基于集成学习的风洞马赫数混合模型建模方法的研究

The Mach number of the test section is an important performance index for wind tunnel measurements, and its stability has a significant effect on the quality of the airflow. The improvement of control precision for the Mach number is determined by advanced control algorithms which generally depend on high-accuracy models of controlled objects. However, the mass, high-dimensional, noisy, redundant and unbalanced data of wind tunnel measurements make it difficult to build a high-efficiency and high-accuracy Mach number model. Consequently, oriented to these difficulties simultaneously, the research on the modeling methods of wind tunnel Mach number mixture models based on the ensemble learning is carried out in this project. Aiming at the high-dimensional data, the FSE (Feature Subsets Ensemble) algorithm is employed; for the unbalanced data, the FSE Mach number mixture models based on different working conditions will be proposed; the TSE-FSEs and the FSE-TSEs learning algorithms will be proposed to deal with mass data, where the TSE denotes the Tree-Structure Ensemble; to handle the noisy and redundant data, the FSE model with the low-pass-filter and redundancy-removal algorithm will be proposed. The research has great significance on improving the accuracy and speed of the Mach number predictions, boosting the precision and level of the Mach number controlling, and reducing the cost and energy consumption of the measurements in the wind tunnel, etc.

试验段马赫数作为风洞试验的一个重要性能指标，它的稳定性对风洞流场品质有着重要影响。马赫数控制精度的提升取决于先进的控制算法，通常先进控制算法依赖于高精度的被控对象模型。然而风洞试验的海量、高维、噪声、冗余、不均衡数据成为了建立高效率、高精度马赫数模型的主要难点。因此本项目同时面向这些难点，研究基于集成学习的风洞马赫数混合模型建模方法。针对高维数据，采用FSE(Feature Subsets Ensemble，特征子集集成)算法；针对不均衡数据，拟提出基于不同工况的FSE马赫数混合模型；针对海量数据，拟提出TSE-FSEs和FSE-TSEs学习算法，其中TSE(Tree-Structure Ensemble)为树结构集成；针对噪声及冗余数据，拟提出带低通滤波去冗余算法的FSE模型。这些研究对于提高马赫数的预测精度及速度、提升马赫数的控制精度和水平、降低风洞试验成本及能耗等具有重要意义。

试验段马赫数是风洞系统中最关键的变量之一。马赫数的高精度实时预测可以更好地帮助系统控制气流。然而，风洞大数据带来的挑战使之变得困难，这些挑战包括实时计算、大量样本、高维变量、不平衡的数据分布以及强噪声。为了同时解决这些挑战，我们进一步探索了风洞大数据，设计了两阶段时间序列在线混合(Two-stage time-Series Online Mixture, TSOM)马赫数模型。TSOM分为混合学习(Online Mixture, OM)学习和单一工况下的两阶段时间序列(Two-stage time-Series, TTS)建模两种策略。OM学习解决了不同工况下数据分布不均衡问题，TTS解决了相同工况下的数据分布不均衡问题。TSOM还可以实现不同工况数据的分流和相同工况数据的分段。其次，为了实现冲压阶段的马赫数建模，我们提出了具有FRO(FSE ReMD-OSELM)的Pseudo-NARX模型。Pseudo-NARX是NARX的改进版本，通过引入回归树对特征进行选择。为了拟合Pseudo-NARX模型的非线性函数，我们提出了FRO算法，它逐批递归学习数据，无需耗时的梯度优化过程。为了实现稳定阶段的马赫数建模，我们提出了带有TFL(TSE-FSEs Fixed-size LS-SVMs)的NARX模型。TFL中的TSE-FSEs具有双重集成结构，FSE局部模型建立在由TSE构造的非重叠样本子集上。当达到时间阈值时，冲压阶段的马赫数模型将自动切换到稳定阶段的马赫数模型。实验证明了TSOM马赫数模型在处理风洞大数据时的有效性，它能够满足工程上精度和速度的要求。