Performance enhancement of convolutional neural network for ultrasonic flaw classification by adopting autoencoder

Abstract The industrial application of deep neural networks to automate the ultrasonic weldment flaw classification system has some limitations. The major problem that affects the classification performance of deep neural networks is the noise in the ultrasonic signals. So, in this article, a deep neural network, also known as autoencoder is investigated to remove noise from ultrasonic signals before feeding them to deep learning classifiers. A database was generated from specimens that were closely resembled with pipe weldment geometry having counterbore and weldment defects. Those signals were, later on, corrupted with noise to mimic industrial applicability. An autoencoder was then employed to remove noise from counterbore, planer and volumetric weldment defect signals. The classification performance of the convolutional neural network (CNN) was evaluated in three different ways. At first, without employing the autoencoder, secondly, on the denoised outputs of the autoencoder and on third CNN was trained with the noiseless signals but was tested on the denoised outputs of the autoencoder. The results demonstrate that the autoencoder can successfully remove noise from the ultrasonic weldment defect signals, which consequently improve the defect classification accuracy of the artificially intelligent deep learning classifiers.