Introduction
In the mid-1980s, automated diagnostic systems typically operated under a "single-fault assumption"—the premise that only one component in a machine is broken at any given time. While this simplifies diagnostic logic, it fails in real-world environments where multiple parts might fail due to age, stress, or a shared catastrophic event. The primary obstacle to diagnosing multiple faults is the "combinatorial explosion": a device with $n$ components has $2^n$ possible states of health. For a modest circuit with 32 components, this results in over 4 billion possibilities.
This paper introduces the General Diagnostic Engine (GDE), a framework designed to identify any number of simultaneous faults efficiently. By treating component functions as "assumptions" and focusing on the minimal sets of components that could cause an observed discrepancy, GDE avoids the exponential search space typically associated with multi-fault diagnosis.
The GDE Methodology
The GDE framework is built upon a model-based approach that relies on four key pillars:
- Assumptions and Conflicts: The system initially assumes every component is functioning correctly. It uses constraint propagation to predict expected outputs. When a measurement contradicts a prediction (a "symptom"), the set of components involved in that prediction is identified as a conflict.
- Minimality: To keep the search space manageable, GDE only tracks minimal conflicts (the smallest set of components that must contain at least one failure) and minimal candidates (the smallest set of failures that explains all observed symptoms).
- Separation of Concerns: GDE maintains a sharp distinction between the diagnostic logic and domain-specific behavior prediction. This modularity makes it a "general" engine applicable to various technical domains.
- Incremental Processing: As new evidence is gathered, GDE refines its list of candidates incrementally rather than restarting the diagnostic process.
Navigating the Search Space
A central innovation of GDE is the visualization of the search space as a subset-superset lattice. By identifying the "minimal" boundary of conflicts, the engine can implicitly represent billions of possibilities without exhaustive enumeration.
This conflict-directed search turns diagnosis into a data-driven process. One of the most significant insights is that GDE does not require "fault models"—descriptions of how parts behave when they break. It only requires a model of how components are supposed to work; any deviation from this expected behavior makes a component a candidate for blame.
Case Study: The "Familiar Circuit"
The authors evaluate GDE's logic using a benchmark known as the "Familiar Circuit," which consists of three multipliers (M1, M2, M3) and two adders (A1, A2).
In this example, GDE demonstrates how it can narrow down the 32 possible states to a handful of likely candidates using just a few measurements. For instance, if two different multipliers are broken simultaneously, GDE uses the observed outputs at F and G to isolate the minimal sets of components that could be responsible.
Key Results and Limitations
GDE successfully handles the $2^n$ search space by focusing on minimal candidates, demonstrating both efficiency and logical soundness. The framework ensures that while the inference engine might be "incomplete" (missing some subtle conflicts), it will never mistakenly eliminate a valid candidate.
However, the approach has limitations. Managing dependencies using an Assumption-based Truth Maintenance System (ATMS) can be computationally expensive for extremely large systems. Additionally, GDE primarily focuses on "snapshot" diagnosis; future work is required to handle time-varying signals and dynamic systems. The authors also point toward information-theoretic approaches—minimizing entropy—as the next step for optimizing the selection of subsequent measurements during troubleshooting.
Summary updated