Application of Attention Mechanism Models in the Identification of Oil–Water Two-Phase Flow Patterns

Accurate identification of oil–water two-phase flow patterns is essential for ensuring the safety and operational efficiency of oil and gas extraction systems. While traditional methods using empirical models and sensor technologies have provided basic insights, they often struggle to capture the nonlinear features of complex operational conditions. To address the challenge of data scarcity commonly found in experimental settings, this study employs a data augmentation strategy that combines the Synthetic Minority Over-sampling Technique (SMOTE) with Gaussian noise injection, effectively expanding the feature space from 60 original experimental nodes. Next, a physics-constrained attention mechanism model was developed that incorporates a physical constraint matrix to effectively mask irrelevant feature interactions. Experimental results show that while the standard attention model (83.88%) and the baseline BP neural network (84.25%) have limitations in generalizing to complex regimes, the proposed physics-constrained model achieves a peak test accuracy of 96.62%. Importantly, the model demonstrates exceptional robustness in identifying complex transition regions—specifically Dispersed Oil-in-Water (DO/W) flows—where it improved recall rates by about 24.6% compared to baselines. Additionally, visualization of attention scores confirms that the distribution of attention weights aligns closely with fluid-dynamic mechanisms—favoring inclination for stratified flows and flow rate for turbulence-dominated dispersions—thus validating the model’s interpretability. This research offers a novel, interpretable approach for modeling dynamic feature interactions in multiphase flows and provides valuable insights for intelligent oilfield development.