Author Response
Thank you for taking the time to read and review our manuscript. Please view the attachment for our full response.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have made revisions but not marked by red or yellow shadows. It took the reviewer a long time to distinguish the differences between two versions.
However, the authors described a lot on the methodology of the present models. The application part and results is rare. I would like to say the application part is still missing.
I am not satisfied with this revision.
Author Response
We thank both reviewers for their additional comments and remarks. In the following we will address the remaining concerns of the second review.
First, we would like to apologise for not highlighting changes in the previous revision. In the new revision, we are highlighting all changes, the ones relating to the first revision are marked in orange, the new changes relating to the second revision are marked in blue.
The remaining concern of the second review relates to clearly stating an application part.
- However, the authors described a lot on the methodology of the present models. The application part and results is rare. I would like to say the application part is still missing.
We thank the reviewer for insisting on this point. In fact, we agree that our applications were intertwined with methodological discussions. We have now clearly separated methodology from application, and moved concerning methodological parts from section 4 to section 3. Section 4 is now exclusively addressing the application of our proposed framework within an industrial setting, hence we renamed the section to “An Application of the Framework: Use Case of Cold Rolling” to make that clear. This section 4 introduces the cold rolling use-case and demonstrates an application of our framework. The use case is based on real world industrial data from the steel industry. Section 4 provides context about the industrial data and how it was used to train a random forest model. This includes a subsection which explains how the random forest is converted into semantic format, as well as how reasoning is applied to achieve rule-base reasoning, clearly stating the validation and limitations of the proposed methods. We also explain how domain expert knowledge is acquired and integrated as part of the framework to provide decision support for steel operators, which provides a validation of the framework based on the use case using qualitative methods with domain experts.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThank you for addressing most of the comments proposed. There is always room for improvement. In its current form, the article would be suitable for publication. I appreciate the efforts you put into the work.
Regards,
The reviewer.
Author Response
We thank both reviewers for their additional comments and remarks. In the following we will address the remaining concerns of the second review.
First, we would like to apologise for not highlighting changes in the previous revision. In the new revision, we are highlighting all changes, the ones relating to the first revision are marked in orange, the new changes relating to the second revision are marked in blue.
The remaining concern of the second review relates to clearly stating an application part.
- However, the authors described a lot on the methodology of the present models. The application part and results is rare. I would like to say the application part is still missing.
We thank the reviewer for insisting on this point. In fact, we agree that our applications were intertwined with methodological discussions. We have now clearly separated methodology from application, and moved concerning methodological parts from section 4 to section 3. Section 4 is now exclusively addressing the application of our proposed framework within an industrial setting, hence we renamed the section to “An Application of the Framework: Use Case of Cold Rolling” to make that clear. This section 4 introduces the cold rolling use-case and demonstrates an application of our framework. The use case is based on real world industrial data from the steel industry. Section 4 provides context about the industrial data and how it was used to train a random forest model. This includes a subsection which explains how the random forest is converted into semantic format, as well as how reasoning is applied to achieve rule-base reasoning, clearly stating the validation and limitations of the proposed methods. We also explain how domain expert knowledge is acquired and integrated as part of the framework to provide decision support for steel operators, which provides a validation of the framework based on the use case using qualitative methods with domain experts.