Ultrasonic signal enhancement for coarse grain materials by machine learning analysis

This paper aims at dealing with the dilemma of examining the existence of a defect in ultrasonic detection of coarse grain materials. In such cases, defect echoes can be drowned in a strong noise background resulting from intricate coarse grain scattering, that is, grain noise. To this end, we develop an innovative signal reconstruction methodology from polluted measurements which combines basic statistical analysis with a series of machine learning algorithms. The proposed methodology analyzes abundant information from numerous raw signals to distinguish the desired signal from grain noise, avoiding the limitation of information provided only by a single signal. The technique is achieved by collecting similar signals together through a clustering algorithm and subsequently inputting these similar signals to a denoising autoencoder to suppress the grain noise. It is successfully employed to ultrasonic signals obtained from an as-cast stainless steel specimen with coarse equiaxed grains, a stainless steel specimen with relatively homogeneous dendrite fabricated by additive manufacturing and a stainless steel weld with heterogeneous columnar grains having variation of grain sizes in various locations. The influence of material microstructure and probe frequency on denoising performance is investigated in detail. Based on this, the proposed methodology is applied to defect detection. Desired A-scan results and B-scan imaging are achieved by the proposed method, where defects are well revealed. The experimental results demonstrate the developed methodology has stable excellent performance and superior denoising capabilities for defect detection with respect to conventional techniques, especially in the case where the noise is almost the same as the desired signal.