The present study proposes a new feature extraction method based on non-stationary conductivity probe signals. Two types of discriminative network models, i.e., the probabilistic neural network (PNN) and nonlinear support vector machine (SVM), are established for flow regime identification using small sample sets. The eigenvectors are composed of 16 feature quantities obtained by wavelet packet decomposition (WPD) and 8 feature quantities in the time-domain derived from the reconstructed low-frequency signals. The 8 features include maximum, minimum, standard deviation, arithmetic mean, kurtosis, peak factor, impulse factor and margin factor. The signals are normalized based on features rather than samples before flow regime identification. In the current study, WPD results show that the conductivity probe signals in two-phase flow are mostly in low frequency. The identification accuracy of the nonlinear SVM is 90.47%, which is better than 83.33% by the PNN method. This study verifies the superiority of nonlinear SVM in solving small samples and nonlinear flow regime classification problems. However, the accuracy of flow regime identification near flow regime transitional boundaries still remains questionable and needs further improvement.