Device Representation and Reasoning with Affective Relations

James M. Crawford, Daniel Dvorak, Diane J. Litman, Anil Mishra, Peter F. Patel-Schneider

Introduction

In the mid-1990s, the challenge of maintaining and diagnosing massive industrial systems like telephone switches reached a critical point. Traditional diagnostic methods were split between two extremes: heuristic-based expert systems and deep model-based reasoning (MBD). While heuristic systems were easy to implement initially, they often devolved into unmanageable "spaghetti code" as system complexity grew. Conversely, MBD systems, which reason from "first principles" by simulating internal behaviors, proved computationally prohibitive for systems as vast as the AT&T 4ESS switch.

This paper proposes a "middle ground" through the introduction of affective relations. This framework allows for high-level abstraction that models how components influence one another diagnostically, bypassing the need for exhaustive behavioral simulation.

The Affective Relations Framework

The core of the proposed approach is an ontology that defines how devices interact at a diagnostic level rather than a physical or functional level. By focusing on "affective" properties—how the state of one component affects the perceived state of another—engineers can build generalizable rules that apply across various hardware types.

The ontology of classes and affective relations used to model device interactions.

The ontology includes several key relations:

  • Structural Hierarchy: Defined by part-of and subpart relations.
  • Functional Dependency: The depends-on relation indicates that if a primary component fails, its dependent components are likely to show faults.
  • Redundancy: The standby relation models components that can take over for one another.
  • Diagnostic Signaling: self-alarm links a device to its internal error signals.

By using these relations, diagnostic principles can be written as general axioms. For instance, a rule might state that if a device's standby unit is non-functional, any alarm from the primary unit should be escalated immediately.

Implementation: The AT&T 4ESS Switch

The researchers applied this framework to the AT&T 4ESS long-distance telephone switch. Given the switch's immense scale, a static model was impractical. Instead, the system utilized Dynamic Model Construction. When an alarm is triggered, the system instantiates only the relevant portion of the device hierarchy to perform reasoning.

A schematic representation of the 4ESS switch model, showing dependencies and standby relations between controllers and clocks.

The system was implemented using R++, a C++ extension that integrates forward-chaining rules directly with object-oriented data structures. This allowed the diagnostic engine to operate efficiently on real-time data from the switch.

Key Insights and Results

The shift from simulation to abstraction yielded several significant advantages:

  1. Leveraging Self-Checking Hardware: Modern industrial hardware is often designed with internal monitoring. The AI's role is not to deduce the fault from scratch but to correlate existing "self-alarms" across the system's structural and functional hierarchy.
  2. Isomorphism: The AI model's structure mirrors the physical device, making it intuitive for human experts to verify.
  3. Rule Efficiency: General principles replaced dozens of specific heuristics. The researchers successfully reconstructed 10% of a legacy expert system's functionality with a much smaller, more maintainable rule set.
  4. Knowledge Acquisition: Technicians found it easier to describe systems in terms of affective relations (e.g., "this depends on that") than in terms of deep behavioral equations.

Conclusion and Future Work

While the prototype demonstrated the viability of affective relations for large-scale diagnosis, challenges remain in scaling the model to the entire 4ESS system. Future work includes refining how symptoms propagate through the hierarchy and developing a universal ontology of affective relations that could be applied to other domains, such as power plants or complex manufacturing lines. The approach represents a significant step toward diagnostic systems that are both principled and practical for real-world industrial applications.

Summary updated

BibTeX
@inproceedings{crawford1995device,
  author = {James M. Crawford and Daniel Dvorak and Diane J. Litman and Anil Mishra and Peter F. Patel-Schneider},
  title = {{Device Representation and Reasoning with Affective Relations}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
}
Figure