复杂过程系统虚拟样本生成方法研究

在大数据时代，许多领域经常由于数据获取成本较高、数据重复或发生概率较小等原因，面临着有用数据有限，即小样本问题。无论是数据驱动还是领域知识与数据融合驱动建模都需要一定量的有效样本数据，因此，解决小样本问题刻不容缓。虚拟样本生成是增加样本数、改善样本分布提高样本多样性和减少样本信息间隔的有效途径。本项目针对高维无标签数据小样本问题，研究提出基于变分自编码器和生成对抗网络融合的虚拟样本生成方法，并引入强化学习等策略，形成更为高效和鲁棒的无监督式虚拟样本生成方法；针对高维标签数据小样本问题，提出基于协同训练和VAE-GAN强化结构的半监督式虚拟样本生成方法；为有效克服维度灾难问题和空空间现象，研究提出基于蒙特卡洛模拟的虚拟样本生成方法。最后，研究提出以信息熵为优化目标的最佳虚拟样本数量确定方法，并提出虚拟样本有效性评价方法，以石化过程数据驱动应用需求为导向，形成一套过程工业虚拟样本生成方法。

In the era of big data, many areas are often due to high cost of data acquisition, data duplication or low probability of occurrence, faced with limited useful data, that is, the small sample problem. Both data-driven and domain knowledge and data fusion-driven modeling require a certain amount of valid sample data. Therefore, it is urgent to solve the problem of small sample size. Virtual sample generation is an effective way to increase the number of samples, improve sample distribution and diversity, and reduce the sample information interval. In this project, a method of virtual sample generation based on the fusion of VEA (Variational Auto-Encoder) and GAN (Generative Adversarial Network) is proposed to solve the problem of small samples of high-dimensional unlabeled data. Integrating reinforcement learning and other learning strategies, a more efficient and robust unsupervised virtual sample generation method is developed. Aiming at the problem of small samples of high-dimensional label data, a semi-supervised virtual sample generation method based on collaborative training and VAE-GAN reinforcement structure is proposed. At the same time, In order to effectively overcome the dimension disaster problem and empty space phenomenon，the methods of virtual sample generation based on the Monte Carlo simulation is proposed. Finally, the methods to determine the optimal number of virtual samples with information entropy as the optimal objective is proposed, and the evaluation method of virtual sample validity is also proposed. A set of virtual sample generation methods for process industry is formed based on the demand of data-driven application in petrochemical process.