Data-Driven Modeling of a Controlled Orthotropic Plate Using Machine Learning

Y. Kim, A. Zuyev 1,, F. Pellicano, A. Zippo · 2026

Read Data-Driven Modeling of a Controlled Orthotropic Plate Using Machine Learning by Y. Kim, A. Zuyev 1,, F. Pellicano, A. Zippo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-20931.txt.

Introduction

Kim and co-authors learn a data driven model of a controlled vibrating composite plate. Exact modeling is hard because the plate is orthotropic and its damping is unknown. They build surrogate models from measured input and output data. Three model types are tried: linear regression, a multilayer perceptron, and a gated recurrent unit. The perceptron gives an acceptable match to the real system. It captures nonlinear effects in the plate's input and output behavior. The study shows machine learning can stand in for a full physical model here.

Catalog record

CREATORY. Kim, A. Zuyev 1,, F. Pellicano, A. Zippo
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29656
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.20931

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Data-Driven Modeling of a Controlled Orthotropic Plate Using Machine Learning Y. Kim, A. Zuyev 1,, F. Pellicano, A. Zippo Abstract We study the problem of learning the input-output map of a controlled vibrating plate with a composite structure from experimental measurements. Analytical modeling of this control system faces challenges due to the essential orthotropy and unknown damping characteristics of the material. Surrogate models based on linear regression, multilayer perceptrons, and gated recurrent units are constructed from the available sampled data. Through comparative analysis, we show that the multilayer perceptron model provides an acceptable approximation of this dynamical system, capturing the potentially nonlinear phenomena in its input-output behavior. \msc 93B15, 68T05 \novelty • Development and experimental comparison of surrogate models for an orthotropic controlled elastic plate, where anisotropic properties, nonlinear actuation characteristics, and complex boundary conditions play a significant role in the dynamics. • The comparison of linear regression models with multilayer perceptrons (MLPs) and gated recurrent unit (GRU) techniques shows that nonlinearity plays a significant role in the system’s behavior. • It is justified that a moderate artificial neural network depth ( h=4 or 6 ) is sufficient to effectively leverage the experimental input–output behavior, and that the pretrained MLPs have acceptable approximation ability with the suggested time history length of s=200 samples. keywords orthotropic plate, vibration control, input-output map, machine learning 1 Introduction Modeling engineering systems from first principles and using surrogate data-driven approaches represent two different paradigms. Purely numerical surrogate methods often fail to provide sufficient insight into the influence of mechanical parameters and the effects of time-varying input actions, since they are typically trained as input–output mappings and therefore lack explicit physical structure or interpretability (see, e.g., [ 12 , 14 ] and references therein). Although analytic modeling based on variational formulations provides valuable insight into the physics of vibrating flexible structures [ 2 , 16 , 13 ] , an accurate representation of damping, structural nonlinearities, and actuator–sensor dynamics typically requires additional analysis and often relies on system identification techniques and experimental data [ 3 , 7 , 10 , 1 , 9 ] . The goal of this paper is to determine the range of applicability of surrogate modelling techniques for a specific flexible structure – controlled orthotropic plate – and to identify machine learning models that achieve efficient performance and have potential for future applications in control design. High-fidelity data obtained from an experimental setup at the Enzo Ferrari Department of Engineering, University of Modena and Reggio Emilia, serve as the basis for a comparative analysis of different machine learning approaches. The experimental setup is shown in Fig. 1 and described in Section 2 . To provide further insight into the mechanical characteristics of the vibrating system and to highlight modelling challenges in the context of an analytical input–output description, we present a previously developed linear PDE-based model of the plate in Section 3 . Figure 1 : Controlled orthotropic plate at the Enzo Ferrari Department of Engineering, University of Modena and Reggio Emilia. To address possible nonlinearities in the input-output behavior of the considered mechanical system and to present computationally feasible validation procedures, we summarize the construction of surrogate models for this setup in Section 4 . Section 5 presents extensive numerical results comparing the input–output behavior of the orthotropic plate using different machine-learning-based surrogate models. The principal outcomes of this work are summarized in the Conclusion. 2 Experimental Setup The experiments were conducted on a rectangular composite sandwich plate with orthotropic behavior, having in-plane dimensions of 1000\times 500\,mm (see Fig. 1 ). The plate consisted of two outer carbon-fiber/epoxy laminated skins and a 3\,mm -thick NOMEX honeycomb polymeric core; the carbon composite was identified as GG285P(T700)-DT120-40, and the fibers were woven in a 0/90 configuration. The specimen was mounted on a dedicated support frame by means of thin nylon wires passing through four holes drilled at the plate corners and connected to elastic bands fixed to the frame members. The suspension geometry and the tensioned lengths of the elastic elements were made as uniform as possible in order to approximate free-boundary conditions and to limit the influence of the supporting structure on the measured dynamics. The plate was excited by an electrodynamic shaker installed behind the specimen and suspended from the supporting structure in order to reduce parasitic mechanical constraints. The shaker was connected to the plate through a slender stinger so as to provide an approximately pointwise transverse excitation while minimizing additional structural constraints. Similar experimental strategies for vibrating plates and flexible structures have been adopted in previous investigations on orthotropic structures and data-driven modeling approaches. In the present study, particular attention was devoted to the calibration of the actuator–sensor chain in order to ensure consistency between the applied input and the measured response. The excitation force generated by the shaker was monitored by a load cell placed along the excitation path, between the shaker rod and the plate surface, allowing the input signal u(t) to be reconstructed from the measured voltage output of the force sensor after appropriate signal conditioning, amplification, scaling, and calibration. The acquisition channels were organized so that the load-cell signal represented the input and the piezoelectric response represented the output. The structural response of the plate was measured through a bonded piezoelectric transducer used as a strain gauge, whose voltage output was taken as the experimental observable. The sensor was installed at a non-collocated location with respect to the actuator, thus providing a sensitive measurement of the local curvature of the plate and enabling the detection of bending vibrations associated with the orthotropic elastic properties of the composite structure. The setup documentation identifies the transducer as a Macro Fiber Composite manufactured by Smart Materials (MFC P1 type, model M8557-P1), with an active area of 85\times 57\,mm , overall dimensions of 103\times 64\,mm , capacitance of 12.84\,nF , free strain of 1350\,ppm , and blocking force of 693\,N . Owing to its very small thickness, the transducer could be surface-bonded to the plate using epoxidic resin with negligible modification of the structural properties, while still ensuring sensitive measurement of the local structural deformation. The actuation and measurement chain was integrated with a digital acquisition and control architecture based on sSPACE and MATLAB/Simulink , ensuring accurate synchronization between excitation and response measurements. The sampling frequency was set to 30\,kHz in order to capture the relevant dynamic content of the plate vibrations and avoid aliasing effects in the recorded time histories. In the broader setup, the piezoelectric transducer is driven through a TREK PA05039 high-voltage amplifier, which converts an input in the range from -2.5\,V to 7.5\,V into an output from -500\,V to 1500\,V , with nominal input impedance of 25\,k\Omega , output impedance below 0.1\,k\Omega , and output noise below 50\,mV rms. The electromagnetic shaker was powered by a TIRA BAA 60 amplifier, capable of supplying up to 2.5\,A with a maximum 5\,V input signal. During the experiments, the structure was excited by short voltage pulses applied to the shaker at regular time intervals. This excitation strategy generated transient responses followed by freely decaying vibrations governed by the intrinsic structural and material damping of the composite plate. Such a procedure provides a rich dataset containing both the forced

Read the complete text