摘要
机械系统存在的外部环境干扰、变工况条件以及无法直接测量等因素,导致获取的数据常常不满足传统机器学习的两个前提:训练与测试数据分布相同以及目标诊断数据量充足,从而影响诊断模型的泛化能力。针对上述问题,提出一种基于辅助数据的增强型最小二乘支持向量机(LSSVM)迁移学习策略,用于数据量不足时的轴承故障诊断。其中利用递归定量分析(RQA)提取非线性特征并与传统时域特征相结合以提高诊断精度。诊断分类器通过改进传统LSSVM模型,在原目标函数和约束条件中分别增加辅助集的惩罚函数和约束条件,最终得到加入辅助集的函数估计,从而将该算法推广至迁移学习。此外,类内类间距离指标用于描述特征区分性,并提出4种辅助数据集的使用方法,从而构建迁移学习为框架的诊断模型。球形轴承的振动信号试验结果表明,相比传统机器学习,在目标振动数据较少条件下所提模型在轴承故障诊断时性能提升显著。
Due to the problem of the environmental interference,various operating conditions and inevitable indirect measurement,it is difficult to obtain abundant training and testing data from a mechanical system that follow the same underlying distribution,which will influence the generalization ability of fault diagnosis model based on traditional machine learning. A novel approach utilizing transfer learning for classification is presented in this paper,which aims at improving the bearing fault diagnostic performance in case of insufficient labeled samples. The Recurrence Quantification Analysis（ RQA） is used to extract the non-linear features that characterize the underlying dynamics of the mechanical system. These nonlinear features are then combined with the time domain statistical parameters to form a feature vector to improve the diagnostic accuracy. By adding penalty function and constraint condition of auxiliary data to the original objective function and constraints of the least square support vector machine（ LSSVM）,it can be extended to implement a transfer learning strategy. Besides,the internal and interval distance criterion is utilized to evaluate the differentiability of various features. Experimental studies indicate that the transfer learning strategy can improve the diagnostic accuracy as compared to the traditional machine learning strategy when the target data set is small.
出处
《仪器仪表学报》
EI
CAS
CSCD
北大核心
2017年第1期33-40,共8页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金(51575102)项目资助
关键词
轴承故障诊断
递归定量分析
迁移学习
最小二乘支持向量机
bearing fault diagnosis
recurrence quantification analysis
transfer learning
least squares support vector machine（LSSVM）
