2. Literature review

3. Methodology

3.1. Systematic framework for integrating LLMs into FMEA

Building on the advantages of integrating LLMs into FMEA, we propose a comprehensive process model and an information system model. This framework is designed to streamline the FMEA process and enhance its effectiveness. The framework was developed based on a comprehensive literature review and the authors’ experience. The proposed process model comprises the following steps:

1. 1. Data collection: Relevant data is gathered from a variety of sources, including design data, historical failure records and other contextual information. This data forms the foundation for training AI algorithms and provides critical insights for risk analysis.

2. 2. Data pre-processing: The collected data undergoes thorough pre-processing to ensure its quality and compatibility with the LLM. This stage may involve data cleaning, normalization, feature extraction and handling of missing values or outliers. The aim is to automate pre-processing with computerized tools to ensure efficiency and accuracy.

3. 3. Model training: Various subsets of data (such as previous FMEAs, external reviews, etc.) are labeled with expected outputs (failure modes, effects, risk assessments, corrective actions, etc.).

   In this step, we can use prompt engineering, fine-tuning or RAG. In this work, we used prompt engineering techniques to help the LLM generate relevant results. To do this, we had to formulate precise prompts and queries to extract the necessary information from the model. By carefully crafting these prompts, we wanted to improve the model’s ability to understand and accurately respond to the specific context and details of the FMEA. Through this process, we evaluated the effectiveness of the prompt engineering in achieving satisfactory accuracy. The goal was to determine if the prompt engineering method alone could make the necessary improvements or if additional steps were needed in the future.

   Had fine-tuning or RAG been applied, this would have required formal training of the model, including the use of a loss function (e.g. cross-entropy loss) to guide the optimization process and a parameter tuning strategy (e.g. grid search) to determine the best hyperparameters, like learning rate and batch size. However, since prompt engineering was used in this work, these training-specific components were not required.

   After several iterations, the prompt for GPT, shown in Figure 4, was designed. The prompt was the same for all LLMs, but different APIs require some format changes.

4. 4. Application – extraction of specific information for FMEA: Once the system is trained, it is applied to the entire dataset. The LLM then suggests failure modes, effects, risk assessment calculations and corrective actions based on the data.

5. 5. Integration of LLMs tools into the FMEA process; regular data analyses, improvement of the FMEA process and decision support on the system level: The generated FMEA information is incorporated into PDP and the Knowledge Management system (KMS). The LLMs tools can produce FMEA reports, visualizations, comprehensive summaries and trends for supporting decision-making. Beyond FMEA, a good overview of data can contribute to the company’s quality assurance system, helping to identify competency needs, recurring failure modes, process bottlenecks and more.

The information system model for integrating LLM tools includes data collection, extraction, knowledge management and application within an industrial setting, as depicted in Figure 1. It leverages company-specific knowledge extracted from key documents in the product lifecycle management (PLM) system, such as previous FMEAs, engineering changes (ECs), and product history files (Tavčar, Benedičič & Žavbi Reference Tavčar, Benedičič and Žavbi2019). In addition, AI analysis incorporates external sources related to the company’s production program and technology-specific information. This is particularly important when developing new products, as external knowledge sources are often crucial. Selected documents are analyzed by AI, and the extracted information is reused in FMEA activities and other PDPs. Systematic data analysis is conducted at regular intervals and leads to the definition of corrective actions for PDP, EC management (ECM), and the FMEA process (Figure 1). LLM tools are exceptionally powerful in analyzing large amounts of data, as they can identify patterns and insights that human analysts might missed. By implementing this framework, organizations can significantly enhance the efficiency and effectiveness of their FMEA processes and leverage the strengths of LLMs to improve risk analysis and decision-making.

Figure 1. Information system model for LLM application in FMEA (upgraded from El Hassani et al. Reference El Hassani, Masrour, Kourouma, Motte and Tavčar2024).

The implementation of the process model and the information system model within the proposed framework must align with the specific phase of the product life cycle. The maturity level of a product plays a crucial role in determining the optimal support required for the FMEA process. For new products where the company has limited experience, external sources of knowledge are prioritized. These sources provide valuable insights and data that are essential for identifying potential failure modes and accurately assessing risks.

However, for more mature products that have several years of development, manufacturing and sales experience, internally generated knowledge sources become more important. This internally collected data, including historical FMEA reports, ECs and product history files, provides a rich information base that can be used to improve the FMEA process. Utilizing this internal knowledge allows for more accurate and informed risk assessments, as it is based on the company’s own experience and lessons learned over time.

