Testing Physical Systems

Peter StrussAAAI 1994

Abstract

This paper presents a formal framework for model-based testing of physical systems characterized by continuous variables. By employing model abstraction and relational behavior models, the proposed approach generates finite, deterministic test sets capable of confirming correct system operation through the systematic refutation of all known fault modes. The methodology is validated using a thyristor case study, demonstrating the transition from infinite continuous domains to manageable qualitative abstractions.

Introduction: The Challenge of Continuous Domains

In digital systems, testing is relatively straightforward due to the discrete, Boolean nature of variables. However, physical systems—such as semiconductor switches—operate in continuous domains where variables like voltage ($\Delta v$) and current ($i$) can take on an infinite range of values. This presents two primary challenges:

  1. Infinite Search Space: Establishing confidence in a component's correctness based on a finite set of observations is difficult when the potential behavior space is infinite.
  2. Confirmation vs. Refutation: While diagnosis focuses on identifying faults, manufacturing testing aims to confirm correctness. True confirmation requires proving that no defined fault mode could have produced the observed behavior.

Methodology: Relational Models and Abstraction

The proposed framework extends model-based diagnosis into test generation by defining behaviors as Relational Behavior Models. These models represent behaviors as sets of possible value tuples (relations) between system variables.

Model Abstraction

To bridge the gap between continuous physical reality and computational feasibility, the author introduces "Representational Transformations." This process maps continuous domains into qualitative, finite abstractions (e.g., mapping voltage ranges to discrete states). The paper formally proves that a test confirming behavior in an abstract model remains valid when mapped back to the original continuous system, justifying the use of qualitative reasoning for quantitative problems.

Algorithmic Framework

The core of the framework is an algorithm designed to generate a "hitting set" of observations. A single well-designed test can refute multiple fault modes simultaneously. The algorithm identifies deterministic test inputs that force the system into states where correct behavior is observably distinct from faulty behavior.

The proposed algorithm for generating a set of deterministic test inputs (TI-SET).

Case Study: The Thyristor

The methodology was validated using a thyristor (a semiconductor switch) as a qualitative case study. The system's behavior is modeled through various relations representing both correct operation and known fault modes, such as "blocked," "punctured," or "reduced breakover voltage."

Qualitative behavior models for a thyristor, showing correct (a) and various faulty (b, c, d) states.

In this case study, despite an infinite number of potential "reduced breakover voltage" faults, the algorithm successfully generated a finite set of just two tests to confirm correct operation.

Key Insights and Results

  • Exhaustive Refutation: Confirmation of correctness is achieved only when tests are diverse enough to rule out every defined fault mode.
  • Efficiency through Hitting Sets: The framework allows for the generation of a minimal set of "confirming tests," optimizing the testing process.
  • Formal Rigor: The paper provides a mathematical foundation for testing continuous systems, moving beyond traditional ad-hoc engineering methods.

Limitations and Future Directions

While the framework is theoretically robust, its practical application depends on high-quality models of both correct and faulty behaviors, which can be difficult to construct for complex dynamic systems. Current research is exploring the use of Binary Decision Diagrams (BDDs) for more compact model representation and investigating "Design for Testability" to optimize sensor placement during the design phase.

Summary updated

BibTeX
@inproceedings{struss1994testing,
  author = {Peter Struss},
  title = {{Testing Physical Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure