The metering of gas-liquid flows is difficult due to the non-linear relationship between flow regimes and fluid properties, flow orientation, channel geometry, etc. In fact, a majority of commercial multiphase flow meters have a low accuracy, limited range of operation or require a physical separation of the phases. We introduce the inference of gas-liquid flowrates using a neural network model that is trained by wire-mesh sensor (WMS) experimental data. The WMS is an experimental tool that records high-resolution high-frequency 3D void fraction distributions in gas-liquid flows. The experimental database utilized spans over two orders of superficial velocity magnitude and multiple flow regimes for a vertical small-diameter pipe. Our findings indicate that a single network can provide accurate and precise inference with below a 7.5% MAP error across all flow regimes. The best performing networks have a combination of a 3D-Convolution head, and an LSTM tail. The finding indicates that the spatiotemporal features observed in gas-liquid flows can be systematically decomposed and used for inferring phase-wise flowrate. Our method does not involve any complex pre-processing of the void fraction matrices, resulting in an evaluation time that is negligible when contrasted to the input time-span. The efficiency of the model manifests in a response time two orders of magnitude lower than the current state-of-the-art.

中文翻译：

---

使用神经网络推断气液流量

由于流态与流体性质、流向、通道几何形状等之间的非线性关系，气液流量的计量很困难。 事实上，大多数商用多相流量计精度低，操作范围有限或需要物理分离相。我们使用由线网传感器 (WMS) 实验数据训练的神经网络模型来介绍气液流量的推断。WMS 是一种实验工具，可记录气液流中的高分辨率高频 3D 空隙率分布。使用的实验数据库跨越两个数量级的表观速度大小和垂直小直径管道的多种流态。我们的研究结果表明，单个网络可以提供低于 7 的准确和精确的推理。所有流态的 5% MAP 误差。性能最好的网络结合了 3D 卷积头部和 LSTM 尾部。这一发现表明，在气液流动中观察到的时空特征可以被系统地分解并用于推断相向流速。我们的方法不涉及对空隙率矩阵进行任何复杂的预处理，因此与输入时间跨度相比，评估时间可以忽略不计。该模型的效率表现在响应时间比当前最先进的技术低两个数量级。这一发现表明，在气液流动中观察到的时空特征可以被系统地分解并用于推断相向流速。我们的方法不涉及对空隙率矩阵进行任何复杂的预处理，因此与输入时间跨度相比，评估时间可以忽略不计。该模型的效率表现在响应时间比当前最先进的技术低两个数量级。这一发现表明，在气液流动中观察到的时空特征可以被系统地分解并用于推断相向流速。我们的方法不涉及对空隙率矩阵进行任何复杂的预处理，因此与输入时间跨度相比，评估时间可以忽略不计。该模型的效率表现在响应时间比当前最先进的技术低两个数量级。