Combining Conceptual Clustering and Principal Component Analysis for State Space Based Process Monitoring

Multivariate statistics and unsupervised machine learning have recently been studied by many researchers for process monitoring and fault diagnosis. These approaches often depend on calculating a similarity or distance measure to group data sets into clusters. Apart from giving predictions, they are not able to give causal explanations on why a specific set of data is assigned to a particular cluster. In this work, a conceptual clustering approach is presented for designing state space based monitoring systems, which is able to generate conceptual knowledge on the major variables which are responsible for clustering, as well as projecting the operation to a specific operational state. A critical issue in this approach is how to conceptually represent dynamic trend signals. For this purpose, principal component analysis is used for concept extraction from real-time dynamic trend signals. The method is introduced using a continuous stirred tank reactor as a case study. Application of the approach to a refinery methyl tert -butyl ether process is also presented.