首页 关于本刊 编 委 会 期刊动态 作者中心 审者中心 读者中心 下载中心 联系我们 English
 自动化学报  2018, Vol. 44 Issue (5): 915-921 PDF

Regression GAN Based Prediction for Physical Properties of Total Hydrogen in Crude Oil
ZHENG Nian-Zu, DING Jin-Liang
State Key Laboratory of Synthetical Automation for Process Industies, Northeastern University, Shenyang 110819
Manuscript received : August 31, 2017, accepted: December 23, 2017.
Foundation Item: Supported by National Natural Science Foundation of China (61590922, 61525302) and the Research Funds by the Ministry of Education of China (N160801001, N161608001)
Corresponding author. DING Jin-Liang  Professor at the State Key Laboratory of Synthetical Automation for Process Industies, Northeastern University. His research interest covers modeling and operation optimization control of complex industrial process, computational intelligence and its application. Corresponding author of this paper
Recommended by Associate Editor TAN Ying
Abstract: In view that generative adversarial network (GAN) is not applicable to prediction of physical properties of crude oil, a novel regression GAN (RGAN) framework is proposed in this study, which consists of a generator G, a discriminator D and a regression model R. Through adversarial learning between discriminator D and generator G, D extracts a series of latent features of 1H nuclear magnetic resonance spectroscopy (1H NMR) of crude oil. The first layer of latent features is shallow representation of the data space, which helps to solve the regression task. The regression model R is established using the first layer of latent features, which improves the accuracy and stability of the prediction. At the same time, the MSE loss function of the regression model R is applied to estimate the lower bound of the mutual information between conditional variables and generated samples, therefore generator G can produce more realistic samples. Experiment results demonstrate that RGAN can improve the prediction accuracy and stability of physical properties of total hydrogen in crude oil efficiently, and also improve the convergence speed of the generator as well as the quality of spectra generation.
Key words: Regression generative adversarial network (RGAN)     prediction of crude oil properties     generative adversarial nets (GAN)     1H nuclear magnetic resonance spectroscopy (1H NMR)

1 基本原理与方法 1.1 GAN的基本理论

 \begin{align} &\mathop{\min }\limits_G \mathop{\max }\limits_D {\mathbb{E}_{x \sim {p_{\rm data}}(x)}}\left[ {\log D\left( x \right)} \right] + \notag\\ &\qquad {\mathbb{E}_{z\sim p(z)}}\left[ {\log \left( {1 - D\left( {G\left( z \right)} \right)} \right)} \right] \end{align} (1)

 \begin{align} \mathop {\min }\limits_\theta \mathop {\max }\limits_w {\mathbb{E}_{x \sim {p_{\rm data}}(x)}}\left[ {{f_w}\left( x \right)} \right] -{\mathbb{E}_{z\sim p(z)}}\left[ {{f_w}\left( {{g_\theta }(z)} \right)} \right] \end{align} (2)

1.2 RGAN模型

 图 1 RGAN模型结构示意图 Figure 1 Diagram of model structure of RGAN

1.2.1 RGAN的目标函数

2.3 RGAN的实现

3 讨论与分析 3.1 回归模型R与生成模型G的影响

RGAN引入一个额外的超参数$\lambda$, 目的是使得回归模型估计的互信息下界值可以有效作用于生成模型中, 并且超参数$\lambda{}$的取值对谱图生成与回归预测十分重要.当超参数$\lambda=0$, RGAN退化为WGAN, 而回归模型R仅仅利用了判别模型的首层潜在特征, 而无法作用于生成模型G.当超参数$\lambda{}$不为零时, 回归模型R对G的条件变量与生成样本的互信息估计并最大化, 使得生成模型G生成与条件变量相关且类似于真实样本, 同时由于对抗性质, G迫使判别模型D提高性能, 并使得D特征层能够与条件变量相关, 从而利于基于D首层潜在特征建立的回归模型R, 因此在R, G与D形成相互促进过程中, 各自模型的性能得到优化, 不仅增强了生成模型G的稳定性, 而且提高了回归网络R的预测性能.

 图 4 超参数$\lambda{}$对生成模型G的影响 Figure 4 Effect of hyper parameter$\lambda{}$ on generative model G

 图 5 超参数$\lambda{}$对回归模型R的影响 Figure 5 Effect of hyper parameter$\lambda{}$ on regression model R
3.2 回归模型R的预测精度

3.3 NMR谱图的生成

 图 6 超参数$\lambda{}$对NMR谱图生成的影响 Figure 6 Effect of hyper parameter $\lambda{}$ on generation of $^{1}$H nuclear magnetic resonance spectrum
4 结论