FMEA Challenges and Solving it Using AI - FMEA Series Chapter 3

Challenges while performing an FMEA

While the FMEA is a powerful technique to identify gaps in design, engineers frequently encounter challenges during its creation. Some of these challenges include 

1. spending too much time creating an FMEA form sheet from an architecture, often extending to a few weeks​; 

2. spending too much time spent in repetitive tasks like filling out effects, cause and mitigation for similar failure modes; 

3. establishing consistency in severity, occurrence and detectability rating across similar failure modes​.

In summary, engineers spend too much time filing out the form sheets rather than engaging in analytical tasks. 

Recommended reading: Read Chapter 2 - How to Perform an FMEA? It provides an explanation, as well as an example on how an FMEA is performed

How can AI and LLMs be used in FMEA

Large Language Models (LLMs) are a type of artificial intelligence trained on massive amounts of data. Through this training, they learn language patterns, factual knowledge, reasoning, and problem-solving skills. This ability to learn from vast datasets makes LLMs valuable tools for supporting and automating tasks such as Failure Modes and Effects Analysis (FMEA). 2 ways how LLMs can help

1. Risk Assessment Support

Traditional FMEA risk assessments often rely on subjective judgment when assigning severity, occurrence, and detection ratings. LLMs can enhance this process by analyzing historical failure data (from the internet or local databases) and provide evidence-based recommendations. LLMs can process maintenance records, warranty claims, and industry databases to establish statistical baselines for failures.

In addition, LLMs can evaluate the effectiveness of proposed detection measures. This enables teams to make informed decisions about risk levels based on documented performance rather than expert opinion alone. 

2. Information Management

Organizations face two related challenges: inconsistent documentation quality across FMEA projects and difficulty using institutional knowledge about failures and mitigation strategies. LLMs can address both by serving as intelligent information management systems. 

For documentation, LLMs can standardize reports by generating comprehensive analyses that adhere to organizational templates and industry standards. They ensure that risk assessments include proper justification, recommended actions are specific and measurable, and analytical logic flows clearly from failure mode identification to mitigation strategy. 

For knowledge management, LLMs make historical FMEA data searchable and accessible across the organization. Engineers can query LLMs to identify relevant risks from previous analyses, access proven mitigation strategies for similar failure modes, and understand the reasoning behind past design decisions. This capability breaks down knowledge silos and enables teams to learn from experiences across different divisions and projects. 

How can correctness of LLMs be ensured?

1. Human Oversight 

The most critical requirement for LLM implementation in FMEA is maintaining qualified human oversight throughout the analysis process. LLMs must function as analytical tools that augment rather than replace human expertise, particularly in safety-critical applications where incorrect analysis of errors can have severe consequences. 

Organizations should establish clear criteria for review adequacy, document validation rationale, and implement escalation procedures for uncertain cases. This ensures that LLM efficiency gains do not compromise the analytical rigor that FMEA requires for effective risk management. 

2. LLM Optimization

Ensuring reliable LLM performance requires both proper input design and model selection. Organizations should develop standardized prompting templates that specify analytical frameworks, required detail levels, and instructions for uncertainty handling. Effective prompts include sufficient context for operating conditions and system specifications to ensure relevant outputs. 

When resources permit, domain-specific models trained in engineering and safety data demonstrate an improved understanding of technical vocabulary and industry-standard failure modes. This specialization reduces irrelevant suggestions and improves overall output quality compared to generic LLMs.

3. Continuous Validation

LLM response validation should be treated as an iterative process where initial outputs serve as starting points for refinement through multiple review cycles. Organizations should implement systematic evaluation methods that assess output quality based on technical accuracy and relevance to the system under analysis. 

Tracking patterns in LLM performance enables continuous improvement of both applications and validation processes. This approach recognizes that LLM integration represents an ongoing optimization opportunity rather than a one-time implementation