A Score-Guided Regularization Strategy-Based Unsupervised Structural Damage Detection Method
Round 1
Reviewer 1 Report
The authors have used unsupervised learning for damage detection in structure SHM on a structure available in the literature.
-First of all the paper has a good mathematical framework and the finding to are in line with the damage detection techniques. The difference between undamaged and damaged response is compared.
-Some references where anomaly scores are defined must be discussed in the literature. Defining the threshold for SHM indicators is still a problem in civil eng. AUthors can refer to the paper vibration-based damage detection of structures employing Bayesian data fusion coupled with TLBO optimisation algorithm and study/discuss the criteria for defining threshold values in damage detection.
-In conclusion, they should summarise the finding of their work. Right now its is very generic and must be modified.
-in Fig 14, why SVC result is very different from other techniques? Can authors explain their findings? What made them choose those unsupervised learning methods in their research.
-Fig. 13 graphs are the best run or average of many runs?
-What were the criteria for choosing anomaly score as a possible damage indicator. Are other indicators also available as the problem has been solved previously in several research papers which have used the same structure?
-3.2.2. Experiment setup write in a way that you are using other paper data. If I am not wrong, the sentences seem that the authors have done the experiment in lab. please charity if not correct.
-Also some literature comparison with SHM benchmark problem is needed. This is major as without it the claims are not substantiated.
-What about false positives. Can authors comment on that?
-Some literature review papers of SHM+ML can be referred to in the literature.
ML techniques for SHM of heritage buildings and ML for structural engineering etc. So that the ground for SHM + ML is clear.
-Just study the papers where the same benchmark problem is used so that the methods are clear. If your approach to unsupervised ML has salient features in comparison to previous ones.
The authors have done good work and I think with these revisions the paper can be accepted.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper presents a method based on an unsupervised technique aimed to detect structural damage of structures during SHM phase. The method is well written and presented, the aim of the study is well justified and, in this Reviewer's opinion, the paper can be accepted for publication. Few comments are in the attached PDF that authors can follow to improve the quality of their work.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The improvement is marginal. Authors should address the comments properly paying attention to each comment. The changes made are minor and do not address the comments made!!
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 1 Report
This time the introduction; discussion of results part is modified; relevance of threshold on damage detection results is written and conclusions improved. Hence it can be accepted from my end.
