Decomposition-Based Bayesian Network Structure Learning Algorithm for Abnormity Diagnosis Model for Coal Mill Process

Round 1

Reviewer 1 Report

In the production process of the coal pulverizing some abnormalities may occur; forecasting the abnormalities is crucial for avoiding the reduction of efficiency in the process.

Related data to be analysed for the diagnosis are increasingly in quantity and variety so the research of a big-data analysis tool has become essential. Due to the learning ability and the possibility to count for the expert knowledge, Bayesian networks are largely applied as a method for the abnormalities’ diagnosis. To overcome the issue of a small scale used in such a kind of applications, some hybrid methods of learning are usually adopted. The hybrid method is made of 4 phases that are: drafting, decomposition, subgraph learning and reunion. In the first phase a large BN is sketched; then the draft is decomposed based on the related moral graph; each subgraph is learnt and, finally, recombined in a single graph.

Despite the existing of some contributions in the literature for applying this strategy, none of them may guarantee the proper learning for the diagnosis of the neighbourhoods of abnormity nodes, i.e. the central topic of such a kind of diagnostic issue. The paper, thus, offers an alternative to existing methods for improving the procedure of learning considering the learning purpose of realizing the reasoning and diagnosis for the abnormity nodes and other important nodes.

The paper deals with a good computational topic for a relevant application but I believe that the theoretical part, representing the basic for the meaning, should improve. Moreover, authors should take much more care of the understanding of the text, by introducing, for instance, the proper notation in the proper sections.

Revisions

- Section 2.1 is synthetic. I suggest explaining better the statistical tool.

- Please, insert a bibliographic reference for BN and align the terminology with it. In fact, I can catch a difference in terminology between the statistical treatment of the BN and the computer science one; also, the interpretation of independency if the graph seems to be not the same. For instance, you speak about dependency whilst the BN is a map of conditional independencies among nodes. Moreover, please connect the text with the figure 1 by some examples. For instance, “d is a child of c”. Explain that arcs are denoted by E.

- Please, specify the two souls of a BN (qualitative and quantitative parts).

- Please, explain the inferential engine power in propagating evidence by the net.

- Section 2.1 and 2.2 are not directly connected. To make the reader at his/her ease, I suggest explaining why the maximum mutual information coefficient is discussed and, therefore, to argue the structural learning phase. Distinction between score and search methods of learning an constraint based algorithms is fully absent.

- In section 3.1:

o   In equation (7) introduce the meaning of the notation II

o   I suggest to specify the conditional independency concept and its notation

o   the algorithms procedures introduce the notation G(D,E) never commented before. Please, specify the notation in section 2.1

o   explain nb_cand, sep, nb. It’s clear for a skilled reader but it may be hard for readers non expert of BNs.

o   Which type of CI test is implemented in the algorithm?

- In section 4: please specify better the simulation process: do you have generating data from the datasets? If so:

o   are you discussing average results in tables 2, 3 and 4?

o   boxplots may help in arguing the SHD variability, also

o Moreover, are the golden net representative of the starting real issue that concerns diagnosis features?

o What do you mean with “sufficient” in the row 536?

o Check capital letter at the beginning of the sentence 540, 545

o I would expect more comments of simulation results and in the application

-  Conclusion: change the word “chapter” with “paper”

-  The bibliography is not alphabetically ordered. Please, verify if required by the journal. Moreover, it seems to me that the reference Don et al. 2019 is not cited in the text. Please, verify and add it in the text or remove it from the list.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes a decomposition-based Bayesian network structure learning algorithm for abnormity diagnosis model for coal mill process. The topic in the paper is interesting and novel. Some improvements should be made. 

1. There are too many long sentences used in the paper. It is hard to read it and catch the main idea. Please polish the English. 

2. It is unclear to present Algorithms 1-3. Some necessary explanation should be provided. Moreover, these algorithms are more like description on the results. Please give some form that can be used to compute the corresponding issues. 

3. The novelties of the paper need to highlighed. The authors only provide the algorithms. However, it is unclear why these algorithms are introduced. It is necessary to compare with existing algorithms. 

4. The approach comparison should be provided. 

5. The representation of the paper should be further improved. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This work proposes a Bayesian network structure learning method for the fault diagnosis of coal mill process. The method is well presented, but the description of the problem and motivation is limited. Some comments are summarized below.

1. The title and the first sentence in introduction is about coal mill process, so this paper may focus on the application of a specific system. But in the abstract, the motivation only focuses on the limitation of the method. The authors should clarify whether the proposed approach is also suitable for other objects, and if so, the description needs to be revised. If it is only applicable to the coal mill process, authors need to highlight the problem in the diagnosis of coal mill process and why the proposed method can solve the problem better than other methods.

2. The author should provide more description for the “abnormity nodes” and “other important nodes” in the introduction. It is better to provide an example to explain these two types of nodes.

3. The author states that “However, although these methods can achieve good results based on open data sets, they fail to consider the above problems in algorithm design, and the neighborhoods of these important nodes are often the main split objects in the process of subgraph decomposition, so that subgraph learning cannot really reflect the expected needs of important nodes such as abnormity nodes” on page 2. What is the “above problems”? Please also elaborate on how this limitation affects the diagnosis of the problem.

4. More details of the coal mill process dataset should be given. Is it the real data or public data? It is better to provide a data reference. And it will be better to add a figure to show the coal mill process. The nodes should also be explained in detail with the actual situation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

In this paper, the authors proposed a decomposition-based Bayesian network structure learning algorithm for abnormity diagnosis model for coal mill process. This proposed method is verified to public data sets. However, there are several major drawbacks of this paper.

(1)   There are too many English grammar mistakes, which make it hard to evaluate the work of this research. The grammar of the whole paper should be carefully checked and revised.

(2)   The state-of-the-art of BN learning related abnormity diagnosis methods should be fully investigated in the introduction part.

(3)   In section 3, the flowchart of the proposed method should be provided.

(4)   The main contributions and the innovation points of this paper should be summarized.

(5)   In section 4, the experimental details should be depicted in the paper.   

(6)   The abstract and the conclusion part should be rewritten.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

The revised version can be accepted.

Author Response

Thank you for your positive comments.

Reviewer 4 Report

The quality of the revised manuscript has been improved much. However, the following problems should be clarified.

(1)    I am not satisfied with the English writing of this paper. The grammars of the paper should be checked and the writing should be improved.

(2)    The main contribution and the innovation points should be further improved.

(3)    What are the connections between the first experiment and the second experiment?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 4 Report

The quality of the paper has been improved. The paper may can be accepted. However, the languages and the expressions of the paper should be further polished. Moreover, the main contributions of the paper should be further clarified.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf