An Intelligent Framework for Implementing AIAG–VDA FMEA and Action Priority (AP) Assessment

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript addresses a relevant topic related to the digitalization of quality management and the implementation of the AIAG–VDA FMEA standard. The paper is well organized and clearly written, and the case study helps readers understand the workflow. However, the work currently reads more as an industrial implementation report than as a research article, and the scientific contribution needs to be more clearly articulated.

1. The proposed framework essentially digitalizes the established AIAG–VDA FMEA procedure. Although automation is valuable from an engineering perspective, the manuscript does not demonstrate a new risk assessment model, decision-making methodology, or theoretical development.
2. The literature review is lengthy and largely descriptive, repeating well-known concepts of FMEA. A more critical discussion of recent research on intelligent risk assessment and digital quality management would improve the academic value of the paper.
3. It is unclear whether the proposed system learns from new operational data or improves its performance over time. The manuscript does not describe any training procedure, feedback loop, or knowledge updating mechanism based on actual manufacturing outcomes.
4. The effectiveness of the framework is not quantitatively validated. The reduction of all risks to a low level after optimization is reported, but this appears to result from managerial actions rather than from the intelligent capabilities of the system. The study should include measurable indicators such as defect rate reduction, accuracy of recommendations, or comparison with human expert assessments.
5. The validation is based on a single 3D laser cutting process. This is insufficient to demonstrate general applicability. The authors should either provide additional cases or justify theoretically why the proposed approach can be extended to other manufacturing processes.
6. The manuscript does not compare the proposed system with other advanced FMEA approaches such as fuzzy FMEA, Bayesian FMEA, or machine learning-based risk assessment methods. Without benchmarking, it is difficult to evaluate whether the proposed framework offers advantages beyond workflow automation.

Author Response

Dear Reviewer,

Thank you for your pertinent and constructive comments, which will significantly improve the quality of our paper.

Below, we provide our detailed responses to each of your observations.

 

The manuscript addresses a relevant topic related to the digitalization of quality management and the implementation of the AIAG–VDA FMEA standard. The paper is well organized and clearly written, and the case study helps readers understand the workflow. However, the work currently reads more as an industrial implementation report than as a research article, and the scientific contribution needs to be more clearly articulated.

1. The proposed framework essentially digitalizes the established AIAG–VDA FMEA procedure. Although automation is valuable from an engineering perspective, the manuscript does not demonstrate a new risk assessment model, decision-making methodology, or theoretical development.

 In the Introduction section, we have added the following information. (yellow).

The objective of this paper is to present an AI-based solution (leveraging a Large Language Model) for managing risks in a real manufacturing process – 3D LASER cutting – with the aim of generating FMEA reports in accordance with modern standards. The purpose of this solution is to ensure full traceability of all required operations carried out by the relevant departments (management, quality assurance, engineers) involved in risk management within the production process. At the same time, it seeks to reduce response time when nonconformities occur during production and to facilitate FMEA report generation through full automation and digitalization of the solution.

The solution is demonstrated through a practical case study – 3D LASER cutting – but it can be extended to other processes as well. The framework can be adapted to additional manufacturing processes by expanding the underlying knowledge base.

The study resulted in a fully operational pilot system that is currently used in a real production environment. The key innovative elements introduced by this research are as follows:

• The use of a pre-processing stage before involving the AI component. In this approach, a group of experts (engineers, management, quality specialists) defined the risk classifications and the corresponding indicators for each failure mode, as is typically done in a traditional risk analysis. Based on this classification, a platform was implemented that allows users to input the number and type of defects. Depending on their classification, the system provides an appropriate response and, when possible, resolves the issue without engaging the higher-level AI analysis component.
• The involvement of the AI component for medium and high risks. Special attention was given to the reliability of the AI-generated responses, considering that the system operates in a real production environment. Hallucinations were mitigated by restricting the AI to a knowledge base developed and validated by experts; responses outside this knowledge base are not accepted. This approach makes the solution practical and applicable in real industrial settings, not just theoretical.
• The implementation of a complete framework covering the entire workflow: from defining the production process and recording defects at batch level to the automatic generation of the FMEA report. To the best of our knowledge and based on the literature review presented below, there is currently no equally comprehensive solution covering the entire process end-to-end.

It is important to emphasize that this solution does not aim to replace FMEA or introduce a new risk assessment methodology. In industry, FMEA must be performed in accordance with established quality methodologies and standards (AIAG & VDA 2019) and must be approved by the customer. Instead, the proposed solution aims to improve the efficiency of the FMEA generation process by enhancing traceability, automation, and response time — from the initial identification of risks, through the occurrence of batch-level defects, to the submission of FMEA documentation to the customer for validation.

The solution consists of a centralized platform connecting all stakeholders involved — engineers, quality specialists, and management — thus ensuring significantly improved communication and coordination. In addition, the system supports continuous data collection, contributing to the expansion of the knowledge base for future defects or even future production processes.

 

Also in section Material and Method we inserted (yellow):

Any production process must be accompanied by the development of an FMEA, following the steps outlined above (see Figure 3). Naturally, each production process has a strong technical component when preparing the FMEA. This requires the involvement of specialists directly engaged in the respective manufacturing process (engineers), members of the quality team, as well as representatives from management — all stakeholders actively involved in the production flow.

The solution proposed in this paper is designed to optimize the FMEA generation workflow. It does not replace specialists; rather, it supports and enhances communication among them, ensures information traceability, and automates the report generation process. As a result, the FMEA can be produced faster and in a more structured and complete manner compared to the traditional approach.

Therefore, our study did not focus on performing the risk evaluation itself for a specific production process. Instead, we concentrated on managing information, facilitating communication among the involved parties, and analyzing data — including the use of AI to support FMEA generation. From this perspective, we examined the entire FMEA workflow described above: from the initial risk assessment stage, through the occurrence of batch-level defects, to the final structure that the FMEA document must follow in order to be validated by the end customer.

Based on this analysis, we designed and implemented a comprehensive workflow that necessarily integrates the AI component. This framework will be presented in detail in the next section.

 

2. The literature review is lengthy and largely descriptive, repeating well-known concepts of FMEA. A more critical discussion of recent research on intelligent risk assessment and digital quality management would improve the academic value of the paper.

In Literature review section we have added to the revised manuscript some discussion regarding existing studies (which we already referred) and comparison with our study (our solution) – see green color background text in Literature review section:

The paper proposes its own approach to risk evaluation, starting from simulations based on different failure probabilities. This approach can serve as a tool that may be integrated into the data flow proposed in our study, and it also illustrates the possibility of introducing additional analytical layers in the FMEA generation process.

Our approach is somewhat different and is more closely aligned with the practical applicability of FMEA analysis. The FMEA data — including the risk assessment component — are collected during the initial stage of the workflow (generated by experts and, where applicable, by secondary analytical tools such as the one presented in the previously mentioned paper). These data form a structured knowledge base that is directly applicable to the conditions under which defects occur during the production process.

…

The paper proposes a method through which two of the indicators involved in risk assessment — namely severity and occurrence frequency — can be inferred or updated by integrating additional blocks into the data flow. These blocks collect real-time data directly from the production process and act as predictors of the defect rate. In this sense, the paper is closely aligned with the approach we present in this article.

In our case, however, we focus on designing a solution in which we do not intervene in the risk evaluation itself. The risk-related data are considered fixed and are provided by a team of experts, just as in the traditional FMEA approach. Instead, our contribution targets the way the FMEA is generated, emphasizing key stages within the data flow: identifying the risk level based on issues occurring during production (such as complaints and nonconformities), intelligently extracting root causes and optimization actions, and automatically generating the final report.

The solution we propose addresses the practical need to bring together all stakeholders involved in FMEA generation within a single platform, supervised by an AI module that ensures efficient management and use of the available data.

…

The paper presents a study with clear practical implications regarding how FMEA can also be applied in a non-manufacturing environment — specifically, in the process of configuring logistics for an electronic manufacturing services provider. The approach responds to emerging “on-demand” industry trends, where flexibility and rapid configuration are essential. In this context, FMEA generation represents one stage within the broader logistics infrastructure configuration chain.

As in our approach, the classical FMEA components (risk indicators, causes, and effects) are included, along with newer elements related to risk control and mitigation. However, unlike our solution, this study places greater emphasis on improving risk evaluation itself and reducing subjectivity in the assessment process.

In our case, as previously mentioned, the focus is on optimizing the FMEA generation workflow. The improvement of both the analysis and the generation flow is achieved primarily by introducing tools that facilitate the identification, classification, and extraction of relevant data, rather than by modifying the fundamental rules of risk evaluation.

 

…

The method delivers two concrete outcomes: a reduction in analysis time through the integration of an evaluation and FMEA generation tool — precisely the objective we pursue in our own paper — and an increase in accuracy by introducing a new risk evaluation approach (based on fuzzy logic). It is important to note that our proposed solution introduces specific indicators to measure how quickly the FMEA report is generated. The framework we present represents a concrete and comprehensive example of a data flow for FMEA analysis, without neglecting any stage of the process: from the acquisition of input data — such as the number of defects, their impact, and their evolution — to the final generation and publication of the FMEA document.

Another fuzzy-based risk assessment solution is presented in paper [31], which proposes a method for reducing subjectivity in FMEA analysis through the use of fuzzy logic. This further illustrates how integrating analytical tools into the data flow can optimize the process. That study includes a case application in the production of an RFID system for the automotive industry.

Therefore, there are existing studies that address the improvement of FMEA analysis by intervening at the data flow level — an approach that is consistent with our own work. The key contribution of our study lies in the complete presentation and implementation of the entire data flow, as described in detail below.

The workflow we propose integrates “classic” data acquisition components via a web-based form, non-AI analysis and classification modules designed to reduce response time, and advanced analytical components leveraging LLM technology. As previously emphasized, the objective is not to modify the risk evaluation methodology itself — which remains grounded in traditional approaches — although future developments may include the integration of methods such as fuzzy logic to enhance risk evaluation. Rather, our primary goal is to improve the data flow leading to the rapid and efficient generation of the FMEA report.

 

3. It is unclear whether the proposed system learns from new operational data or improves its performance over time. The manuscript does not describe any training procedure, feedback loop, or knowledge updating mechanism based on actual manufacturing outcomes.

 

In section 4.2 we have added to the revised manuscript (yellow):

For each request sent to the AI (LLM) component, the necessary knowledge base is also transmitted to ensure that responses are generated strictly based on validated information. This knowledge base was developed through the direct involvement of specialists — engineers, quality representatives, and management. When the production process was initiated, the initial risks were defined as part of the “classic” FMEA analysis stage.

These risks were then structured in JSON format so they could be systematically transmitted to the LLM. It is essential to clearly emphasize this aspect: such information cannot be randomly generated or independently produced by the AI tool. The data must be precise, process-specific, and formally defined by domain experts. Only specialists can provide this validated input.

In the data flow presented below, the exact stage at which the knowledge base is transmitted to the LLM is explicitly indicated.

 

In the Data flow presented in that section we already presented the knowledge base format and when is transmitted to LLM.

