Batch bioprocess monitoring using multiway localized discriminant embedding approach

Abstract

We propose a new batch bioprocess monitoring approach which combines the Gaussian mixture model (GMM) with the multiway analysis in localized discriminant embedding subspace for detecting and classifying different types of faults. Due to the inherent process multimodality within abnormal points, the capability of traditional multimay locality preserving projection (MLPP) is degraded. For the routine bioprocess operation with abnormal events, GMM is used to partition the training set with various types of faults into different clusters. Three localized distance matrices reflecting the relationships of nearby points are then computed. The extracted leading discriminant embedding directions can not only separate the normal and faulty data by maximizing the distance among nearby data points from different clusters, but also preserve the intrinsic geometrical structure within the multiple faulty clusters by minimizing the small distance and penalizing the large distance among neighboring data from the same cluster. The newly developed multiway localized discriminant embedding (MLDE) approach is applied to two test scenarios in the fed-batch penicillin fermentation process and compared with the conventional MLPP method. The results demonstrate that the MLDE approach is superior to the MFDA method in detecting abnormal events and classifying different types of faults in fed-batch process with higher accuracy and sensitivity.