Smarter System Safety: AI and Knowledge-Based FMEA

Author: Denis Avetisyan

This review explores how integrating artificial intelligence and formal knowledge representation is reshaping Failure Mode and Effects Analysis for more reliable and adaptable systems.

The BEQIC team formalized a four-step Failure Management Analysis (FMA) framework, acknowledging that even meticulously designed systems inevitably accrue failure modes which will, in production, challenge theoretical elegance and necessitate pragmatic, iterative refinement.

The article examines current developments and future directions in applying AI and ontology to enhance FMEA within a Model-Based Systems Engineering framework.

As engineered systems increase in complexity, traditional Failure Mode and Effects Analysis (FMEA) struggles to keep pace with demands for robust reliability assessment. This review, ‘AI- and Ontology-Based Enhancements to FMEA for Advanced Systems Engineering: Current Developments and Future Directions’, synthesizes recent advances in leveraging artificial intelligence and formal knowledge representation via ontologies to transform FMEA into a more dynamic and intelligent process. By integrating these technologies with Model-Based Systems Engineering, this work demonstrates pathways towards automated failure prediction, improved knowledge extraction, and enhanced system resilience. Will these developments pave the way for truly adaptive and self-learning safety-critical systems?

The Inevitable Complexity: A Failure Foretold

Contemporary systems, from global financial networks to sophisticated aerospace engineering, are characterized by an escalating level of intricacy that fundamentally challenges conventional failure prediction techniques. These systems aren’t simply collections of components; they exhibit emergent behaviors arising from the interactions of numerous interconnected elements, making linear, reductionist analyses increasingly inadequate. Traditional methods, often reliant on analyzing individual components or simplified models, struggle to account for these nonlinear dynamics and cascading failures. The sheer scale of these systems-often involving millions of lines of code and countless physical parts-compounds the problem, creating a combinatorial explosion of potential failure modes that exceed the capacity of even the most powerful computational resources. Consequently, predicting system behavior with sufficient accuracy to prevent critical failures requires a paradigm shift towards more holistic, model-based approaches capable of capturing the inherent complexity and interdependencies within these modern infrastructures.

Addressing system failures after they occur carries substantial costs, extending beyond immediate repair to encompass downtime, lost productivity, and potential damage to reputation. In critical applications – encompassing sectors like aerospace, healthcare, and energy infrastructure – a reactive stance is simply unacceptable due to the high stakes involved; a failure can lead to catastrophic consequences, including loss of life. The economic burden of post-failure fixes often dwarfs the investment in preventative measures, yet the true cost is frequently underestimated as it fails to account for long-term implications like eroded public trust or legal liabilities. Consequently, a shift towards predictive maintenance and proactive system health monitoring is becoming increasingly vital, allowing for the identification and mitigation of potential issues before they escalate into full-blown failures and their associated repercussions.

Ensuring system reliability in the face of increasing complexity demands a shift from simply reacting to failures, to actively predicting and preventing them. A proactive, model-based approach utilizes computational simulations and analytical techniques to anticipate potential issues before they manifest in real-world operation. However, the efficacy of such systems hinges on the development of robust methodologies – those capable of accurately capturing system behavior, accounting for uncertainties, and scaling to accommodate the intricate interdependencies within modern designs. These methodologies must move beyond simplistic representations, incorporating advanced modeling techniques like agent-based simulation, Bayesian networks, and machine learning to provide a sufficiently detailed and nuanced understanding of system dynamics. Without these rigorous foundations, model-based approaches risk producing inaccurate predictions and ultimately failing to deliver the promised improvements in reliability and safety.

This diagram illustrates how system-level failures originate from underlying root causes, as detailed in the referenced failure analysis handbook.

MBSE: A Paper Shield Against the Inevitable

Model-Based Systems Engineering (MBSE) establishes a formalized approach to system development that centers on the creation and maintenance of system models from initial concept through retirement. These models, encompassing various views and levels of abstraction, serve as a single source of truth for system requirements, design, and behavior. By constructing and analyzing these models throughout the entire lifecycle – including requirements elicitation, architectural design, detailed design, implementation, testing, and maintenance – potential inconsistencies, ambiguities, and design flaws can be identified and addressed earlier in the process. This proactive approach reduces the risk of costly rework, delays, and performance issues that often arise when relying solely on document-based systems engineering methods. The use of modeling tools and standardized modeling languages further facilitates communication, collaboration, and traceability throughout the project.

