摘要
针对滚动轴承寿命特征提取与寿命阶段智能识别问题,提出加噪样本扩展深度稀疏自编码神经网络的滚动轴承寿命阶段识别方法。稀疏自编码具有非监督自动学习数据内部结构特征的能力,但属浅层网络,特征提取能力有限且不具备分类能力。因此,将多个稀疏自编码堆栈并添加分类层构建出集寿命特征自动提取与识别功能于一体的深度稀疏自编码神经网络,通过无监督逐层自学习与有监督微调,完成寿命特征的自动提取与表达,并实现寿命阶段智能识别。同时,为解决寿命样本量不足导致的网络过拟合,对原训练样本进行加噪扩展来训练网络,以抑制网络过拟合并提高网络的鲁棒性。通过工程应用,证明了所提方法的可行性和有效性。
Aiming at solving the problems of life state feature extraction and intelligent recognition for rolling bearing, a life state recognition method based on deep sparse auto-encoder neural network with sample expansion by adding noise is proposed. The sparse auto-encoder is an unsupervised learning methods, it can learn the internal structure information of datasets auto-matically. However, its feature extraction ability is limited because it is a shallow network and does not have the ability of clas-sification. Therefore, the multiple auto-encoders are stacked and the classification layer is added to construct the deep sparse auto-encoder neural network which can automatically learn the inherent life feature and intelligently recognize the life state. Thus, the life feature can be automatically extracted and expressed to realize the intelligent recognition through individual unsu-pervised learning layer by layer and supervised fine tuning. Moreover, in order to solve the overfitting because of insufficient training samples, the sample expansion method by adding noise is proposed to inhibit the overfitting and improve the robust-ness. The feasibility and validity of the proposed method are verified by its application to the recognization of the life state of rolling bearings.
出处
《振动工程学报》
EI
CSCD
北大核心
2017年第5期874-882,共9页
Journal of Vibration Engineering
基金
机械传动国家重点实验室开放基金资助项目(SKLMT-KFKT-201710)
国家自然科学基金资助项目(51305471
51775065)
中国博士后科学基金资助项目(2014M560719)
关键词
故障诊断
滚动轴承
寿命阶段
稀疏自编码
神经网络
fault diagnosis
rolling bearing
life state
sparse auto-encoder
neural network