Hybrid deep learning approach for invasive identification of gas-liquid two-phase flow patterns in horizontal pipes

Horizontal pipelines are widely used in petroleum, natural gas, and chemical industries, where gas-liquid two-phase flow is a common phenomenon. Certain inherently unstable dynamic flow patterns generate pressure surges, vibrations, and safety-critical phenomena, directly impacting pipeline reliability. Precise flow identification is vital for operational safety and efficiency. This paper proposes a hybrid deep learning method for intelligent, non-intrusive recognition of horizontal pipeline flow patterns. Our framework extracts multimodal features encompassing time and frequency domains from conductivity signals via dynamic hard thresholding and Fast Fourier Transform (FFT), followed by Principal Component Analysis (PCA) dimensionality reduction. To address severe turbulent fluctuations and noise under high gas-liquid velocities, a novel architecture integrates multi-scale enhancement using wavelet-based convolution and attention mechanisms. A transformer-based multi-branch structure classifies features into four flow patterns. Bayesian optimization tunes hyperparameters, and ablation studies demonstrate each module's effectiveness in enhancing classification performance. Comparative evaluation against six state-of-the-art models confirms superior accuracy and robustness. The method shows strong potential for practical deployment in intelligent multiphase flow monitoring and non-destructive evaluation.