A Design Based Approach to Constructing Computational Solutions to Diagnostic Problems

Dmitry Volovik, Imran A. Zualkernan, Paul E. Johnson, Charles E. MatthewsAAAI 1990

Introduction

In the high-stakes environment of digital hardware manufacturing, troubleshooting complex devices presents a significant challenge. Traditional diagnostic methods often rely on "first principles" simulations, which become computationally prohibitive as device complexity grows. Furthermore, rapid product lifecycles frequently render specialized expert systems obsolete before they are fully deployed. This paper proposes a novel diagnostic theory that bridges the gap between general troubleshooting strategies and the rich internal representations created by designers.

Methodology: Leveraging Design Knowledge

The proposed approach is built upon three conceptual pillars:

  1. Weak Methods: Device-independent strategies, such as path following or subdivision, that human experts use when specific diagnostic rules are unavailable.
  2. Design Knowledge: Abstract representations including functional specifications, block diagrams, and test requirements that are often discarded after the design phase.
  3. The Principle of Locality: The strategy of identifying a search space where a fault is "local" and thus more easily identifiable. A fault invisible in physical wiring may be glaringly obvious in a functional specification.

The core of the system is a hierarchical mapping that links observed symptoms (test failures) back to functional specifications, which are then mapped to physical modules. This allows the system to automate the "hunch" of a human expert by identifying shared high-level functional units likely to contain the fault based on which specific tests are failing.

Hierarchical mapping between modules, signals, tests, and functional specifications.

Experimental Setup

To validate this theory, the researchers utilized a custom benchmark: an Interrupt Control Coprocessor consisting of 254 components and approximately 10,000 signals. Twenty-seven simulated faults were introduced into a digital model of the board, including:

  • Stuck-at faults: Inputs or outputs held at a constant high/low value.
  • Gate failures: Total component failure (e.g., an inverter acting as a straight wire).
  • Bridge faults: Short circuits between signals.

Key Results and Analysis

The "Focus of Attention" method demonstrated significant effectiveness, correctly identifying the faulty functional area in 77.8% of the cases. The system was particularly adept at identifying bridge faults (6 out of 7) and gate failures (8 out of 10).

Beyond accuracy, the "Power" of the method—defined by the reduction in the search space—was a key metric. On average, the method narrowed the search area down to 27.9% of the total components. In the most successful cases, the search space was reduced to just 8.8%, drastically lowering the manual effort required for physical fault isolation.

Search space remaining after applying the diagnostic method across various scenarios.

Conclusion and Future Work

This research highlights that the "missing link" in automated diagnostics is not necessarily more processing power, but the utilization of design abstractions. By shifting the search space from physical wiring to functional specifications, complex problems become more manageable. While this study focused on single-fault scenarios in moderately complex designs, future work will focus on integrating additional "weak methods" to handle multiple faults and the multi-billion transistor scales of modern hardware.

Summary updated

BibTeX
@inproceedings{volovik1990design,
  author = {Dmitry Volovik and Imran A. Zualkernan and Paul E. Johnson and Charles E. Matthews},
  title = {{A Design Based Approach to Constructing Computational Solutions to Diagnostic Problems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1990},
}
Figure