Qualitative Reasoning with Modelica Models

Matthew Evans Klenk, Johan de Kleer, Daniel G. Bobrow, Bill JanssenAAAI 2014 · DOI

Introduction

In early-stage engineering, designers face the challenge of exploring a vast "design space" to understand potential system behaviors. While industry-standard tools like Modelica provide robust quantitative simulations, they require precise numerical parameters—values that are often unknown or variable during the initial design phase. Consequently, identifying rare failure modes or edge cases becomes a matter of trial and error.

This research bridges the gap between quantitative engineering simulation and Qualitative Reasoning (QR). By establishing a sound mapping from Modelica’s Hybrid Differential Algebraic Equations (DAEs) to qualitative abstractions, the authors enable the automatic exploration of all possible system behaviors—including failures—without requiring exact numerical inputs.

The conceptual framework for mapping quantitative Modelica models to qualitative envisionments.

Methodology: From Hybrid DAEs to Qualitative Constraints

The core contribution of this work is a framework that translates Modelica models into qualitative constraints. Unlike traditional QR, which often struggles with the complexity of hybrid systems (combining continuous physics with discrete events), this approach introduces several key innovations:

  • Declarative Event Handling: The algorithm automatically computes the consequences of discrete "events" (e.g., a switch opening or a brake locking) directly from the Modelica code, removing the need for manual state-change mapping.
  • Non-Discontinuous Variables (NDV): To maintain physical realism during instantaneous transitions, the system identifies variables that must remain constant across an event, preventing the generation of physically impossible "jumps."
  • Constraint Inference for Pruning: To combat "state explosion"—the generation of thousands of impossible (spurious) behaviors—the authors identify three classes of additional constraints to prune the results:
  • Continuity and Compatibility: Redundant physical laws (like Kirchhoff’s laws) are added to filter out inconsistent states.
  • Higher-order Derivatives: Constraints ensure that derivatives (e.g., velocity and acceleration) remain consistent with their integrals (e.g., position).
  • Partially Ordered Landmarks: The system automatically infers relationships between different thresholds, such as knowing that a maximum friction limit is inherently greater than sliding friction.

Evaluation and Key Results

The framework was evaluated using three representative engineering models: a car brake system with complex "stuck/sliding" transitions, an RC-ladder circuit, and an oscillatory Adby circuit.

The results demonstrate that the inclusion of additional qualitative constraints is not just helpful but essential for effective reasoning. By making implicit physical information explicit, the authors achieved an exponential reduction in the number of states the system had to analyze.

State space reduction across three test models when applying additional qualitative constraints.

As shown in the table above, the state space for the RC-ladder dropped from 241 states to just 28, while the Adby circuit saw a reduction from 5,858 states to 646. This dramatic pruning ensures that the resulting "envisionment" (the map of all possible behaviors) remains useful and interpretable for engineers.

Conclusion and Future Work

This work demonstrates that "redundancy is a virtue" in qualitative reasoning. While numerical solvers often strip away redundant equations to improve speed, these same equations serve as vital filters in the qualitative domain, eliminating physically impossible trajectories.

The primary remaining challenge is "ambiguous branching," where the algorithm identifies multiple possible successor states during event transitions. Current efforts require some manual specification of non-discontinuous variables to resolve these. Future work aims to integrate symbolic equation solvers to automate this process entirely, further streamlining the path from high-level Modelica designs to comprehensive qualitative analysis.

Summary updated

BibTeX
@inproceedings{klenk2014qualitative,
  author = {Matthew Evans Klenk and Johan de Kleer and Daniel G. Bobrow and Bill Janssen},
  title = {{Qualitative Reasoning with Modelica Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8876},
}
Figure