Hybrid Case-Based Reasoning for the Diagnosis of Complex Devices

Michel P. Féret, Janice I. GlasgowAAAI 1993

Abstract

This paper introduces a hybrid diagnostic framework that integrates Model-Based Reasoning (MBR) with Case-Based Reasoning (CBR) to diagnose complex mechanical devices. By utilizing the structural output of an MBR model as an index for a library of past experiences, the system maintains diagnostic accuracy even when the underlying device model is incomplete or erroneous. Experimental results on robotic and nuclear ventilation systems demonstrate that the hybrid approach provides graceful degradation and significantly reduces failure rates as the casebase grows, even when model specifications are flawed.

Introduction

Diagnosing complex mechanical systems, such as robotic fairing servicing subsystems or nuclear ventilation units, presents a significant challenge. Traditional Model-Based Reasoning (MBR) relies on high-fidelity "first-principles" models. However, creating perfect models is often unfeasible due to human design errors, overlooked rare fault patterns, or incomplete specifications. Conversely, standalone Case-Based Reasoning (CBR) systems struggle with efficient indexing in large databases. This research proposes a hybrid methodology where the model provides the logical search structure, and the casebase fills the gaps left by model imperfections.

Methodology

The proposed system, the Automated Data Management System (ADMS), operates in two primary phases:

  1. Structural Isolation (MBR Phase): The device is modeled as a hierarchy of components. Real-time sensor data is used to traverse this hierarchy and prune branches representing correctly functioning parts, resulting in a set of "potential diagnoses."
  2. Case-Based Retrieval (CBR Phase): The results from the MBR phase serve as an index. Cases—both successful (where the model was correct) and failure (where the model was incorrect but corrected by an operator)—are stored under the components identified by the model.
Figure 1: The hybrid architecture showing how structural decomposition of the device model is used to index and retrieve cases from the casebase.

A similarity measure then compares current sensor values and the structural path taken to identify the most relevant past experience. This allows the system to "overrule" a flawed model if historical data suggests a component the model's logic initially dismissed (e.g., a failure case pointing to a different component).

Experimental Results

The framework was evaluated using simulators for a robotic Fairing Servicing Subsystem and a Reactor Building Ventilation System. To test robustness, the authors intentionally introduced errors into the model's sensor functions, simulating real-world modeling inaccuracies.

Key findings include:

  • Graceful Degradation: Unlike pure MBR systems that fail catastrophically when the model is wrong, the hybrid system's performance degrades linearly with the number of model errors.
  • Impact of Casebase Size: As the casebase accumulates more experiences, the overall failure rate of the diagnostic system decreases significantly. This holds true even for highly degraded models (represented by the different numbered lines in the figure below).
Figure 5: Reduction in diagnostic failure rate as the casebase size increases, shown for various levels of model degradation.
  • Efficiency: The indexing schema ensures that the number of potential diagnoses remains manageable even as the casebase grows, with minimal computational overhead because it leverages the work already performed by the structural isolation process.

Conclusion

The hybrid MBR-CBR approach effectively addresses the "knowledge acquisition bottleneck" of pure model-based systems. By treating the model as a heuristic index for experience, the system achieves a synergy where logic guides the search and history corrects logical gaps. Future work will address sensor reliability and cross-component generalization to further enhance the power of the casebase in multi-fault scenarios.

Summary updated

BibTeX
@inproceedings{feret1993hybrid,
  author = {Michel P. Féret and Janice I. Glasgow},
  title = {{Hybrid Case-Based Reasoning for the Diagnosis of Complex Devices}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure