Flow pattern identification of horizontal two-phase refrigerant flow using neural networks

Abstract

In this work, electrical capacitance tomography (ECT) and neural networks were used to automatically identify two-phase flow patterns for refrigerant R-134a flowing in a horizontal tube. In laboratory experiments, high-speed images were recorded for human classification of liquid-vapor flow patterns. The corresponding permittivity data obtained from tomograms was then used to train feedforward neural networks to recognize flow patterns. An objective was to determine which subsets of data derived from tomograms could be used as input data by a neural network to classify nine liquidvapor flow patterns. Another objective was to determine which subsets of input data provide high identification success when analyzed by a neural network. Transitional flow patterns associated with common horizontal flow patterns were considered. A unique feature of the current work was the use of the vertical center of mass coordinate in pattern classification. The highest classification success rates occurred using neural network input which included the probability density functions (in time) for both spatially averaged permittivity and center of mass location in addition to the four statistical moments (in time) for spatially averaged permittivity data. The combination of these input data resulted in an average success rate of 98.1% for nine flow patterns. In addition, 99% of the experimental runs were either correctly classified or misclassified by only one flow pattern. NomenclatureA i Individual pixel area [m 2 ] A T Sum of all pixel areas [m 2 ] D Tube diameter [m] D b Bubble diameter [m] ECT Electrical capacitance tomography L Length [m] ____________________

CFD Numerical and Experimental Investigation of Two-Phase Flow Development After an Expansion Device in a Horizontal Pipe

“…The most appropriate choice is influenced by the flow regime to be simulated, and four categories identified: gas-liquid or liquid-liquid flows, gas-solid flows, liquid-solid flows and three-phase flows [21]. In the gas-liquid flow regime, there are different flow patterns such as bubbly flow, droplet flow, slug flow, stratified wavy flow and stratified flow [22]. Many numerical simulations of gas-liquid two-phase flow have used the Eulerian model in different geometries including sudden expansion [23] and flow in a horizontal tube [24] because it has proven to be more accurate than the VOF and Mixture models [23].…”

Computational and Experimental Investigation of using an Extractor in the Vertical Gravitational Flash Tank Separator

“…The Eulerian model is appropriate for the gas-liquid two-phase flow [18]. Many numerical simulations of gas-liquid two-phase flow have used the Eulerian model in different geometries including sudden expansion [19] and flow in a horizontal tube [20], [21] because it is more accurate than the Volume of Fraction (VOF) and Mixture models [19]. In the Eulerian approach, the liquid phase and vapour phase are both treated as a continuous phase by using the volume fraction for each phase [22].…”

Non-intrusive classification of gas-liquid flow regimes in an S-shaped pipeline riser using a Doppler ultrasonic sensor and deep neural networks

“…Flow regime prediction using neural networks has proved capable of objectively classifying two-phase flow regimes, clustering features for large numbers of samples when training processes are properly implemented [14,15]. Over the years, different types of neural networks have been used by many researchers for flow regime identification: namely, Feed-Forward Neural Networks (FFNN), Self-Organizing Neural Networks (SONN), and Probabilistic Neural Networks (PNNs).…”

“…Over the years, different types of neural networks have been used by many researchers for flow regime identification: namely, Feed-Forward Neural Networks (FFNN), Self-Organizing Neural Networks (SONN), and Probabilistic Neural Networks (PNNs). Also, different statistical parameters such as skewness, the mean and standard deviation of void impedance signals [15], the void impedance signals of the Probability Density Function (PDF) [16], the void fraction Cumulative Probability Density Function (CPDF) [17], the local pressure variation's Power Spectral Density (PSD) [18], and the local and global bubble chord length CPDF [19] have been used as input in the neural net systems. An integral parameter, CPDF is found to be more reliable and stable than other statistical parameters.…”