基于小样本下改进ChaosNet的轴承故障诊断
Research on Intelligent Diagnosis of Bearing Faults Based on Improved ChaosNet with Small Samples
李天昊 1李志星 1王衍学1
作者信息
- 1. 北京建筑大学 机电与车辆工程学院,北京 100044;北京建筑大学 城市轨道交通车辆服役性能保障重点实验室,北京 100044
- 折叠
摘要
为解决在训练样本不足条件下,轴承故障特征提取困难的问题,提出一种基于改进神经混沌学习(neurochaos learning +AdaBoost,NL-AdaBoost)的轴承故障诊断新方法.首先,对时域振动信号进行快速傅里叶变换(fast fourier transform,FFT)提取频域特征,拼接时频域信号获得一维特征样本;其次,输入信号产生对混沌GLS神经元的激励,形成ChaoFEX 特征,馈送至集成学习分类器(AdaBoost);随后,选取轴承故障特征样本,对样本集做k折交叉验证,获得模型最优超参数值,将其应用于测试集进行模型分类能力验证;最后,在小样本对比实验中,与 4 种常见深度学习算法比较模型的macro F1-score.实验结果证明,在低训练样本条件下,NL-AdaBoost模型具有良好的准确性和泛化能力.
Abstract
A new method of bearing fault diagnosis based on improved neurochaos learning(neurochaos learning +AdaBoost,NL-AdaBoost)is proposed.First,the fast fourier transform(FFT)of the time-domain vibration signal is performed to extract the frequency-domain features,and the one-dimensional feature sam-ples are obtained by splicing the time-frequency-domain signals;then,the input signal generates excitation to the chaotic GLS neurons to form ChaoFEX features,which are fed to the integrated learning classifier(Ada-Boost);subsequently,the bearing fault feature samples,and do k-fold cross-validation on the sample set to obtain the optimal hyperparameter values of the model,which are applied to the test set for model classifica-tion capability validation;finally,in a small-sample comparison experiment,the macro F1-score of the model is compared with four common deep learning algorithms.The experimental results demonstrate that under the low training sample condition,NL-AdaBoost,the model has good accuracy and generalization ability.
关键词
小样本训练/神经混沌学习/滚动轴承/故障诊断Key words
small sample training/neural chaos learning/rolling bearing/fault diagnosis引用本文复制引用
基金项目
国家自然科学基金(51875032)
国家自然科学基金青年基金(51805275)
北京建筑大学青年教师科研能力提升计划课题(X21053)
出版年
2024
NETL
NSTL
维普
万方数据