Abstract

ABSTRACT

A Technique for Real-Time Detection of Defects in Composite Structure using Carbon nanotubes, Machine Learning Including Transfer Learning

Farzad Kashefinishabouri

Fiber-reinforced polymer composites have garnered interest in a range of industrial applications due to their outstanding mechanical characteristics and lightweight. Monitoring the health of polymer composite structures in real-time is one of the most challenging issues in the practical use of composites due to their susceptibility to many types of damage. Conventional non-destructive tests (NDT) methods such as X-ray tomography and ultrasonic may be used to assess composite materials but the drawback of conventional NDT techniques is that they cannot be implemented when the composite part is in use.
Composite plates’ electrical characteristics can be to do real-time health monitoring. Carbon nanotubes (CNTs) because of their excellent conductivity characteristics, are used in composites to enhance the conductivity of resins within composites. It is shown that by monitoring the electrical behavior of composites with CNT embedded resins, defects can be detected. The issue with this approach is that numerous wires and connections are required. The weight of composite structures is greatly influenced by the number of wires and connections, which also makes the system more prone to errors because numerous connections must function well for the system to respond as intended.
To tackle this problem, in this study, the goal is to reduce the number of required probes and connections by limiting the probes to the edges of composite plates rather than throughout the plate. By having the probes on the edges of the plate, there may not be a direct correlation between defects at different locations within the plates to the measurements as it was in the previous cases. There are two approaches to tackle this problem:1. To develop a physics-based model that can precisely model the electrical behavior of composite with CNTs embedded in the resin. 2. To develop a data-driven model that can relate the measurements to the location of the defect. As the first approach is expensive and time-consuming the second approach is picked in this study. Neural network (NN) is used to find the pattern between measurements on the edges and the location of the defect. The problem with using neural network (NN) models is that they require numerous numbers of labeled examples. To tackle this problem Transfer Learning (TL) and data augmentation is used. TL is used to reduce the number of labeled data points required for the training process as in composites it is too expensive and time-consuming to generate a huge number of data points. In the TL method that is used, the training is done in two stages, first stage the training is done based on the data generated from a similar problem that the data can be abundant, and the second stage of training is done using experimental data of the exact problem. Data augmentation is used on experimental data to increase the data points for training NN. The performance of the trained neural network in locating defects by having the probes only on the edges of the samples is promising (accuracy of 78.57% on the test set).
Also, the performance of the neural network models for different precisions and sample sizes were studied. The precision is defined as the area in which the defects can be located within.

KEYWORDS
Real-time defect detection, Neural Network, Transfer learning, Data augmentation