Hybrid model development methodology for industrial soft sensors

Abstract

Soft sensors are essentially on-line models that provide an estimate of a desired process variable that is not easily measured directly, on the basis of other process variables that are directly measurable and are continuously available. The paper describes a novel methodology for the development of robust sensors that integrates various techniques: stacked analytical neural networks (SANN), support vector machines (SVM), and genetic programming (GP). Advantages of this hybrid approach include: (a) direct implementation in a distributed control or process information system; (b) explicit input/output relationships and thus easier interpretation; and (c) robustness and reliability due to the built-in performance indicators. The proposed approach is of special interest to transition control.