The fusion of deep learning and acoustic emission response methods for identifying solid particles in annular multiphase flows

Annular flows carrying sand are common flow patterns in high-production gas-bearing wells. The real-time monitoring of sand particle information in the annular flows of wellheads is critical for efficient commercial production. In this study, an experiment was designed to monitor sand production in annular multiphase flows, and methods were proposed to identify sand using empirical mode decomposition (EMD), the Hilbert–Huang transform (HHT), statistical analysis, and deep learning methods. Corresponding sand migration behaviours near pipe walls were observed by acoustic emission (AE); the behaviours included sand carried by the gas core (IMF1), forward liquid film (IMF2) and reverse liquid film (IMF3). Furthermore, relationships between the AE response and gas superficial velocity (14–18 m/s), liquid superficial velocity (0.00366–0.01351 m/s), and mean particle size (150–380 μm) were proposed, and the AE responses of different sand migration patterns were verified. Finally, CNN, LSTM, and CNN-LSTM deep learning models were constructed to identify particle sizes based on the optimized sand-carrying information. The accuracy of the CNN-LSTM model was 6.44% and 18.9% higher than that of the CNN model and the LSTM model, respectively, which significantly improved the accuracy of particle size identification in annular particle flows. Therefore, this research provides an efficient method for the intelligent identification of sand in multiphase annular flows.