基于核判别分析的特征约简方法在故障诊断中的应用

针对故障诊断中传统的线性多变量统计分析方法不能解决线性不可分问题,提出了一种基于核判别分析的非线性特征约简方法。在核判别分析中,核函数决定了其非线性映射的能力;为此,提出了一种基于K均值聚类的核函数参数优化方案,并将该方法应用于滚动轴承的故障诊断中。结果表明,与主成分分析及线性判别分析相比,核判别分析能够更有效地区分轴承的4种状态,适用于故障诊断中的非线性特征约简。

Application of kernel discriminant analysis based feature reduction method to fault diagnosis

Since the traditional linear methods based on multivariate statistical analysis cannot solve the linearly inseparable problems,a novel nonlinear feature reduction method based on kernel discriminant analysis(KDA) was presented for fault diagnosis.The performance of KDA depended on the choice of kernel functions in KDA.Therefore,a parameter optimization scheme for kernel functions was proposed using K-means clustering.Finally,the proposed method was applied to fault diagnosis of the rolling bearing.The results showed that the proposed method outperformed principle component analysis(PCA) and linear discriminant analysis(LDA) to discriminate the four conditions of bearings and thereby was suitable for nonlinear feature reduction in fault diagnosis.