摘要
为了提高发动机故障诊断精度,利用集成学习异构集成机械学习模型,并使用注意力机制提高模型准确度。利用集成学习理论训练异构的随机森林、卷积神经网络、深度神经网络诊断模型,再将各个子异构学习器的输出值通过注意力调节特征权重,作为输入数据通过决策树训练,得到完整的Stacking异构模型;解决各个神经网络在故障诊断过程中特征值提取不准确或者部分故障诊断不精确的问题。方法采用凯斯西储大学轴承数据集为数据依托,与传统的Stacking集成学习故障诊断模型进行对比,注意力机制的加入,使诊断精度提高5.6%。
Aiming at improving the accuracy of engine fault diagnosis,the combination of heterogeneous ensemble learning and attention mechanism is proposed.Firstly,the heterogeneous random forest,CNN and DNN mechanical learning models are trained by using the ensemble learning theory.Secondly,the output value of each sub heterogeneous learner is adjusted by attention,which is used as the input data for training through the decision tree.Finally,the complete stacking heterogeneous model is obtained.The proposed model solves the problem of inaccurate extraction of single eigenvalue or inaccurate partial fault diagnosis in the process of fault diagnosis.The proposed method uses the Case Western Reserve University bearing data set as the data support.Compares with the traditional stacking ensemble learning fault diagnosis model,with the attention mechanism is added,the diagnostic accuracy is improved by 5.6%.
作者
宋高腾
刘君强
曹斯言
左洪福
SONG Gao-teng;LIU Jun-qiang;CAO Si-yan;ZUO Hong-fu(Nanjing University of Aeronautics and Astronautics,Nanjing 211000,China)
出处
《航空计算技术》
2022年第1期50-54,共5页
Aeronautical Computing Technique
基金
国家自然科学基金项目资助(U1533128)
中央高校基本科研业务费项目资助(NS2020050)。
关键词
故障诊断
发动机轴承
注意力
集成学习
fault diagnosis
engine bearing
attention
ensemble learning
作者简介
宋高腾(1997-),男,江苏徐州人,硕士研究生。
