A Framework for Decentralized Qualitative Model-Based Diagnosis

Luca Console, Claudia Picardi, Daniele Theseider DupréIJCAI 2007

Introduction

Modern industrial systems, such as aircraft or complex cloud-based software services, are increasingly designed as "systems of systems." These architectures are composed of multiple subsystems, often developed by different suppliers, which interact to achieve a global goal. While this modularity aids design and manufacturing, it creates significant challenges for traditional model-based diagnosis (MBD). Traditional MBD typically relies on a monolithic global model, which is often impractical due to proprietary privacy concerns, computational complexity, and the dynamic nature of modern system configurations.

This paper introduces a supervised, decentralized framework for diagnosing faults in such complex environments. By utilizing a "Supervisor" to coordinate "Local Diagnosers" (LDs) through a standardized interface, the approach maintains subsystem privacy and handles dynamic reconfigurations while avoiding the computational explosion associated with centralized models.

The Challenges of Decentralized Diagnosis

The authors identify three primary hurdles that traditional diagnostic methods fail to address effectively:

  1. Privacy: Suppliers often treat their internal diagnostic models as proprietary intellectual property. They are willing to provide a "black box" diagnostic service but refuse to share the underlying model with a central authority.
  2. Complexity: As systems grow, creating and maintaining a single, massive global model becomes computationally overwhelming and prone to errors.
  3. Dynamism: Modern systems are not static. Components may be replaced, or interaction paths (especially in software-defined services) may shift at runtime, rendering static global models obsolete.

Proposed Supervised Decentralized Architecture

The core of the proposed framework is a hierarchical structure consisting of Local Diagnosers and a high-level Supervisor.

  • Local Diagnosers (LDs): Each subsystem is equipped with its own LD. These LDs possess the internal qualitative model of their respective subsystem. Crucially, these models remain private and are never shared.
  • Supervisor: The Supervisor acts as a high-level coordinator. It does not possess knowledge of how individual subsystems function internally; instead, it only understands how they are connected at any given time.

The EXTEND Interface

The primary innovation of this framework is the EXTEND interface. Rather than sharing models, the Supervisor and LDs communicate through this single operation. The Supervisor passes a partial hypothesis (e.g., "The input voltage is low") to an LD. The LD then checks if this hypothesis is consistent with its local observations. If consistent, it returns minimal admissible extensions—the smallest set of additional facts (such as "the pump is stuck") required to explain the observed behavior.

Qualitative Modeling

To handle the inherent uncertainty and complexity of physical systems, the framework employs "deviation models." Instead of solving complex differential equations, it uses sign abstractions (-, 0, +) to represent whether a value is lower than, equal to, or higher than its expected nominal value.

Case Study: Aircraft Landing Gear

The authors validate their framework using a running example of an aircraft landing gear system. This system (Figure 1) integrates various subsystems, including DC power sources, an engine-driven power transmission, and the Hydraulic Extraction Subsystem (HES).

Figure 1: High-level architecture of the aircraft landing gear system example.

In a typical diagnostic scenario, a fault such as the landing gear failing to extract is detected. The Supervisor coordinates with the LDs of the various subsystems to narrow down the cause. For instance, the HES (Figure 2) contains internal components like pumps, pipes, and valves. The LD for the HES can determine if an internal failure, like a stuck valve, explains the lack of extraction pressure without revealing the internal logic of the valve to the Supervisor.

Figure 2: Internal structure of the Hydraulic Extraction Subsystem (HES).

Key Results and Insights

  • Soundness and Completeness: The authors provide theoretical proof that their decentralized algorithm is guaranteed to produce the same set of diagnoses as a traditional centralized system, ensuring no loss in diagnostic accuracy.
  • Complexity Management: By partitioning the system into smaller, manageable subsystems, the local cost of diagnosis remains bounded. While finding minimal extensions is computationally intensive (NP^NP), the authors propose a "RELAXED-EXTEND" variant for scenarios where speed is prioritized over absolute minimality.
  • Statelessness and Runtime Discovery: LDs are designed to be stateless, with the Supervisor managing the overall diagnostic session. This allows for "plug-and-play" components where the Supervisor can discover new connections at runtime and immediately begin diagnosing the reconfigured system.

Limitations and Future Work

Currently, the framework is "atemporal," meaning it analyzes snapshots of system state rather than temporal sequences or feedback loops. Future research aims to extend the EXTEND interface to handle temporal models. Additionally, the authors acknowledge that while decentralized computation is efficient, the increased communication overhead between the Supervisor and LDs could become a bottleneck in high-latency environments.

Summary updated

BibTeX
@inproceedings{console2007framework,
  author = {Luca Console and Claudia Picardi and Daniele Theseider Dupré},
  title = {{A Framework for Decentralized Qualitative Model-Based Diagnosis}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure