Improving the reliability of industrial reactors by using differential neural network architecture in ultrasonic tomography

Ultrasonic tomography (UST) represents a powerful non-invasive diagnostic technique for monitoring and analyzing internal processes within industrial reactors. Despite its potential, UST-based reconstructions are often challenged by the ill-posed nature of the inverse problem, limited measurements, and the presence of noise. To address these limitations, this study introduces a novel differential neural network architecture that enhances conventional deep learning models by incorporating a specialized differential layer. This layer processes two parallel input streams and operates on their residuals, thereby amplifying subtle variations in the data critical for accurate tomographic reconstructions. This study aims to empirically validate the concept of the efficacy of differentiated architecture. Reconstruction performance was evaluated using established quantitative metrics. Results demonstrate that models incorporating the differential layer consistently outperform their standard counterparts, delivering higher resolution, and superior noise robustness.