Methods for Plant Data-Based Process Modeling in Soft-Sensor Development

Abstract

Original scientific paper There has been an increased use of soft-sensors in process industry in recent years. These soft-sensors are computer programs that are used as a relatively cheap alternative to hardware sensors. Since process variables, which are concerned with final product quality, cannot always be measured by hardware sensors, designing the appropriate soft-sensor can be an interesting solution. Additionally, a soft-sensor can be used as a backup sensor, when the hardware sensor is in fault or removed due to maintenance or replacement. Soft-sensor is based on the mathematical model of the process. Since industrial processes are generally quite complex, a theoretical modeling approach is often impractical, expensive or sometimes even impossible. Therefore, process model building is based on measured data. This approach significantly gets complicated if only plant data, taken from the process database, are available. In this paper the most popular methods for plant data-based modeling that appeared in the last two decades are summarized and briefly explained. Apart from giving a short survey of the most important papers, tips about choosing the appropriate methodology for process model building are also provided.

Adaptive Estimation of Difficult-to-Measure Process Variables

“…Rapid expansion of artificial neural networks and other intelligent methods led to the expansion of different hybrid methods [3][4][5]. Estimators based on these (static) models usually have fixed parameters.…”

“…In prediction model building, the parameters are usually determined by regression in which all model parameters are estimated based on minimization of the output Adaptive Estimation of Difficult-to-Measure Process Variables D. Slišković, R. Grbić, Ž. Hocenski error of approximation. However, this approach generally fails when plant data are available for modeling due to low plant data informativity as already stated in Section 1 [4]. Additionally, to ensure necessary robustness of the softsensor, a great number of ETM variables are used in modeling which results in a high dimensional and highly correlated input space [17].…”

“…process model) that approximates the unknown natural functional dependence between process input-output variables. This procedure involves different steps (see [1,3,4]) and it is often performed in several iterations.…”

“…Apart from the selection of appropriate functions, this model structure requires the definition of two separate criteria for parameter estimation [17]. There are a number of different linear or nonlinear methods available for model building which differ in the selected functions and criteria that are used in the model (2) [4,17].…”

“…The most economic way to build a process model is to use measured data which can be taken from the plant database. Various statistical methods and/or computational learning methods can be used in model developing [3][4][5]. mativity.…”