Register or Log In

International Society for Industrial Process Tomography

12th World Congress on Industrial Process Tomography

Enhanced Image Reconstruction in Ultrasonic Process Tomography Using a Differential LSTM Architecture

M. Kulisz1 , G. Kłosowski1, T. Rymarczyk2,3*, K. Niderla3, D. Wójcik2,3

1Lublin University of Technology, Lublin, Poland
2Netrix S.A., Research & Development Centre, Lublin, Poland
3WSEI University, Lublin, Poland

*Email: tomasz@rymarczyk.com


ABSTRACT

This study investigates the effectiveness of a differential neural network architecture based on long short-term memory models for ultrasonic tomographic image reconstruction in industrial applications. The proposed model features two parallel long short-term memory branches that process identical input data and integrate a custom NegationLayer, which multiplies the output of one branch by –1. This design enables element-wise subtraction of the learned representations, allowing the network to extract residual information particularly relevant to the inverse nature of tomographic reconstruction. To evaluate the performance of the proposed architecture, a synthetic dataset of time-of-flight measurement samples was generated using finite element simulations of a water-filled reactor with multiple submerged objects. The data were used to train and test both the standard and differential models. The results show that the differential model consistently outperforms the standard architecture across all test cases. It produces sharper and more accurate reconstructions, with improved alignment to reference patterns and higher scores across all quality metrics. These findings confirm the potential of differential architectures as a generalizable enhancement for neural models in inverse imaging problems and highlight their suitability for real-time, non-invasive monitoring of industrial processes.


Keywords: Differential Neural Network, Industrial Process Tomography, Ultrasonic Tomography,

Industrial Application: Chemical Processing, Petrochemical Industry, General

Sign-in to access the full text