4. The effectiveness of the framework is not quantitatively validated. The reduction of all risks to a low level after optimization is reported, but this appears to result from managerial actions rather than from the intelligent capabilities of the system. The study should include measurable indicators such as defect rate reduction, accuracy of recommendations, or comparison with human expert assessments.

We would like to clarify that the reduction of all risks to the Low level after the optimization stage is the result of the effective implementation of corrective and preventive measures defined within the FMEA analysis, in accordance with the AIAG & VDA (2019) methodology. In industrial practice, risk reduction is achieved through concrete technical and managerial interventions within the production process, and not through a mere formal reassessment of indicators.

To eliminate any possible ambiguity, we have revised Section 4.1 and the Results section of the manuscript to explicitly state that the decrease in Occurrence (O) and Detection (D) values reflects real process improvements implemented by the multidisciplinary team. In its revised form, the manuscript clearly distinguishes between the optimization actions applied in the production environment and the analytical framework used to structure and document the FMEA process (see with purple font color) .

5. The validation is based on a single 3D laser cutting process. This is insufficient to demonstrate general applicability. The authors should either provide additional cases or justify theoretically why the proposed approach can be extended to other manufacturing processes.

In introduction we have added to the revised manuscript (yellow)

The objective of this paper is to present an AI-based solution (leveraging a Large Language Model) for managing risks in a real manufacturing process – 3D LASER cutting – with the aim of generating FMEA reports in accordance with modern standards.

In Conclusion we have added to the revised manuscript (green):

The solution proposed in this paper has been implemented — in the form of a pilot system — for a specific production process (3D LASER cutting). However, it can be extended to other manufacturing processes as well.

First, many failure modes may be similar across different production processes. Second, the knowledge base can be adapted accordingly — as shown, it is a parameter transmitted to the LLM with each request. This makes the framework flexible and scalable, allowing it to be customized for various industrial contexts simply by adjusting the expert-defined knowledge base.

After that in Conclusions we already have (in initial version of paper):

Future research directions involve completing the knowledge base to support more production processes and expanding the industrial domain in which it can be used.

6. The manuscript does not compare the proposed system with other advanced FMEA approaches such as fuzzy FMEA, Bayesian FMEA, or machine learning-based risk assessment methods. Without benchmarking, it is difficult to evaluate whether the proposed framework offers advantages beyond workflow automation.

In Literature review section we have added to the revised manuscript some discussion regarding existing studies (which we already referred). We have paper [18] with Monte Carlo risk assessment, paper 20 where are presented different assessments methods including fuzzy and ML and paper [21] where an assessment method is presented based on fuzzy. To be clearer we have included in the revised version another reference with fuzzy risk assessment (paper [31] ) and we have included a comparison between our solution and others solutions.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

1. Please clarify the scientific novelty of the framework beyond system integration of AIAG–VDA FMEA, AP assessment, and digital tools.
2. The LLM component is described conceptually, but no quantitative evaluation is provided. Consider reporting accuracy, consistency, or comparison with human-generated FMEA entries.
3. The case study is limited to a single 3D laser cutting process. Adding a comparison with a conventional (non-AI) FMEA workflow would help demonstrate the incremental benefit.
4. The AP thresholds and non-AI classification rules appear heuristic. Some justification or basic sensitivity discussion is recommended.
5. Claims regarding reduced subjectivity and improved efficiency are not supported by measurable indicators such as analysis time reduction or inter-user agreement rates.
6. Please clarify whether the LLM knowledge base is static or dynamically updated during operation.
7. Hardware and software specifications, including server setup and LLM configuration, should be reported more explicitly for reproducibility.
8. Add a concise overview diagram summarizing the workflow from PFMEA through AP evaluation, non-AI filtering, and LLM assistance.
9. Improve figure readability by increasing label sizes and simplifying dense layouts.
10. Separate methodological description more clearly from case-study results.
11. Minor grammatical and stylistic issues are present throughout the manuscript and should be corrected through careful English proofreading.

Author Response

Dear Reviewer,

Thank you for your pertinent and constructive comments, which will significantly improve the quality of our paper.

Below, we provide our detailed responses to each of your observations.

 

1. Please clarify the scientific novelty of the framework beyond system integration of AIAG–VDA FMEA, AP assessment, and digital tools.

In the Introduction section, we have added the following information. (yellow).

The objective of this paper is to present an AI-based solution (leveraging a Large Language Model) for managing risks in a real manufacturing process – 3D LASER cutting – with the aim of generating FMEA reports in accordance with modern standards. The purpose of this solution is to ensure full traceability of all required operations carried out by the relevant departments (management, quality assurance, engineers) involved in risk management within the production process. At the same time, it seeks to reduce response time when nonconformities occur during production and to facilitate FMEA report generation through full automation and digitalization of the solution.

The solution is demonstrated through a practical case study – 3D LASER cutting – but it can be extended to other processes as well. The framework can be adapted to additional manufacturing processes by expanding the underlying knowledge base.

The study resulted in a fully operational pilot system that is currently used in a real production environment. The key innovative elements introduced by this research are as follows:

• The use of a pre-processing stage before involving the AI component. In this approach, a group of experts (engineers, management, quality specialists) defined the risk classifications and the corresponding indicators for each failure mode, as is typically done in a traditional risk analysis. Based on this classification, a platform was implemented that allows users to input the number and type of defects. Depending on their classification, the system provides an appropriate response and, when possible, resolves the issue without engaging the higher-level AI analysis component.
• The involvement of the AI component for medium and high risks. Special attention was given to the reliability of the AI-generated responses, considering that the system operates in a real production environment. Hallucinations were mitigated by restricting the AI to a knowledge base developed and validated by experts; responses outside this knowledge base are not accepted. This approach makes the solution practical and applicable in real industrial settings, not just theoretical.
• The implementation of a complete framework covering the entire workflow: from defining the production process and recording defects at batch level to the automatic generation of the FMEA report. To the best of our knowledge and based on the literature review presented below, there is currently no equally comprehensive solution covering the entire process end-to-end.

It is important to emphasize that this solution does not aim to replace FMEA or introduce a new risk assessment methodology. In industry, FMEA must be performed in accordance with established quality methodologies and standards (AIAG & VDA 2019) and must be approved by the customer. Instead, the proposed solution aims to improve the efficiency of the FMEA generation process by enhancing traceability, automation, and response time — from the initial identification of risks, through the occurrence of batch-level defects, to the submission of FMEA documentation to the customer for validation.

The solution consists of a centralized platform connecting all stakeholders involved — engineers, quality specialists, and management — thus ensuring significantly improved communication and coordination. In addition, the system supports continuous data collection, contributing to the expansion of the knowledge base for future defects or even future production processes.

2. The LLM component is described conceptually, but no quantitative evaluation is provided. Consider reporting accuracy, consistency, or comparison with human-generated FMEA entries.

In table 8 we already make that comparison with the previously existing solution.

We changed the header of the table for more clarification and added the following text near the table (green):

The previously existing solution relied on a conventional approach to FMEA generation. At the start of the production process, or whenever nonconformities occurred, the involved parties (engineers, quality representatives, and management) communicated via email groups and held meetings to establish an action plan.

The proposal of optimization measures, as well as the preparation of the FMEA document, was only partially automated, mainly through the use of predefined templates that were manually completed by members of the quality team.

 

3. The case study is limited to a single 3D laser cutting process. Adding a comparison with a conventional (non-AI) FMEA workflow would help demonstrate the incremental benefit.

In table 8 we already make that comparasion with the previously existing solution.

We changed the header of the table for more clarification and added the following text near the table (green):

The previously existing solution relied on a conventional approach to FMEA generation. At the start of the production process, or whenever nonconformities occurred, the involved parties (engineers, quality representatives, and management) communicated via email groups and held meetings to establish an action plan.

The proposal of optimization measures, as well as the preparation of the FMEA document, was only partially automated, mainly through the use of predefined templates that were manually completed by members of the quality team.

4. The AP thresholds and non-AI classification rules appear heuristic. Some justification or basic sensitivity discussion is recommended.

We would like to clarify that the AP thresholds were not defined heuristically. As presented in Table 4 and further explained in Section 4 of the revised manuscript, the Action Priority levels are directly linked to measurable operational indicators, including the monthly number of internal nonconformities, defect rate expressed in ppm, number of external nonconformities, customer complaints, and the defect trend evolution.

The Low, Medium, and High categories are associated with clearly defined quantitative ranges and explicit customer-impact criteria (such as functional impact, line stoppage, sorting, return, or 8D request). This ensures that the classification reflects the actual technical and operational risk exposure of the analyzed process.

In addition, the Severity (S), Occurrence (O), and Detection (D) ratings were assigned strictly in accordance with the AIAG & VDA 2019 evaluation grids. The scores were determined based on structured criteria, including functional impact, historical defect frequency, process parameter stability, and the effectiveness of preventive and detection controls.

To improve methodological transparency, an explicit explanation of the AP thresholds and the S–O–D assignment approach has been added in Section 4 of the revised manuscript.

We believe these clarifications demonstrate that the classification rules are grounded in measurable industrial criteria and practical process data rather than heuristic assumptions (see with purple font color).

5. Claims regarding reduced subjectivity and improved efficiency are not supported by measurable indicators such as analysis time reduction or inter-user agreement rates.

In the Results section we have added some comments in text which clarify the results from table 8 (green):

The previously existing solution relied on a conventional approach to FMEA generation. At the start of the production process, or whenever nonconformities occurred, the involved parties (engineers, quality representatives, and management) communicated via email groups and held meetings to establish an action plan.

The proposal of optimization measures, as well as the preparation of the FMEA document, was only partially automated, mainly through the use of predefined templates that were manually completed by members of the quality team.

Compared to the previously used digitalized method based on a database, the pilot solution demonstrated superior performance, shown in Table 8.

 

So in the table 8 we have a comparison with previous solution which was implemented.  Also in table 9 we already have a comparison with a full AI solution (without preprocessing module) to show the advantages of the non -AI preprocessing analysis. So the indicators are users interaction, time reduction, number of failure cases reduced. Also, to be connected to academic environment we have added to the revised version a comparison table (table 10) between our solution and others solutions presented in literature review with comments on it (green):

In the table below (Table 10), we provide a comparison with other studies cited in the literature review section. In addition to the quantitative evaluations presented in Tables 7, 8, and 9 — regarding response time, number of users involved, improvement in risk levels, and the number of resolved defects — we also include a qualitative comparison focused on the components of the data flow. We believe that, from a practical perspective, this data flow has been comprehensively addressed in our proposed solution.

Accordingly, the table highlights how input data are collected and identifies their sources. As can be observed, in all the analyzed articles, the data originate from real industrial environments. The comparison also includes how the data are processed and how the results are ultimately generated and published.

 

6. Please clarify whether the LLM knowledge base is static or dynamically updated during operation.

We have added to the revised manuscript in section 4 the following:

The data must be precise, process-specific, and formally defined by domain experts. Only specialists can provide this validated input. The knowledge base is updated to each failure mode at the beginning of the production process. So it is populated statically with the information for the experts and data from a failure mode can be used to different production processes.

7. Hardware and software specifications, including server setup and LLM configuration, should be reported more explicitly for reproducibility.

In section 5 to table 5 we already have the hardware specification of the pilot and to table 6 we already have the software specifications including LLM version. Also we have a cost of the solution near the table.

8. Add a concise overview diagram summarizing the workflow from PFMEA through AP evaluation, non-AI filtering, and LLM assistance.

At the beginning of sub-section Data flow section 4 – 4.2 we have added to the revised manuscript a block diagram of data flow (new figure 7)  and a short text explication before to details each step:

In figure 7 is presented a block diagram of data flow.

The flow begins with introduction of data (faults cases, mode, effects, controls and actions – in the initial stage of the production process or with introduction of number of non-conformities and their impact taking in account the complaints from the customer. The first analysis is non-AI analysis – here is established if the risk is low, medium or high based of the number of non-conformities, the evolution of the number for last two months, the type and the impact. If the risk is low then we just publish (display) the result – info – else we perform AI LLM Analysis. The results in that case can be publication of the optimizations to the screen (actions to decrease the risk factor) or even print of the updated FMEA report.

Below are presented the stages in details.

 

9. Improve figure readability by increasing label sizes and simplifying dense layouts.

The resolution of the figures was increased.

10. Separate methodological description more clearly from case-study results.

In the Results section we have added to revised version a table (table 10) with  comparison between different approaches in others studies and our approach. Also we have added the following explanation (green after the table):

As previously mentioned, the existing solutions presented in other studies primarily focus on modifying the way risk evaluation is performed. Such a feature has not been implemented in our solution, although it may represent a future direction for further development. Instead, our solution implements the entire data flow required for FMEA generation — from data input and collection to the preparation of the FMEA report and its publication in a printable format. As can be observed, the stages of the process are addressed at the research-results level in all the solutions discussed in other articles. In our case, however, each individual stage is implemented as a functional component within a continuous, automated workflow with direct practical applicability. The AI component is used for the extraction and classification of the data necessary for generating the FMEA report. In essence, we present a solution in which the AI LLM tool is an integral part of a broader application framework with multiple practical usage perspectives.

 

11. Minor grammatical and stylistic issues are present throughout the manuscript and should be corrected through careful English proofreading.

The manuscript has been carefully revised to address minor grammatical and stylistic issues. The text has undergone thorough English proofreading to improve clarity, consistency, and overall readability.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The article proposes an intelligent framework for implementing AIAG–VDA FMEA and Action Priority (AP), enhanced by Industry 4.0 digitalisation and the use of LLM. The case study (3D laser cutting) is relevant and well documented. The hybrid non-AI + AI (LLM) approach is interesting and operational. However, there is much to modify and discuss. The article is more of an application than a novelty; it is a work that has applied the approach with the AI algorithm. I wonder what is special about the work in terms of research.

- The plagiarism rate is 11% without references and the AI writing rate is 23%. These rates must be reduced.

- The paper lacks structure at the end of the introduction and the objectives of the work carried out.

- The scientific contribution is not clearly positioned in relation to the state of the art.

- The review is rich but very descriptive. Where exactly are the limits of existing solutions? How does your work position itself in relation to these limits?

- The S–O–D and AP criteria are presented, but the justification for the values assigned remains somewhat informal.

- Why didn't you think of applying fuzzy logic to calculate RPN, given that one of the major limitations of the FMEA method is the uncertainty and inaccuracy of parameter values? There is a paper entitled ‘Addressing Uncertainty in Digital Risk Evaluation Using a Fuzzy-FMEA Methodology’ which deals with the same topic using fuzzy logic. I invite you to read it and take it into consideration in your discussion.

- Furthermore, there is a lack of robust scientific validation (generalisation, comparison, quantitative metrics) in your work. Please review this point.

- Furthermore, the results obtained, ‘100% low risk after optimisation’, are unrealistic without discussion of biases.

- Lack of qualitative assessment of the quality of AI recommendations.

- The discussion is very weak; the authors should discuss and compare with existing work. In fact, there is a great deal of work that has dealt with FMEA. However, the references presented are very limited.

- I invite you to review the conclusion; it is weak and lacks perspective.

 

Author Response

Dear Reviewer,

Thank you for your pertinent and constructive comments, which will significantly improve the quality of our paper.

Below, we provide our detailed responses to each of your observations.

 

The article proposes an intelligent framework for implementing AIAG–VDA FMEA and Action Priority (AP), enhanced by Industry 4.0 digitalisation and the use of LLM. The case study (3D laser cutting) is relevant and well documented. The hybrid non-AI + AI (LLM) approach is interesting and operational. However, there is much to modify and discuss. The article is more of an application than a novelty; it is a work that has applied the approach with the AI algorithm. I wonder what is special about the work in terms of research.

1- The plagiarism rate is 11% without references and the AI writing rate is 23%. These rates must be reduced.

We reformulate the text to reduce the AI writing rate (see blue font color text) -  we perform a second check with Turnitin application and we found similarities only to names, standards etc – see attached report.

2- The paper lacks structure at the end of the introduction and the objectives of the work carried out.

In the Introduction section, we have added the following information. (yellow).

The objective of this paper is to present an AI-based solution (leveraging a Large Language Model) for managing risks in a real manufacturing process – 3D LASER cutting – with the aim of generating FMEA reports in accordance with modern standards. The purpose of this solution is to ensure full traceability of all required operations carried out by the relevant departments (management, quality assurance, engineers) involved in risk management within the production process. At the same time, it seeks to reduce response time when nonconformities occur during production and to facilitate FMEA report generation through full automation and digitalization of the solution.

The solution is demonstrated through a practical case study – 3D LASER cutting – but it can be extended to other processes as well. The framework can be adapted to additional manufacturing processes by expanding the underlying knowledge base.

