基于最小信息熵损的KLPP算法在化工监控中的应用

Chemical process monitoring using KLPP based Minimum Entropy Loss

针对化工过程的非线性以及过程的动态特征,本文开发出了一种基于最小信息熵损的核局部保留算法(MEL-KLPP)。算法优点:①能够有效提取过程中的信息,建立准确的统计模型②在降维过程中考虑了样本之间的关联信息,所得模型更加符合实际。将算法应用于润滑油重质过程以检验其故障检出能力,结果表明MEL-KLPP算法的误报率和KLPP相近,低于KPCA,故障检出率(81.30%)高于KLPP(3.25%)和KPCA(69.7%)。将过程收集的数据根据工艺知识进行分块建模后,KLPP算法的故障检出率显著提高,MEL-KLPP检出率变化不大,表明KLPP算法对强噪声的复杂数据并不适用,MEL-KLPP算法对数据质量的要求不高,算法鲁棒性好,具有更广阔的应用前景。

A new technique named Kernel Locality Preserving Projection based Minimum Entropy Loss （MET-KLPP） is developed to deal with the complex and nonlinear problem for chemical process monitoring,Kernel function and the minimum entropy loss was introduced to the The traditional methods named Locality Preserving Projection（LPP） at the same time. Comparing with other statistical process monitoring methods, MET-KLPP has two advantages in the process of dimension reduction .First, MET-KLPP method considers both the transition matrix eigenvalue and eigenvector, this can be more effective to reveal the essence of data feature and extract more effective information from the data .Second The relationship between samples are considered, so the model was developed using this method is more conform to the actual process. MET-KLPP was test using industry data from a lubricating replacement process to check its effectiveness,the lubricating including furfural refining process and ketone-benzol dewaxing process, the results are compared with other two methods Kernel Locality Preserving Projection （KLPP） and Kernel Principal Component Analysis （KPCA）. Fault alarm rate of them is very similar with KLPP （4.31 %）, MEL-KLPP （3.68%） and KPCA （6.40%）. The fault detection of MET-KLPP is 81.30%, which is higher than KLPP （3.25%） and KPCA （69.7 %）. In order to monitor the process and test the methods better the industry data was divided and modeling separately based on the technology knowledge , the fault detection rate of KLPP increased a lot, with 79.06 % for furfural refining process and 13.01 % for ketone-benzol dewaxing process, while MET-KLPP is almost the same with 81.30%for furfural refining process 6.10%for ketone-benzol dewaxing process, the fault detection rate of KPCA is not change two much but it is more lower than MET-KLPP, this means MET-KLPP have a good robustness and a wide application prospect.