基于拉普拉斯正则化概率主元分析的故障检测

概率主元分析(PPCA)及其扩展方法用于过程监测时,只提取了过程数据的全局特征,并未考虑数据的局部结构.当数据的流形结构复杂时,传统的全局建模方法难以获得准确的预测效果.提出了一种基于拉普拉斯正则化的概率主成分(LapPPCA)模型,将数据的流形结构引入到传统概率模型的似然函数中,使得LapPPCA能够同时提出数据的全局和局部特性.同时提出了基于LapPPCA的过程监测模型,并在田纳西-伊斯曼(TE)过程上验证了该方法的有效性.

Laplacian Regularized PPCA for Fault Detection

When the traditional probabilistic principal component analysis(PPCA) and its extended methods were used for process monitoring, the global characteristics of the process data were extracted, while, the local structure of the data was not taken into account. When the manifold was complex, the local information needed to be incorporated into the traditional model so that the model prediction could be more accurate. An Laplacian regularized PPCA (LapPPCA) model was proposed for containing both global and local information of the data. Using graph Laplacian, the manifold was introduced into the likelihood of the conventional probabilistic model and EM algorithm. The process monitoring schemes based on LapPPCA were also developed and the case study on TE benchmark indicated that the proposed method was effective for both feature extraction and fault detection.