基于卷积神经网络的撞击流反应器浓度场混合特性

doi:10.16085/j.issn.1000-6613.2022-0744

摘要/Abstract

摘要：

基于PLIF测试技术结合卷积神经网络技术提出混合性能预测方法，分析水平对置撞击流反应器浓度场混合特性，能准确预测其内部浓度场的混合均匀度及混合时间。基于卷积神经网络构建了混合性能预测模型，利用水平对置撞击流反应器浓度场实验数据对构建的模型进行有监督地训练并进行预测，预测结果显示对混合均匀度的预测准确率达95%，计算效率提高了99.99%。为更好地理解混合性能预测模型对混合均匀度的预测机理，本文对其卷积层输出进行可视化处理，通过功率谱分析卷积核的响应给出了撞击流反应器浓度场特征提取的物理解释。最后利用预测模型搭建混合均匀度快速获取系统并应用于撞击流混合特性研究。所提出的基于卷积神经网络的预测模型可以有效分析水平对置撞击流反应器的混合特性，预测模型可靠、适用范围广，为深度学习算法应用于撞击流领域提供了方案经验。

关键词: 撞击流反应器, 卷积神经网络, 混合, 浓度场, 预测

Abstract:

In this paper, a mixing performance prediction method based on PLIF test technology and convolutional neural network technology was proposed to analyze the mixing characteristics of the flow field in the impingement flow reactor, which could accurately predict the mixing uniformity and mixing time of the concentration field in the impingement flow reactor. Mixed performance prediction model was constructed based on CNN, using the impinging stream reactor concentration field experimental data for the construction of the model of supervised training and forecasting. The prediction results showed that the mixing uniformity of prediction accuracy was 95%, and the calculation efficiency increased by 99.99%. In order to better understand the prediction mechanism of the mixing performance prediction model on the mixing uniformity, the output of the convolution layer was visualized, and the physical interpretation of the concentration field feature extraction of impingent flow reactor was given by analyzing the response of the convolution kernel through power spectrum analysis. Finally, the prediction model was used to build a fast acquisition system of mixing uniformity and apply it to study the mixing characteristics of impingement flow. The proposed prediction model based on CNN could effectively analyze the mixing characteristics of impingement flow reactor, and the prediction model was reliable and applicable to a wide range of applications, providing scheme experience for the application of deep learning algorithm in the field of impingement flow.

Key words: impingement flow reactor, convolutional neural network, mixing, concentration field, prediction

中图分类号:

TP389.1

张建伟, 许蕊, 张忠闯, 董鑫, 冯颖. 基于卷积神经网络的撞击流反应器浓度场混合特性[J]. 化工进展, 2023, 42(2): 658-668.

ZHANG Jianwei, XU Rui, ZHANG Zhongchuang, DONG Xin, FENG Ying. Mixing characteristics of concentration field in impingement flow reactor based on convolutional neural network[J]. Chemical Industry and Engineering Progress, 2023, 42(2): 658-668.