Function modelling, utilizing frameworks such as Functional Block Diagrams (FBS), systematically decomposes a system into its constituent functions and their interrelationships. This process defines not only what a system does – its purpose – but also how it achieves that purpose through defined behaviors and a logical structure. FBS, specifically, represents functions as black boxes with defined interfaces – inputs and outputs – allowing for analysis of information flow and dependencies. The resulting model details functional hierarchies, enabling engineers to understand the system’s architecture, identify potential redundancies, and assess the impact of changes to individual functions on overall system behavior. This structured approach is crucial for both the design and verification phases of system development, providing a clear and unambiguous representation of system intent.

While model-based systems engineering (MBSE) and function modeling establish a structured basis for system understanding and preemptive problem identification, these methodologies are limited by their reliance on manual analysis. The data generated through these models requires significant human effort to interpret and derive actionable intelligence. Current MBSE practices do not natively incorporate automated reasoning or learning capabilities. Consequently, opportunities to identify complex interdependencies, predict emergent behaviors, or optimize system performance based on large datasets are unrealized without the integration of artificial intelligence and machine learning techniques to augment the analytical process.

Ontology plays a key role in supporting Model-Based Systems Engineering (MBSE) by formalizing domain knowledge into a structured and machine-readable format. This involves defining concepts, properties, and relationships within a specific domain – such as aerospace, automotive, or medical devices – using a formal language. The resulting ontology serves as a shared vocabulary and understanding, enabling consistent interpretation of system models and facilitating automated reasoning. Specifically, ontologies allow for the validation of model consistency, the automated derivation of system requirements, and the support of knowledge reuse across different projects and engineering teams. This formal representation is crucial for enabling interoperability between various MBSE tools and promoting accurate communication of system information throughout the lifecycle.

The Function-Behaviour-Structure (FBS) framework illustrates a process of translating desired functions into behaviors and ultimately realized structures.

AI-FMEA: Trading Manual Labor for Algorithmic Hope

Failure Mode Effects Analysis (FMEA) is a structured, proactive technique for identifying potential failure modes in a design, process, or system. The application of Artificial Intelligence (AI) to FMEA augments this process by enabling automation of tasks traditionally performed manually. This includes automated data collection, pattern recognition within historical failure data, and the suggestion of potential failure modes based on system characteristics. AI-driven FMEA aims to improve the thoroughness of the analysis, reduce the time required to complete an FMEA, and minimize subjective biases inherent in manual assessments. The integration of AI doesn’t replace the expertise of FMEA practitioners but provides a supportive tool for more effective risk identification and mitigation.

The automation of Failure Mode Effects Analysis (FMEA) through Artificial Intelligence (AI) centers on the application of Machine Learning (ML) and Semantic Reasoning techniques. ML algorithms can be trained on historical failure data to identify correlations and predict potential failure modes, accelerating the identification process. Semantic Reasoning allows the system to understand the relationships between components and functions, enabling the automated propagation of failure effects. This approach aims to minimize human error and inconsistencies inherent in manual FMEA processes, leading to more reliable risk assessments and improved product designs. The objective is not to replace human expertise entirely, but to augment it by handling repetitive tasks and flagging potentially critical areas for focused review, thereby increasing both efficiency and accuracy.

Automated Failure Mode Effects Analysis (FMEA) utilizes Artificial Intelligence algorithms to analyze historical data, engineering specifications, and operational parameters to detect correlations indicative of potential failure modes. These AI systems, employing techniques like machine learning, can identify subtle patterns and anomalies within complex datasets that may not be readily apparent through manual review. This capability extends beyond simple fault detection to predictive analysis, allowing for the anticipation of failures before they occur, based on the identified patterns and their correlation to specific operating conditions or component behaviors. The result is a more comprehensive and proactive approach to identifying risks, potentially uncovering failure modes overlooked by traditional, human-driven FMEA processes.

