Batch process monitoring with Gaussian mixture model in neighborhood preserving embedding subspace

An improved GMM (Gaussian mixture model) based batch process monitoring approach is proposed in this article to handle batch processes with multiple operating phases. GMM is an effective tool to construct monitoring models by estimating separate probability density functions of the nominal batch data. However, the existing GMM based monitoring method has the following disadvantages: (1) GMM utilize all the observed variables for online monitoring, which are computationally intensive for complex processes with dozens of variables. (2) Different measure units of variables will impact the monitoring results significantly since there is no auto-scaling procedure. (3) The trajectory of faulty is likely to fall within normal areas of other Gaussian components, which will leads to obvious false negatives. To overcome these deficiencies, an NPE (neighborhood preserving embedding) algorithm is introduced to generate an enhanced monitoring subspace, which not only facilitates the computational burden of training and utilizing GMMs, but also improves sensitivity to incipient fault symptoms. The efficiency of the proposed method is verified through a simulated fed-batch penicillin fermentation process.