摘要
准确评估涡轮叶片的损伤状态对于指导其大修/更换行为具有重要的意义,但由于涡轮叶片结构及服役环境的复杂性,现有技术手段难以模拟真实服役状态下的叶片损伤情况,而基于真实服役叶片的损伤数据进行损伤预测又存在数据采集成本高、样本量小的客观限制。为此,针对小样本条件下服役涡轮叶片的损伤状态评估需求,提出一种基于元学习的损伤参数预测方法,在有限的服役数据基础上,实现对涡轮叶片损伤参数的有效预测。首先制备涡轮叶片不同叶身高度的切片试样,并通过场发射扫描电镜获取切片试样典型部位的微观图片,使用图像处理技术提取不同部位损伤数据,并根据图片所处部位的不同将损伤数据划分为不同训练任务数据;然后提出一种基于MAML-LSTM模型的涡轮叶片损伤参数预测方法,增强了服役条件与叶片损伤参数之间的相关性,建立了服役参数与损伤参数之间的有效映射。利用本文所提出的MAML-LSTM模型对测试集数据进行预测,预测结果的平均绝对百分比误差为7.55%,对比BP、RNN、LSTM、Bi-LSTM等神经网络的预测结果,所提出的模型在测试集上的平均绝对误差下降了至少52.37%,均方误差下降了至少76.98%。
Accurately evaluating the damage state of turbine blades is of great significance for guiding their overhaul/replacement behavior. However, due to the complexity of turbine blade structure and service environment, it is difficult for existing technical means to simulate the blade damage under real service conditions, and the damage prediction based on the damage data of real service blades has the objective limitation of high data acquisition cost and few samples. Therefore, a damage parameter prediction method based on meta-learning is proposed to evaluate the damage state of turbine blades in service under small sample conditions. Based on the limited service data, the damage parameters of turbine blades are predicted. Firstly, the slice samples of different blade heights of turbine blades were prepared, and the microscopic images of typical parts of the slice samples were obtained by field emission scanning electron microscopy. The image processing technology was used to extract the damage data of different parts, and the damage data were divided into different training task data according to the different parts of the picture. Then, a prediction method of turbine blade damage parameters based on MAML-LSTM is proposed. Metalearning is carried out on the basis of different training tasks, which enhances the correlation between service conditions and blade damage parameters, establishes an effective mapping between service parameters and damage parameters, and realizes the prediction of turbine blade damage parameters. Finally, the MAML-LSTM model proposed in this paper is used to predict the test set data. The average absolute percentage error of the prediction results is 7. 55%. Compared with the prediction results of BP, RNN, LSTM, Bi-LSTM and other neural networks, the average absolute error of the model proposed in this paper on the test set is reduced by at least 52. 37%, and the mean square error is reduced by at least 76. 98%.
作者
黄渭清
李宁
刘开霖
付志忠
纪鹏飞
张生良
董立伟
孙燕涛
HUANG Weiqing;LI Ning;LIU Kailin;FU Zhizhong;JI Pengfei;ZHANG Shengliang;DONG Liwei;SUN Yantao(School of Mechanical Engineering,Beijing Institute of Technology,Beijing 100081,China;Beijing Aeronautical Technology Research Center,Beijing 100076,China)
出处
《航空学报》
EI
CAS
CSCD
北大核心
2024年第18期270-281,共12页
Acta Aeronautica et Astronautica Sinica
基金
国家自然科学基金(5210050392)。
关键词
元学习
损伤预测
涡轮叶片
神经网络
小样本
微组织损伤
meta-learning
damage prediction
turbine blade
neural network
small sample
microstructure damage
