Abstract
Identifying the underlying physics of dynamical systems typically requires full state measurements, which are often unavailable in real-world structural monitoring. This paper presents sparse Approximate Bayesian Computation (s-ABC), a framework that discovers governing equations using only acceleration time-series data. By integrating spike-and-slab priors with likelihood-free inference, the method identifies parsimonious models without requiring displacement or velocity measurements.
The Challenge of Acceleration-Only Data
In structural health monitoring, "equation discovery" aims to learn physical laws directly from sensor data. While methods like SINDy (Sparse Identification of Nonlinear Dynamics) are effective, they generally necessitate high-fidelity measurements of all state variables, including displacement and velocity. In practice, engineers primarily use accelerometers. Integrating acceleration to obtain displacement and velocity introduces significant drift and error, creating a bottleneck for traditional discovery algorithms. The s-ABC framework overcomes this by working directly with the acceleration data actually available in the field.
Methodology: Sparse Approximate Bayesian Computation (s-ABC)
The s-ABC framework is built upon three core pillars designed to handle the intractability of acceleration-only likelihoods and the need for model simplicity.
- Approximate Bayesian Computation (ABC): Since the likelihood function for acceleration-only data is often intractable, ABC uses simulations. It compares data generated from candidate models to observed data using a distance metric ($\rho$), bypassing the need for an explicit likelihood.
- Spike-and-Slab Prior: To ensure the discovered model is parsimonious (simple and interpretable), the authors employ a spike-and-slab prior. This mechanism forces most terms in a functional library to zero (the "spike") while allowing significant physical terms to take non-zero values (the "slab").
- Nested Sampling & Reinitialization: The algorithm iteratively refines a population of candidate models ("particles"). A Gaussian Mixture Model (GMM) samples new candidates, while a reinitialization step prevents convergence to local optima, balancing exploration of new structures with parameter fine-tuning.
Experimental Validation
The authors tested s-ABC on four benchmark systems: a damped pendulum, a linear oscillator, a Duffing oscillator, and an oscillator with quadratic viscous damping. Synthetic data included 2% Gaussian white noise to mimic realistic sensor conditions.
Key Results
The s-ABC algorithm demonstrated exceptional accuracy in both model selection and parameter estimation:
- Model Selection: Inclusion Probability (IP) analysis confirmed that the algorithm consistently identified the correct terms from the functional library, effectively distinguishing between physical signals and noise.
- Parameter Accuracy: The Mean Squared Relative Error (MSRE) was remarkably low across all test cases. For instance, in the damped pendulum case, the error was as low as $2.68 \times 10^{-6}$.
- Predictive Power: Discovered models were integrated to predict future states. The resulting displacement and velocity trajectories closely matched the ground truth, despite the algorithm never having access to those states during the discovery phase.
Discussion and Future Directions
While s-ABC is robust, it is computationally intensive due to the thousands of simulations required by the ABC process. Additionally, the method's performance depends on the predefined functional library and manual tuning of hyperparameters like the sparsity factor $\beta$. Future research aims to automate hyperparameter selection and accelerate the sampling process to enhance the framework's applicability for large-scale engineering systems.
Summary updated