Abstract
Structural health monitoring (SHM) has gained significant popularity in the last decade. This growing interest, coupled with new sensing technologies, has resulted in an overwhelming amount of data in need of management and useful interpretation. Acoustic emission (AE) testing has been particularly fraught by the problem of growing data and is made the focus of this dissertation. The dissertation is divided into three studies, the first of which attempts to identify the computing resources necessary for the acquisition, management, and analysis of AE datasets. A computing framework capable of managing AE data is designed and implemented using the methods described in the first study. With a computing framework in place, the second study addresses the problem of unwanted signals in AE; these signals form a large part of most AE datasets and must be removed before a meaningful analysis can be performed. A semi-supervised data mining scheme for detecting and characterizing unwanted AE signals is proposed. The scheme is demonstrated on a synthetic dataset, and applications are presented for pencil lead-break and single-edge tension (SE(T)) datasets. This study suggests that underlying rules can be systematically derived from raw AE datasets. Finally, an artificial neural network (ANN) framework for crack-growth prediction is proposed. The ANN takes AE absolute energy and crack mouth opening displacement (CMOD) data from two SE(T) specimens as an input in order to forecast future values for each of these parameters. The predicted values are then input into linear-elastic fracture mechanics (LEFM) models in order to estimate the long-term crack evolution. The study concludes that ANNs can adequately model the complex relationships between non-destructive measurement parameters and the crack-length evolution in structural members.