Local linear regression for soft-sensor design with application to an industrial deethanizer

Soft-sensors for estimating in real-time important quality variables are a key technology in modern process industry. The successful development of a soft-sensor whose performance does not deteriorate with time and changing process characteristics is troublesome and only seldom achieved in real-world setups. The design of soft-sensors based on local regression models is becoming popular. Simplicity of calibration, ability to handle nonlinearities and, most importantly, reduced maintenance costs while retaining the requested accuracy are the major assets. In this paper, we introduce several approaches for defining an appropriate locality neighborhood and we propose a recursive version of local linear regression for soft-sensor design. To support the presentation, we discuss the results in designing a soft-sensor for estimating the ethane concentration from the bottom of a full-scale deethanizer.