Lightweight real-time network for multiphase flow patterns identification based on upward inclined pipeline pressure data

Abstract

Within the realm of oil and gas pipelines, multiphase flow phenomena are universal existence. To mitigate the various potential hazards caused by multiphase flow, accurate recognition of flow patterns is crucial. Current research on flow pattern identification typically employs data such as images and Doppler ultrasound signals. These types of data have obvious drawbacks: image data have poor universality and are only applicable to transparent pipelines, while Doppler ultrasound signal acquisition is complex. Therefore, this article uses upward inclined pipeline pressure data that are more generally applicable and easier to collect. Machine learning methods represented by deep learning have strong feature extraction capabilities. But the deep learning models proposed in current research have poor real-time performance and large parameter size, making them unsuitable for deployment on resource-constrained devices which are widely used in industrial sites. In response to the above problems, this paper proposes a novel lightweight YOLOv8_1D network model based on the YOLOv8 lightweight model. This model has fewer parameters, higher real-time performance, and utilizes universally applicable one-dimensional pressure data for training, aiming for precise identification of various flow patterns. During the data preprocessing stage, z-score standardization is applied to prevent gradient explosion and enhance model performance. To further improve model performance, the Empirical Wavelet Transform, an adaptive filter banks generation method, is introduced during the data preprocessing phase. The final experimental results show that EWT can effectively extract features from pressure data, enabling the YOLOv8_1D model to achieve the accuracy of 97.37 % which is higher than other contrast models and meet real-time requirements (63 samples/second). The source code of this paper is publicly available at https://github.com/JiuYu77/flow_identification.