TL;DR
This paper introduces and evaluates "process-based modeling" (PBM), a novel gray-box system identification method that leverages domain-specific knowledge to discover nonlinear system equations. PBM successfully reconstructs both the structure and parameters of a cascaded water-tank system from synthetic and real-world data, demonstrating its potential for building interpretable and accurate models.
What Problem Does This Paper Address?
System identification aims to build mathematical models of dynamic systems from observed data. While black-box methods can achieve high accuracy, they often lack interpretability, failing to reveal the underlying system structure. Gray-box methods, on the other hand, strive for interpretable models by incorporating prior knowledge. However, existing gray-box approaches like symbolic regression can be overly complex and difficult to guide with domain expertise. This paper addresses the challenge of effectively integrating domain-specific knowledge into the process of discovering nonlinear system equations, aiming for models that are both accurate and understandable.
Methods
The core of this work is the Process-Based Modeling (PBM) approach, implemented in the ProBMoT software. PBM works by:
- Formalizing Domain Knowledge: Users define a library of "entity" (static components) and "process" (dynamic interactions) templates. These templates represent generic building blocks of a system, including their properties, parameters, and mathematical relationships.
- Computer-Aided Structure Identification: Given a specific modeling scenario (e.g., a system with two tanks and a pump) and the PBM library, ProBMoT instantiates these templates into specific model components. It then generates candidate model structures by combining these components.
- Parameter Estimation: For each candidate structure, ProBMoT estimates the unknown constant parameters using a numerical optimization task, minimizing the discrepancy between the model's output and the measured data.
- Model Ranking: The discovered models are ranked based on their performance on a validation dataset.
A key innovation is the explicit mapping between the model structure and the physical system structure, achieved through the modular, component-based nature of PBM. The paper also explores two modes of operation: multi-stage (where the modeler intervenes between stages) and single-stage (fully automated).
Datasets
The paper uses two types of data:
- Synthetic Data: Generated by simulating a known cascaded water-tank model with varying levels of added Gaussian noise (variances of 0.01, 0.02, 0.05, 0.1, and 0.2).
- Measured Data: A standard benchmark dataset for a cascaded water-tank system, consisting of 2500 measurements of the input signal and the water levels of two tanks. This data was split into training (1000 points), validation (500 points), and testing (1000 points) sets.
Key Insights and Ideas
- Domain Knowledge as Building Blocks: PBM treats domain knowledge not as a rigid set of rules, but as a library of reusable components (entities and processes) that can be assembled and instantiated. This makes the modeling process more flexible and guided.
- Explicit Structure-Model Correspondence: The component-based nature of PBM ensures that the structure of the discovered equations directly reflects the hypothesized structure of the physical system, enhancing interpretability.
- Automated Exploration of Model Space: ProBMoT can automatically explore a vast space of potential model structures, guided by the provided knowledge library, without requiring manual intervention for every possible combination.
- Handling Implicit Nonlinearities: When the initial domain knowledge is insufficient to capture observed behavior (as seen with measured data), PBM can be extended by weakening constraints or adding new process alternatives, demonstrating its adaptability.
Key Results
- Synthetic Data: Both multi-stage and single-stage ProBMoT accurately identified the correct model structure (S-S) across various noise levels. Parameter estimation was also highly accurate, with relative errors below 5% and often below 1% for the single-stage mode.
- Measured Data:
- Initially, ProBMoT struggled to definitively identify the single "correct" structure among the top candidates, with the L-S model performing slightly better on validation than the expected S-S model.
- However, simulation responses for the top models (S-S and L-S) were nearly indistinguishable on test data.
- By extending the domain knowledge to include a "Power" process (allowing for sub-linear dynamics), ProBMoT successfully identified a model (P-P) that significantly outperformed S-S and L-S, achieving a validation error of 0.238 and a test error of 0.164. This highlights the importance of comprehensive domain knowledge.
Limitations and Future Work
- Limitations:
- The initial experiments with measured data showed ProBMoT struggling to perfectly distinguish between the top candidate structures, suggesting potential limitations in the initial domain knowledge provided.
- The paper acknowledges that the performance improvement with the "Power" process was due to changing the model structure, not just adding parameters.
- Future Work:
- Extend the evaluation to other standard nonlinear benchmark systems.
- Compare PBM to other state-of-the-art gray-box modeling approaches in terms of accuracy and model comprehensibility.
- Further explore the application of PBM to different domains beyond the water-tank system.
Summary updated