基于无监督学习的健康指标构建方法研究
Research on Unsupervised Learning-based Health Indicator Construction Methods
俎海东 1焦晓峰 1张万福 2李春2
作者信息
- 1. 内蒙古电力科学研究院分公司,内蒙古呼和浩特 010020
- 2. 上海理工大学能源与动力工程学院,上海 200093
- 折叠
摘要
针对设备健康指标构建方法依赖专家经验的问题,本文结合基于情境化编码器-解码器架构的多尺度残差卷积神经网络与平方欧式距离(MSR-CNED-SE)提出了一种基于无监督数据集的健康指标构建方法,利用深度学习模型实现剩余寿命预测,并在轴承全寿命数据集上验证方法的可靠性.此外,还研究了不同相似度与深度学习网络对健康指标的影响.结果表明:基于平方欧式距离作为相似度度量构建的指标更容易找到退化的起始点,而Bi-LSTM网络在不同预测场景下表现出更好的稳定性和可靠性.
Abstract
In response to the problem of relying on expert knowledge in health indicator construction methods,this paper proposed a method for constructing health indices based on unsupervised database,integrated with multi-scale residual convolutional neural network with contextualized encoder-decoder ar-chitecture and squared Euclidian distance(MSR-CNED-SE).Furthermore,it utilized deep learning models to predict remaining useful life.The reliability of the proposed method was validated on a compre-hensive bearing life dataset.Additionally,the effects of different similarity and deep learning networks on health indicators were studied.The results indicate that health indicators constructed using the squared Euclidean distance as a similarity measure are more effective in identifying the onset of degradation.Mo-reover,the Bi-LSTM network exhibits better stability and reliability under different prediction scenarios.
关键词
健康指标/无监督学习/剩余寿命预测/平方欧式距离/Bi-LSTM网络Key words
health indicator/unsupervised learning/remaining useful life prediction/squared Euclide-an distance/Bi-LSTM network引用本文复制引用
基金项目
国家自然科学基金(52006148)
内蒙古电力科学研究院2022年自筹项目(510241220009)
出版年
2024
NETL
NSTL
万方数据