Date of Award
Master of Science (MS)
Oliver J Myers
The characterization of the damage state of a system provides insight into its performance and safety during operation. In composite materials, specifically, fiber-reinforced polymers, delaminations form from the evolution of cracks in a matrix that leads to adhesion failure between adjacent laminae. Nondestructive evaluation (NDE) seeks to characterize the state of a material or system during non-operational times. A previously proposed NDE method employs embedded magnetostrictive particles between laminae of carbon fiber reinforced polymer (CFRP) for damage sensing. The phenomenon of magnetostriction couples the mechanical state of a material with its magnetic state so that a change in the local stress field alters its magnetic susceptibility. The change in magnetic susceptibility is measured using an induced sensing voltage.
This work aims to provide a preliminary exploration of machine learning to predict the presence of a delamination using the embedded magnetostrictive particle NDE method with CFRP laminates. Machine learning algorithms' ability to decipher and develop relationships among input features attracted its use since the visual examination of the experimental induced sensing voltage plots yielded inconsistent delamination predictions. This work investigated the feasibility of an analytical model based on the Euler-Bernoulli beam theory to generate data. This model utilized functional relationships to characterize the nonlinear behavior of the magnetostrictive material Terfenol-D. Fourier series relationships reduced the error in the characterization of the properties over the previously proposed functions. The previously proposed derived model failed to converge for the calculation of stress within the magnetostrictive material.
Eight machine learning algorithms were employed using a Python script to classify the presence of a delamination within a unidirectional HexPly AS4/3501-6 CFRP embedded with Terfenol-D particles. The maximum accuracy achieved was approximately 80, whereas the average accuracy for all the models was just below 71%. The multi-layer perceptron (MLP) models, a neural network algorithm, produced the highest prediction accuracy for this two-class classification problem because of their ability to account for nonlinear relationships. A parametric study involving the architecture and activation function was necessary for the MLP models because of the range of obtained prediction accuracies. A direct relationship was observed between the number of hidden layers and the accuracy outliers. Despite the accuracy of the examined machine learning models' being less than that of other NDE applications, this preliminary investigation demonstrated that machine learning could be paired with previously indiscernible experimental data to detect delaminations.
Nelon, Christopher, "Detection of Delaminations in Carbon Fiber Reinforced Polymers Embedded with Terfenol-D Particles Using Machine Learning" (2020). All Theses. 3445.