Two-phase flow regime identification in a horizontal pipe was realized based on the liquid phase velocity information and the machine learning method. Ultrasound Doppler velocimetry was employed to measure the liquid velocity. Statistical features extracted from the velocity time series data, such as mean, root mean square, and power spectral density, were used to realize real-time flow regime identification. In addition, two novel parameters—maximum velocity ratio and maximum velocity difference ratio—were proposed to identify plug and decaying slug flow. Different classification algorithms were employed to achieve a high identification accuracy. Moreover, transient flow regime identification with a fast response was realized based on two classification algorithms—long–short term memory and convolutional neural network. The results show that the accuracy of real-time flow regime identification based on a flow regime map can reach up to 93.1% using support vector machine, the maximum velocity ratio and maximum velocity difference ratio are effective in identifying plug and decaying slug flow, and transient flow regime identification under slug flow condition can be realized with an accuracy of 94% based on a convolutional neural network (CNN). Decaying slugs with long lengths confuse the CNN and are responsible for the error in identification. The results presented herein are expected to expand the available knowledge on two-phase flow regime identification.

中文翻译：

---

基于液相速度信息和机器学习的两相流态识别

基于液相速度信息和机器学习方法，实现了水平管内两相流态识别。采用超声多普勒测速仪测量液体速度。利用从速度时间序列数据中提取的均值、均方根和功率谱密度等统计特征来实现实时流态识别。此外，提出了两个新参数——最大速度比和最大速度差比——来识别塞流和衰减段塞流。采用不同的分类算法来实现高识别精度。此外，基于两种分类算法——长短期记忆和卷积神经网络，实现了具有快速响应的瞬态流态识别。结果表明，基于流态图的实时流态识别使用支持向量机的准确率可达93.1%，最大速度比和最大速度差比能有效识别塞流和衰减段塞流，并且基于卷积神经网络 (CNN) 可以实现段塞流条件下的瞬态流态识别，准确率可达 94%。长度较长的腐烂弹头会混淆 CNN，并导致识别错误。预计本文提出的结果将扩展有关两相流态识别的可用知识。最大速度比和最大速度差比能有效识别塞流和衰减段塞流，基于卷积神经网络（CNN）可以实现段塞流条件下瞬态流态识别的准确率达到94%。长度较长的腐烂弹头会混淆 CNN，并导致识别错误。预计本文提出的结果将扩展有关两相流态识别的可用知识。最大速度比和最大速度差比能有效识别塞流和衰减段塞流，基于卷积神经网络（CNN）可以实现段塞流条件下瞬态流态识别的准确率达到94%。长度较长的腐烂弹头会混淆 CNN，并导致识别错误。预计本文提出的结果将扩展有关两相流态识别的可用知识。