基于深度学习的气液固三相反应器图像分析方法及应用

气液固三相反应器中复杂的颗粒背景给流动参数的图像检测带来巨大挑战。发展了一种基于深度学习的气液固三相反应器图像分析方法，包括采集图像、制作训练集、建立图像识别模型和提取流动参数四个步骤。采用全卷积神经网络，在学习率为0.005、训练次数为2000次、训练集大小超过400张图像的条件下，图像识别误差小于5%。利用该方法可以获取三相反应器中局部相含率（气相分数和液相分数）及其空间分布、时间序列等信息，再采用时域分析、频率分析、小波分析等分析方法提取二次参数，可用于流型识别、压降预测和气液分布的均匀性判别等。将该方法用于涓流床中流动参数的检测，结果表明，局部液相分数的时间序列信号及其功率谱、概率密度分布均能清晰地区分涓流、脉冲流、鼓泡流等典型流型；时间序列信号的均值、标准差、极差和概率密度分布曲线半峰宽等特征参数可用于确定流型边界；平均液相分数可以用于预测涓流区的压降，计算值与实验测量值的平均相对偏差约为15%；液相分数空间分布的标准差可用于表征涓流床中不同流型的气液分布均匀性。该方法为气液固三相反应器的研究提供了新的工具。

Development and application of image analysis method based on deep-learning in gas-liquid-solid three-phase reactor

In gas-liquid-solid three-phase reactors, the great challenges for image detection of flow parameters have been caused by the complex background with particles. An image analysis method of gas-liquid-solid three-phase reactor based on deep-learning is developed, which includes four steps: collecting image, making training set, establishing image recognition model and extracting flow parameters. The full convolutional neural network algorithm was chosen to build the identification model, and when the learning rate was set as 0.005, the number of epochs was set as 2000 and the training set size was set as 400 images in this model, the identification error of the model is less than 5%. The flow parameters such as local phase fractions (gas fraction and liquid fraction) and its spatial distribution, time series signals were obtained by the method. Then time domain analysis, frequency analysis, wavelet analysis and other analysis methods were used to extract the quadratic parameters from the spatial distribution and time series signals. And the quadratic parameters could be used to identify flow regimes, predict pressure drop and detect the uniformity of gas-liquid distribution. Furthermore, the method was applied to the detection of flow parameters in a trickle bed. The results showed that the time-domain variation, power spectrum density and probability density distribution of liquid fractions were used to distinguish the trickle flow, pulse flow and bubble flow. Meanwhile, the flow regime boundary was detected by characteristic parameters such as mean, standard deviation, range and half-width of probability density distribution curve of liquid fractions. The pressure drop in the trickle flow regime was predicted by the average liquid fraction, and the average relative deviation between the theoretical and experimental measurements was about 15%. In addition, the standard deviation of liquid fractions in the spatial distribution was used to detect the uniformity of gas-liquid distribution of different flow regimes in the trickle bed. This method provides a new tool for the research of gas-liquid-solid three-phase reactors.