Open Access
Subscription or Fee Access
Unsupervised Damage Detection in Time Varying Environmental Conditions Using an Autoencoder and Support Vector Data Description
Abstract
Active guided-wave-based structural health monitoring techniques have been widely studied for inspecting large structures. The vast majority of such methods use baseline comparison as a measure to detect damage on the structure. Temperature changes between the times of baseline and test signal acquisition affects wave propagation in the structure, and changes the sensor signals. Such changes, not related to damage, confound damage detection algorithms. Baseline temperature compensation methods are often used to mitigate the effects of temperature but residual compensation errors may still cause difficulties for damage detection and characterization. An alternative is to use a machine learning algorithm and appropriate signal features to discriminate between damaged and undamaged states from sensor signals. Such methods require data acquired on the damaged structure to train the machine. In this paper, we present a novel damage detection algorithm that is only trained on data from an undamaged structure. Results of experimental data analysis demonstrating that the the method of this paper can reliably detect notches greater than 1 cm in length on an aluminum plate under varying environmental conditions with accuracies greater than 99% are presented
DOI
10.12783/shm2019/32496
10.12783/shm2019/32496