几何损伤导致涡轮静叶气动衰减的CNN预测  

CNN Prediction of Aerodynamic Degradation of Turbine Vanes Due to Geometric Damage

作　　者：何庆富 迟重然[1] 臧述升[1] HE Qingfu;CHI Zhongran;ZANG Shusheng(School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学机械与动力工程学院,上海200240

出　　处：《工程热物理学报》2022年第12期3219-3224,共6页Journal of Engineering Thermophysics

基　　金：工信部重大专项基础研究项目(No.2017-I-0007-0008)。

摘　　要：燃气轮机运维工作中,需根据涡轮变形损伤情况制定合理的维修策略,以更低的成本恢复燃机性能。对此,本文基于卷积神经网络(CNN),发展了一种根据几何损伤情况快速预测叶栅气动性能影响的方法。研究中,将涡轮叶片损伤分布用二维矩阵描述,生成了样本数为670的损伤叶片数据集,通过CFD仿真得到了气动性能样本。训练了CNN损伤叶片性能预测模型,并对模型进行了敏感性分析,归纳了损伤尺寸和位置对性能的影响,表明吸力面的损伤对性能的影响最显著;讨论了训练样本数对CNN精度的影响,对未来基于CNN的几何-气动关联模型构建给出了建议。In the operation and maintenance of gas turbines,it is necessary to formulate a reasonable maintenance strategy according to the deformation and damage of the turbine to restore the performance of the gas turbine at a lower cost.In this regard,based on convolutional neural network(CNN),this paper develops a method to rapidly predict the impact of cascade aerodynamic performance according to geometric damage.In the study,a damaged blade dataset with a sample size of 670 is generated,in which the turbine blade damage is described by a two-dimensional matrix,and aerodynamic performance samples are obtained through CFD simulations.A CNN model for predicting the performance of damaged blades is trained and the model is subjected to sensitivity analysis.The effect of damage size and location on performance is summarized,and the results show that the damage on the suction side has the most significant effect on performance.The influence of the training sample size on the accuracy of CNN is discussed,and suggestions are given for the construction of CNN-based geometry-aerodynamic correlation models in the future.

关 键 词：涡轮叶片 损伤 性能退化 卷积神经网络 深度学习 

分 类 号：TK471[动力工程及工程热物理—动力机械及工程]