Studying Failure Modes of GFRP Laminate Coupons Using AE Pattern-Recognition Method

Abstract

Acoustic emission (AE) monitoring has been identified as one popular technique applicable for damage discrimination in composites. However, because the damage is not directly identified from the AE data, other nondestructive techniques are usually needed to further evaluate the damage. Moreover, there is no general guideline for AE monitoring in composites because of the unique AE signatures of each defect and the diverse properties due to varied material configurations. Therefore, this study aimed to develop a reliable method using advanced pattern-recognition techniques on AE data to identify damage modes exhibited in the glass fiber reinforced polymer (GFRP) coupons and composite systems with similar configurations. Three sets of GFRP laminate samples originated from a new fiber-reinforced polymer (FRP)–balsa wood composite bridge were tested to study critical failure modes such as fiber breakage, matrix cracking, delamination, and debonding. A methodology based on unsupervised pattern recognition was then developed. First, in order to eliminate uncorrelated AE features, the complete link hierarchical clustering algorithm and principal component analysis were used to preprocess the data and select a suitable subset of AE features for the clustering task. The unsupervised k-means clustering analysis was then applied to separate the date into several clusters and correlate each cluster to its corresponding failure mechanisms. After a reliable AE database was built for a typical sample of each test set, neural networks (NNs) based on multilayer perceptron (MLP) and support vector machine (SVM) algorithms were developed. Finally, the trained NNs were used for pattern recognition in samples with unknown damage modes. Most results conformed to the visual observation and thus led to the NN models with good performance. The developed network system was also applied on a full-scale GFRP bridge deck with a similar configuration, and the performance achieved further confirms the potential of the developed methodology for damage identification in full-scale structures.