3.2. Mapping FMEA challenges to framework solutions

The proposed framework systematically addresses the key challenges in the FMEA process as highlighted in the literature review. Table 1 lists the challenges and the corresponding solutions provided by the framework.

Table 1. Mapping FMEA challenges to solutions.

The proposed framework addresses key challenges identified in existing research by using LLMs to automate data pre-processing, failure mode identification and risk analysis to reduce manual effort and minimize human error. To tackle inefficient reuse of knowledge, KMS was integrated to systematically capture and reuse findings from previous analyses. In addition, the framework uses LLMs to effectively process unstructured data and improve data processing and analysis. Scalability and contextual understanding have been improved through prompt development and fine-tuning to ensure adaptability and efficiency in various FMEA tasks. These improvements directly overcome the limitations highlighted in previous research and provide a more robust approach.

4. Practical case study and evaluation

The applicability of the proposed framework is validated by a case study using publicly available data from the automotive industry, in particular data from vehicle reviews by private individuals. This type of data is less structured compared to company FMEA reports and other product-specific documents and therefore poses a greater challenge. This also allows the first steps of the framework process model to be implemented and tested without the security issues that would arise when using proprietary company data. Although this approach has limitations (e.g. the incomplete representation of the company context and the inability to extract corrective actions as it is limited to the design FMEA), it still provides a valuable opportunity to evaluate various aspects such as automatic data pre-processing, model training and information extraction.

In this case study, we aimed to test two key aspects where LLMs can be beneficial for FMEA:

• • Accuracy of information extraction with training:

The aim was to assess how accurately an LLM can extract relevant information from the data when the model has been trained specifically for the task. For this, we used a model for extracting negative reviews and identifying associated components (the part finder model) as described in Section 4.2 and achieved an accuracy of 98–99%.

• • Accuracy of information extraction without training:

Here we evaluated the performance of the LLMs in extracting information from the data without specific training. Table 2 shows a summary of the results of a semantic comparison between different LLMs using only prompt engineering. Human-identified error modes (gold standard) were compared with GPT-4o, GPT-4 and Gemini 1.5 FLASH. The comparison included 100 reviews and focused on the similarity scores.

Table 2. Results of semantic comparison between different LLMs and human analysis.

These aspects were tested using the five-step framework process model. First, the preprocessing of the data was tested, then the extraction of the failure modes and finally the relevance of the effects suggested by the LLM. The detailed process flow applied is shown in Figure 2 and the result in Figure 3.

Figure 2. Process flow of the automated FMEA framework.

Figure 3. Example of an automatically generated FMEA table.

Figure 3 shows an example centered on design FMEA (DFMEA), which is the specific focus of our case study. While DFMEA is a type of FMEA that deals specifically with design-related failures, the underlying principles of risk identification, analysis and mitigation apply broadly to other types of FMEA, including process FMEA (PFMEA). Therefore, the DFMEA example effectively demonstrates the utility of LLMs in systematically analyzing failure modes, which is representative of FMEA activities in general. We believe that this example provides valuable insights that are applicable to a wider range of FMEA scenarios.

For the case study, we initially used GPT-3.5 as a proof of concept to evaluate the feasibility of the approach (El Hassani et al. Reference El Hassani, Masrour, Kourouma, Motte and Tavčar2024). After validating this concept, we moved on to using GPT-4, GPT-4o and Gemini 1.5 FLASH to enhance the performance. We compared these models based on their accuracy and overall effectiveness. The summarized results can be found in Table 2.

The steps involved in the case study are as follows:

1. 1. Data collection

   We used a dataset of car reviews from the Kaggle platform (www.kaggle.com), which included fifty different datasets of customer reviews for 50 vehicle brands (AnkurJain Reference AnkurJain2019), totaling about 227,000 reviews.

2. 2. Data pre-processing

   This process was carried out in two main steps:

   a) Filtering out reviews without identified parts (part finder model)

First, we performed a string comparison to filter out reviews without identified parts. The data was cleaned, formatted and merged into a single dataset. As the FMEA focused on vehicle parts, reviews that contained no references to parts had to be excluded. To facilitate this, we compiled a list of parts by scouring websites such as Wikipedia (“List of auto parts,” 2023) and List Explained (Kan Reference Kan2022). Using a string comparison, ratings without identified parts were removed. However, this method had its limitations (e.g. spelling errors, incomplete data and use of different languages). Therefore, we developed a model using LLMs to extract reviews with parts more accurately. The model was trained with GPT-3 (text-avinci-002) and achieved 98–99% accuracy, including reviews in French and Spanish. This refined dataset included about 100,000 reviews.

To perform fine-tuning, we first formatted our data according to the requirements for fine-tuning a text-davinci-003 model. Our dataset was streamlined into two columns: a “Prompt” column containing the reviews and a “Completion” column containing the names of the extracted parts determined via string comparison. We added special tokens “- >” at the end of each prompt to indicate that the prompt has concluded and that the model can begin the completion and “END” to signal that the completion is finished and that the model should stop.

Subsequently, we converted our dataset into a JSONL file, which is the required format for fine-tuning an OpenAI model. After preparing the JSONL file, we uploaded it to OpenAI’s fine-tuning platform using the OpenAI CLI. We configured specific training parameters, such as the number of epochs and batch size, optimizing them to prevent overfitting. During the fine-tuning process, the model adapted to our dataset, enhancing its performance.

Finally, we evaluated the fine-tuned model by testing it on a separate validation set, measuring its accuracy and observing its behavior on unseen prompts. Upon completion of the fine-tuning process, we obtained a results file from the OpenAI API containing detailed evaluation metrics, including loss, accuracy and other key indicators. These metrics enabled us to assess the effectiveness of the fine-tuning, determining how well the model had adapted to the dataset and whether it was suitable for the task or required further adjustments.

b) Extracting negative reviews (sentiment analysis model)

Next, we focused on identifying negative reviews, as they often contain valuable information about potential problems and concerns related to automotive components and systems. The reviews were divided into negative and positive ratings using an existing labeled dataset (Maas et al. Reference Maas, Daly, Pham, Huang, Ng, Potts, Lin, Matsumoto and Mihalcea2011) for training. Several deep learning algorithms were tested. TensorFlow’s CNN (Abadi et al. Reference Abadi, Barham, Chen, Chen and Davis2016), which uses bidirectional short-term memory (BiLSTM), achieved an accuracy of 87%. Fine-tuning with GPT-3 (Curie model) improved the accuracy to 97%. The model obtained by GPT-Curie was then used to classify the reviews into negative and positive categories. Only the negative reviews were used to extract failure modes for the FMEA. This pre-processed dataset included about 18,000 reviews, all of which were negative and contained names of automotive parts.

For training the TensorFlow models, we split the dataset into training, testing and validation sets. Specifically, 80% of the data was used for training, with 20% of the training set reserved for validation, while 20% was allocated for testing. We employed cross-validation to optimize hyperparameters. Throughout training, we monitored loss curves to compare validation against training, mitigating overfitting. For certain models, TensorFlow’s EarlyStopping method was implemented to stop training early, preventing overfitting or excessive adjustment to the training data.

1. 3. Model training

   In this step, we could use prompt engineering, fine-tuning or RAG. In this work, we used prompt engineering techniques to help the LLM generate relevant results. This required us to formulate precise prompts and queries to extract the necessary information from the model. By carefully crafting these prompts, we wanted to improve the model’s ability to understand and accurately respond to the specific context and subtleties of the FMEA. Through this process, we evaluated the effectiveness of the prompt engineering in achieving satisfactory accuracy. The goal was to determine if the prompt engineering method alone could make the necessary improvements or if additional steps were needed in the future.

   After several iterations, the prompt for GPT, shown in Figure 4, was designed. The prompt was the same for all LLMs, but different APIs require some format changes.

   

   Figure 4. Prompt for search of failure modes and effects, which is formatted for GPT-4 and GPT-4o.

2. 4. Extraction of specific information for FMEA and validation of results

   A total of 100 reviews were randomly selected using the sentiment analysis model and the part finder model: 20 reviews each for the parts “door,” “tire,” “seats,” “wheel” and “window.” The LLM was prompted to extract failure modes as accurately as possible, while also making suggestions for other FMEA information elements. To create the files, we need structured responses (in JSON format) so that we can programmatically retrieve the necessary values and insert them into the Excel file.

