U-NET autoencoder for amplitude recovery on binarized Full Matrix Captures

Abstract

Full Matrix Capture (FMC) and the Total Focusing Method (TFM) are instrumental techniques in ultrasonic nondestructive testing (NDT) in industries such as aerospace, oil and gas, and manufacturing, and allow efficient defect detection by capturing all possible transmitter–receiver pairs and generating highly resolved images on a predefined pixel grid. The use of dense linear or matrix probes presents significant challenges in data storage and transfer but also in the complexity of the acquisition system’s electronics. In this context, binary acquisition steps in as an attractive alternative for simplifying acquisition equipment and reducing data size. However, binary formats carry the drawback of amplitude information loss. To address this, the present study explores the application of a U-NET autoencoder neural network to reconstruct amplitude data from binarized FMC signals. The autoencoder’s U-NET architecture is particularly suited for this task due to its effectiveness with limited datasets, a common issue in NDT. Finite element simulations were used to generate training and validation datasets. Experimental tests were then conducted on steel samples containing various defects, such as Electrical Discharge Machining (EDM) cracks, side-drilled holes (SDH), and a realistic fatigue crack in a steel bar. The reconstructed FMC data were evaluated using TFM images and Structural Similarity Index Measure (SSIM), showing that the neural network accurately reconstructed FMCs. Notwithstanding the presence of minor amplitude errors, the spatial positioning of defects remained precise, demonstrating the method’s viability for practical NDT applications.