An Intelligent Generation Method for Building Fire Protection Maintenance Work Orders Based on Large Language Models

1. Summary

Thank you very much for your constructive and detailed comments. We have carefully studied your suggestions and made corresponding modifications. Specifically, we have:

1. The mapping from specification clauses to work orders is explained in the "Related Work" section. A diagram is added to compare the "clause-work order" mapping, and the maintenance work order in this study is defined as a structured six-tuple;

2. The incorrect usage of "Few-Shot Fine-tuning" is corrected to "In-Context Learning (ICL)" in the paper;

3. In Part IV of the paper, it is compared with other LLMs (GPT-4o, Llama-3, Qwen-2.5) for a more comprehensive comparison.

Detailed responses are as follows.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The paper fails to clearly define the core object of study: what exactly is a “maintenance work order” in the context of fire protection? The authors do not explain its purpose, typical structure, or real-world format (e.g., tabular, checklist, or structured digital record). A concrete example of a work order would greatly improve accessibility.

Response 1: Agree. We apologize for the lack of clarity in the initial definition. We have revised the "Materials and Methods" section to explicitly define the research object.

Revisions:

1. In Section 3, we formally defined the "Maintenance Work Order" not as a simple task list, but as a structured 6-tuple（On page 8, between lines 278 and 279） W={D_type,D_subtype,A,R,F,N} covering device type, subtype, action, requirement, frequency, and number constraint.

2. We added Figure 1 ("Mapping and comparison of clauses to work order information") in Section 2 （On page 4, above line 124）to visually demonstrate how an unstructured regulatory clause is transformed into a structured work order.

Comments 2: Lack of Clear Problem Formalization, it doesn't formally define the input-output of their method. A formal problem statement—following conventions in information extraction or structured prediction literature—would greatly strengthen the methodological foundation.

Response 2: Agree. We have strengthened the methodological foundation by adding a mathematical formalization of the problem.

Revisions:

In Section 3, we added Equation (1) （On page 8, between lines 278 and 279）to formally define the target structured output W. Furthermore, in Section 3.4.1（On page 17, line 556, and line 564）, we introduced Equation (2) and (3) to mathematically define the expansion logic function f_expand(C) used to map a single clause  into multiple independent work order records R, ensuring a rigorous definition of the method's input-output relationship.

Comments 3: Poor Methodological Clarity, the proposed method  is described using dense, domain-specific jargon without sufficient explanation or intuitive examples. 

Response 3: Agreed. We sincerely apologize that the previously submitted manuscript was not clear enough in its methodological description. The revised article framework is now ready and is presented in Section 3（Pages 8 to 20） of this paper.

Revisions:

1. Visual Illustration: We added Figure 2 (Overall Architecture)（Page 9） and Figure 5 (FS-RAG Framework Design) （Page 13）to visually present the technical flow, making the "dense" text easier to follow.

2. Concrete Examples: In Section 3.2.2（Page 11, line 383）, we detailed the structure of the FSKB with Figure 4, showing actual examples of "Knowledge Cards" (e.g., KB-Devices, KB-Actions). This provides an intuitive explanation of the domain-specific terms used in the retrieval process.

3. Text Simplification: We have refined the description of the FS-RAG mechanism in Section 3.3 （From page 13 to page 15）to clearly explain how "retrieval" and "generation" interact, reducing unnecessary jargon.

Comments 4: Misuse of “Few-shot Fine-tuning”, The paper repeatedly refers to a “few-shot fine-tuning strategy,” which is technically inaccurate. In machine learning, fine-tuning implies updating model parameters using gradient descent on labeled data, while few-shot learning typically refers to in-context learning without parameter updates. 

Response 4: Agree. Thank you for correcting this terminology error. We clarify that our method relies on API-based inference without updating model parameters.

Revisions:

We have replaced the term "Few-shot Fine-tuning" with "Few-shot In-Context Learning (ICL)" or "In-Context Learning strategy" throughout the manuscript, including the Abstract, Introduction, and Section 3. We explicitly state in the Introduction that due to the scarcity of labeled data, we adopt an In-Context Learning (ICL) strategy rather than fine-tuning .

