基于位势函数的欠定盲源分离识别诊断方法

对观测器数目小于源信号数目的欠定盲源分离进行了研究,提出了一种基于位势函数的稀疏信号欠定盲源分离方法,该方法从能量的观点出发,通过构造位势函数,将寻找混合信号在直线方向上的聚类问题转化为寻找累积位势函数的局部极大值问题,从而准确的估计出源信号数目和混合矩阵,克服了通常的基于k-means聚类的混合矩阵估计法需预先给定源信号数目的缺点.利用仿真信号检验了该方法的有效性.基于信号频域稀疏性假设,将该方法应用于欠定条件下的滚动轴承振动故障信号的盲分离,较好地分离出了故障信号. 

Fault diagnosis based on underdetermined blind source separation using potential energy function

This paper studied underdetermined blind source separation (UBSS), i.e. the number of mixtures is smaller than that of sources. A blind sparse source separation algorithm based on potential energy function was proposed to estimate both mixing matrix and the number of sources. In the algorithm, the problem of finding mixture signals' linear clustering was transformed to that of finding the accumulative potential energy function's local maximum by constructing potential energy function. The algorithm avoided the disadvantage of k-means clustering algorithm that the number of sources in UBSS should be given in advance. The advantage of the proposed algorithm was verified from simulation signals. Based on the assumption that the signal was sparse in frequency domain, the algorithm was applied to underdetermined blind separation of rolling bearing vibration fault signals. The results show the fault signals are separated effectively.