Aeroelasticity & Structural Dynamics in a Fast Changing World
17 – 21 June 2024, The Hague, The Netherlands
Home Program Author Index Search

Experimental Validation of Strain-Load Neural Network Model on a Slender Hypersonic Vehicle


Go-down ifasd2024 Tracking Number 185

Presentation:
Session: High speed aeroelasticity 2
Room: Room 1.3
Session start: 11:00 Wed 19 Jun 2024

Ana Cristine Meinicke   meinicke@umich.edu
Affifliation: University of Michigan

Carlos Cesnik   cesnik@umich.edu
Affifliation: University of Michigan

Brianna Blocher   bblocher@utexas.edu
Affifliation: The University of Texas at Austin

Aditya Panigrahi   aditya.panigrahi@austin.utexas.edu
Affifliation: The University of Texas at Austin

Jayant Sirohi   sirohi@utexas.edu
Affifliation: The University of Texas at Austin

Noel Clemens   clemens@mail.utexas.edu
Affifliation: The University of Texas at Austin


Topics: - Computational Aeroelasticity (High and low fidelity (un)coupled analysis methods:), - Experimental Methods in Structural Dynamics and Aeroelasticity (Experimental methods)

Abstract:

Recovery of in-flight loads is crucial for guidance, navigation, and control. The harsh aerothermal conditions experienced in hypersonic flight provide additional challenges for conventional sensors typically installed on the outer surface of the structure. This study investigates a novel vehicle-as-a-sensor concept, where internal measurements of the vehicle’s deformed state are used to infer the loading it is subjected to. The proposed inverse model for this problem consists of a neural network, where strain measured through fiber optic sensors characterizes the deformed state and is used as an input to the machine learning algorithm which outputs the load state. An experimental testbed consisting of an aluminum scaled representative version of the IC3X, a slender hypersonic vehicle, is used as a proof of concept. A finite element model is developed and verified against results of a ground vibration test. The testbed is instrumented with fiber optic strain sensors along the length of the vehicle and force is applied through four actuators attached to load cells. Several static loading cases consisting of combinations of the various actuators are used to evaluate discrepancies between the as-built structure’s response and the predictions from the model. Calibration factors are applied to the model strain results to account for manufacturing of the aluminum model and sensor installation uncertainties, such as thickness of the adhesive layer used to attach the optical fibers to the model surface. The neural network is trained with data consisting of numerical load and strain pairs under conditions spanning those of the experiment. The neural network-based inverse model is validated against the experimental data and compared with a Data-Driven Force Reconstruction method that assumes a linear relation between strain and force. Errors on load recovery given the strain measurements are quantified.