Similar to coding methods in qualitative research (e.g. Campbell et al. Reference Campbell, Quincy, Osserman and Pedersen2013), the failure modes were first analyzed manually and evaluated independently by two experts. They compared their results one-to-one and reached a consensus after discussion to ensure accuracy and minimize bias. As the data were user reviews and not technical reports, the failure modes reflected the users’ perspective, which might not always match the technical definitions.

Semantic analysis was used to compare the results of the different LLMs. Semantic analysis is a process in natural language processing (NLP) that focuses on understanding the meaning of words and texts in context. It goes beyond the simple matching of keywords and interprets the actual meaning, relationships and nuances between words, such as synonyms, antonyms or hierarchical relationships. By capturing the context and subtle differences in meaning, semantic analysis helps us to create a more accurate assessment for the FMEA of each LLM.

LLMs outperform traditional NLP methods in semantic analysis because they use deep learning techniques, especially transformer architectures, to dynamically understand context and meaning. Unlike traditional methods that rely on fixed word embeddings and statistical correlations, LLM model words as vectors in a high-dimensional space where their positions are influenced by the words around them, capturing relationships, meanings and nuances more accurately. LLMs achieve this by learning from huge datasets and using attention mechanisms to focus on different parts of a sentence and understand word meanings in context rather than in isolation. This allows them to capture complex language patterns, disambiguate meanings and provide more accurate and flexible interpretations. This is particularly useful when comparing the FMEA provided by each model to the “golden standard,” as this requires a deep understanding of nuanced or contextual information (Xu et al. Reference Xu, Wu, Zhao, Shu, Liu, Liao and Li2024).

We compared the results of different LLMs and the failure modes found by experts. The used prompt for the semantic comparison of failure modes is presented in Figure 5.

Figure 5. Prompt for semantic comparison of failure modes with GPT-4o.

An example of the numerical semantic analysis is shown in Figure 6, which contains input data, numerical results and reasoning for the evaluation.

Figure 6. Example of a semantic comparison with numerical evaluation.

The semantic comparison of the failure modes was carried out separately for each review. The results of the semantic analysis were not deterministic. For the upper example from Figure 6, we could obtain a similarity score of 8/10 or 9/10. The variation is within a predictable range and acceptable according to the authors.

Application of LLM for semantic comparison has many advantages. The results of the comparison are understandable and logical for human criteria. The comparison can be performed on a much larger amount of data, which is not possible for humans.

There are several LLMs on the market. One of the objectives was to compare the results between different models for the specific task – the search of failure modes in reviews. The results of the semantic comparison are summarized in Table 2.

Results of comparison:

Explanation of the semantic comparison parameters:

• • Average similarity level: Represents how close the failure modes identified by LLM are semantically to the “golden standard” created by humans. This results in an overall performance score for each model.

• • Standard deviation: Highlights the variability of semantic similarity level between different reviews, where a lower standard deviation reflects greater consistency.

• • Number of cases (similarity = 1; 2): Indicates instances of significant deviation or poor performance, useful for identifying model weaknesses.

• • Number of cases (similarity = 6; 7; 8; 9; 10): Indicates the ability of the LLM to achieve high semantic relevance, closely matching the failure modes extracted by humans.

The visual differences when comparing the text output of different LLMs are significant. However, if we perform a deeper comparison on a semantic level, the differences are smaller. GPT-4o and GPT-4 provide more similar results compared to Gemini, and this is in line with expectations. In particular, Gemini represents error modes with different words.

Based on recommendations from the literature (Reference Landis and KochLandis and Koch, 1977) and manual comparisons, a similarity level of 6 or more is considered a substantial agreement. GPT-4 and GPT-4o have a degree of similarity with substantial agreement of 91 and 87%, respectively. Gemini 1.5 is still good with 80%.

Table 2 shows that a large percentage of the ratings have no or only a very low similarity with rating 1 or 2. A manual comparison has shown that no similarity is found if the manual search has found a failure mode but the LLM has not or vice versa.

It can be critical if the failure mode is not recognized in the text. For example, GPT-4o did not recognize a failure mode in the sentence: “No way can you fit 3 car seats in the back row,” and the similarity was scored as 1.

On the other hand, in the sentence: “The car does not handle well on rain or snow,” the expert did not identify a failure mode, but the LLM did, and therefore the similarity was rated 1. However, from the end user’s point of view, it is a problem, and we can agree that the LLM is right.

In some cases, a failure mode was not understandable to the LLM, and the answer was N/A. The expert was able to identify a failure mode.

