基于机器学习方法的压气机落后角与总压损失预测代理模型  

Surrogate model for deviation angle and total pressure loss prediction of compressor based on machine learning methods

作　　者：马博文 巫骁雄 于洋 MA Bowen;WU Xiaoxiong;YU Yang(School of Aeronautics,Chongqing Jiaotong University,Chongqing 404100,China)

机构地区：[1]重庆交通大学航空学院,重庆404100

出　　处：《航空动力学报》2023年第7期1675-1690,共16页Journal of Aerospace Power

摘　　要：为了提高压气机特性预测的精度,基于机器学习方法构建预测落后角与总压损失的代理模型。以一台两级压气机为研究对象,基于多个转速工况下的流场实验数据和叶片几何参数建立了基元叶型数据库。通过灵敏度分析方法筛选出对落后角和总压损失影响最大的输入参数,分别采用高斯过程回归和人工神经网络两种机器学习算法建立落后角与总压损失模型,并引入贝叶斯优化算法搜索最佳模型超参数。对于人工神经网络面临的优化问题和泛化问题,调整模型学习率和修正参数梯度以加速收敛,同时采用正则化方法增强模型泛化能力。模型训练过程采用交叉验证策略以降低过拟合风险,并将优化后的代理模型整合到通流程序中对压气机进行特性预测验证。对比表明,低转速工况代理模型的压比特性预测误差显著低于经验模型,其中人工神经网络建模改善最明显,相比经验模型预测误差降低了0.1。通过代理模型横向对比,基于人工神经网络建立的代理模型比基于高斯过程的代理模型预测精度更高且鲁棒性更强。In order to improve the prediction accuracy of compressor performance,a surrogate model of deviation angle and total pressure loss based on machine learning methods was built.A two-stage compressor was used as the research object,and the elementary cascade database was established using experimental data of the flow field and the geometric parameters under multiple rotation speed conditions.The sensitivity analysis method was used to screen out the input parameters that have the greatest impact on the deviation angle and total pressure loss.Two machine learning algorithms,i.e.:Gaussian process regression and artificial neural network,were used to establish the deviation angle and total pressure loss model,and Bayesian optimization algorithm was introduced to search for the best hyperparameters of the model.For the optimization iteration and generalization problems faced by artificial neural networks,the Adam algorithm was introduced to adjust the model learning rate and modify the parameter gradient to accelerate the convergence.At the same time,the regularization method was used to enhance the generalization of the model.The cross validation scheme was used in model training process to reduce the risk of overfitting,and the optimal surrogate models were integrated into the throughflow program.The comparison of the calculation results showed,the total pressure ratio prediction error of the surrogate model on low speed conditions was significantly lower than that of the empirical model,among which the artificial neural network modeling had the most significant improvement,and the prediction error was reduced by 0.1 compared with the empirical model.Through comparison of surrogate models,the surrogate model based on artificial neural network had higher prediction accuracy and stronger robustness than the surrogate model based on Gaussian process.

关 键 词：落后角 损失模型 代理模型 机器学习 人工神经网络 

分 类 号：V231.3[航空宇航科学与技术—航空宇航推进理论与工程]