Abstract
Modern industrial systems, such as offshore wind turbines and spacecraft, require sophisticated diagnostic tools to minimize downtime and maintenance costs. While Model-Based Diagnosis (MBD) provides a rigorous framework for fault localization, its adoption is hindered by high modeling costs and computational complexity. This paper presents a methodology to automate the creation of diagnostic models by mapping existing Failure Mode and Effect Analysis (FMEA) data to propositional Horn logic. This approach enables efficient, polynomial-time fault localization, making MBD feasible for complex industrial applications.
Introduction
As technical systems grow in complexity, identifying the root cause of failures becomes increasingly difficult and expensive. In high-stakes environments like offshore wind farms, a single incorrect diagnosis can lead to astronomical costs due to the specialized logistics—such as ships, cranes, and specialized spare parts—required for repairs.
Model-Based Diagnosis (MBD) offers a formal solution by reasoning from observed symptoms back to possible causes. However, MBD faces two primary hurdles in industrial settings:
- The Modeling Gap: Creating the formal logic models required for MBD is labor-intensive and requires specialized expertise that most maintenance engineers do not possess.
- Computational Complexity: Finding a diagnosis is generally an NP-hard problem, making it difficult to scale for large-scale industrial systems.
This research aims to lower these barriers by repurposing the "safety paperwork" companies already produce—specifically FMEAs—into active, executable diagnostic tools.
Methodology
The proposed framework focuses on Abductive Model-Based Diagnosis, which reasons from observed effects back to possible causes. The core innovation lies in three pillars:
1. Automated Modeling via FMEA Mapping
Instead of building models from scratch, the authors utilize a mapping function ($\mathfrak{M} : 2^{FMEA} \mapsto HC$) to convert standard industrial Failure Mode and Effect Analysis (FMEA) entries into propositional Horn clauses. Since FMEAs already document the relationship between component faults and their implications, this process extracts existing engineering knowledge without requiring a manual modeling phase.
2. Logic Restriction for Efficiency
By restricting the system description to Horn logic—a subset of logic where clauses contain at most one positive literal—the diagnostic problem is transformed from an NP-hard search into one that can be solved in polynomial time. This provides a balance between expressiveness (capturing fault-effect chains) and computational speed.
3. Algorithmic Implementation
The system employs an Assumption-based Truth Maintenance System (ATMS) to compute explanations. Additionally, the authors utilize SAT solvers to recast the abduction problem as a search for conflicting hypotheses, further optimizing performance for real-time applications.
Empirical Evaluation
The methodology was validated using a combination of artificial examples, public FMEA datasets, and real-world industrial data provided by Uptime Engineering GmbH. The results demonstrate significant practical utility:
- Computational Speed: In tests on both artificial and real-world models, the median runtime was below 10 milliseconds.
- Scalability: Even in worst-case scenarios, computation times did not exceed 5 seconds, making the system viable for near-real-time decision support.
- Feasibility: The study demonstrates that abductive models derived from FMEAs are computationally manageable, directly addressing the complexity criticisms often leveled at MBD.
Key Insights
The most significant contribution of this work is the realization that industry does not necessarily need new models; it needs a way to make its existing reliability documentation machine-readable. For domains like offshore wind, the economic value of even slight improvements in diagnostic accuracy is massive. By using "good enough" logic (Horn clauses), the authors provide a framework that is simple enough for millisecond-level computation while remaining expressive enough for industrial diagnosis.
Limitations and Future Work
While promising, the approach is dependent on the quality of the underlying FMEA; a model is only as accurate as the documentation it is based on. Current efforts are focused on empirically verifying these models against real-world turbine behavior.
Furthermore, because diagnosis often yields multiple possible solutions, the authors are developing "discrimination procedures" using probability theory to narrow down root causes. Future research also includes integrating consistency-based diagnosis to identify failures that may not have been explicitly documented in the original FMEA.
Summary updated