The Recognition Algorithm of Two-Phase Flow Patterns Based on GoogLeNet+5 Coord Attention

Area of Science:

• Fluid Dynamics
• Artificial Intelligence
• Microfluidics

Background:

• Two-phase flow in microchannels is crucial in various industrial applications.
• Accurate recognition of flow patterns is essential for process control and optimization.
• Deep learning offers promising approaches for automated flow pattern identification.

Purpose of the Study:

• To enhance the accuracy of deep learning-based flow pattern recognition in microchannel two-phase flows.
• To develop a novel algorithm combining GoogLeNet with a five-layer Coord attention mechanism.
• To validate the model's performance on diverse liquid-liquid and gas-liquid flow datasets.

Main Methods:

• Utilized GoogLeNet with varied convolutional kernels for multi-scale feature extraction.
• Integrated a five-layer Coord attention mechanism to strengthen channel and spatial features.
• Trained and tested the optimized model on datasets of NaAlg-Oil, GaInSn-Water, water-soybean oil, water-lubricating oil, and argon-water flows.

Main Results:

• The combined algorithm achieved 95.09% accuracy in training and 98.12% in testing for liquid-liquid flows.
• The model demonstrated over 97% recognition accuracy for both liquid-liquid and gas-liquid flow patterns.
• The enhanced model effectively identified complex flow regimes in microchannels.

Conclusions:

• The proposed Coord attention and GoogLeNet algorithm significantly improves two-phase flow pattern recognition accuracy.
• This method offers a robust solution for automated monitoring and control in microfluidic systems.
• The validated model shows high potential for real-world applications involving diverse two-phase flows.