Segmented analysis of time-of-flight diffraction ultrasound for flaw detection in welded steel plates using extreme learning machines.

This work investigates the application of extreme learning machine, a fast training neural network model, for an ultrasound nondestructive evaluation decision support system. A novel segmented analysis of time-of-flight diffraction ultrasound signals is proposed in order to produce high flaw detection efficiency and low computational requirements, making it possible to be used in embedded applications. The frequency contents of TOFD signals temporal segments, estimated using the discrete Fourier transform, were used to feed the classification system. The test objects consisted of a set of SAE 1020 welded carbon steel plates, in which occur four types of defects. The obtained experimental results indicate that the proposed method is able to combine high accuracy, fast training and full exploration of the TOFD signal information.