基于ConvLSTM的风机轴承寿命预测
Prediction of Fan Bearing Life Based on ConvLSTM
肖宗朕 1杜浩飞 1王勇 1张超 2张丹丹 1李建军1
作者信息
- 1. 内蒙古科技大学 信息工程学院,包头 014010
- 2. 内蒙古科技大学 机械工程学院,包头 014010
- 折叠
摘要
针对普通滚动轴承寿命预测模型在提取特征过程中存在特征提取不充分、预测误差大等问题,提出了基于双通道的卷积长短时记忆网络(ConvLSTM)风机轴承寿命预测模型.首先,将原始轴承振动信号进行小波阈值去噪,去除振动信号中的噪声干扰;其次,为充分提取特征采用双通道提取振动信号特征,其中一路为轴承振动信号信息,另一路为频域幅值信号;然后,采用ConvLSTM模型进行特征提取,该模型可同时兼顾空间局部特征和时间序列上的依赖关系,具有良好的特征提取能力;最后,将两路特征融合深入到全连接层,输出模型预测结果;此外,为提高模型预测准确率,还对损失函数作了相应改进.实验结果表明,所提模型轴承剩余寿命预测误差百分比均在20%以下,其误差百分比小于其他基于深度学习的模型.
Abstract
A dual channel ConvLSTM wind turbine bearing life prediction model is proposed to address the issues of insufficient feature extraction and large prediction errors in the process of feature extraction in or-dinary rolling bearing life prediction models.First,the original bearing vibration signal is subjected to wave-let threshold denoising to remove the noise interference in the vibration signal;secondly,to fully extract fea-tures,this paper uses two channels to extract vibration signal features,one of which is the bearing vibration signal information,and the other is the frequency domain amplitude signal;then use the ConvLSTM model for feature extraction,which can take into account the dependence of spatial local features and time series at the same time,and has good feature extraction capabilities;finally,the two-way feature fusion is deep into the fully connected layer,and the model prediction results are output;in addition,In order to improve the prediction accuracy of the model,this paper also makes corresponding improvements to the loss function.The experimental results show that the percentage of error in predicting the remaining life of bearings in the proposed model is below 20%,which is smaller than other deep learning based models.
关键词
寿命预测/深度学习/卷积长短时记忆网络/振动信号/特征提取Key words
life prediction/deep learning/convolutional long short term memory network/vibration signal/feature extraction引用本文复制引用
基金项目
国家自然科学基金(51965052)
国家自然科学基金(62066036)
内蒙古自治区自然科学基金(2022MS06009)
内蒙古自治区高等学校青年科技英才计划人才项目(NJYT22074)
出版年
2024
NETL
NSTL
万方数据