A hybrid deep learning model for ECT image reconstruction of cryogenic fluids

Compared to general room-temperature fluids, the characteristics of cryogenic fluids, as well as the complexity of the cryogenic environment, pose greater challenges for reconstruction algorithms for Electrical Capacitance Tomography (ECT). Based on deep learning, a hybrid model is proposed for cryogenic fluid ECT image recon-struction in this study. The multi-head self-attention mechanism is employed to initially establish the mapping of capacitance to the image, and then an improved U-net-like convolution neural network is presented to perform deep feature extraction and image reconstruction. The ConvNeXt block is adopted for multi-level feature extraction, and a separate downsampling layer is used to replace the pooling layer. A dataset covering a variety of two-phase typical flow patterns and irregular flow patterns is built for training. A capacitance vector and an image of phase distribution are included in each sample. Extensive numerical experiments are carried out on the trained model. The results show that the model can accurately predict phase distribution and produce a clear interface. Finally, the model was successfully applied in cryogenic experiment to obtain the phase distribution image of liquid nitrogen stratified flow.

Automated summary

In this study, a hybrid model based on deep learning is proposed for image reconstruction of cryogenic fluid electrical capacitance tomography (ECT). The model utilizes a multi-head self-attention mechanism to establish the mapping between capacitance and image, and an improved U-Net-like convolutional neural network for deep feature extraction and image reconstruction. Experimental results demonstrate that the model accurately predicts phase distribution and produces clear interfaces.