Introduction
In model-based reasoning, engineers use digital representations of physical systems to predict behavior or diagnose faults. To make these systems scalable, they often utilize libraries of generic "model fragments." However, these fragments frequently present a "Goldilocks" problem: they are either too detailed, making reasoning computationally expensive, or too coarse, missing distinctions necessary for specific tasks. This paper addresses the challenge of Automated Qualitative Domain Abstraction, proposing a method to automatically simplify high-fidelity models into the coarsest possible sets that still allow for accurate task resolution.
Methodology: Task-Interchangeability and AQUA
The core of the proposed approach is a generalization of the concept of interchangeability. In standard constraint satisfaction, values are interchangeable if swapping them preserves the global solution set. This research introduces task-interchangeability: two values are interchangeable if swapping them does not change the target results, given the observable inputs.
The authors implemented this in a prototypic system called AQUA (Automated Qualitative Abstraction). AQUA employs structural decomposition to transform complex constraint networks into acyclic versions, significantly increasing the efficiency of calculating these abstractions for large-scale systems.
Case Study: Pedal Position Sensor
The theory is evaluated using a pedal position sensor model, a standard automotive component. This system includes a potentiometer and an idle switch that provide data to an Electronic Control Unit (ECU).
By applying the AQUA system, the continuous voltage domains of the sensors are "crunched" into discrete qualitative intervals. This transformation allows the diagnostic system to focus only on transitions between meaningful states, such as moving from an idle to a pressed position.
Key Insights: Induced Abstractions
A central finding is that abstraction must be task-dependent; a distinction vital for design may be irrelevant for diagnosis. The authors define an abstraction as "adequate" if it never yields a different answer than the original model, and "simplest" if merging any further values would introduce errors.
As illustrated in the induced abstraction for the pedal sensor, the domains of $v_{pot}$ and $v_{switch}$ are partitioned into discrete regions (1-4). This "induced" abstraction allows for the pre-compilation of simplified models based on sensor granularity, enabling efficient real-time reasoning.
Results and Conclusion
The AQUA system was tested on a real-world turbo control subsystem involving hundreds of variables. The qualitative abstraction was instrumental in meeting real-time requirements, allowing the diagnostic system to handle high-rate sensor data by ignoring qualitatively identical fluctuations.
While the theory applies to both finite and infinite domains, efficiently deriving these abstractions for continuous, real-valued models remains an area for future research. Other future directions include iterative refinement—starting with coarse models and adding detail only when necessary—and exploiting mathematical properties like monotonicity to find domain boundaries more rapidly.
Summary updated