GPT-4o recognized a failure mode in the sentence: “There is no window bar between the front and back window,” but the expert did not.

A general conclusion is that different LLMs provide useful results. More advanced LLMs such as GPT-4o generally provide better results. However, the results sometimes differ from the expert’s pragmatic assessment. This is the reason why the similarity of GPT-4o is worse than the similarity of GPT-4.

The conclusion may be that the search with LLMs is not perfect. However, the search enables a computer-aided search in large datasets. So LLMs enable something that is simply not possible for humans. We can expect better results when searching in professional documents with a more consistent document structure and more precise language.

Finally, the system was asked to generate potential causes, current controls, severity, occurrence and detection for a given failure mode (examples can be found in Figure 3, right). Several suggested values were consistent with the reviews, but no quantitative analysis was performed as there was no specific data to justify the proposed numbers.

5. Integration of LLMs tools into the FMEA process, regular data analyses, improvement of the FMEA process and decision support.

5.1 Integration of the LLM application into the FMEA process.

The supporting tools and methods must be well integrated into the key processes. The data elements obtained from the sentiment analysis model and the parts finder model were systematically organized to facilitate the subsequent FMEA. Using a well-formulated prompt, the LLM generated a comprehensive table of relevant information and an FMEA table for each review. The extracted FMEA information is shown in Figure 3. This illustrates the capabilities of the LLM as an integral part of a KMS.

In the case study, we investigated the identification of failure modes from individual review texts. However, in a more realistic scenario, we may need to analyze failure modes for a given component in multiple reviews simultaneously. This more complex task leads to several considerations regarding the approach and response time.

In the context of searching for all potential failure modes associated with a particular component in a large set of reviews, two possible approaches can be considered:

• • Extraction of all data with subsequent filtering: In this method, the system first extracts all relevant failure modes from the entire set of reviews. Once the data is collected, it is filtered by component within the FMEA table. While this approach ensures comprehensive coverage of all possible failure modes, it can significantly increase processing time as large amounts of data are processed before filtering by component.

• • Component-specific prompts for targeted extraction: Another, more targeted approach is to design prompts that explicitly request the extraction of failure modes for a specific component across all documents. By specifying the component in the initial prompt, the LLM can limit its search to the relevant data, which could improve response time and reduce unnecessary processing. This approach could increase efficiency when working with large datasets as it directly targets the information that is relevant to the component under investigation.

There are other strategies that can be used to optimize the process:

• • Batch processing of reviews: Instead of processing all reviews at once, the data could be split into smaller batches, with each batch focusing on a specific component. This would distribute the computational load and could lead to faster response times while maintaining accuracy.

• • Pre-classification of reviews by component: A pre-processing step could classify reviews based on the components they mention. This classification would allow the LLM to focus only on the subset of reviews that are relevant to a particular component, further reducing the data load and speeding up the extraction process.

The data used in the case study – end-user reviews – is not reliable data to determine effects, failure mechanisms and probability of severity, occurrence and detection. We were not able to assess the quality of the predictions. However, according to the experts, the answers were appropriate for the given context. At this point, we would like to highlight the technical possibilities of integrating applications and the next steps for industrial implementation.

5.2 Periodic data analysis (LLM) for decision making and corrective actions.

The integration of FMEA into the PDP using LLMs can significantly improve decision support by streamlining information extraction and risk management. In this proposed framework (Figure 1), a knowledge base serves as a central repository for the systematic organization and storage of FMEA results and related documents. This integration provides comprehensive access to information on failure modes and associated risk mitigation strategies, supporting different phases of the product development cycle.

Effective knowledge management is based on continuous learning and system improvement. Regular reviews of product development data and FMEA results by senior engineers are crucial, with corrective action taken where necessary (Figure 1). Obtaining key information can be time-consuming but is essential for informed decision-making and the implementation of effective corrective actions. This is where LLMs offer a powerful solution. Their advanced NLP capabilities enable fast and accurate data extraction, significantly reducing the time and effort required for these tasks.

Although regular data analysis using LLMs is beyond the scope of this paper, we have conducted an example of data analysis to demonstrate the practical application of these tools and illustrate the steps described in the framework (Figure 1). Figure 7 shows a visual representation of the statistical data generated by LLM tools and illustrates their efficiency in processing large volumes of documents. The visual representation of data, such as the frequency of parts involved in customer reviews or the frequency of particular failure modes, can be of great help in long-term decision-making. Only systematic and statistical processing of data can identify trends and show the big picture that is not visible in day-to-day work. Traditional manual methods for searching, collecting and summarizing large datasets are not only time-consuming but also costly. LLMs can effectively automate these tasks, enabling continuous data analysis and the improvement of processes and products based on this data.