Regarding the "In-Context Learning (ICL) strategy", we have provided a detailed overview in Section 3.4, "In-Context Learning Strategy Based on Prompt Engineering".（Page 15, line 508）

Comments 5: Minor but Notable Issues,

1.A typo appears in Section 6 (“Disscussion” → should be “Discussion”).

2.In line 374, the variable n (as in “expand the n devices into n separate lines”) is written in plain text rather than as a mathematical symbol, reducing clarity.

3.the experiment is lack of other baselines, maybe some other LLMs could be considered as a baseline.

Response 5:

1.We sincerely apologize; we have corrected this typo in the newly submitted manuscript.

Revisions:

Please note that in the revised manuscript, the Discussion is now located in Section 5（Page 25）, and the spelling has been corrected to "Discussion". Section 6 is now "Conclusions".

2.Agree. We have corrected the formatting to improve clarity.

Revisions:

In Section 3.4.1（Page 16）, we have updated the formatting for the variable n. It is now presented as a mathematical symbol (e.g., "n devices", "n records") in the description of the decoupling constraint logic.

3.We agree with your suggestion. We introduced other baseline models into our experiments to conduct performance comparisons.

Revisions:

In Section 4.1（Page 21, line 684）, we added Table 8 ("Performance analysis of the baseline LLM in stages B4"). We compared our method (DeepSeek-V3.1) with GPT-4o, Qwen-2.5-72B, and Llama-3.1-70B. The results demonstrate that our selected model achieves SOTA performance (F1=90.42%) with a better cost-performance ratio compared to these baselines.

1. Summary

Thank you very much for taking the time to review this manuscript. We have carefully studied your constructive comments and made extensive revisions. Specifically, we have:

1. Detailed the construction and dynamic update mechanism of FSKB in Section 3 and 5.2;

2. Added a comprehensive robustness test (Section 4.2) against typos, formatting errors, and semantic conflicts;

3. Clarified the model selection and Top-k rationale. We believe these revisions significantly strengthen the paper's theoretical depth and practical value.

Detailed responses are as follows.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The introduction's review of existing related work is not comprehensive enough, failing to clearly compare the core differences between this method and LLM+RAG methods in other fields, and failing to fully highlight the targeted optimization advantages of this research in the fire protection field. It is recommended to expand the comparative analysis of related work.

Response 1: Agree. We have rewritten Section 2 (Related Work) to clarify the theoretical positioning of our method.

Revisions:

1. We added a comparative analysis between Naive RAG (prone to noise in engineering contexts) and our FS-RAG, highlighting our advantage in handling strict engineering constraints and logical dependencies.

2. We emphasized that unlike open-domain QA（Page 13 to 15）, our method focuses on "Structured Information Extraction (SIE)" with zero tolerance for hallucinations, which is achieved through the structured constraints of FSKB.

Comments 2: The construction process of the FSKB knowledge base is vaguely described, failing to explain the specific sources of knowledge cards, the manual editing process and quality control methods, and the knowledge base update mechanism. It is recommended to add explanations of relevant content and a description of the knowledge base update mechanism.

Response 2: Thank you for your valuable feedback; we fully appreciate it. We have reorganized Section 3.2（Page 10 to 13） in the new version of the manuscript. The knowledge cards in the text are specifically derived from the Multi-source normative standard and domain knowledge, and are formed into the FSKB through manual construction. We discussed the quality control methods and knowledge base updates in Section 5（Page 25 to 26）, Discussion.

Comments 3: The model robustness testing is insufficient. It does not test the work order generation effect under actual scenarios such as typos, non-standard formatting, and conflicting clauses in the specification text, nor does it analyze the impact of clauses of different lengths and complexities on model performance. It is recommended to add abnormal scenario testing.

Response 3: Agree. We've added an experimental section to verify the model's robustness.

Revisions:

1. New Section: We added Section 4.2 (Robustness verification)（Page 23） and constructed an "Adversarial Test Set" containing 100 clauses with three types of noise: Character-level (Typos/OCR errors), Format-level (No linebreaks), and Semantic-level (Conflicting clauses).

2. Quantitative Analysis: We added Table 9（Page 24）, which quantifies the performance degradation under noise. The results show that our B4 method maintains high stability (F1 drops only by 2.3% under typo noise), significantly outperforming the baseline (B0), which dropped by 13.3%.

