Session: 01-13-01: Congress-Wide Symposium on NDE & SHM: Computational Nondestructive Evaluation and Structural Health Monitoring Count

Paper Number: 94850

94850 - A Machine Learning Framework for Physics-Based Multi-Fidelity Modeling and Health Monitoring for a Composite Wing 

Aircraft being highly complex structures along with high operational reliability and safety needs, the employed health monitoring systems need to be more accurate, robust, and reliable than ever before. In the simulation space, the studies deploy finely meshed models to get an accurate representation of the model in absence of the physical structure. Even though there has been a huge growth in computational capabilities during the past few years, the simulations based on first-principles models are still very expensive. Because of this, the majority of designers have to rely on a coarse-mesh model to make predictions about the overall system and its properties. Although these coarse-grained models are cheaper to evaluate, they involve several simplifying assumptions. In design optimization, multi-fidelity optimization is used extensively for its ability to scale model fidelity as necessary. It uses a rapidly evaluated low order model to close in on the approximate area of focus (damage or optimal area of operation), and a computationally expensive, more accurate model to make corrections to the low-fidelity model. This study is focused on (1) the Application of neural networks for making corrections for a healthy and damaged wing under a variety of loading conditions and (2) for detection and damage localization on the wing.

 

The low-fidelity and high-fidelity models utilized in this study are derived from a wing for an unmanned aerial vehicle (UAV) and modeled using Siemens NX. The high-fidelity model possesses distinct geometric features for the wing’s spars, ribs, and skin. Custom materials are created to model the carbon fiber composite layups for the skin and spars and the epoxy for
the ribs. Each of the bodies in the high-fidelity model is meshed using two- dimensional mapped meshes with CQUAD4 type elements, where the skin and spars are modeled using the laminate method and the ribs are represented using the PSHELL method. In total, the high-fidelity model comprises nearly 220,000 elements. The low-fidelity model is developed from the high-fidelity model, and two different approaches are explored for its creation. The first version of the model consists of one-dimensional CBAR elements, while the other employs a similar procedure as the high-fidelity model with two-dimensional shell elements. The former has a total of 580 elements, and the latter has about 2,000 elements. Damage is modeled in both the low fidelity and high-fidelity models by selecting six geometric regions on the wing and varying the local material stiffness in these regions. This, therefore, allows for the effect of damage size, location, and the extent to be investigated.

 

 Artificial neural networks are a popular choice in machine learning applications because of their ability to model any arbitrary function, as proven by the universal approximation theorem. An algorithm is developed that makes use of the error between low-fidelity data and high-fidelity data, and the correlation between the two models for the learning process. Next, all the data is used to train the network for the detection and damage localization process.

In the full paper, the algorithm will be discussed comprehensively and the effects of using correlation would be shown on the prediction capabilities of neural networks. The results will be discussed for a healthy wing followed by corrections in the damaged wing. Finally, the results for inverse modeling using a global model will be discussed and used for the detection and localization of damage on the outer skin of the wing.

Presenting Author: Gaurav Makkar Rensselaer Polytechnic Institute

Presenting Author Biography: I am a Ph.D. Student at Rensselaer Polytechnic Institute. My work focuses on developing algorithms for model corrections for eVTOL aircraft and structures using learning methods. My interest lies at the intersection of physics-based machine learning, controls, dynamics, and monitoring.

Authors:

Gaurav Makkar Rensselaer Polytechnic Institute
Cameron Smith Rensselaer Polytechnic Institute
George Drakoulas FEAC Engineering
Fotis Kopsaftopoulos Rensselaer Polytechnic Institute
Farhan Gandhi Rensselaer Polytechnic Institute