Monitoramento de defeitos em dutos rígidos longos por parâmetros de emissão acústica e redes neurais

Abstract

Monitoring equipments in real time is of fundamental importance in production systems that pose risks to safety and the environment. The Acoustic Emission method stands out among nondestructive tests for the possibility of analyzing the integrity of equipments during its use. Using neural networks to analyze acoustic emission signals from cracks can be a way to monitor equipment failures in real time and without the need for expert analysis. Crack propagation can be classified into several phases, and it is important to stop production for repair before a failure of the equipment enters in the unstable propagation phase. The objective of this work is the use of supervised and unsupervised neural networks to perform an automatic division of the crack propagation phase into 40 meter pipelines and then train supervised neural networks to classify the data previously categorized by the developed automatic methodology . Two sets of data from different hydrostatic tests were used in this research. For one of them, the supervised neural networks achieved, on average, approximately 82% accuracy in the classification of the crack propagation phases for the test data set and, for the other, around 90%.