AI-driven residual strength diagnostics of composites using their electrical behavior under low-stress cyclic loading

Abstract

A novel non-destructive testing (NDT) method was developed to predict the residual strength of composites, with unknown histories of fatigue damage, using their electrical behavior during a low stress cyclic loading test. Ninety-five samples, representing a wide range of fatigue damage levels, were prepared and subjected to a low-stress cyclic loading test, as a diagnostic test, while their electrical behavior was monitored. The samples then underwent quasi-static loading until failure to measure their corresponding residual strengths. To establish a relationship between the electrical behavior of samples during the diagnostic test and their corresponding residual strengths, various machine learning techniques were implemented. K-nearest neighbor (KNN), Decision Tree (DT), Random Forest, Extreme Gradient Boosting, Support Vector Regressor (SVR), and Feedforward Artificial Neural Networks were employed in two different approaches: as standalone predictors, and in an ensemble learning approach. The analysis demonstrated that a KNN meta-model, incorporating DT, SVR, and KNN as base models, in an ensemble framework, achieved the best performance, with a mean absolute percentage error (MAPE) of 5.7 % in predicting residual strength. This significant performance underscores the potential of our low-stress diagnostic test for predicting the residual strength of composites, even when the fatigue damage history is unknown.