The study resulted in a fully operational pilot system that is currently used in a real production environment. The key innovative elements introduced by this research are as follows:

• The use of a pre-processing stage before involving the AI component. In this approach, a group of experts (engineers, management, quality specialists) defined the risk classifications and the corresponding indicators for each failure mode, as is typically done in a traditional risk analysis. Based on this classification, a platform was implemented that allows users to input the number and type of defects. Depending on their classification, the system provides an appropriate response and, when possible, resolves the issue without engaging the higher-level AI analysis component.
• The involvement of the AI component for medium and high risks. Special attention was given to the reliability of the AI-generated responses, considering that the system operates in a real production environment. Hallucinations were mitigated by restricting the AI to a knowledge base developed and validated by experts; responses outside this knowledge base are not accepted. This approach makes the solution practical and applicable in real industrial settings, not just theoretical.
• The implementation of a complete framework covering the entire workflow: from defining the production process and recording defects at batch level to the automatic generation of the FMEA report. To the best of our knowledge and based on the literature review presented below, there is currently no equally comprehensive solution covering the entire process end-to-end.

It is important to emphasize that this solution does not aim to replace FMEA or introduce a new risk assessment methodology. In industry, FMEA must be performed in accordance with established quality methodologies and standards (AIAG & VDA 2019) and must be approved by the customer. Instead, the proposed solution aims to improve the efficiency of the FMEA generation process by enhancing traceability, automation, and response time — from the initial identification of risks, through the occurrence of batch-level defects, to the submission of FMEA documentation to the customer for validation.

The solution consists of a centralized platform connecting all stakeholders involved — engineers, quality specialists, and management — thus ensuring significantly improved communication and coordination. In addition, the system supports continuous data collection, contributing to the expansion of the knowledge base for future defects or even future production processes.

3- The scientific contribution is not clearly positioned in relation to the state of the art.

In Literature review section we have added to the revised manuscript some discussion regarding existing studies (which we already referred) and comparison with our study (our solution) – see green color background text in Literature review section:

The paper proposes its own approach to risk evaluation, starting from simulations based on different failure probabilities. This approach can serve as a tool that may be integrated into the data flow proposed in our study, and it also illustrates the possibility of introducing additional analytical layers in the FMEA generation process.

Our approach is somewhat different and is more closely aligned with the practical applicability of FMEA analysis. The FMEA data — including the risk assessment component — are collected during the initial stage of the workflow (generated by experts and, where applicable, by secondary analytical tools such as the one presented in the previously mentioned paper). These data form a structured knowledge base that is directly applicable to the conditions under which defects occur during the production process.

…

The paper proposes a method through which two of the indicators involved in risk assessment — namely severity and occurrence frequency — can be inferred or updated by integrating additional blocks into the data flow. These blocks collect real-time data directly from the production process and act as predictors of the defect rate. In this sense, the paper is closely aligned with the approach we present in this article.

In our case, however, we focus on designing a solution in which we do not intervene in the risk evaluation itself. The risk-related data are considered fixed and are provided by a team of experts, just as in the traditional FMEA approach. Instead, our contribution targets the way the FMEA is generated, emphasizing key stages within the data flow: identifying the risk level based on issues occurring during production (such as complaints and nonconformities), intelligently extracting root causes and optimization actions, and automatically generating the final report.

The solution we propose addresses the practical need to bring together all stakeholders involved in FMEA generation within a single platform, supervised by an AI module that ensures efficient management and use of the available data.

…

The paper presents a study with clear practical implications regarding how FMEA can also be applied in a non-manufacturing environment — specifically, in the process of configuring logistics for an electronic manufacturing services provider. The approach responds to emerging “on-demand” industry trends, where flexibility and rapid configuration are essential. In this context, FMEA generation represents one stage within the broader logistics infrastructure configuration chain.

As in our approach, the classical FMEA components (risk indicators, causes, and effects) are included, along with newer elements related to risk control and mitigation. However, unlike our solution, this study places greater emphasis on improving risk evaluation itself and reducing subjectivity in the assessment process.

In our case, as previously mentioned, the focus is on optimizing the FMEA generation workflow. The improvement of both the analysis and the generation flow is achieved primarily by introducing tools that facilitate the identification, classification, and extraction of relevant data, rather than by modifying the fundamental rules of risk evaluation.

 

…

The method delivers two concrete outcomes: a reduction in analysis time through the integration of an evaluation and FMEA generation tool — precisely the objective we pursue in our own paper — and an increase in accuracy by introducing a new risk evaluation approach (based on fuzzy logic). It is important to note that our proposed solution introduces specific indicators to measure how quickly the FMEA report is generated. The framework we present represents a concrete and comprehensive example of a data flow for FMEA analysis, without neglecting any stage of the process: from the acquisition of input data — such as the number of defects, their impact, and their evolution — to the final generation and publication of the FMEA document.

Another fuzzy-based risk assessment solution is presented in paper [31], which proposes a method for reducing subjectivity in FMEA analysis through the use of fuzzy logic. This further illustrates how integrating analytical tools into the data flow can optimize the process. That study includes a case application in the production of an RFID system for the automotive industry.

Therefore, there are existing studies that address the improvement of FMEA analysis by intervening at the data flow level — an approach that is consistent with our own work. The key contribution of our study lies in the complete presentation and implementation of the entire data flow, as described in detail below.

 

4- The review is rich but very descriptive. Where exactly are the limits of existing solutions? How does your work position itself in relation to these limits?

