Identification of defect-propagation stage in rigid pipes by means of acoustic-emission data in streaming format and neural networks

Abstract

Destructive and non-destructive tests are basic for understanding physical properties of materials. For instance, defects in a pipe under pressure can propagate until failure; therefore, a proper identification and analysis of defects is of the greatest practical importance. A wide class of non-destructive tests exploits the fact that materials under pressure emit acoustic waves, and data from acoustic emissions are analysed by a specialist. This thesis proposes the use of feedforward neural networks to automate the process of data analysis. To achieve this goal, acoustic emissions are divided into three classes according to their “pattern”: no propagation (NP), stable propagation (SP) and unstable propagation (UP). The ability of correctly classifying the acoustic emissions generated by a defect permits to classify which level of risk the system is undergoing. A classification rate higher than 85% was achieved using two distinct datasets, showing that such methods have a potential for practical applications.