图/表 14

参考文献 29

1	胡立舜, 王兴军, 于广锁, 等. 撞击流气液吸收特性及用于甲醇合成应用研究[J]. 高校化学工程学报, 2009, 23(3): 404-410.
	HU Lishun, WANG Xingjun, YU Guangsuo, et al. Study on absorption characteristics of impinging stream and its application in methanol synthesis[J]. Journal of Chemical Engineering of Chinese Universities, 2009, 23(3): 404-410.
2	白净, 李凌乾, 刘风莉, 等. 机械扰动强化气体水合物快速生成研究进展[J]. 化工进展, 2018, 37(1): 60-67.
	BAI Jing, LI Lingqian, LIU Fengli, et al. Progress of rapid formation of gas hydrate by mechanical disturbance[J]. Chemical Industry and Engineering Progress, 2018, 37(1): 60-67.
3	盛磊, 李培钰, 牛宇超, 等. 微尺度过程强化的结晶颗粒制备研究进展[J]. 化工学报, 2021, 72(1): 143-157.
	SHENG Lei, LI Peiyu, NIU Yuchao, et al. Progresses in the preparation of micro-scale process-enhanced crystalline particles[J]. CIESC Journal, 2021, 72(1): 143-157.
4	王苏炜, 肖磊, 胡玉冰, 等. 纳米单质含能材料制备及其应用现状[J]. 火炸药学报, 2021, 44(6): 705-734.
	WANG Suwei, XIAO Lei, HU Yubing, et al. A review on the preparation and application of nano-energetic materials[J]. Chinese Journal of Explosives & Propellants, 2021, 44(6): 705-734.
5	WANG Xianghu, WANG Xuehua， MENG Alan，et al. Rapid, continuous, large-scale synthesis of ZnO/Ag hybrid nanoparticles via one-step impinging stream route for efficient photocatalytic and anti-algal applications[J]. Materials Today Communications, 2022, 30: 103121.
6	SRISAMRAN Charinrat, DEVAHASTIN Sakamon. Numerical simulation of flow and mixing behavior of impinging streams of shear-thinning fluids[J]. Chemical Engineering Science, 2006, 61(15): 4884-4892.
7	LI Chang, YUE Song, LI Ming. Numerical simulation of the drying characteristics of a high-moisture particle in a dynamic asymmetric impinging stream reactor[J]. Chemical Papers, 2022, 76(1): 41-56.
8	罗燕, 周剑秋, 郭钊, 等. 液体连续相撞击流反应器中流体混合时间的数值模拟[J]. 化工学报, 2015, 66(6): 2049-2054.
	LUO Yan, ZHOU Jianqiu, GUO Zhao, et al. Numerical simulation of fluid mixing time in liquid impinging streams reactor[J]. CIESC Journal, 2015, 66(6): 2049-2054.
9	LIANG Lina, LIU Youzhi, ZHANG Qiaoling,et al. Liquid nitrogen freezing method for exploring the mixing characteristics of liquid/liquid heterogeneous in outer packing cavity of Impinging Stream-Rotating Packed Bed[J]. Chemical Engineering and Processing-Process Intensification, 2020, 150: 107906.
10	HU Hui, CHEN Zhiming, JIAO Zhen. Characterization of micro-mixing in a novel impinging streams reactor[J]. Frontiers of Chemical Engineering in China, 2009, 3(1): 58-64.
11	张航. 撞击流反应器内混合及界面反应机理的实验研究与数值模拟[D]. 上海: 华东理工大学, 2021.
	ZHANG Hang. Experimental and numerical study on mechanism of mixing and interfacial reaction in impinging jets reactors[D]. Shanghai: East China University of Science and Technology, 2021.
12	张航, 张巍, 李伟锋, 等. T型反应器内流动、混合及界面反应特征[J]. 化工学报, 2021, 72(10): 5064-5073.
	ZHANG Hang, ZHANG Wei, LI Weifeng, et al. Characteristics of flow, mixing and interfacial reaction in T-jet reactor[J]. CIESC Journal, 2021, 72(10): 5064-5073.
13	张建伟, 闫宇航, 沙新力, 等. 撞击流强化混合特性及用于制备超细粉体研究进展[J]. 化工进展, 2020, 39(3): 824-833.
	ZHANG Jianwei, YAN Yuhang, SHA Xinli, et al. Advances on impinging stream intensification mixing mechanism for preparing ultrafine powders[J]. Chemical Industry and Engineering Progress, 2020, 39(3): 824-833.
14	张建伟, 张学良, 冯颖, 等. 浸没对置撞击流的液相混合行为研究[J]. 高校化学工程学报, 2016, 30(2): 311-317.
	ZHANG Jianwei, ZHANG Xueliang, FENG Ying, et al. Study on liquid mixing behavior within a submerged impinging stream mixer[J]. Journal of Chemical Engineering of Chinese Universities, 2016, 30(2): 311-317.
15	STROFER Carlos Michelen, WU Jinlong, XIAO Heng, et al. Data-driven, physics-based feature extraction from fluid flow fields using convolutional neural networks[J]. Communications in Computational Physics, 2019, 25(3): 625-650.
16	SHEN Chong, ZHENG Qibo, SHANG Minjing, et al. Using deep learning to recognize liquid-liquid flow patterns in microchannels[J]. AIChE Journal, 2020, 66(8): 625-650.
17	DU Meng, YIN Hongyi, CHEN Xiaoyan, et al. Oil-in-water two-phase flow pattern identification from experimental snapshots using convolutional neural network[J]. IEEE Access, 7: 6219-6225.
18	XU Zhuoqun, WU Fan, YANG Xinmeng, et al. Measurement of gas-oil two-phase flow patterns by using CNN algorithm based on dual ECT sensors with venturi tube[J]. Sensors, 2020, 20(4): 1200.
19	SEAL M K. Machine learning classification of in-tube condensation flow patterns using visualization[J]. International Journal of Multiphase Flow, 2021, 143: 103755.
20	BRANTSON Eric Thompson, ABDULKADIR Mukhtar, AKWENSI Perpetual Hope, et al. Gas-liquid vertical pipe flow patterns convolutional neural network classification using experimental advanced wire mesh sensor images[J]. Journal of Natural Gas Science and Engineering, 2022, 99: 104406.
21	李欣铜, 陈志冰, 魏志强, 等. 具有输入数据注意力机制的卷积神经网络用于氟化工产品质量预测[J]. 化工进展, 2022, 41(2): 593-600.
	LI Xintong, CHEN Zhibing, WEI Zhiqiang, et al. Convolution neural network with attention mechanism of input data for quality prediction of fluorine chemical products[J]. Chemical Industry and Engineering Progress, 2022, 41(2): 593-600.
22	于长东, 毕晓君, 韩阳, 等. 基于轻量化深度学习模型的粒子图像测速研究[J]. 光学学报, 2020, 40(7): 142-149.
	YU Changdong, BI Xiaojun, HAN Yang, et al. Particle image velocimetry based on a lightweight deep learning model[J]. Acta Optica Sinica, 2020, 40(7): 142-149.
23	LI Gaoyang, GUO Yuting, MABUCHI Takuya, et al. Prediction of the adsorption properties of liquid at solid surfaces with molecular scale surface roughness via encoding-decoding convolutional neural networks[J]. Journal of Molecular Liquids, 2022, 349: 118489.
24	徐日辛, 周骛, 张翔云. 基于深度学习的显微离焦图像法颗粒深度测量[J]. 化工进展, 2021, 40(12): 6499-6504.
	XU Rixin, ZHOU Wu, ZHANG Xiangyun. Particle depth position measurement using microscopic defocused imaging method based on deep learning[J]. Chemical Industry and Engineering Progress, 2021, 40(12): 6499-6504.
25	张建伟, 张学良, 冯颖, 等. 水平对置撞击流的POD分析及混合特性[J]. 过程工程学报, 2016, 16(1): 26-33.
	ZHANG Jianwei, ZHANG Xueliang, FENG Ying, et al. POD analysis and mixing characteristics of impinging streams from two opposite nozzles[J]. The Chinese Journal of Process Engineering, 2016, 16(1): 26-33.
26	胡银玉, 刘喆, 杨基础, 等. 偏心搅拌反应器内的液相混合行为[J]. 化工学报, 2010, 61(10): 2517-2522.
	HU Yinyu, LIU Zhe, YANG Jichu, et al. Liquid mixing in eccentric stirred tank[J]. CIESC Journal, 2010, 61(10): 2517-2522.
27	王红霞, 周家奇, 辜承昊, 等. 用于图像分类的卷积神经网络中激活函数的设计[J]. 浙江大学学报(工学版), 2019, 53(7): 1363-1373.
	WANG Hongxia, ZHOU Jiaqi, GU Chenghao, et al. Design of activation function in CNN for image classification[J]. Journal of Zhejiang University (Engineering Science), 2019, 53(7): 1363-1373.
28	ZOU G Y. Confidence interval estimation for the Bland-Altman limits of agreement with multiple observations per individual[J]. Statistical Methods in Medical Research, 2013, 22(6): 630-642.
29	CARKEET Andrew. Exact parametric confidence intervals for Bland-Altman limits of agreement[J]. Optometry and Vision Science, 2015, 92(3): e71-e80.