Comments 4: The FS-RAG retrieval mechanism lacks detail, failing to explain the specific selection of the text embedding model and the rationale for choosing Top-k=30. This manuscript clarifies the specific selection and version of the text embedding model, explaining the selection criteria; it also demonstrates the impact of different Top-k values on model performance through experiments, providing a reasonable basis for Top-k=30.

Response 4: Agree. We have provided the technical specifications and experimental basis for these parameters.

Revisions:

1. FS-RAG retrieval mechanism: In section 3.3.1（Page 13,14）, we reorganized the context-accurate matching mechanism based on knowledge cards and reorganized the text content of this part.

2. Top-k Analysis: In Section 3.3.2, we added a Top-k Sensitivity Analysis. We tested k∈{10,20,30,40,50} and presented the results in Figure 7（Page 15）,. The experiment demonstrates that F1-score peaks at k=30; lower values lead to low recall, while higher values introduce noise.

Comments 5: The discussion section lacks sufficient analysis of the method's limitations, failing to mention potential problems in practical applications.

Response 5: Agree. We have expanded the discussion to include critical self-reflection.

Revisions:

In Section 5.3（Page 26）, we re-examined the limitations of this study. In order to better guide fire protection maintenance work, we will break down the barriers between the work order database constructed by the method in this paper and BIM information in future work, in order to achieve customized fire protection maintenance management.

Comments 6: More lastest research should be considered for related work. e.g. “Video saliency prediction via single feature enhancement and temporal recurrence”.

Response 6: Yes. We surveyed the latest relevant research, and by studying cross-disciplinary studies, we provided theoretical support for the methods presented in this paper.

Comments 7: The reference style is inconsistent; for example, the publication year is bolded in some references but not in others. It is recommended to check the overall typesetting and formatting of the manuscript.

Response 7:Agree. We sincerely apologize for this; we have carefully proofread the new version of the manuscript.

Revisions:

We have standardized the citation format in the References section. All publication years, author lists, and journal names now strictly follow the journal's template requirements. We also unified the font sizes and indentation throughout the document.

Comments 8: The placement of figures and tables is inconsistent, making the layout appear disorganized.

Response 8: Agree. We have re-layout the manuscript.

Revisions:

We ensured that all Figures and Tables are placed immediately after they are first mentioned in the text. We also standardized the caption styles and ensured that no tables are split across pages unnecessarily to improve readability.

Comments 9: The usage of abbreviations and full terms is inconsistent. For instance, “FS-RAG (Fire Services-Retrieval-Augmented Generation)” places the abbreviation first, while in other places the full term appears first followed by the abbreviation in parentheses.

Response 9: Agree. We have conducted a global check for abbreviations.

Revisions:

We standardized the usage: The full term is used at the first mention, followed by the abbreviation in parentheses (e.g., "Fire Services-Retrieval-Augmented Generation (FS-RAG)"), and the abbreviation is used consistently thereafter. We checked terms like FS-RAG, FSKB, LLM, and ESS.

Comments 10: The experimental dataset is relatively limited, which raises concerns about the generalizability of the results.

Response 10: Thank you very much for your feedback. We take your question about the "dataset size" very seriously. It's important to understand that there are limited standard and regulatory documents in Chinese related to fire protection maintenance. While descriptions of core maintenance content may vary slightly across different regions, we tested the model's generalization ability in Section 4.3（Page 24）, validating the generalization capability of our proposed method.

Comments 11: The conclusion section is somewhat brief; future work could be described in more concrete terms.

Response 11: Agree. After careful review, we have revised our conclusions.

Revisions:

In Section 6 (Conclusions)（Page 26,27）, we summarized the quantitative improvements more specifically (F1 increased by 12.62%). By further integrating multimodal information with agent-based intelligent technology, the fire protection maintenance work order database is expected to bridge information with the BIM model, further promoting the construction of smart fire protection.