Intelligent Systems Engineering (ISE) builds upon traditional Failure Mode Effects Analysis (FMEA) by incorporating knowledge-based reasoning techniques. This integration allows the system to move beyond identifying potential failure modes based solely on historical data or expert opinion; instead, it leverages a codified body of engineering knowledge, including design principles, material properties, and physical laws, to infer potential failures and their effects. The system utilizes knowledge representation methods – such as ontologies and semantic networks – to reason about complex system interactions and predict failure propagation paths. This reasoning capability enables a more comprehensive analysis, uncovering potential failure modes that might be overlooked in conventional FMEA and providing deeper insight into the root causes and consequences of system failures.

This example Design FMEA table for a window lifter system illustrates a structured approach to identifying potential failure modes and their effects during the design phase.

From Root Cause to Algorithmic Accountability

Determining why failures occur is paramount to preventing them, and a robust root cause analysis, when integrated with AI-powered Failure Mode and Effects Analysis (FMEA), offers a powerful diagnostic approach. This synergy moves beyond simply identifying potential failure points; the AI algorithms can sift through vast datasets – including historical maintenance records, sensor data, and operational logs – to uncover the complex chains of events leading to a problem. By analyzing these interconnected factors, the system pinpoints not just the immediate cause, but the fundamental, underlying reasons for the failure. This detailed understanding enables proactive interventions, targeting the source of the issue rather than merely treating the symptoms, and ultimately bolstering system reliability and safety.

Modern risk assessment increasingly leverages artificial intelligence to navigate the sheer volume and intricacy of data inherent in complex systems. Traditional methods often struggle with datasets encompassing numerous variables and subtle interdependencies; however, AI algorithms excel at identifying patterns and correlations that might otherwise remain hidden. This capability allows for a more nuanced prioritization of risks, moving beyond simple probability-impact matrices to consider cascading failures and systemic vulnerabilities. By processing historical data, real-time sensor readings, and even unstructured information like maintenance logs, AI models can dynamically adjust risk scores and flag potential issues before they escalate. The result is a proactive approach to risk management, enabling organizations to allocate resources more efficiently and bolster resilience against unforeseen events, ultimately shifting from reactive problem-solving to preventative safeguarding.

The precision of risk prediction is notably enhanced through the integration of ontology-informed machine learning. This approach moves beyond traditional data analysis by incorporating explicitly defined relationships between components, failure modes, and effects-essentially, a structured knowledge base. By grounding the machine learning models in this semantic understanding, the system can identify subtle patterns and dependencies often missed by purely data-driven methods. This not only boosts the accuracy of predictions but also provides a level of interpretability previously unattainable; instead of simply flagging a risk, the system can articulate why a particular scenario is deemed problematic, referencing the underlying relationships defined within the ontology and fostering greater confidence in its assessments.

The effective implementation of artificial intelligence in risk mitigation hinges not simply on predictive accuracy, but on the capacity to elucidate why a particular recommendation is made. Explainable AI, therefore, becomes paramount; it moves beyond a ‘black box’ approach, offering transparency into the reasoning behind risk assessments and proposed solutions. This interpretability is essential for fostering trust amongst stakeholders, allowing them to validate the system’s logic and understand the factors driving its conclusions. Without such clarity, even highly accurate predictions risk being dismissed or misinterpreted, hindering informed decision-making and potentially leading to suboptimal outcomes. The ability to trace the pathway from data input to risk assessment empowers users to confidently integrate AI insights into their strategies, ensuring both effective risk management and responsible technological adoption.

The Illusion of Control: Towards Truly Resilient Systems

A Digital Twin functions as a virtual replica of a physical system, constantly updated with real-time data to mirror its operational state. This isn’t merely a static model; it’s a dynamic representation built upon the foundations of Model-Based Systems Engineering (MBSE). MBSE provides the rigorous framework for defining system components and their interactions, while the Digital Twin leverages this structure to ingest live sensor data, simulating the system’s behavior under various conditions. Consequently, it allows for continuous monitoring of performance, identification of potential anomalies, and predictive analysis of future states. This capability extends beyond simple diagnostics; it facilitates proactive interventions, enabling operators to anticipate failures, optimize performance, and ultimately enhance system resilience by testing ‘what-if’ scenarios without impacting the physical asset.

