Detecção de anomalia aplicado ao monitoramento da integridade estrutural em um sistema de tubulações.

Abstract

In this work, two semi-supervised algorithms were compared in relation to anomaly detection in electromechanical Impedance-based Structural Health Monitoring (ISHM) in a piping system. The experiment consisted of building a low-cost piping system with a Pb-lead Zirconate Titanate (PZT) patch attached to the main surface using a clamp, modeled and printed on a 3D printer, capable of wrapping around the pipe to improve the interaction between the PZT pellet and the structure. To simulate failure situations, three types were considered: clogging using an epoxy putty, damage caused by scratches on the inside of the pipe and minor fouling caused by a filter dimensioned on the inside of the pipe. The techniques chosen to classify the natural state and damage were Local Outlier Factor (LOF) and Support Vector Machine (SVM), each configured in a semi-supervised way. Using the evaluation metrics, confusion matrix and ROC curve, it was possible to see that the two methods were able to correctly identify all the classes in the dataset.