基于改进核主元分析的故障检测

针对电主轴系统特点,提出基于改进核主元分析(KPCA)的故障检测方法,引入混合核函数的定义,将多项式核和径向基核的混合核方法与主元分析方法(PCA)相结合,解决采用单一核函数诊断故障时的高误诊率问题.首先对数据进行预处理,然后使用混合核函数对数据矩阵进行映射,映射到高维特征空间,使非线性数据变量变为线性数据变量,并使用PCA提取变量数据的高维空间相关特征确定主元个数,最后根据混合非线性主元特征计算出的T2和Q统计量,实现在线故障检测.该方法改进传统核函数的选取方法,充分考虑工业过程中的非线性,更精确地描述工业过程特性,可以准确、有效地检测出电主轴系统故障.对田纳西-伊斯曼(TE)过程以及电主轴系统的应用实例证明该方法的可行性.

Fault Detection Based on Improved Kernel Principal Component Analysis

According to the characteristics of electron spindle, a fault detection approach based on improved kernel principal component analysis (KPCA) was proposed, definition of mixed kernel function was introduced by combing RBF kernel and polynomial kernel with PCA, which was aimed at solving the higher misdiagnosis rate problem of single kernel function. Firstly, the data matrix was preprocessed, then, mapping the input sample data into a kernel feature space by mixtures of kernels, and then linear PCA in the nonlinearly mapped feature space was performed to find the principal component feature vectors for diagnosis. The fault could be detected on-line by monitoring T2 and squared prediction error (Q) which were calculated by mixtures nonlinear relative PCs. The proposed method could extract effectively nonlinear feature of industrial process by improving traditional selection method of kernel function, and fully consider the nonlinear feature of industrial process so that industrial process characteristics could be described accurately. The fault of electron spindle could be detected precisely by the improved method. Tennessee Eastman (TE) process and motorized spindle working process were applied to validate the practicability and feasibility of the improved method