A comparison of machine learning algorithms for assessment of delamination in fiber-reinforced polymer composite beams

Abstract

Structural health monitoring techniques based on vibration parameters have been used to assess the internal delamination damage of fiber-reinforced polymer composites. Recently, machine learning algorithms have been adopted to solve the inverse problem of predicting delamination parameters of the delamination from natural frequency shifts. In this article, a delamination detection methodology is proposed based on the changes in multiple modes of frequencies to assess the interface, location, and size of delamination in fiber-reinforced polymer composites. Three types of machine learning algorithms including back propagation neural network, extreme learning machine, and support vector machine algorithm were adopted as inverse algorithms for assessment of the delamination parameters, with a special focus on the interface prediction. A theoretical model of fiber-reinforced polymer beam with delamination under vibration was constructed to learn how the frequencies are affected by the delaminations (“forward problem”) and to generate a database of “frequency shifts versus delamination parameters” to be used in machine learning algorithms for delamination prediction (“inverse problem”). Multiple carbon/epoxy fiber-reinforced polymer beam specimens were manufactured and measured by a laser scanning Doppler vibrometer to extract the modal frequencies. Numerical and experimental verification results have shown that support vector machine has the best prediction performance among the three machine learning algorithms, with high prediction accuracy and only requiring a small number of samples. For predicting the interface of delamination which is a discrete variable, support vector machine classification has observed better prediction accuracy and requiring less running time than regression. This study is one of the first to prove the applicability of support vector machine for structural health monitoring of delamination damage in fiber-reinforced polymer composites and has the potential to improve the prediction capability of machine learning algorithms. Another significant outcome of the study is that the interface of delamination has been predicted accurately with support vector machine.

Autonomous Assessment of Delamination Using Scarce Raw Structural Vibration and Transfer Learning

“…Moreover, from the test confusion matrices, it can be observed that the correct prediction rate from using the pretrained model to autonomously extract features, regardless of the model used, is reasonably high and show similar behavior to that shown by the training confusion matrices. The results for the increasing size of delamination and the ease in its detectability in the current work are supported by the research finding of the articles in the references [ 50 , 53 ].…”

“…The autonomous features were processed via a quadratic support vector machine using 10-fold cross validation and a one-vs-all training strategy for the detection, quantification, and localization of delamination in laminated composites. The mathematical details of SVM for the assessment of discriminative features and classification results can be found in the references [ 50 , 51 , 52 ].…”

Classification of Textile Polymer Composites: Recent Trends and Challenges

“…He et al proposed a delamination detection approach for the detection of location, size and interfacial bonding of delamination in fiber-reinforced polymer composites. This method was based on frequency changes in multiple modes [ 97 ]. They employed a combination of different algorithms i.e., support vector machine, extreme learning machine and back propagation neural network for the detection of delamination parameters.…”

Static strain-based identification of extensive damages in thin-walled structures

“…SVMs were proposed by Vapnik 41 and are a powerful tool for classifying linearly separable data, which has recently been applied to SHM applications as well. 42 The guiding principle of SVMs used for binary classification is finding the hyperplane which divides the two groups of data in such a way that the distance to the hyperplane from the nearest point from each group is maximized. Assuming that the two classes can be labeled as either y i = 1 or y i = −1, finding this hyperplane can be expressed as the following optimization problem …”