An artefact suppression framework in ultrasonic phased array tomography for metal welds

Abstract

Artefacts that resemble defects can lead to inaccurate quality assessments in metal welds. Reducing residual artefacts while preserving the integrity of defect signals presents a significant challenge. To address this, a framework for artefact suppression based on a 3D denoising autoencoder is proposed, comprising two key developments: (1) The artefact rather than defects is reconstructed for the gradient vanishing issue in 3D denoising autoencoders when processing high-dimensional data, which results in excessive residual artefacts. Subsequently, a network framework based on multilayer bidirectional long short-term memory is proposed for tomographic image processing, enhancing reconstruction accuracy. (2) A cross-modal temporal-spatial attention module is developed to assist 3D autoencoders in identifying latent patterns of artefact. Particularly, their periodic differences are captured, regarded as the intrinsic distinction between defects and artefacts in tomographic detection. Experimental results demonstrate that the proposed framework effectively suppresses the side-lobe artefact and those caused by multiple reflections and mode conversions. While this approach is primarily designed for metal welds, it also shows promise for artefact suppression in metal castings and potential applications in medical imaging.