Investigation of deep neural network with drop out for ultrasonic flaw classification in weldments

• Published in: Journal of mechanical science and technology Vol. 32; no. 7; pp. 3073 - 3080
• Main Authors: Munir, Nauman, Kim, Hak-Joon, Song, Sung-Jin, Kang, Sung-Sik
• Format: Journal Article
• Language: English
• Published: Seoul Korean Society of Mechanical Engineers 01-07-2018; Springer Nature B.V; 대한기계학회
• Subjects: Artificial neural networks; Classification; Control; Defects; Dynamical Systems; Engineering; Feature extraction; Industrial and Production Engineering; Mechanical Engineering; Neural networks; Pattern recognition; Regularization; Signal classification; Signal processing; Vibration; Weldments; 기계공학; Weldment flaws classification; Deep neural network; Ultrasonic testing; Drop out
• ISSN: 1738-494X; 1976-3824
• DOI: 10.1007/s12206-018-0610-1

Ultrasonic signal classification of defects in weldment, in automatic fashion, is an active area of research and many pattern recognition approaches have been developed to classify ultrasonic signals correctly. However, most of the developed algorithms depend on some statistical or signal processing techniques to extract the suitable features for them. In this work, data driven approaches are used to train the neural network for defect classification without extracting any feature from ultrasonic signals. Firstly, the performance of single hidden layer neural network was evaluated as almost all the prior works have applied it for classification then its performance was compared with deep neural network with drop out regularization. The results demonstrate that given deep neural network architecture is more robust and the network can classify defects with high accuracy without extracting any feature from ultrasonic signals.