In Literature review section we have added to the revised manuscript some discussion regarding existing studies (which we already referred) and comparison with our study (our solution) – see green color background text in Literature review section:

The paper proposes its own approach to risk evaluation, starting from simulations based on different failure probabilities. This approach can serve as a tool that may be integrated into the data flow proposed in our study, and it also illustrates the possibility of introducing additional analytical layers in the FMEA generation process.

Our approach is somewhat different and is more closely aligned with the practical applicability of FMEA analysis. The FMEA data — including the risk assessment component — are collected during the initial stage of the workflow (generated by experts and, where applicable, by secondary analytical tools such as the one presented in the previously mentioned paper). These data form a structured knowledge base that is directly applicable to the conditions under which defects occur during the production process.

…

The paper proposes a method through which two of the indicators involved in risk assessment — namely severity and occurrence frequency — can be inferred or updated by integrating additional blocks into the data flow. These blocks collect real-time data directly from the production process and act as predictors of the defect rate. In this sense, the paper is closely aligned with the approach we present in this article.

In our case, however, we focus on designing a solution in which we do not intervene in the risk evaluation itself. The risk-related data are considered fixed and are provided by a team of experts, just as in the traditional FMEA approach. Instead, our contribution targets the way the FMEA is generated, emphasizing key stages within the data flow: identifying the risk level based on issues occurring during production (such as complaints and nonconformities), intelligently extracting root causes and optimization actions, and automatically generating the final report.

The solution we propose addresses the practical need to bring together all stakeholders involved in FMEA generation within a single platform, supervised by an AI module that ensures efficient management and use of the available data.

…

The paper presents a study with clear practical implications regarding how FMEA can also be applied in a non-manufacturing environment — specifically, in the process of configuring logistics for an electronic manufacturing services provider. The approach responds to emerging “on-demand” industry trends, where flexibility and rapid configuration are essential. In this context, FMEA generation represents one stage within the broader logistics infrastructure configuration chain.

As in our approach, the classical FMEA components (risk indicators, causes, and effects) are included, along with newer elements related to risk control and mitigation. However, unlike our solution, this study places greater emphasis on improving risk evaluation itself and reducing subjectivity in the assessment process.

In our case, as previously mentioned, the focus is on optimizing the FMEA generation workflow. The improvement of both the analysis and the generation flow is achieved primarily by introducing tools that facilitate the identification, classification, and extraction of relevant data, rather than by modifying the fundamental rules of risk evaluation.

 

…

The method delivers two concrete outcomes: a reduction in analysis time through the integration of an evaluation and FMEA generation tool — precisely the objective we pursue in our own paper — and an increase in accuracy by introducing a new risk evaluation approach (based on fuzzy logic). It is important to note that our proposed solution introduces specific indicators to measure how quickly the FMEA report is generated. The framework we present represents a concrete and comprehensive example of a data flow for FMEA analysis, without neglecting any stage of the process: from the acquisition of input data — such as the number of defects, their impact, and their evolution — to the final generation and publication of the FMEA document.

Another fuzzy-based risk assessment solution is presented in paper [31], which proposes a method for reducing subjectivity in FMEA analysis through the use of fuzzy logic. This further illustrates how integrating analytical tools into the data flow can optimize the process. That study includes a case application in the production of an RFID system for the automotive industry.

Therefore, there are existing studies that address the improvement of FMEA analysis by intervening at the data flow level — an approach that is consistent with our own work. The key contribution of our study lies in the complete presentation and implementation of the entire data flow, as described in detail below.

 

Also we have added the following:

The workflow we propose integrates “classic” data acquisition components via a web-based form, non-AI analysis and classification modules designed to reduce response time, and advanced analytical components leveraging LLM technology. As previously emphasized, the objective is not to modify the risk evaluation methodology itself — which remains grounded in traditional approaches — although future developments may include the integration of methods such as fuzzy logic to enhance risk evaluation. Rather, our primary goal is to improve the data flow leading to the rapid and efficient generation of the FMEA report.

 

5- The S–O–D and AP criteria are presented, but the justification for the values assigned remains somewhat informal.

In the revised manuscript, we have clarified that the AP thresholds are directly linked to measurable operational indicators (internal nonconformities, ppm rates, external nonconformities, customer impact, and defect trends), as presented in Table 4.

We have also explicitly stated that the Severity (S), Occurrence (O), and Detection (D) values were assigned through a structured mapping to the AIAG & VDA 2019 evaluation scales. Each rating was aligned with the corresponding level descriptors in the standard matrices, based on functional impact, historical defect frequency, and the effectiveness of preventive and detection controls.

These clarifications have been added in Section 4 to strengthen the methodological transparency of the scoring approach. (see with purple font color)

 

6- Why didn't you think of applying fuzzy logic to calculate RPN, given that one of the major limitations of the FMEA method is the uncertainty and inaccuracy of parameter values? There is a paper entitled ‘Addressing Uncertainty in Digital Risk Evaluation Using a Fuzzy-FMEA Methodology’ which deals with the same topic using fuzzy logic. I invite you to read it and take it into consideration in your discussion.

We already do that to paper [20] where are presented different assessments methods including fuzzy and ML and paper [21] where an assessment method is presented based on fuzzy. To be clearer we have included in the revised version another reference with fuzzy risk assessment (paper [31] ) and we have included a comparison between our solution and others solutions. Also we have included some comments in Literature reviews (green) to more clearly highlight the technologies and methods used in the referenced articles.

7- Furthermore, there is a lack of robust scientific validation (generalisation, comparison, quantitative metrics) in your work. Please review this point.

In the Results section we have added some comments in text which clarify the results from table 8 (green):

The previously existing solution relied on a conventional approach to FMEA generation. At the start of the production process, or whenever nonconformities occurred, the involved parties (engineers, quality representatives, and management) communicated via email groups and held meetings to establish an action plan.

