A Technique for Real-Time Defect Detection of Composite Material Using Machine Learning and Transfer Learning

Abstract

In this study, the aim is to develop a method that is capable of locating the defects in composite plates by using a minimum number of wires and contact points. The reason is that previously there was not a method that can be used to detect defects with a reasonable number of wires and connections. To do so, measurements are restricted to be done only on the edges of the composite plates to predict the location of the defect within the plate. In this work, Single-walled carbon nanotubes (SWNT) were mixed with epoxy resin. By doing so, electrical conductivity was introduced to the resin. The prepared resin is then incorporated with woven fiberglass fabrics to make nanocomposite plates. Probes to measure the electric resistances were placed on the edges of the plates. The 4-probe method was used to measure the electrical resistance between the probes. Due to the complexity of the electrical behavior of composites and having multiple numbers of inputs, Machine Learning (ML) and Deep Learning (DL) methods were used. Data Augmentation (DA) was used to increase the number of labeled data examples that are needed to train the Neural Network (NN) model. Coupled with DA, Transfer Learning (TL) was used. A simulation model of the grid of resistors was developed in LTspice software. In the simulation, the defect was represented by increasing the resistors’ value of the defected cell. The data from the simulation was used for the initial training of the NN and then the experimental data is used in the second stage of Transfer Learning to train and test the final NN. The model's performance in locating the defects is 78.57% while the precision is 2â€x2â€.