Deep model based on mode elimination and Fisher criterion combined with self-organizing map for visual multimodal chemical process monitoring

Abstract

Multimode feature is widely adopted in complex continuous chemical processes to meet changes in market demand. However, conducting effective process monitoring in a multimode chemical process is challenging because data usually have multimodal distribution. In this study, DMF, a novel model based on a deep network, is proposed for learning new feature spaces; the multimodality of the mix data is eliminated, the same faults from all operating modes are assembled, and different faults are separated from each other. The proposed model uses a three-layer stacked autoencoder to extract features from the mixed data, uses a mode elimination term to remove data multimodality, and adopts a Fisher criterion term to separate the normal and fault states. Subsequently, DMF is combined with a self-organizing map (DMF–SOM) for visual process monitoring. DMF extracts discriminative features from the original data, and SOM visualizes these features such that the normal and fault states are distinguished on the two-dimensional output plane. The effectiveness of the DMF–SOM in process monitoring is verified by a study on the multimodal Tennessee Eastman process.