Automatic classification of weld defects using simulated data and an MLP neural network

Abstract

An effective weld defect classification algorithm has been developed using a large database of simulated defects. Twenty-five shape descriptors used for the classification were studied and an optimal set of nine descriptors with highest discriminative capability was selected using a statistical approach. A multi-layer perceptron (MLP) neural network was trained using shape parameters extracted from the simulated images of weld defects. By testing on 60 unknown simulated defects, the optimised set of nine shape descriptors gave the highest classification accuracy of 100%. Defect classification on 49 real defects from digitised radiographs produced maximum overall classification accuracy of 97.96%.