Gas-liquid vertical pipe flow patterns convolutional neural network classification using experimental advanced wire mesh sensor images

Abstract

Flow pattern identification for simultaneous gas-liquid flow in pipes is a central problem in the petroleum industry. However, there are many conventional methods used for flow patterns identification that come with some form of restrictions limited to specific operating conditions. Although previous studies made efforts to make flow patterns identification free from biases (subjectivity), objective predictions are yet to be fully explored. For the first time, this study employs a 4-layer convolutional neural network (CNN) architecture and threshold segmentation algorithm to classify industrial working fluids of air/silicone oil flow patterns in a vertical pipe. Also, an advanced wire mesh sensor (WMS) instrumentation was used to obtain the cross-sectional frame images from a series of 52 experimental runs. This study made an attempt to use the sequential cross-sectional frames original data obtained from the experimental WMS because it is the crossing wires of the WMS that interact with the flow field and also capture the main flow pattern transitional zones. The obtained experimental testing results indicate that the supervised CNN model has an accuracy of 99.90% better than the seven-benchmarked supervised machine learning classification models used in recognizing bubbly, slug and churn flow patterns. The CNN model experimental test accuracy also outwits other related flow patterns identified in the literature when compared, thereby affirming the model's robustness. Since the WMS provide high spatial and temporal resolutions data, the average pixel intensity values from the segmented images were used as the flow indicator in identifying the transition zones within the main flow patterns. Furthermore, the Locally Interpretable Model-agnostic Explanation (LIME) algorithm for the first time was used to explain and interpret features in the WMS flow images that were contributing to the overall CNN classification scores. Finally, the CNN trained model was able to classify the main flow patterns and segmented transitional zones with certainty and confidence without subjectivity which is inherent in the direct conventional methods.