The proposal of optimization measures, as well as the preparation of the FMEA document, was only partially automated, mainly through the use of predefined templates that were manually completed by members of the quality team.

Compared to the previously used digitalized method based on a database, the pilot solution demonstrated superior performance, shown in Table 8.

 

So in the table 8 we have a comparison with previous solution which was implemented.  Also in table 9 we already have a comparison with a full AI solution (without preprocessing module) to show the advantages of the non -AI preprocessing analysis. So the indicators are users interaction, time reduction, number of failure cases reduced. Also, to be connected to academic environment we have added to the revised version a comparison table (table 10) between our solution and others solutions presented in literature review with comments on it (green):

In the table below (Table 10), we provide a comparison with other studies cited in the literature review section. In addition to the quantitative evaluations presented in Tables 7, 8, and 9 — regarding response time, number of users involved, improvement in risk levels, and the number of resolved defects — we also include a qualitative comparison focused on the components of the data flow. We believe that, from a practical perspective, this data flow has been comprehensively addressed in our proposed solution.

Accordingly, the table highlights how input data are collected and identifies their sources. As can be observed, in all the analyzed articles, the data originate from real industrial environments. The comparison also includes how the data are processed and how the results are ultimately generated and published.

8- Furthermore, the results obtained, ‘100% low risk after optimisation’, are unrealistic without discussion of biases.

We would like to clarify that achieving “100% Low risk after optimization” is not intended to imply the complete elimination of risk. In the context of the AIAG & VDA 2019 methodology, the objective of the FMEA optimization stage is to reduce all identified failure modes to an acceptable risk level through the implementation of appropriate corrective and preventive actions.

The classification of all failure modes as Low after optimization reflects that the identified risks have been mitigated to a level considered acceptable according to defined AP thresholds and industrial quality criteria. It does not mean that failures can no longer occur, but rather that the manufacturer has implemented all technically and operationally reasonable measures to prevent, detect, and control the identified causes.

In industrial practice, an FMEA is considered complete and acceptable (including from the customer perspective) only when high and medium risks have been addressed and reduced to an acceptable level. The revised manuscript clarifies this distinction to avoid any possible misinterpretation.

We believe this explanation better contextualizes the reported result and aligns it with standard FMEA practice.

We have also updated the Results section to explicitly clarify that “100% Low risk after optimization” reflects the achievement of an acceptable risk level according to the FMEA methodology, and not the complete elimination of risk. (see with purple font color)

9- Lack of qualitative assessment of the quality of AI recommendations.

To address this point, we have added a qualitative clarification in the revised manuscript (Results and Conclusions sections) regarding the quality of the generated recommendations. We explained that the actions proposed for High-risk failure modes (FMEA revision and Control Plan update) and for Medium-risk cases (mandatory optimization measures) are based on the structured FMEA analysis and on validated process knowledge defined by the multidisciplinary team.

The recommendations were reviewed and confirmed by the responsible process and quality engineers during the pilot implementation phase, ensuring their technical relevance and practical applicability in the real production environment. We also clarified that the system supports expert decision-making rather than replacing professional engineering judgment.

We believe these additions strengthen the qualitative validation of the recommendations and clearly demonstrate their alignment with industrial practice and expert assessment. (see with purple font color).

10- The discussion is very weak; the authors should discuss and compare with existing work. In fact, there is a great deal of work that has dealt with FMEA. However, the references presented are very limited.

In Literature review we insert new references and comments/comparison with our study (green background font).

 

11- I invite you to review the conclusion; it is weak and lacks perspective.

We have improved the conclusions sections with an elaborated future trends and extended explanation regarding the results (see purple and green texts):

The application of the FMEA method according to AIAG & VDA 2019 to the 3D laser cutting process enabled the identification of the dominant failure modes (missing holes and an unsatisfactory cutting edge surface appearance) and their associated causes, with a direct impact on assembly and customer acceptance. The initial assessment showed a risk distribution of 67% Low, 29% Medium, and 4% High, while the implementation of corrective and preventive measures led to the reduction of all failure modes to a Low risk level (100%) after reassessment.

Beyond the operational efficiency improvements, the study confirms that the proposed framework provides technically sound and industrially applicable recommendations. The actions suggested for High and Medium risks were consistent with expert evaluations and aligned with established quality management practices. This highlights that the system supports engineering decision-making rather than replacing professional expertise.

The presented solution is based on two processing levels: one, non AI in which the risk is established depending on the number of non-conformities and their type (internal, external) and an AI level based on LLM in which a knowledge base is used to generate risk analysis and propose optimizations. In this way, the response time of the solution is improved. The solution proposed in this paper has been implemented — in the form of a pilot system — for a specific production process (3D LASER cutting). However, it can be extended to other manufacturing processes as well. First, many failure modes may be similar across different production processes. Second, the knowledge base can be adapted accordingly — as shown, it is a parameter transmitted to the LLM with each request. This makes the framework flexible and scalable, allowing it to be customized for various industrial contexts simply by adjusting the expert-defined knowledge base.. Future research directions involve completing the knowledge base to support more production processes and expanding the industrial domain in which it can be used. Future developments may also include advanced risk assessment tools aimed at further reducing subjectivity and enhancing the overall performance of FMEA analysis. It is certainly an Industry 4.0 tool that can be successfully used in all branches of production.

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have invested considerable effort in addressing all of the issues raised in the previous round of review, which has significantly improved the quality of their paper. 

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have carefully adressed my concerns. I have no further comments and recommend the manuscript for acceptance in its current form.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have taken all comments into consideration. The paper is ready for publication.