摘要
在实际工程应用中,传感器收集的振动信号常常包含噪声等异常值,这会严重影响从观测信号中分离源信号的准确性。混合矩阵估计在欠定盲源分离中起着重要作用,直接决定了源信号恢复的效果。因此,这里提出了一种结合层次聚类和K-means的两阶段聚类方法,以提高混合矩阵估计的可靠性。具体而言,该方法旨在解决K-means算法中的两个主要问题,即初始聚类中心的随机选择和算法对异常值的敏感性。首先,通过层次聚类对观测信号进行聚类,以获得聚类中心。然后,利用余弦距离剔除偏离聚类中心的异常值。接着,通过计算剩余聚类的均值来获得初始聚类中心。最后,采用改进的K-means算法估计混合矩阵,并利用最小二乘法恢复源信号。以港口装卸机械设备中的滚动轴承故障仿真以及实验结果验证了所提方法的有效性。
In practical engineering applications,vibration signals collected by sensors often contain outliers such as noise,which seriously affects the accuracy of separating source signals from observed signals.Mixed matrix estimation plays an important role in underdetermined blind source separation,which directly determines the effect of source signal recovery.Therefore,a two-stage clustering method combining hierarchical clustering and K-means is proposed in this paper to improve the reliability of mixed ma-trix estimation.Specifically,the method aims to solve two major problems in K-means algorithms,namely the random selection of initial clustering centers and the sensitivity of the algorithm to outliers.Firstly,the observed signals are clustered through hierarchi-cal clustering to obtain the clustering center.Then,the cosine distance is used to eliminate outliers that deviate from the cluster cen-ter.Then,the initial cluster center is obtained by calculating the mean of the remaining clusters.Finally,the improved K-means algorithm is used to estimate the mixed matrix,and the least square method is used to recover the source signal.The effectiveness of the proposed method is verified by simulation and experiment results of rolling bearing faults in port handling machinery.
作者
阮百强
张海燕
RUAN Baiqiang;ZHANG Haiyan(Advanced Manufacturing Industry Department of Zhanjiang Technician College,Guangdong Zhanjiang 524037,China;School of Intelligent Manufacturing,Zhanjiang University of Science and Technology,Guangdong Zhanjiang 524031,China)
出处
《机械设计与制造》
北大核心
2025年第11期42-48,共7页
Machinery Design & Manufacture
基金
广东省人力资源和社会保障厅省级教学研究立项课题(KT2022074)。
关键词
欠定盲源分离
混合矩阵估计
K-MEANS
滚动轴承
特征提取
Underdetermined Blind Source Separation
Mixed Matrix Estimation
K-means
Rolling Bearing
Feature Extraction
