Defect detection in composites by deep learning using solitary waves. (1st February 2023)

• Record Type: Journal Article
• Authors: Yoon, Sangyoung; (Song-Kyoo) Kim, Amang; Cantwell, Wesley J.; Yeun, Chan Yeob; Cho, Chung-Suk; Byon, Young-Ji; Kim, Tae-Yeon
• Is Part Of: International journal of mechanical sciences. Volume 239(2023)
• Publication Date: 2023-02-01
• Subjects: AS4/PEEK -- Composite -- Convolutional neural networks -- Defect detection -- Granular crystal -- Machine Learning -- Non-destructive evaluation -- Solitary wave -- Artificial intelligence; Mechanical engineering -- Periodicals; Génie mécanique -- Périodiques; Mechanical engineering; Maschinenbau; Mechanik; Zeitschrift; Periodicals; 621.05
• DOI: 10.1016/j.ijmecsci.2022.107882
• Languages: English
• ISSNs: 0020-7403

Abstract

Highlights: Developed the CNN-based deep learning algorithm for a real-time detection of the existence and location of delamination in laminated composites using HNSW signals generated from a granular crystal sensor in a non-destructive manner. Investigated the influences of the hidden layer and other CNN parameters such as learning rate, activation function, dropout, input image pixel size, batch size, and filter size to improve the accuracy of the deep learning algorithm. Furthermore, a general fitting curve (see Eq. (5) ) that can be used for the optimal choices of the input pixel and batch sizes was developed. Investigated the efficiency and accuracy of three different types of the input signals, i.e., original (raw) without pre-processing and two pre-processed signals (i.e., time-sliced and time-sliced noise-cutting signals), for real-time detection of detects using HNSWs. Moreover, we provided mathematical formulations to obtain time-sliced signals in Eq. (1) and time-sliced noise-cutting signals in Eq. (2) from pre-processing of the original HNSW signals. Developed a multiple mode testing scheme, classifying defects using multiple HNSW signals instead of using a single HNSW signal, to improve the classification accuracy of the deep learning algorithm. Abstract: This paper proposes a real-time non-destructive evaluation technique to detect defects in laminated composites by deep learning using highly nonlinear solitary waves (HNSWs). HNSW data are collected by conducting experiments using a granular crystal sensor composed of a vertical array of steel beads directly contacting an AS4/PEEK composite plate. Using HNSW data, a deep learning algorithm based on the convolution neural networks (CNN) is trained and tested for the identification of delamination in AS4/PEEK composites. The influence of the number of hidden layers and various CNN parameters is investigated for improved classification accuracy of the deep learning algorithm. A general curve fit is presented in order to facilitate the correct choice of the input pixel and batch size. Moreover, a multiple mode testing scheme, classifying defects using multiple HNSW signals, is introduced to improve the accuracy of the algorithm. The efficiency and accuracy of using three different types of the input signal (i.e., original (without pre-processing) and time-sliced/time-sliced noise-cutting signals (with pre-processing)) are examined for the real-time detection of defects. Mathematical formulations are established to obtain time-sliced and time-sliced noise-cutting signals from the original HNSW signals. It was found that accuracy could be improved by increasing both the number of hidden layers and the input pixel size, reducing the learning rate, and by using a batch normalization process and RELU activation function. For all three input signals, accuracy levels of over 90% were achieved in identifying the existence and location of delamination in AS4/PEEK composites, highlighting the possibility of using the proposed deep learning algorithm for the real-time detection of defects in laminated composites. Graphical abstract: Image, graphical abstract