Defect detection during pulse-echo ultrasonic testing (UT) is challenging when defects are located in a dead zone where the echoes from the defects are overshadowed by disturbances from the initial ringing signal of the UT transducer. The time-gate method is one of the most widely used approaches in UT to filter out such unwanted components, but defects in the dead zone are virtually impossible to detect using conventional methods. This paper proposes an autoencoder-based end-to-end ultrasonic testing method to detect defects within the dead zone of a transducer. The autoencoder is designed to predict the normal behavior of ultrasonic signals including disturbances, thus enabling the identification of even subtle deviations made by defects. To advance the performance of the autoencoder further with a limited amount of training data, a two-step training procedure is presented, involving training using pure normal signals measured from a defect-free specimen and re-training using pseudo-normal samples identified by the autoencoder with a smart thresholding strategy. This two-step procedure enables us to develop an adaptive autoencoder model that can be effectively employed to process the newly measured ultrasonic signals. For a demonstration of the proposed method, UT-based B-scan inspections of aluminum blocks with near-surface defects are conducted. The results suggest that the proposed method outperforms the conventional gate-based inspection approach with regard to its ability to identify the sizes and locations of near-surface defects.
Keywords: Autoencoder; Dead zone; Deep learning; Near surface; Ultrasonic testing.
Copyright © 2021 The Authors. Published by Elsevier B.V. All rights reserved.