基于互信息与核熵成分分析的故障检测算法

Fault detection algorithm based on mutual information and kernel entropy component analysis

现有故障诊断方法对样本数据分布要求过于苛刻且难以提取高阶的统计信息,故本研究提出一种基于互信息与核熵成分分析的故障检测算法(MIKECA)。首先根据训练样本建立核熵成分分析(KECA)模型,然后通过其残差矩阵与互信息求得一种新型统计量—基于互信息的平方预测误差(MISPE),利用核密度估计计算统计量的控制限;最后把测试样本投影在KECA模型上,计算测试样本的统计量并与控制限对比,统计量超出控制限的样本即被识别为故障样本。将这种方法用于TE(Tennessee Eastman)过程的故障检测,对比传统核主成分分析、传统核熵成分分析算法,具有比较明显的优势。

The existing fault detection methods require too strict sample data distribution and have difficulty in extracting high order statistical information.This paper proposed a fault detection algorithm based on mutual information and kernel entropy component analysis(MIKECA).Firstly,a kernel entropy component analysis(KECA)model was built based on the training samples.Secondly,a new type of statistic,namely,the squared prediction error based on mutual information(MISPE),was obtained through its residual matrix and mutual information and the kernel density estimation was used to calculate the control limit of the statistic.Finally,the test samples were projected on the KECA model and the calculated statistic of the test samples was compared with the control limit.The samples whose statistic exceeded the control limit were identified as fault samples.The proposed method,when applied to the fault detection of Tennessee Eastman(TE)process,shows obvious advantages over the traditional kernel principal component analysis and kernel entropy component analysis algorithm.