Flaw Detection of Industrial Parts By Using of Artificial Neural Network

Present work evaluates the application of artificial neural networks for weld beads pattern recognition using pulse-echo ultrasonic techniques. In this study pattern classifier is an MLP artificial neural network implemented in MATLAB. The ultrasonic signals, acquired from pulse-echo, are separately introduced to the neural network with and without preprocessing. The preprocessing objective is smoothing the signal that improves the classification. Five conditions of weld beads are evaluated: lack of fusion (LOF), lack of penetration (LOP), excess penetration (EX.P), slag inclusion (SL) and non-defect (ND). The defects are intentionally inserted in a weld bead of AISI 1020 steel plates of 20 mm thickness and confirmed using radiographic test. The results obtained show that it is possible to classify ultrasonic signals of weld joints by the pulse-echo techniques using artificial neural networks and the success rate of 78.82% can be achieved. The preprocessing role is significant in this respect.