A Hybrid HSIC-ACO Algorithm for Variable Selection in Process Engineering

Recently, data mining and machine learning techniques have been increasingly applied in process engineering. Various successful applications include fault detection and development of data driven models. While fault detection is useful for steady operation of the plant, data driven models can be employed for robust prediction of structure activity relationships. Many of these models require nonlinear classification techniques. The success of these techniques relies on the integration of informative domain knowledge to the concerned methods. In this study, we propose a hybrid Ant Colony optimization (ACO) based variable selection approach in conjunction with Support Vector Machines (SVM) to determine informative subsets of process variables that may help detect faults efficiently, making the fault detection model more robust in the process. In addition, we employ a Hilbert Schmidt Independence Criterion (HSIC) based variable ranking heuristic to guide ACO towards better search spaces. Performance testing of HSIC-ACO was carried out on the benchmark Tennessee Eastman Process challenge and large scale QSAR prediction data collected from relevant sources. Our results demonstrate improved fault detection and structure-activity prediction capabilities using the HSIC-ACO algorithm.