Gas-Liquid Two-Phase Stratified Flow Interface Reconstruction With Sparse Batch Normalization Convolutional Neural Network

Two-phase stratified flow is ubiquitous in industrial processes, monitoring its phase interface is important for improving the safety and efficiency of process. Electrical Resistance Tomography is a promising non-intrusive visualization technique for monitoring two-phase flow. However, some electrodes could lose their contact with the liquid in stratified flow, aggravating the under-determined image reconstruction solution and resulting in low-quality reconstructed image in traditional methods. In this paper, a sparse batch normalization convolutional neural network (SBN-CNN) method is proposed for accurate and rapid gas-liquid interface reconstruction. Up-sampling and normalization were used for preprocessing of measurements. A novel network structure including convolution, pooling and batch normalization layers was designed to extract features from sparse measurements and the images were reconstructed by deep fully connected layers. After batch-based training, the proposed SBN-CNN achieves a superior performance to the-state-of-art methods in terms of convergence rate, imaging accuracy, noise resistance and generalization ability, through experimental validation.