编号	类型	参数	数据维度	可学习参数	可学习参数总数
1	图像输入		150×150×3		0
2	卷积层	卷积核大小：5×5 卷积核数量：64 步幅：1×1 激活函数：Relu	150×150×64	权重：5×5×3×64 偏置：64	4864
3	池化层	池化核大小：4×4 步幅：4×4	38×38×64		0
4	卷积层	卷积核大小：5×5 卷积核数量：256 步幅：1×1 激活函数：Relu	38×38×256	权重：5×5×64×256 偏置：256	409856
5	池化层	池化核大小：4×4 步幅：4×4	10×10×256		0
6	卷积层	卷积核大小：3×3 卷积核数量：8 步幅：1×1 激活函数：Relu	10×10×8	权重：3×3×256×8 偏置：8	18440
7	池化层	池化核大小：4×4 步幅：4×4	3×3×8		0
8	卷积层	卷积核大小：3×3 卷积核数量：64 步幅：1×1 激活函数：Relu	3×3×64	权重：3×3×8×64 偏置：64	4672
9	池化层	池化核大小：4×4 步幅：4×4	1×1×64		0
10	扁平层		64		0
11	全连接层	神经元数量：256 激活函数：Relu	256	权重：64×256 偏置：256	16640
12	输出层	神经元数量：1 激活函数：Elu	1	权重：256×1 偏置：1	257

