Classification Functions and Optimization Algorithms for Debonding Detection in Adhesively Bonded Lap-joints through Ultrasonic Guided Waves

This study compares three classification functions and optimization algorithms for debonding detection in an adhesively bonded aluminum lap-joint. This comparison is in terms of accuracy of the prediction, and the time required to execute the algorithms. For this purpose, a lap-joint specimen with two different sizes of artificial debonding is created. The data acquired from applying ultrasonic guided waves through angle beam transducers and by a B-scan along the overlap region several times. Using signal processing techniques, different features are extracted from received signals and binary classifications are carried out on these features. Logistic loss, smoothed hinge loss and regularized least square loss are three different functions used for classifications. To find the optimal parameters of the two former loss functions, three optimization algorithms are used including stochastic gradient descent with momentum (SGDM), stochastic variance-reduced gradient (SVRG) and fast iterative shrinkage-thresholding algorithm (FISTA).