Model-Based Diagnosis of Hybrid Systems Using Satisfiability Modulo Theory

Alexander Diedrich, Alexander Maier, Oliver NiggemannAAAI 2019 · DOI

Introduction

This paper introduces a novel model-based diagnosis (MBD) framework for hybrid systems—those combining discrete logic and continuous physical processes. By using Satisfiability Modulo Theory (SMT) to bridge the gap between dynamic state-space models and symbolic logic, the authors provide a way to automatically identify the "root cause" of faults in complex industrial plants without the need for manual, rule-based expert systems.

The Challenge of Hybrid Systems

In the era of Industry 4.0, Cyber-Physical Production Systems (CPPSs) are becoming increasingly complex and distributed. Identifying the exact source of a fault, such as a sticking valve or a leaking pipe, is a significant challenge. Traditional diagnosis methods generally fall into two categories:

  1. Heuristic/Phenomenological: These rely on machine learning or expert rules to map symptoms to faults but often struggle with complex causalities and require extensive labeled "faulty" data.
  2. Model-Based Diagnosis (MBD): These use a "digital twin" to predict healthy system behavior. Deviations from the model indicate a fault.

The primary hurdle for MBD in industrial settings is hybrid behavior. Most existing MBD tools are designed for static, discrete systems. Real-world plants, however, involve continuous variables (e.g., pressure, flow) alongside discrete events (e.g., pump activation). Current solutions often over-discretize time and values, leading to lost accuracy or computational bottlenecks.

Proposed Methodology

The authors propose a three-step pipeline that decouples physical processes from logical diagnosis:

  1. Dynamic State-Space Modeling: The system is represented using differential equations (approximated via the Euler method). This "observer" predicts the next state of the system based on current inputs and outputs.
  2. Residual Generation: The system calculates "residuals"—the numerical difference between sensor observations and model predictions.
  3. SMT-LRA Logic: This core innovation uses Satisfiability Modulo Theory with Linear Arithmetic (SMT-LRA) to translate numerical residuals into symbolic logic. The SMT solver evaluates whether current observations are consistent with the healthy model of each component.
  4. Diagnosis via Reiter’s Lattice: If inconsistencies are found, the system employs Reiter’s diagnosis lattice to identify the minimal cardinality diagnosis—the smallest set of components whose failure explains the anomalies.
The proposed MBD framework architecture, decoupling dynamic physical models from static logic-based diagnosis.

Crucially, these models can be automatically derived from standard engineering artifacts like Piping and Instrumentation Diagrams (P&IDs) and AutomationML files, reducing manual effort.

Evaluation and Datasets

The approach was evaluated using two primary simulations:

  • Four-Tank Model: A classic benchmark for hybrid systems consisting of four interconnected water tanks, pumps, and valves.
  • Tennessee Eastman Process: A complex simulation of a real-world chemical process involving 20 different types of injected faults.
Schematic of the Four-Tank Model used for benchmarking the diagnosis algorithm.

Key Insights

  • Decoupling Modeling and Diagnosis: Using residuals as an interface allows the diagnosis algorithm to operate without direct knowledge of the underlying differential equations.
  • Weak Fault Models (WFMs): By focusing on defining healthy behavior, the system can flag any component that deviates from its expected state as a potential fault candidate.
  • Precision without Manual Discretization: Leveraging SMT-LRA allows the system to handle real-valued residuals directly, preserving the precision of continuous physical signals.

Results and Limitations

The algorithm demonstrated soundness on the Four-Tank Model, identifying injected faults immediately upon occurrence. In the Tennessee Eastman Process, it successfully isolated all faults with identifiable component causes, such as pipe feed losses and reactor cooling failures.

However, the authors noted that success depends heavily on the sampling rate; low granularity can cause symptoms to propagate, obscuring the root cause. Additionally, indirect faults (e.g., changes in input feed ratios) remain difficult to detect without direct sensors at the inputs.

Conclusion

This SMT-based framework offers a robust path forward for diagnosing complex hybrid systems. Future work will focus on increasing model expressiveness with set-points and validating the framework on real-world industrial hardware.

Summary updated

BibTeX
@inproceedings{diedrich2019model,
  author = {Alexander Diedrich and Alexander Maier and Oliver Niggemann},
  title = {{Model-Based Diagnosis of Hybrid Systems Using Satisfiability Modulo Theory}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33011452},
}
Figure