Troubleshooting: When Modeling Is the Trouble

Philippe Dague, Olivier Raiman, Philippe DevesAAAI 1987

TL;DR

This paper introduces DEDALE, an expert system that tackles the "modeling bottleneck" in analog circuit troubleshooting. By moving beyond simple signs (+/-) to a formal system of order-of-magnitude reasoning, the authors demonstrate how to identify faulty components in complex circuits even when precise numerical models are unavailable or too complex to use.

The Analog Troubleshooting Challenge

Troubleshooting analog circuits is notoriously difficult for traditional model-based diagnosis. Unlike digital systems where components have discrete, well-defined states, analog systems are highly interdependent. A perfectly healthy component might behave radically differently (e.g., a transistor switching from an amplifier to an open circuit) simply because a neighboring component is failing.

The authors identify three main hurdles:

  1. Multiple Correct Behaviors: A component's state depends on its environment, making prediction complex.
  2. Lack of Numerical Models: Non-linearities and time-dependencies make numerical simulation computationally expensive or impossible due to missing parameters.
  3. Lack of Measurements: While voltage is easily measured, measuring current—the lifeblood of circuit analysis—is often impossible without breaking the circuit.

Standard qualitative reasoning, which only looks at whether a value is positive, negative, or zero, is too "shallow" to catch these subtle shifts. As the authors put it, "Modeling becomes the trouble."

Methods: Qualitative Modeling with FOG

The core innovation is the application of Order of Magnitude Reasoning through a formal system called FOG. Instead of exact numbers, DEDALE reasons using three key operators:

  • $A \ll B$: $A$ is negligible in comparison with $B$.
  • $A \approx B$: $A$ is close to $B$ (their difference is negligible).
  • $A \sim B$: $A$ has the same order of magnitude as $B$.
Formal FOG notation for transistor current and voltage relations in a specific state.

DEDALE uses these relations to build a Library of Qualitative Models for generic components like transistors and resistors. It then employs a "Consistency Checker" to find contradictions between observed behavior and the presumed "correct" models of components.

Qualitative models for various transistor states (no, open, on) defined by order-of-magnitude relations.

The system follows a three-pronged strategy:

  • Top-Down: Focuses on functional blocks where "deviations" (significant changes from nominal behavior) are observed.
  • Horizontal: If a block looks faulty, it checks linked blocks to see if the error is being propagated from elsewhere.
  • Bottom-Up: Ensures that local consistency at the component level actually scales up to a consistent global state.

Key Insights and Results

The system was evaluated on real-world analog hybrid circuits in a factory environment at Electronique Serge Dassault. The specific example detailed in the paper is a "voltage follower" circuit consisting of transistors and resistors.

  • Deviation as a Heuristic: Rather than trying to predict exactly what a circuit should do, DEDALE looks for where it deviates from its designed (nominal) behavior. This "deviation $\rightarrow$ focusing" rule acts as a powerful search heuristic.
  • Implicit Fault Discovery: Unlike "shallow" expert systems that need a database of known faults, DEDALE can discover new types of faults (like a specific short-circuit) simply by finding which qualitative model is no longer consistent with the physics of the circuit.
  • Success Rate: In factory experiments on complex circuits (around 100 components), DEDALE successfully identified defects in approximately 75% of investigated failures.
  • Diagnostic Depth: The system proved capable of identifying not just which component failed, but the nature of the failure (e.g., identifying a base-emitter short-circuit) without having a pre-defined model for that specific fault.

Limitations and Future Work

The remaining 25% of failures were not due to faulty components but to components working at the very limits of their design. In these cases, there are no "significant" order-of-magnitude deviations, so DEDALE’s core assumption fails. The authors suggest specific heuristics are needed for these "limit-of-design" cases.

Additionally, the current model assumes a steady electric state and reproducible observations. It does not account for intermittent failures, which remain a major challenge. As circuit density increases, even more complex hierarchical strategies may be required to maintain efficiency.

Summary updated

BibTeX
@inproceedings{dague1987troubleshooting,
  author = {Philippe Dague and Olivier Raiman and Philippe Deves},
  title = {{Troubleshooting: When Modeling Is the Trouble}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure