Deep learning meta architecture to detect spatially coherent coarse grain regions in ultrasonic data

This study introduces a novel methodology for detecting coarse grain volumina in ultrasonic data. The approach is based on the premise that the grain structure in metals produces ultrasonic grain noise, which can then be utilized to extract information about the grain structure. To achieve the detection of coarse grain fully embedded in finer grain material, three distinct deep learning models based on convolutional neural networks, recurrent neural networks and transformers, all derived from a shared meta architecture, were implemented and trained using ultrasonic Full-A-scan data. The proposed method was applied to a substantial dataset of synthetically manufactured coarse grain volumina, collected using a conventional single probe ultrasonic immersion system. Additionally, two baseline approaches were implemented and compared against the deep learning methods for experimental evaluation. The uncertainty of the deep learning approach was quantified using the model agnostic Monte Carlo Dropout technique, enabling the classification of predictions into high- and low-confidence categories. This research highlights the potential of deep learning in enhancing safety-critical systems and emphasizes the necessity for explainability in such domains. • Specimen manufactured from material derived from a real turbine disk. • Conventional ultrasonic testing is unable to detect artificial inclusions. • Meta-architecture from which multiple deep learning model instances were derived. • Two trained instances outperformed best-performing baseline. • Monte-Carlo Dropout was employed to enhance the output of a deep learning model.