Table 1.
Flow pattern types and characteristics.
Fig 1.
Schematic diagram of the experimental setup: (1) Water tank, (2) Centrifugal pump, (3) Valve, (4) Turbine flowmeter, (5) Air compressor, (6) Air buffer tank, (7) Needle valve, (8) Gas flowmeter, (9) Check valve, (10) Gas-liquid mixer, (11) Observation section, (12) Gas-liquid separation water tank, (13) High-speed camera, (14) Computer.
Fig 2.
Results after preprocessing.
Fig 3.
Transitional flow pattern images.
Fig 4.
ECA module structure.
Fig 5.
Working principle of dilated convolution.
Fig 6.
Receptive fields under different dilation rates.
Fig 7.
DenseNet network architecture diagram.
Fig 8.
Dense Block structure diagram.
Fig 9.
Dense Layer structure diagram.
Fig 10.
Schematic diagram of ED-DenseNet model structure.
Fig 11.
Schematic diagram of flow pattern recognition via transfer learning.
Table 2.
Ablation study results of the ED-DenseNet model.
Fig 12.
(The loss and accuracy curves in Figs 11 and 12 remain unchanged, as the additional transitional flow dataset is only used in the test phase and is not included in training or validation).
Fig 13.
Accuracy change curve.
Fig 14.
Confusion matrix for flow pattern classification.
Table 3.
Classification performance of primary and transitional flow patterns.
Fig 15.
Deep learning model flow pattern recognition accuracy.
Table 4.
Performance and training efficiency of different models.
Fig 16.
Model accuracy curves.
Fig 17.
Model loss curves.
Table 5.
Comparison of different optimizers on ED-DenseNet.
Table 6.
Comparison of different optimizers on DenseNet121, MobileNetV3, and ConvNeXt.
Fig 18.
Two-phase flow patterns for nitrogen condensation.
Table 7.
Generalization experiment results.