基于长短期记忆神经网络的多级涡轮过渡态叶尖间隙预测  

Prediction of multi-stage turbine transient tip clearance based on long short-term memory neural network

作　　者：杨超[1] 毛军逵[1] 杨悦 王飞龙 邵发宁 毕帅 YANG Chao;MAO Junkui;YANG Yue;WANG Feilong;SHAO Faning;BI Shuai(College of Energy and Power Engineering,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)

机构地区：[1]南京航空航天大学能源与动力学院,江苏南京210016

出　　处：《推进技术》2025年第2期248-257,共10页Journal of Propulsion Technology

基　　金：国家科技重大专项(2017-Ⅲ-0010-0036)。

摘　　要：为了解决多级涡轮模型在高维度变量的复杂空间耦合效应下向高效、高精度过渡态叶尖间隙预测提出的挑战,本文搭建了基于贝叶斯优化和多任务学习算法的长短期记忆神经网络(BO-MTLLSTM)多级涡轮过渡态叶尖间隙智能预测模型,以实现过渡态叶尖间隙高效、高精度预测。在BOMTL-LSTM模型中,通过高效的长短期记忆神经网络(Long Short-Term Memory,LSTM)模型对基于有限元分析方法得到的高精度过渡态叶尖间隙时序信息进行学习,并在LSTM模型的基础上,引入多任务学习(Multi-Task Learning,MTL)用于多个叶尖间隙预测任务之间的信息共享,以缓解高维度变量复杂空间耦合作用的影响。同时,结合贝叶斯优化(Bayesian Optimization,BO)对神经网络模型超参数进行全局自动优化,提升预测精度与训练效率。结果表明,相比于传统计算模型,BO-MTL-LSTM模型在同等预测精度下,能够在秒量级时间内完成一个完整发动机历程的多级涡轮过渡态叶尖间隙的预测。此外,相比常规的BO-LSTM模型,BO-MTL-LSTM模型的均方根误差和平均绝对误差分别降低了84.39%和89.21%,模型训练时间缩短了30%,该模型可以实现多级叶尖间隙的高效、精准预测。To address the challenges posed by multi-stage turbines toward efficient and high-precision transient tip clearance prediction under complex spatial coupling effects of high-dimensional variables,this paper builds a multi-stage turbine transient tip clearance prediction model based on Bayesian optimization and multitask learning algorithm with long short-term memory neural network(BO-MTL-LSTM),aiming to achieve effi⁃cient and high-precision prediction of transient tip clearance.In the BO-MTL-LSTM model,the efficient long short-term memory(LSTM)neural network learns high-precision transient tip clearance chronological informa⁃tion obtained from the finite element analysis based method.Additionally,multi-task learning(MTL)is incorpo⁃rated based on the LSTM model to share information among multiple tip clearance prediction tasks,alleviating the complex spatial coupling effects of high-dimensional variables.Simultaneously,the Bayesian optimization(BO)algorithm is applied to automatically optimize the hyperparameters of the neural network model globally,enhancing prediction accuracy and training efficiency.The results demonstrate that compared to the conventional computational model,the BO-MTL-LSTM model efficiently predicts multi-stage turbine transient tip clearance for a complete engine history within a second-scale time frame while maintaining the same prediction accuracy.Furthermore,in comparison to BO-LSTM model,the BO-MTL-LSTM model achieves an 84.39%reduction in root-mean-square error and an 89.21%reduction in average absolute error.The model training time is also short⁃ened by 30%,enabling efficient and accurate prediction of multi-stage tip clearance.

关 键 词：多级涡轮 叶尖间隙预测 多任务学习 长短期记忆神经网络 贝叶斯优化 

分 类 号：V231[航空宇航科学与技术—航空宇航推进理论与工程]