Session: 17-01-01: Research Posters

Paper Number: 149641

149641 - Identification of Damage State in Carbon Fiber Reinforced Composites Using Transfer Learning 

In recent years, composite materials have been implemented in many applications starting from micro-level to large-scale structures. With the increase in applications, the need to understand their behavior under various stress conditions has also become equally important. Multiple studies are being conducted in order to develop a non-destructive framework to understand the damage in composites. Using artificial intelligence and new tools for automating such detection is significantly important considering the complexities of the composite materials. Due to the nature of composite materials, they undergo sudden damage without showing any sign, unlike metals. Cracks and delamination are two important failure modes in composites that cannot be detected by visual inspections. These occur in between the layers of composites and can travel along the thickness as well as the length of the component. This sudden damage can prove to be catastrophic in applications like wind turbines where the blades can fall off without a warning causing loss of life at times. In order to avoid sudden catastrophes, identifying the damage level and creating a maintenance schedule is equally important. Thus, it is important to conduct a non-destructive inspection from time to time to identify the defects in the components and create a maintenance schedule to maintain the integrity of the components. Even if the repair is not possible, identification of critical areas can save economic losses and future catastrophes. Thus creating a framework that can effectively use the nondestructive inspection data and create a framework to identify the damage accurately is significantly important. In this study, we are using a public dataset of carbon fiber-reinforced polymer composite which contains the Lamb wave signals of composite specimens at multiple damage levels. The energy of the Lamb waves dissipates when it comes in contact with irregularities so the reduction in amplitude can be related to the amount of damage. However, the amplitude in some cases may not be significantly different from each other causing wrong predictions. Therefore, identifying the damage based on these signals becomes nearly impossible. For that reason, we are proposing to incorporate deep learning algorithms based on transfer learning. The signals were converted into images and data augmentation tools were used for increasing the size of the dataset. After this multiple transfer learning algorithms such as VGG16, incpetionV3, and LSTM-based algorithms were implemented and compared to each other. A comparison of transfer learning vs regular CNN algorithm was also performed. Results showed that the transfer learning showed better accuracy of damage state identification. A sensitivity analysis based on the number of images was also performed. It was observed that the sensitivity analysis is significantly important in optimizing the hyperparameters. This research is an attempt to provide a robust framework with significantly high accuracy to detect damage in composite materials.

Presenting Author: Shweta Dabetwar University of Arkansas at Little Rock

Presenting Author Biography: Dr. Dabetwar is a mechanical engineering professor at the University of Arkansas at Little Rock.

Authors:

Hannah Jones University of Arkansas at Little Rock
Shweta Dabetwar University of Arkansas at Little Rock