一种高泛化能力的神经网络气动力降阶模型

发布时间：2018-05-01 02:22

本文选题：非定常气动力 + 降阶模型　；参考：《空气动力学学报》2017年02期

【摘要】：神经网络模型的泛化能力是指模型对于新样本的适应能力,而当前基于神经网络的非线性气动力模型往往泛化能力不足。针对这一局限,发展了一种基于计算流体力学的带验证信号气动力降阶模型,用于跨声速气动力预测。采用递归径向基函数神经网络,通过引入差分进化算法,对模型中隐含层神经元宽度进行优化,优化目标为验证信号的均方根误差最小。训练信号采用结构大幅振荡的位移和气动力响应,用小扰动下微幅运动的信号作为验证信号。算例表明建立的降阶模型不仅具有比全阶数值模拟更短的计算时长,并且由于采用了小幅运动的验证信号,使模型在不同频率和振幅下的泛化能力得到增强。
[Abstract]:The generalization ability of neural network model refers to the adaptability of the model to new samples, but the current nonlinear aerodynamic model based on neural network is often lack of generalization ability. In order to overcome this limitation, a reduced order model of aerodynamic forces with validated signals based on computational fluid dynamics is developed, which is used to predict transonic aerodynamic forces. Using recurrent radial basis function neural network and introducing differential evolution algorithm, the hidden layer neuron width in the model is optimized. The optimization target is to verify the minimum root mean square error of the signal. The displacement and aerodynamic response of the large oscillation of the structure are used in the training signal, and the signal of micro-amplitude motion under small disturbance is used as the verification signal. Numerical examples show that the proposed model not only has a shorter computational time than the full-order numerical simulation, but also enhances the generalization ability of the model at different frequencies and amplitudes due to the use of small motion verification signals.
【作者单位】：西北工业大学翼型叶栅空气动力学国家重点实验室;布朗大学工程学院;
【基金】：国家自然科学基金(11572252);国家自然科学基金优秀青年科学基金(11622220) 国家高等学校学科创新引智计划(“111”计划)(B17037)
【分类号】：V211.3

【相似文献】