Multi-task learning for gas-liquid two-phase flow through optical carrier microwave interferometry and autoencoder-multilayer perceptron architecture

Gas-liquid two-phase flow is commonly encountered in energy systems, but accurate characterization remains challenging. A multi-parameter joint measurement method that integrates optical fiber sensing with machine learning-based signal demodulation is presented. A sensing platform based on optical carrier microwave interferometry (OCMI) is constructed, employing cascaded intrinsic Fabry-Perot interferometers as the sensing units. The mechanical coupling mechanism between the fluid and the optical fiber is theoretically analyzed, enabling the fluid properties to be reflected through variations in the cavity length and refractive index, thereby laying a foundation for the precise characterization of two-phase flow characteristics. A neural network architecture is developed by integrating an autoencoder (AE) for feature extraction, whose compression dimension is determined by a genetic algorithm (GA), with a multilayer perceptron (MLP) with multi-task outputs, enabling joint learning for simultaneous flow pattern classification and gas-liquid flow rates prediction. Experimental results demonstrate that the proposed method accurately distinguishes between the elongated bubble flow and slug flow, with flow rates errors of 4.39 % for the gas and 6.17 % for the liquid, while improving training efficiency without compromising measurement accuracy based on the decoupled training strategy.