Open Collections

Yun, Hongguang

University of British Columbia

Composite materials are widely used in aerospace, wind energy, and shipbuilding, offering significant benefits such as superior strength-to-weight ratios. However, they are prone to damage from impacts, necessitating advanced methods for damage localization to ensure safety and reduce maintenance costs. Deep learning-based ultrasonic guided waves (UGW) methods are increasingly recognized for their effectiveness in locating damage in composites. This study aims to develop deep neural networks (DNN) models for damage localization, focusing on two main challenges: limited dataset sizes and the integration of physical knowledge. Three specific application scenarios were identified: independent and identically distributed ($i.i.d.$) damage localization, domain generalized damage localization, and out-of-distribution (OOD) damage localization. Firstly, to achieve accurate damage localization, incorporating transducer coordinates with UGW signals is essential. This study developed a multi-dimensional data fusion DNN model that integrates a convolutional neural network-based encoder for high-dimensional wave signals and a novel random Fourier projection head for transducer coordinates. The proposed method outperformed existing DNN-based methods, achieving an average absolute distance error of approximately 2 millimeters in damage localization on plate-like composite structures. Secondly, distribution discrepancies often occur between training and testing data, leading to reduced model performance. This study proposed a domain generalization method to mitigate this issue in a multi-class classification task. A feature distance maximization mechanism was introduced to improve model’s resistance to such distribution discrepancies. The proposed method outperformed leading DNN-based models, achieving an average accuracy of around 90%, compared to less than 85% for other methods. Thirdly, this study proposes a physics-informed neural network model to address the challenging OOD damage localization. The damage localization process was divided into two stages: extracting the time-of-flight from UGW, and estimating the location. The proposed method outperformed existing DNN-based and traditional methods, achieving relative prediction errors of 0.08 meters, compared to over 0.1 meters for other models. Overall, this study introduces three methodologies that are designed for different levels of data availability and the embedding of physical knowledge. Each method contributes to improving the robustness and accuracy of damage localization, paving the way for effective and efficient damage assessment in composite materials.

Text

2024-05-30

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/88389

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/