Adaptive data generation framework for machine learning models in nonlinear structural dynamics

RWTH Publications (RWTH Aachen)

Abstract

Growing demand for structural vibration control devices, such as tuned mass dampers (TMDs), has increased the need for efficient, data-driven models that enhance design, analysis, and real-time applications including digital twins. Machine learning methods enable fast and accurate modeling of nonlinear dynamic systems, but require high-quality training data that is challenging and costly to obtain in structural dynamics. Vibration control devices undergo mandatory laboratory testing before deployment, generating extensive experimental data currently used only for verification rather than developing predictive models. This dissertation presents a novel framework for testing-integrated modeling that simultaneously verifies device performance and generates high-fidelity data-driven models through adaptive experimental design. The core contribution is the Adaptive Data Generation (ADaGen) framework, which formulates experimental data collection as an optimization problem in the space of experimental parameters. Unlike conventional testing protocols, ADaGen iteratively selects experimental conditions to maximize model accuracy using derivative-free optimization. This addresses the dynamic sampling problem inherent in testing devices with auxiliary mass, where desired states cannot be directly specified but must be reached through unknown system dynamics. Numerical validation using a Duffing oscillator demonstrates that ADaGen generates an informative dataset with 19 experiments compared to 35 experiments required by conventional space-filling methods. Experimental validation employs a tuned liquid column damper prototype with Neural Ordinary Differential Equations, including a novel architecture extension that simultaneously identifies system dynamics and unknown modal mass parameters without direct force measurements. Furthermore, the presented extensions to the baseline framework address real-world constraints through interpretable machine learning initialization for reduced computational cost, cumulative data generation for multi-degree-of-freedom systems, and reference-free adaptive sampling that eliminates the need for comprehensive test datasets.

Authors 1

1. Pavle Milicevic corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen