Machine learning enabled damage classification in composite laminated beams using mode conversion quantification

摘要

Damage identification and classification in composites help in determining the remaining useful life of structures and components. Structural health monitoring systems utilizing guided waves have gained importance especially with the capability of providing prediction of the remaining useful life. Determination of damage types and severity often involves extensive signal processing techniques that often require high sampling rate, which is challenging due to the hardware limitations. It also includes the associated numerical error, which leads to a high false alarm rate or low detection capabilities. We propose a model-assisted machine learning algorithm to identify damage types and severity based on mode converted wave strength. Machine learning techniques are employed to develop classification models assisted by the finite element simulation models. Finite element simulation models provide the training data for various cases of damage and severity involving common types of damages in composites like delamination, transverse crack and material degradation. Damage classification models based on mode conversion strength versus frequency curves of the participating four wave modes: fundamental axial, fundamental flexural, higher-order shear and higher-order thickness contraction have been developed with experimental validation. For damage classification, several statistical machine learning models have been studied. Classification models like Naive Bayes, Support Vector Machine (SVM), Random Forest (RF), k nearest neighbor (KNN), and linear discriminant analysis (LDA) classifiers are implemented on the dataset by extracting different meaningful features from the signals obtained from COMSOL. The dataset has been randomly divided into training sets and test sets and the classification models have been implemented on both the original datasets and the reduced datasets after applying principal component analysis (PCA). Results show that the classification models worked better on the original dataset and the error rate is minimum for the SVM and Random Forest among all the models. The optimum model has been established by comparing the results obtained using the existing models.