Evolutionary sparse data-driven discovery of multibody system dynamics

Ehsan Askari, Guillaume Crevecoeur

Introduction

Understanding the dynamics of multibody systems—such as robots, vehicles, or industrial machinery—is critical for effective control and monitoring. Traditionally, engineers rely on physics-based models derived from first principles. However, these models often suffer from bias or fail to capture real-world complexities like non-linear friction and environmental interactions.

Conceptual overview of the hybrid modeling approach combining physics-based and data-driven methods.

Data-driven system identification (SI) offers an alternative, yet existing methods like SINDy (Sparse Identification of Nonlinear Dynamics) face significant hurdles. They typically require a predefined "dictionary" of candidate functions; if the true physical law involves a complex rational function or a specific constant not in that dictionary, the method fails. Furthermore, many systems exhibit hybrid dynamics, switching modes (e.g., a walking robot making ground contact) that are difficult to identify.

Methodology: Evolutionary Symbolic Sparse Regression

To address these limitations, this research introduces a hybrid "evolutionary symbolic sparse regression" approach. This method "distills" parsimonious mathematical models directly from raw data, capable of uncovering complex rational functions and hidden physical constants even in switching systems.

The evolutionary symbolic sparse regression workflow for discovering equations from time-series data.

The core of the approach is the Evolutionary Symbolic Sparse Regression module, which employs a "best of both worlds" strategy:

  • Genetic Programming (GP): Instead of a fixed dictionary, GP evolves the structure of equations using tree-based representations. This allows the algorithm to "invent" complex functional forms, including rational functions.
  • Sparse Regression: Embedded within the GP, a hard-thresholding ridge regression (or Lasso/Elastic Net) calculates optimal coefficients for evolved terms, ensuring the final model is "parsimonious"—using the fewest terms possible to explain the data.
  • Adaptive Fitness Function: A multi-objective fitness measure balances accuracy (minimizing residual sum of squares) with simplicity (penalizing the number of terms and operators) to prevent overfitting.
  • Hybrid Identification: For systems with switching modes, K-nearest neighbors (KNN) clustering is used to identify sub-models for different state-space regions and determine switching boundaries.

Experimental Validation

The methodology was evaluated using three simulated multibody benchmarks designed to test specific capabilities:

  1. Two-DoF Spring-Mass System: A standard benchmark for forced vibrations.
  2. Crank-Slider Mechanism: Chosen for its complex rational equations of motion, which are notoriously difficult to identify via standard methods.
  3. Sliding Mass with Stribeck Friction: A hybrid system used to test mode-switching detection between different friction regimes.

Robustness was tested by adding Gaussian white noise (SNR from 15dB to 40dB) to the datasets.

Key Results and Insights

The algorithm successfully identified complex rational governing equations for the crank-slider mechanism, a significant advancement over standard sparse regression. Key insights include:

  • Evolving Constants: The method can estimate real-valued constants (e.g., frequencies) inside function arguments via a specific mutation rule.
  • Parsimony as a Physical Prior: By rewarding simplicity, the algorithm mimics scientific preference for elegant laws, helping it ignore noise and find underlying physical truths.
  • Data Efficiency and Extrapolation: The method discovered accurate models with as few as 20 sampling points. Because the output is a symbolic law rather than a "black box," it demonstrated excellent extrapolative capability for conditions outside the training set.
  • Noise Robustness: Total Variation Regularization proved superior for pre-processing noisy time-series data, allowing the algorithm to maintain structural accuracy even at 20dB noise levels.

Limitations and Future Work

While promising, the method shows sensitivity in rational functions when one term in a denominator dominates others. Future work will focus on applying the approach to real-world experimental data and extending it to systems described by redundant coordinates and Differential Algebraic Equations (DAEs).

Summary updated

BibTeX
@inproceedings{askari2023evolutionary,
  author = {Ehsan Askari and Guillaume Crevecoeur},
  title = {{Evolutionary sparse data-driven discovery of multibody system dynamics}},
  booktitle = {Multibody},
  year = {2023},
  doi = {10.1007/s11044-023-09901-z},
}

Loading related content...

Figure