Multimode Operating Performance Visualization and Nonoptimal Cause Identification

Round 1

Reviewer 1 Report

This paper presents the development steps of a new method for the process control which is capable to distinguish the transient and the steady-state stages. The results of the application of this model are supported by older experimental simulations such as Tenesee Eastman process. This paper could be published as it is. 

Author Response

Thank you very much for your comments and suggestions.

Reviewer 2 Report

The manuscripts describes a classification method to distinguish different operating conditions of a process, including a self-organizing map to visualize the result of the clustering.

The paper addresses an important topic and presents interesting results. Before recommending it for publication, I have the following two concerns:

i) The review of methods that are being used is quite poor. For example, the distinction between transition and steady-state regimes can be done by time-series analysis, for example change-point detection. Another example is the authors' comment on PCA and PLS (page 1, lines 38-42). These methods are unsupervised learning methods that aim at extracting features from data, i.e. looking for latent variables in a lower dimensional space - originally not designed for fault detection, as the authors claim. Actually, the feature-extraction algorithm the authors present starting with Eq. (4) is VERY close to what PCA is doing, resp. SVD (singular value decomposition). The authors should comment on this.

ii) The authors present the substractive clustering method, starting with Eq. (1). Could the authors describe in how far this clustering method compares to known ones, like K-Means, regression clustering or SVMs?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The authors proposed a method of multimode operating performance visualization and nonoptimal cause identification to reduce human influence and eliminate transition data during multimode process online assessment. Their method used the subtractive clustering algorithm (SCA), the multi-space principal component analysis (MsPCA), and a self-organizing map (SOM) methods with verification on the Tennessee Eastman (TE) process, in which is a subject of interest. However, the main issue is that the novelty and contribution of this paper is not clear since all used methods (i.e., SCA, MsPCA, and the SOM) are already known and extensively used. In addition, many major issues are to be carefully considered. Thus, I am obliged to reject this paper.

 

The main issue about this paper is that it hadn’t well-defined its novelty and contribution. The authors used or precisely combine three methods (SCA, MsPCA, and the SOM) with an application to a well-studied process, the TE process, in which all of these methods are well defined and not proposed by the authors. Thus, the main question is what is the novelty and contribution of this paper?

Major issues:

The cited references are very old. More recent references should be considered in the introduction such when introducing the used methods (SCA, MsPCA, and the SOM). With the ones that dealt or applied to the multimode process online assessment All variables should be defined just after they are mentioned. Also, a lot of same variables defined differently in the paper, such as J which is defined in page 4 as the number of process variable and in page 5 as the neuron, but again in page 7 as the number of manipulated variable. This is just an example, where a lot of other variables have the same issue. The authors are encourage to re-check the whole paper notation. All abbreviations should be also defined, and the paper should be self-descriptive, such as TE page 2 (it doesn’t matter if is defined in the abstract but should be defined again in the main text), A, C, D, E in table 2, etc. The same for about the trained and the tested samples in page 9. The authors are considering a 1400 samples for testing with a sample interval of 5, which means 700 samples are token from the first half (first 3500 sample) and the rest 700 samples from the second half (last 3500 sample). The problem is that the first testing 700 samples are token from the first half, which mean 700 for testing from 3500 trained that is big portion of the trained samples are considered for testing and this will for sure increase the accuracy of the results. I recommend the authors to keep considering 3500 samples (let’s say the first half) for training and for testing, they will take the samples (1400 samples) from the second half which was not considered for training. One of the objective of this method, as the authors mentioned, is to reduce the human influence, which was not clearly explained in the paper and there was not specific equation or way of how the process knowledge (optimal, average, or poor) was defined automatically without human interference, as shown in Fig. 3, and afterward.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors answered my questions and thus improved the manuscript.

One minor issue I would like to point out refers to section 2.1, where the subtractive clustering algorithm is described. The algorithm depends on a couple of hyperparameters (r_a, \epsilon,  \delta). The authors should comment on how these parameters were chosen or refer the reader to the section in the paper where this choice is addressed.

After this being done, I am happy to recommend the manuscript for publication in "Processes".

Author Response

Please see the attachment.

Author Response File: Author Response.docx