Scalable Knowledge Representation for Fault Diagnosis of Cyber Physical Systems: a Systematic Literature Review

I have carefully reviewed the revised manuscript and the authors have adequately addressed all of my previous comments and concerns. I am satisfied with the revisions made and believe that the manuscript has been significantly improved. Therefore, I recommend that the paper be accepted for publication.

This article is a major revision of a version I previously reviewed. The authors have done an excellent job at taking into account my earlier feedback, and the documented changes in the supplementary materials are very much appreciated. There remain some fundamental design choices that could not easily be changed, but the limitations flowing from these choices are now appropriately documented in the article as limitations. Justifications were added where they were previously lacking. A major improvement is the consistency of the survey result tables; the current format would make it much easier for someone to use the survey to find and pick relevant techniques and associated publications. I do still think that some of the wording when presenting the need for future work could be adjusted to further clarify that they apply specifically to fault diagnosis in CPS research; the survey filters out such techniques when they are not found to be used within this particular application domain. I would also recommend the authors take another look at their references because some are quite strange; I sometimes wondered whether a reference was used simply because a term was mentioned. In those cases, I would prefer to instead get the primary source. I have added a list of minor issues including typos below. My recommendation is a minor revision to tidy up the article before acceptance.

Minor comments:
- p2: Artificial Intelligence is sometimes capitalised and sometimes kept lower-case.
- p3: "precise mathematical equations" -> What does this mean?
- p3: "Dimention KG" -> Should be 'Dimension KG', but this is also a company's product and sales pitch. Given that there are plenty of others, why is it relevant here?
- Related Work: I usually prefer to have the author names when directly referencing a specific paper.
- Figure 1: Is phase 2.1 actually three steps?
- Table 1: 'methods of' -> 'methods for'; 'faults can' -> 'can faults'; 'underspecified' is just a type of incomplete.
- p6: "fault\footnote{}related knowledge" -> "fault-related\footnote{} knowledge"
- Figure 3: Opening quotes are the wrong font size for Semantics box
- p9: "Fig. 5 shows the application domains of the 304 papers after abstract screening and highlights the selected domains retained in this step." -> Should probably come earlier, alongside Fig 4.
- p12: Strange reference [87] for BNs.
- p12: Strange reference [25] for PCC.
- p12: "paths paths" -> "paths"
- p12: Strange reference [89] for dependency parsing. Also missing page number (since it's a book).
- p13: "ontology/ schema" -> "ontology/schema"
- p20: Definition of KR too restrictive; exists outside of fault diagnosis as well.
- p20: What is a 'semantic KR model'? Do you mean a 'KR technique/method'?
- p20: Definition of ontology is inaccurate and does not match the one used in your reference. An ontology can be seen as a formal representation of concepts and relations between concepts. It does not have to be specific to a domain (see to-level ontologies).
- p22: "several" -> "Several"
- p22: "In addition, highly structured ontological models may reduce flexibility when dealing with evolving systems, incomplete data, or unforeseen fault scenarios." -> This is at least partially handled by the open-world assumption.
- p23: Strange reference [60] for triples.
- Figure 9: "similarity-based" -> "Similarity-based"
- p29: "a summary of the prerequisites, pros and cons" -> Prerequisites not really discussed.
- p31: Strange reference [58] for the Louvain algorithm.
- p31: Strange reference [29] for the PageRank algorithm.
- p33: Broken reference for GNNs.

The authors have taken the time and addressed most of my comments sufficient. After reading the paper in its current form, I think that several things should be further addressed.

The research questions named in the introduction feel a bit unclear and generic, and it is missing how they were derived.

While the contributions are somewhat clear, they are not concretely mentioned in the introduction.

In Table 1, the keywords column is placed before the research question rather than after it.

There is a lack of a clear distinction between acquisition and extraction, making it difficult to see the difference or a clear boundary when surveying papers, especially since many papers likely fall under both categories.

The scope of the work is clarified a bit too late in the paper in Section 3.1.3 rather than in the introduction and abstract.

The discussion about forward and backward snowballing is a bit difficult to understand, and since snowballing is not a systematic approach, there is a high chance of missing papers. This raises the question of why defined keywords or search strings were not used for a more systematic search instead.

The scope of the survey goes back to 1995 but only up to 2024, meaning a lot of work might have been done in the past two years on the topic, leaving it uncertain how up to date the survey actually is. This is one of my main concerns. The authors also acknowledge this. I think that we should publish survey papers that are up to date as much as possible if we want them to be useful. It might be useful to investigate recent developments from the past few years and update the survey accordingly rather than merely pointing them out as something for the reader to consider. Could it be the case that the topic is very niche and no papers have been published since 2024? This is unclear and is not positive for the paper.

It remains a bit unclear why workshop papers have been excluded; while this is up to the authors' decision, many conferences publish workshop papers that usually feature innovative work worth noting. I see this as another limitation in addition to using only Scopus.

In Table 7, the reference column does not include the author names, which goes against good practice for citing.