Abstract
Traditional diagnostic systems for complex mechanical systems often rely on rigid, pre-compiled rules that fail when encountering novel faults or dynamic symptom propagation. This paper presents Draphys (Diagnostic Reasoning About Physical Systems), a framework that treats diagnosis as a search problem within a hierarchical hypothesis space. By employing status abstraction and dual physical representations—functional and proximity-based—Draphys achieves robust performance in identifying complex aircraft engine failures, including those involving physical damage propagation.
Introduction
In the late 1980s, diagnostic systems for high-stakes environments like aviation were often "fragile." These systems typically utilized a flat list of "if-then" rules linking specific symptoms to known faults. While effective for routine maintenance, such systems struggled during "operative diagnosis"—the real-time identification of system status while a vehicle is in flight.
The primary challenges identified include the inability to reason about novel faults, difficulty handling dynamic symptoms as failures spread, and a reliance on single-model representations that ignore physical proximity. For instance, a mechanical break in a fan blade might damage an adjacent hydraulic line, a "bridge fault" that purely functional models cannot predict.
The Draphys Framework
Draphys addresses these limitations by formulating diagnosis as a qualitative simulation and search process. Instead of simple lookup tables, the system generates and tests hypotheses by simulating how faults propagate through the engine.
Status Abstraction and Graceful Degradation
A core innovation of Draphys is its use of hierarchical abstraction. Sensor data is abstracted from specific quantitative values (e.g., "low pressure") to qualitative states (e.g., "abnormal"). Similarly, faults are organized hierarchically. If the system cannot pinpoint a specific cause like an "eroded compressor blade," it can still provide a useful, higher-level diagnosis such as "compressor failure." This allows the system to degrade gracefully rather than failing entirely when faced with uncertainty.
Multiple Physical Representations
To detect "bridge faults," Draphys maintains two distinct models:
- Functional Model: Maps dependencies required for operation (e.g., the compressor feeds the combustor).
- Physical Proximity Model: Maps spatial relationships (e.g., which components are physically adjacent).
By combining these, Draphys can reason about how a mechanical failure in one subsystem might physically impact another, seemingly unrelated subsystem.
Evaluation and Results
The system was evaluated using eight real-world civil transport aircraft accidents reconstructed from NTSB reports, including high-profile cases like United Airlines Flight 232. Four cases were used for development, and four were held back for objective testing.
Draphys successfully diagnosed seven out of the eight cases. Notably, it identified five instances of physical damage that traditional functional models would have missed. Furthermore, by monitoring for "abnormal" trends rather than just "out-of-bounds" thresholds, Draphys often detected symptoms earlier than standard cockpit warning systems.
Conclusion
Draphys demonstrates that treating diagnosis as a dynamic search in an abstracted hypothesis space significantly improves robustness. While the current model does not account for external environmental factors—leading to its failure in a case involving massive water ingestion—the framework provides a powerful template for future operative diagnostic systems. The shift from "how it broke" to "where the failure is" proves critical for ensuring safety in real-time operative environments.
Summary updated