Unsupervised Feature Selection Based on Fuzzy Clustering for Fault Detection of the Tennessee Eastman Process

The large number of components involved in the operation of industrial processes increases its complexity, together with the likelihood of failure or unusual behaviors. In some cases, industrial processes solely depend on plant operator experience to prevent and identify failures. It has been shown that automatic identification of failures within functional states of the process brings support to the operator performance, reducing the incidence of accidents and defective products. However, increasing use of automatic measurement systems generates large amounts of information that hinders fault detection. Obtaining adequate fault identification systems requires the use of the most informative variables to cope with large amounts of data by intelligently removing redundant and irrelevant variables. In this paper, an unsupervised methodology based on fuzzy clustering is applied on fault identification of the Tennessee Eastman process. Results show that an optimal variable subset improves the classification percentages and avoid the use of unnecessary variables.