Experience-Aided Diagnosis for Complex Devices

Michel P. Féret, Janice I. GlasgowAAAI 1994

Abstract

Traditional Model-Based Diagnosis (MBD) often struggles with computational complexity and the inherent imperfections of real-world device models. This paper presents the Experience-Aided Diagnosis (EAD) model, a hybrid framework that integrates the logical rigor of MBD with the practical memory of Case-Based Reasoning (CBR). By combining deduction, abduction, and induction, the EAD model provides a robust diagnostic mechanism that remains effective even when underlying models are incomplete or unfaithful to the physical system.

Introduction

Diagnostic systems for complex engineering environments—such as space systems or nuclear reactors—face two primary challenges. First, the computational complexity of abductive reasoning, required to find the best explanation for symptoms, is NP-hard. Second, the "deep" knowledge models used in MBD are rarely perfect; they often fail to account for unexpected component interactions or contain modeling errors.

The Experience-Aided Diagnosis (EAD) model addresses these hurdles by bridging the gap between pure logic and human-like experience. It orchestrates a multi-modal reasoning process that uses "rule-out" knowledge to prune search spaces and historical cases to bypass logical gaps in the device model.

Methodology: The EAD Framework

The EAD model operates through three distinct yet integrated reasoning phases:

  1. Deduction (The Pruner): This phase utilizes "rule-out" knowledge (rok) to immediately discard impossible hypotheses. By identifying components that are functioning correctly based on sensor data, the system can exponentially reduce the search space ($2^n$) before more intensive reasoning begins.
  2. Abduction (The Logician): This is the standard MBD process. It generates hypotheses by inferring potential causes from observed effects based on the structural and functional model of the device.
  3. Induction/CBR (The Experienced Technician): This component leverages a case base of past failures. If current symptoms align with a historical case, the system suggests a diagnosis through "Abductive Induction." This allows the system to remain functional even when the logical model cannot fully prove a cause-effect relationship.
A diagnostic model for an internal combustion engine, illustrating the causal links between component failures and observable symptoms across different sessions.

As shown in the diagnostic graph above, the system maps observations (OBS) to a causal network. In Session 1 and 2, different sets of symptoms (e.g., vapour, acceleration irregularities) are used to derive positive ($\psi^+$) and negative ($\psi^-$) diagnostic sets, demonstrating how the model handles varying observational data.

Theoretical Framework and Error Taxonomy

A core contribution of this work is the formalization of model quality. The authors distinguish between models that are faithful (all predictions are true) and complete (all truths are predicted). When these conditions are not met, diagnostic errors occur.

The paper provides a rigorous taxonomy of these errors, categorized as Types A, B, and C. These errors are mapped to specific failures in model faithfulness, completeness, and the coverage of the rule-out knowledge.

Taxonomy of diagnostic errors based on model completeness, faithfulness, and the coverage of rule-out knowledge (rok).

This classification allows for a unified theoretical framework that subsumes previous seminal diagnostic works, such as those by Reiter and de Kleer, while providing a more generalized approach for practical application.

Practical Evaluation and Insights

The EAD model was evaluated on two large-scale systems: a robotic Fairing Servicing Subsystem for the Canadian Space Agency and a Reactor Building Ventilation System. The results indicated that the hybrid approach successfully solved "hard" diagnostic problems that typically stymie pure MBD systems.

Key insights from the study include:

  • Efficiency via Pruning: Applying deductive pruning before abductive generation is a critical efficiency gain for real-time systems.
  • CBR as a Safety Net: By treating past cases as a form of induction, the system can override flawed logic. If history shows a component causes a symptom that the current model deems impossible, the CBR component allows for a correct diagnosis.
  • Implicit Generalization: Unlike standard induction that creates explicit rules, CBR generalizes implicitly through similarity, avoiding the computational cost and bias associated with generating new rule sets.

Conclusion and Future Work

The EAD model demonstrates that integrating experience-based reasoning with logical models significantly enhances diagnostic robustness and efficiency. However, challenges remain in defining universal similarity measures for CBR and ensuring the quality of the case base. Future work will focus on more general methods for merging abductive and inductive diagnoses in a device-independent manner.

Summary updated

BibTeX
@inproceedings{feret1994experience,
  author = {Michel P. Féret and Janice I. Glasgow},
  title = {{Experience-Aided Diagnosis for Complex Devices}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure