Grafted Composite Decision Tree: Adaptive Online Fault Diagnosis with Automated Robot Measurements

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Thank you for submitting your paper to Sensors journal. The paper is well written, logically organized, and demonstrates strong technical depth. The authors clearly motivate the problem of integrating robotic inspections with online monitoring systems, and they propose an innovative solution in the form of GCDTs. The methodology is detailed, the case study on a nuclear power plant cooling water system is well designed, and the results convincingly show the advantages of the proposed models over conventional decision trees.

The clarity of explanation, especially regarding the difference between MDT, LDT, and ADT models, helps readers follow the technical contributions. Figures and algorithms are used effectively, and the conclusion properly highlights both contributions and future research opportunities.

SOme questions from my side for better understanding

1. While GCDTs show improved performance on the nuclear power plant case study, how would the models scale to much larger facilities with thousands of measurable variables and complex robot navigation constraints? Have the authors considered computational or training-time limitations in such large-scale settings?

2. The framework assumes reliable access to monitored and robot-measured variables. How robust is the proposed approach if some sensors malfunction or if the robot cannot access certain measurement points due to obstacles or hardware limitations? Could the GCDT framework adaptively re-plan measurement sequences in such cases?

3. The paper mentions that GCDT ideas might be extended to ensemble methods or other ML architectures. Have the authors considered directly comparing GCDTs with popular ensemble models (e.g., Random Forests, Gradient Boosted Trees) or deep learning approaches for fault diagnosis? Such comparisons could help position GCDTs within the broader machine learning landscape.

The authors may consider modifying the paper based on the above questions, either by expanding discussion in the conclusion section or by providing supplementary results.

Or the answers to these questions can also be written in a separate sheet by the authors.

 

Author Response

Please take see the attached document of our answers to your comments. Thank you.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents an online fault diagnosis approach for industrial environments using measurement robots. The authors introduce a grafted composite decision tree (GCDT), which combines a “prior” tree based on continuously monitored variables with grafted subtrees linked to robot-measurable variables. Two variants (LDT and ADT) are compared with a conventional decision tree (MDT) using a digital twin of a nuclear power plant (iFANnpp).

Overall, the article is well written, but several points should be clarified or strengthened:

1. The “Related Studies” section cites many references on feature selection and POMDPs, but it could compare GCDT more thoroughly with other partially observable inference approaches, such as ensemble methods (random forests, gradient boosting) combined with active selection strategies. It would also be helpful to explain why GCDT is preferable to recent reinforcement learning approaches with adaptive planning, despite the known challenges (e.g., sample inefficiency).
2. The ADT algorithm is well explained, but it's still hard for someone who isn't an expert to understand. To improve clarity, it would be helpful to include a simple, easy-to-follow diagram or a concrete example that illustrates how the layers are progressively built. The paper should also provide more detail on how the depth regulation was determined—specifically, how the overfitting thresholds were selected and how these choices influenced the outcomes.
3. The authors should explain how the added Gaussian noise affected the results: why were the specific values of 0.02 and 0.05 selected? A brief sensitivity analysis could highlight GCDT’s robustness to different noise levels. It would also strengthen the evaluation to include additional performance metrics, such as the F1-score or ROC AUC, alongside accuracy and impurity. Finally, the discussion of computational complexity could be clearer—reporting, for instance, the average number of subtrees created and the time required to update each one would make the resource demands more transparent.
4. The conclusion briefly addresses prospective extensions to ensemble or non-tree models; this might be elaborated to clarify the next stages. Also, examine the limits of assuming data stationarity during robot movements: in a real-world environment, dynamic changes may occur. Develop a paragraph on the generalizability of GCDT to other industries (e.g., aerospace, renewable energy) and the challenges of multi-robot environments.
5. In the introduction, add recent figures or examples illustrating the growing use of robots for industrial diagnostics to strengthen the motivation.
6. In the Results section, define the meaning of the colors and gray bars in the legend and explain why six features were picked.
7. In Section 3.1, define whether the robot’s speed (16.67 × 10⁻³ sec⁻¹) reflects a physically accurate number or simply a normalized unit.

Comments on the Quality of English Language

Address the minor typographical errors and word breaks present in the manuscript.

Author Response

Please take see the attached document of our answers to your comments. Thank you.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

After careful consideration of the authors' responses and the revised version of the manuscript, I consider that all of my comments have been addressed satisfactorily. I therefore recommend acceptance of the manuscript for publication in its revised form.