1. Summary

Thank you very much for taking the time to review this manuscript. We greatly appreciate your recognition of the engineering value and structural organization of our work. We have carefully studied your suggestions regarding practical application, cost analysis, and liability issues. In the revised manuscript, we have: 1) Added a detailed comparison between API and local deployment (Table 2); 2) Clarified the human-in-the-loop workflow (Section 5.1) to address compliance and responsibility concerns; 3) Discussed the dynamic update mechanism (MOC) in Section 5.2.

Detailed responses are as follows.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The experiment is limited to the open fire regulations (GB, DB, etc.), and there is a lack of comparison with the actual building's facility list, maintenance history, and work instructions reflecting the judgment of field engineers, so an experiment using actual maintenance company's operation data is necessary.

Response 1: Agree. We acknowledge that incorporating real-world operational data is crucial. While full access to private facility ledgers is restricted by privacy concerns, we have strengthened the practical validation.In the future, we will bridge the maintenance work order library generated by this research method with the BIM model of the actual building. In Section 5.3（Page 26）, we discuss the limitations of the current work and the prospects for future work.

Revisions:

1. Data Expansion: In Section 3.5.1 and Table 4（Page 18）, we explicitly included 3 Enterprise Standards from maintenance companies in our dataset to reflect real-world operational requirements.

2. Generalization Test: In Section 4.3（Page 24）, we added an experiment using an unseen external standard (Gansu DB62/T 4727-2023). The model achieved an F1 score of 87.76% on this new dataset, demonstrating its capability to handle diverse, real-world regulatory demands beyond the training set.

Comments 2: Rline and F1-Score measure structural and field-level accuracy, but the failure to assess the "practical usefulness of the work instruction" (e.g., whether it is easy for field technicians to understand or not missing items) needs to be supplemented.

Response 2: Agree. We have clarified the definition of "usefulness" in the context of engineering execution.In the future, by combining the location and quantity of fire protection facilities with the information in the ledger (derived from the BIM model of a specific building), we will greatly improve the guidance of maintenance work orders in actual projects.

Revisions:

1. Formal Definition: In Section 3 (Eq. 1)（On page 8, between lines 278 and 279）, we formally defined the work order as a 6-tuple W={Dtype,Dsubtype,A,R,F,N}. This ensures the generated instructions are not vague text but structured, actionable data points (Actionable & Verifiable) that technicians can directly execute.

2. Metric Relevance: We explained in Section 3.5.3（Page 20） that the "Line-level Compliance Rate" is a direct proxy for practical usefulness, as it measures whether the output strictly follows the schema required for field operations. The achieved rate of 97.3% ensures high usability.

Comments 3: FSKB needs to manually define facility type, operation, cycle, trigger term, etc., and since this process requires considerable domain expert labor, it is necessary to suggest a direction to reduce the cost of FSKB construction.

Response 3: Agree. We have elaborated on the cost-effective construction and maintenance strategies.

Revisions:

1. Lightweight Design: In Section 3.2.2（Page 11 to 13）, we described the FSKB as a "lightweight" ontology focusing only on core entities, avoiding the high cost of building complex comprehensive knowledge graphs.

2. Update convenience: In Section 5.2（Page 25）, we argue that updating the FSKB is far cheaper than re-training the model when regulations change.

Comments 4: A structure that relies on the DeepSeek API that needs to provide advantages and disadvantages compared to the combination of on-premise deployment + RAG in lightweight LLMs (e.g., Qwen, Llama) because API failures, speed delays, and cost increases can affect the entire system.

Response 4: Agree. We have added a comprehensive comparison to justify the API choice.

Revisions:

1. New Comparison Table: We added Table 2 in Section 3.1（Page 10, line 344）, comparing DeepSeek-V3.1 (API) with Llama-3-70B and Qwen-2.5-14B (Local) across five dimensions: Parameter Scale, Logical reasoning Ability, Context Window, Hardware Requirement, Cost.

2. Analysis: We concluded that for small-to-medium maintenance firms, the API approach lowers the "Computational Power Threshold" while providing SOTA-level reasoning that lightweight local models cannot match in complex logic extraction.

Comments 5: Discussion of ensuring reliability should be added because API output is inconsistent, or updates can cause results to vary at the same prompt.

Response 5: Agree. We have addressed the consistency issue through parameter settings and architectural design.