Figure 7. The periodic analysis supported by LLM and visual presentation as a tool for corrective actions.

The ability of LLMs to scale FMEA analyzes across multiple tests and components and integrate the results into a knowledge base underscores their potential to revolutionize engineering processes. By automating these aspects of FMEA, companies can significantly speed up the analysis process and improve decision making. This automation reduces the burden of manual data analysis and allows engineers and decision makers to focus on more strategic and challenging tasks. As a result, overall productivity and innovation are increased as more resources are available for exploring new ideas and optimizing product design.

In summary, the use of LLMs in FMEA and product development offers significant benefits. They enable the rapid extraction of relevant information, support decision-making processes and facilitate the continuous improvement of products and systems. By using these tools, companies can make more informed decisions, implement effective risk mitigation strategies and ultimately drive innovation and quality in product development.

6. Conclusion and prospects for future research

This study addresses two major gaps in the literature: the lack of a comprehensive framework for the integration of LLMs into the FMEA process and the limited research on the practical challenges of such integration. The proposed framework combines a process model and an information system model to streamline FMEA activities and demonstrate the benefits of LLMs in improving efficiency, scalability and accuracy.

Key findings

• • Framework contributions:

  • ○ Development of a systematic framework that uses LLMs to automate data pre-processing, failure mode identification and risk analysis to reduce manual effort and minimize human error.

  • ○ Use of a KMS to improve the reuse of knowledge to ensure that past analyses are incorporated into future FMEA processes.

  • ○ Demonstrate the ability to process unstructured data, such as user reviews, and transform it into structured input for FMEA.

  • ○ Improved scalability through the ability of LLMs to process large datasets quickly and efficiently.

• • Insights from the case study:

  • ○ Minimal programming and no fine-tuning were required, with effective results achieved through prompt engineering in just 2–3 iterations.

  • ○ The generated FMEA tables highlighted the potential of LLMs for organizing, documenting and visualizing information to improve risk assessment and decision-making.

  • ○ Working with company-specific FMEA reports could lead to even greater precision in the extraction of failure modes.

• • Limitations:

  • ○ Generalization of outputs: LLMs occasionally extrapolate failure modes beyond the input data, which raises questions about the desired level of detail for specific FMEA tasks.

  • ○ Reliance on SMEs: Outputs had to be validated and refined by SMEs, limiting the scalability of fully automated processes.

  • ○ Domain-specific focus: The study was focused on automotive parts, leveraging LLMs’ pre-existing knowledge in this area. The extension to less common products or industries has yet to be tested.

  • ○ Confidentiality of data: Security concerns, such as risks associated with external LLMs, were not fully addressed in this study.

Future research directions

• • Extension of the framework: Inclusion of additional FMEA elements, such as the selection of corrective actions and the evaluation of risk mitigation strategies.

• • Improve generalization control: Explore training and prompting strategies to minimize undesirable generalizations while retaining the ability to infer relevant context.

• • Dynamic data management: Integrate real-time updates that link extracted information to source documents and dynamically refresh results as new data becomes available.

• • Testing in diverse domains: Apply the framework to lesser-known products and industries to evaluate its robustness and adaptability.

• • Improve security and privacy: Deploy locally hosted LLM infrastructures or explore open access models, such as OpenLLaMA (Geng & Liu Reference Geng and Liu2023), to safeguard sensitive proprietary data.

• • Industrial validation: Conduct large-scale evaluations in real industrial contexts to measure the impact of the framework on efficiency, time savings and accuracy. For example, Dell’Acqua et al. (Reference Dell’Acqua, McFowland, Mollick, Lifshitz-Assaf and Kellogg2023) have shown that LLMs improve the quality of tasks by over 40% and reduce task duration by 25%, which is promising for FMEA applications.

This study illustrates the transformative potential of LLMs in automating labor-intensive and time-consuming FMEA processes. Although the framework offers significant benefits, such as faster pre-processing and better data organization, future efforts should focus on addressing the identified challenges to enable broader industrial application. By refining the framework and expanding its scope, LLMs can unlock their full potential in engineering processes and pave the way for more efficient, accurate and innovative solutions.