基于多尺度自适应注意力网络的剩余寿命预测  

Remaining useful life prediction based on multi-scale adaptive attention network

作　　者：刘斌[1] 许靖 霍美玲 崔学英[1] 谢秀峰[1] 杨栋辉[1] 王嘉 Bin LIU;Jing XU;Meiling HUO;Xueying CUI;Xiufeng XIE;Donghui YANG;Jia WANG(School of Applied Science,Taiyuan University of Science and Technology,Taiyun 030024,China)

机构地区：[1]太原科技大学应用科学学院,太原030024

出　　处：《航空学报》2023年第5期160-170,共11页Acta Aeronautica et Astronautica Sinica

基　　金：国家自然科学基金(11701406,11901134,12061091,11972019);山西省基础研究计划(202103021224274,202203021211204);山西省省筹资金资助回国留学人员科研项目(2022-163);山西省社会经济统计科研课题(KY[2022]73)。

摘　　要：剩余寿命预测是复杂设备系统智能维护决策的关键和前提。不同指标数据间的潜在关系给预测的精度带来了挑战,参数的选择也增加了模型预测的困难。采用多尺度自适应注意力网络训练方法,别从纵向和横向两个维度融合数据间的特征关系,给出分段非线性目标函数,提高了预测的精度,降低了预测的均方根误差。使用自适应机制自动选择卷积核大小,提高了网络训练的效率。通过对涡扇发动机数据集进行实证分析验证了该方法在复杂系统剩余寿命预测中的有效性。Remaining useful life prediction is the key to and premise of intelligent maintenance decision-making of complex equipment systems.The potential relationship among different index data brings challenges to prediction accuracy,while parameter selection also increases model prediction difficulty.We use the multi-scale adaptive attention network method to fuse the feature relationship among data from vertical and horizontal dimensions,respectively,and give the piecewise nonlinear target function to improve the prediction accuracy and reduce the root mean square error.The adaptive mechanism is employed to automatically select the size of convolution kernel,improving the efficiency of network training.Empirical analysis of Company-Modular aero-propulsion system simulation data sets prove the effectiveness of this method in remaining useful life prediction of complex systems.

关 键 词：剩余寿命预测 多尺度学习 自适应注意力 非线性目标函数 复杂系统 

分 类 号：V235.13[航空宇航科学与技术—航空宇航推进理论与工程] O213.2[理学—概率论与数理统计]