基于深度元学习的固体发动机性能预测方法研究  

Research on Solid Motor Performance Prediction Method Based on Deep Meta-Learning

作　　者：崔研 娄碧轩 于鹏程 杨慧欣 Cui Yan;Lou Bixuan;Yu Pengcheng;Yang Huixin(College of Aerospace Engineering,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]沈阳航空航天大学航空宇航学院,沈阳110136

出　　处：《航空兵器》2024年第5期110-114,共5页Aero Weaponry

基　　金：辽宁省属本科高校基本科研业务费专项资金资助。

摘　　要：针对飞行器动力系统中固体发动机性能实验的诸多限制,如高成本、专业设备需求、特定实验环境、高风险性等问题,本文提出了一种基于深度元学习的人工智能方法,用于发动机性能预测。该方法采用模型不可知元学习(Model-Agnostic Meta-Learning,MAML)和深度卷积神经网络(Deep Convolutional Neural Networks,DCNN)模型,首先根据不同实验条件划分推力-时间数据为不同训练任务,通过内循环训练得到各任务最佳模型参数,在外循环中更新模型初始化参数,内外循环迭代优化后,获得了能够高精度预测固体发动机总冲的模型,最后用新任务进行测试。测试结果显示,相较于无元学习的DCNN,该方法在测试集上的误差显著下降,百分比误差最大为2.27%。证明了元学习模型在小样本条件下对固体发动机性能的高精度预测能力。Addressing the numerous limitations in solid motor performance experiments for aircraft power systems,such as high costs,specialized equipment requirements,specific experimental environments,and high risks,this paper proposes an artificial intelligence method based on deep meta-learning for engine performance prediction.This method employs model-agnostic meta-learning(MAML)and deep convolutional neural networks(DCNN)models.Firstly,thrust-time data is divided into different training tasks according to varying experimental conditions.The optimal model parameters for each task are obtained through inner-loop training,and the model initialization parameters are updated in the outer-loop.After iterative optimization of the inner-loop and the outer-loop,a model predicted the total impulse of solid engines with high accuracy is obtained,and finally it is tested for new tasks.The test results demonstrate that compared to DCNN without meta-learning,this method reduces the error on the test set significantly,with a maximum percentage error of 2.27%.This verifies the high-precision prediction ability of the meta-learning model for solid motor perfor-mance under small sample conditions.

关 键 词：固体发动机 发动机性能 元学习 模型不可知元学习 深度卷积神经网络 

分 类 号：TJ760[兵器科学与技术—武器系统与运用工程] V23[航空宇航科学与技术—航空宇航推进理论与工程]