Local Entropy Principal Component Analysis and Its Application for Multimode Process Monitoring

Aiming at the multimode industrial process monitoring problem, this paper proposes a fault detection method based on local entropy principal component analysis (LEPCA) algorithm. Firstly, in order to deal with the multimode characteristic of operating data, k nearest neighbor Parzen window (kNN-Parzen) method is used to estimate each sample's local probability density. Then, a local relative density estimate function is constructed to decrease the sensitivity to window width parameter. Lastly, the local entropies of process data are calculated by using information entropy theory to extract the feature information effectively, and principal component analysis (PCA) is applied to establish the local entropy component statistical model for fault detection. Simulation results on one continuous stirred tank reactor (CSTR) system indicate that LEPCA can provide superior performance in process monitoring.