Identification of two-phase flow patterns based on capacitance data of electrical capacitance tomography with semi-supervised generative adversarial network

Currently, the flow pattern identification algorithms based on ECT (electrical capacitance tomography) technology have low identification accuracy for complex flow patterns and require a large amount of label data for learning. A novel flow pattern identification method based on a semi-supervised generative adversarial network (SGAN) with capacitance data of ECT is proposed. First, the principles of the ECT technique and general GAN are briefly described, and the model parameters, loss function, and training process of the SGAN are explained in detail. Second, a capacitance data sample set of 11 400 random flow patterns is constructed by co-simulations of COMSOL and MATLAB, and then, the SGAN and BP (back propagation) and SVM (support vector machine) network models are trained and validated by the training set. Finally, static experiments are conducted on the self-developed ECT system, and the identification results of different algorithms are compared and analyzed by modifying the label sample size of the training set. The experimental results show that SGAN maintains a higher average identification accuracy under the training condition where the number of label samples of SGAN is ten times smaller than that of the other two algorithms.