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

Abstract

Computational mechanics is a useful tool in the structural health monitoring community for accurately predicting the mechanical performance of various components. However, high-fidelity models simulated through the finite element analysis (FEA) necessitate a large amount of computing power. This paper presents a new approach to develop a multi-fidelity model using artificial neural networks for health monitoring purposes. The proposed framework provides significant savings in computational time compared to a model trained only using high-fidelity data, while maintaining an acceptable level of accuracy. The analysis is conducted using two finite element models, of different fidelity, of an unmanned aerial vehicle (UAV) wing, with damage modeled at six locations, and varying severity. The damage is modeled by changing the stiffness properties of the materials at these locations. The algorithm developed aims at minimizing the number of high-fidelity data points for correcting the outputs of the low-fidelity model. It was observed that the low-fidelity model requires 8 high-fidelity data points to meet the desired error tolerance. This corrected low-fidelity model is then used for locating and quantifying the damage given the strains and frequency by expanding the previously trained network to output damage diagnosis results. The model with applied correction is able to locate the damage with an accuracy of ∼ 94% and quantify the damage with an accuracy of 93%. The performance of the corrected low-fidelity model is compared with a network trained only with high-fidelity datasets and it was observed that the corrected model requires 54% fewer data points as compared to the high-fidelity trained network.