Application of machine learning in electrical process tomography with variable frequency measurement sequences
Grzegorz Kłosowski, Monika Kulisz, Tomasz Rymarczyk, Łukasz Skowron, Paweł Olszewski, Konrad Niderla
Abstract
• Application of frequency-differentiated electric current in impedance tomography. • Use of multi-sequence LSTM network input differentiated by current frequency. • A new way of using LSTM neural networks for multisequential data. • Original algorithmic approach that improves the detection of EIT. This study presents an innovative methodology for reconstructing internal structures of industrial tank reactors using Electrical Impedance Tomography (EIT). The approach leverages a sequence of three measurement vectors, each corresponding to different frequencies (100 kHz, 50 kHz, and 10 kHz), to enhance the accuracy and robustness of EIT reconstructions. By incorporating frequency-specific information into the machine learning model, the study aims to fully exploit the valuable resistive and reactive data provided by EIT measurements, represented as complex numbers. The research involved generating synthetic data with varying levels of Gaussian noise to simulate real-world conditions. A multi-branch Long Short-Term Memory (LSTM) network was employed to process the multi-frequency measurement data. Comparative analyses were conducted between the proposed multi-frequency approach and traditional single-frequency methods using various quality indicators such as MSE, PSNR, SSIM, and ICC. The results demonstrate that the multi-frequency measurement model significantly outperforms the traditional single-frequency model regarding reconstruction quality, even under high noise conditions. This novel approach enhances the reliability and precision of EIT in monitoring industrial processes, contributing to the advancement of measurement science in industrial applications.