The effective management of increasingly complex systems relies heavily on the ability to capture and utilize vast amounts of knowledge, a challenge now being addressed through the synergy of Large Language Models (LLMs) and Knowledge Graphs. LLMs excel at processing and understanding natural language descriptions of system components, behaviors, and relationships, while Knowledge Graphs provide a structured, machine-readable representation of this information. By integrating these technologies, systems engineers can move beyond traditional documentation to create a dynamic, interconnected web of knowledge. This allows for automated reasoning about system performance, identification of potential failure modes, and even the generation of design alternatives. The combination facilitates a deeper understanding of interdependencies, enabling more informed decision-making and ultimately fostering the development of more robust and adaptable systems.

The convergence of artificial intelligence, Model-Based Systems Engineering (MBSE), and Digital Twins is forging a new paradigm for system development, one that moves beyond simple functionality towards genuine resilience and intelligence. By leveraging the predictive power of AI algorithms within the detailed, physics-informed models of MBSE and the real-time data streams of Digital Twins, systems can anticipate and adapt to unforeseen circumstances. This integrated approach isn’t merely about building smarter devices; it’s about creating systems capable of self-diagnosis, self-healing, and continuous optimization. The resulting architecture allows for proactive identification of potential failures, enabling preventative maintenance and minimizing downtime, ultimately leading to more robust and dependable systems across diverse applications – from complex infrastructure to autonomous vehicles and beyond.

A fundamental shift is occurring in systems engineering, moving away from addressing failures after they occur towards anticipating and preventing them during the design phase. This proactive approach leverages the power of artificial intelligence to analyze system behavior and identify potential vulnerabilities before implementation. While precise gains in scalability remain a subject of ongoing research, the integration of AI promises to handle significantly more complex systems than traditional methods allow. This is particularly impactful in tasks like Failure Mode and Effects Analysis (FMEA), where manual effort can be substantially reduced, freeing engineers to focus on innovation and optimization rather than exhaustive, reactive troubleshooting. The ultimate goal is not simply to react to problems, but to design systems inherently capable of withstanding unforeseen challenges and maintaining operational integrity.

The pursuit of intelligent FMEA, as outlined in the paper, inevitably courts the specter of future tech debt. The integration of AI and ontologies – promising a knowledge-driven, adaptive process – feels less like innovation and more like a temporary reprieve from complexity. As Carl Friedrich Gauss observed, “If I were to wish for anything, I should wish for more time.” This rings true; each layer of abstraction, each attempt to ‘solve’ system reliability with clever algorithms, simply pushes the inevitable emergence of unforeseen failure modes further down the line. The paper’s focus on knowledge representation is admirable, but one suspects that production environments will always discover new and inventive ways to optimize – and then re-optimize – even the most carefully constructed systems. Architecture isn’t a diagram; it’s a compromise that survived deployment, at least for a time.

What’s Next?

The promise, of course, is automated FMEA. A system that doesn’t require a seasoned engineer to painstakingly review every potential failure mode. It started with a simple bash script, really – a checklist with some conditional logic. Now it’s ontologies and machine learning. They’ll call it AI and raise funding. The immediate challenge isn’t building the ‘intelligence’, it’s curating the knowledge. Because garbage in, garbage out applies with particular force when you’re modeling system failures. A beautifully crafted ontology is useless if it’s populated with optimistic assumptions or, worse, data copied from marketing materials.

The more fundamental issue, predictably, is human resistance. Engineers, understandably, distrust black boxes telling them their designs are flawed. Expect a lot of ‘explainable AI’ theater, where systems painstakingly justify conclusions already reached by someone with twenty years of experience. The real win won’t be automating the process, but providing tools to augment human expertise, not replace it. The documentation lied again, naturally.

Ultimately, this field will be measured not by the elegance of the algorithms, but by the reduction in field failures. A slight reduction, most likely, and attributed to ‘improved processes’ rather than ‘revolutionary AI’. Tech debt is just emotional debt with commits, after all. And systems, as always, will find new and interesting ways to break.

Original article: https://arxiv.org/pdf/2511.17743.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/