Deep Learning‐Based Gas Holdup Detection in Plunger Lift Systems: A VGG‐Unet Approach

ABSTRACT In the realm of oil and gas engineering, plunger lift systems are vital for mitigating liquid loading in gas wells. The gas‐liquid ratio between the plunger and the inner wall clearance of the oil pipe is a key parameter for accurately analysing the motion state of the plunger. Traditional image processing techniques have weak generalisation ability when detecting this parameter and are easily affected by environmental factors. Therefore, a deep learning‐based approach was introduced in this study to detect gas holdup in the annulus of plunger lift systems upon the VGG‐Unet architecture. By replacing the encoder module of U‐Net with the first 13 convolutional layers of VGG16 and incorporating pre‐trained weights, our model achieved superior segmentation performance with a mean Intersection over Union (mIoU) of 93.19%, which surpassed Mobile‐UNet by 0.28%, Resnet‐UNet by 4.81%, and Segformer by 5.67%. Additionally, the segmented images were processed in MATLAB to calculate bubble area and gas holdup parameters, revealing strong correlations (R 2 = 0.97 for bubble area, R 2 = 0.94 for gas holdup) with manual validation data. This method provides immediate applicability for field‐based quality control and real‐time gas holdup monitoring in operational plunger systems.