Revisions:

In Section 3.5.2（Page 19）, we specified setting the model temperature to 0.0 to minimize randomness and ensure deterministic outputs.

Comments 6: MOC linkage logic, in which work instructions are automatically updated when changing facilities or design, is not covered in the paper, so it needs to be supplemented.

Response 6: Agree. We have added a discussion on system adaptability and change management.

Revisions:

In Section 5.2 (System Adaptability)（Page 25）, we described the logic for handling regulatory changes. The system decouples "knowledge" (FSKB) from "reasoning" (LLM). When standards change, only the specific Knowledge Cards in the FSKB need to be updated. This triggers an immediate update in the generated work orders without the need for model retraining, effectively supporting the Management of Change (MOC) process.

Comments 7: The process of reviewing and revising the LLM-generated work instructions by the field technician (e.g. "AI Generation → Technician Review → Finalization") should be specified in the system.

Response 7: Agree. We have formalized the workflow to emphasize human oversight.

Revisions:

In Section 5.1 (Practical Workflow and Compliance)（Page 25）, we explicitly defined the "AI Proposal - Expert Decision" workflow. The system acts as a "Copilot," generating 95% of the draft content, which is then subject to mandatory review and finalization by certified fire engineers. This ensures that the efficiency of AI is combined with the accountability of human experts.

Comments 8: There is a lack of discussion on whether AI-generated work instructions can be considered to ensure compliance and where the responsibility lies in the event of an error.

Response 8: Agree. We have clarified the liability boundaries.

Revisions:

In Section 5.1（Page 25）, we clearly stated that legal responsibility still rests with the qualified engineers. Artificial intelligence systems are tools to ensure "process traceability," but they cannot replace the legal responsibility of maintenance service providers. The "Line-level Compliance Rate" metric serves as a technological safeguard, while manual review provides the legal guarantee.

Response to Reviewer 2 Comments

1. Summary

Thank you very much for taking the time to review our manuscript again. We sincerely appreciate your constructive comments and suggestions, which have helped us further clarify our motivation, improve the manuscript's readability, and broaden the scope of our related work.

Please find the detailed response below; the corresponding revisions have been made in the manuscript. We have carefully addressed each of your comments, paying particular attention to refining the language, strengthening the research motivation in the introduction, and incorporating the latest relevant research findings into the work.

2. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The format of the article is confused, which affects the reading experience. At the same time, there are some grammatical problems.

Response 1: Agree. Thank you for pointing this out. We strictly followed the journal's standard structure (Introduction, Materials and Methods, Results, Discussion, and Conclusions). To further address your concern, we have conducted a thorough proofreading of the entire manuscript to correct grammatical errors and improve the visual layout. We specifically polished the Abstract, Introduction, Discussion, and Conclusions sections to ensure the language is academically rigorous and the format is clear.

Revisions:

1. We have reviewed the entire text, paying particular attention to refining the abstract, introduction, related work, and conclusion sections to ensure precise and professional language.

2. We have also adjusted the layout and reorganized the tables in the manuscript.

Comments 2: The author does not clearly explain the motivation behind this manuscript. What are the existing problems and why they are crucial should be explained in more detail. I suggest the author further strengthen the relevant parts of the Introduction.

Response 2: Agree.Thank you for pointing this out.We have, accordingly, revised the Introduction section (Paragraph 4) to explicitly articulate the "cognitive load" caused by the multi-source heterogeneity of regulations and to emphasize the critical nature of fire maintenance as the "last line of defense" for public safety.

Revisions:

In the manuscript (Introduction section, page 3 and 4).

Comments 3: More lastest research should be considered for related work. e.g. “ORSI Salient Object Detection via Progressive Interaction and Saliency-Guided Enhancement”.

Response 3: Agree.Thank you for the suggestion. We agree that incorporating these latest studies provides valuable methodological context. In the “Related Work” section, we cite recommended articles that combine feature enhancement and interaction mechanisms to demonstrate how they can be successfully applied to complex data processing tasks, such as visual salient detection, and draw analogies to our approach in text extraction.

Revisions:

In the manuscript (Chapter 2, Related Work section, page 7 and 8).We have cited the relevant research you recommended as references 30 and 31.