编号	类型	参数	数据维度	可学习参数	可学习参数总数
1	图像输入		150×150×3		0
2	卷积层	卷积核大小：5×5 卷积核数量：64 步幅：1×1 激活函数：Relu	150×150×64	权重：5×5×3×64 偏置：64	4864
3	池化层	池化核大小：4×4 步幅：4×4	38×38×64		0
4	卷积层	卷积核大小：5×5 卷积核数量：256 步幅：1×1 激活函数：Relu	38×38×256	权重：5×5×64×256 偏置：256	409856
5	池化层	池化核大小：4×4 步幅：4×4	10×10×256		0
6	卷积层	卷积核大小：3×3 卷积核数量：8 步幅：1×1 激活函数：Relu	10×10×8	权重：3×3×256×8 偏置：8	18440
7	池化层	池化核大小：4×4 步幅：4×4	3×3×8		0
8	卷积层	卷积核大小：3×3 卷积核数量：64 步幅：1×1 激活函数：Relu	3×3×64	权重：3×3×8×64 偏置：64	4672
9	池化层	池化核大小：4×4 步幅：4×4	1×1×64		0
10	扁平层		64		0
11	全连接层	神经元数量：256 激活函数：Relu	256	权重：64×256 偏置：256	16640
12	输出层	神经元数量：1 激活函数：Elu	1	权重：256×1 偏置：1	257

项目	MSE	MAE	LOSS
测试集（1263）	2.55×10^-5	0.0029	1.28×10^-5
验证集（302）	2.82×10^-6	0.0011	—

项目	MSE	MAE	LOSS
测试集（1263）	2.55×10^-5	0.0029	1.28×10^-5
验证集（302）	2.82×10^-6	0.0011	—

操作参数	MSE	MAE
d=8mm（1504）	1.88×10^-6	0.000932
d=10mm（2705）	2.54×10^-6	0.000978
L=3d（3007）	2.49×10^-6	0.000972
L=4d（1202）	1.84×10^-6	0.000933
M=0.8（903）	1.89×10^-6	0.000935
M=0.9（901）	1.30×10^-6	0.000808
M=1.0（902）	4.3×10^-6	0.001095
M=1.2（901）	1.96×10^-6	0.000983
M=1.4（602）	1.97×10^-6	0.000996