Register or Log In

International Society for Industrial Process Tomography

11th World Congress on Industrial Process Tomography

Impact of loss function selection on LSTM network training for industrial reactor monitoring via electrical impedance tomography

G. KÅ‚osowski1*, M. Kulisz1, T. Rymarczyk23

1 Faculty of Management, Lublin University of Technology, Lublin, Poland

2 Institute of Computer Science and Innovative Technologies, WSEI University, Lublin, Poland

3 Research and Development Center, Netrix S.A., Lublin, Poland

*Email: g.klosowski@pollub.pl


ABSTRACT

This study focuses on applying electrical impedance tomography (EIT) to monitor industrial tank reactors. This imaging technique is advantageous due to its non-invasive nature and ability to visualize phase changes within tanks filled with liquids effectively. A neural network with a long short-term memory (LSTM) layer was used to convert the electrical measurements obtained from EIT into corresponding values for the reconstructed 3D images. The main objective of this study was to measure the influence of different loss functions on the effectiveness of the neural network's learning process. Four different types of loss functions were used during the training process, i.e. HMSE (half mean squared error), Huber, L1Loss (L1 loss in regression problems - Mean Absolute Error) and L2Loss (L2 Loss in regression problems - Mean Square Error). The study focused primarily on the loss function, making the exact choice of neural network architecture relatively unimportant. Therefore, a relatively simple LSTM-based neural network model was used. Based on the results, it was found that the HMSE method achieved better results than the other loss functions in all cases in terms of MSE, PSNR, and ICC indicators. However, the L1loss method consistently outperformed the others regarding the SSIM indicator.


Keywords: machine learning, neural networks, electrical impedance tomography, industrial reactor, loss function

Industrial Application: food industry, pharmaceutical industry, chemical and petrochemical industry

Sign-in to access the full text