“Assumed inherent sensor” inversion based ANN dynamic soft-sensing method and its application in erythromycin fermentation process

Abstract

An artificial neural network (ANN) soft-sensing method, based on the “assumed inherent sensor” and its inversion concepts, is proposed and used to estimate some crucial process variables which would be very difficult to be measured directly. For a real biochemical process whose mathematical model is a general nonlinear dynamic system, one may assume that, in its interior, there exists an “inherent sensor” subsystem whose inputs are exactly the process variables to be estimated while whose outputs are the directly measurable ones. To verify this assumption, this paper presents an algorithm to construct the mathematical model of the “assumed inherent sensor” and furthermore presents a global invertibility condition of the “assumed inherent sensor” which guarantees the existence of the inversion of such an “assumed inherent sensor” in theory. The “assumed inherent sensor” inversion consists of a set of nonlinear functions and a series of differentiators and could be treated as the dynamic soft-sensing model because its outputs are capable of reproducing the input variables of the “assumed inherent sensor”, or the process variables to be estimated. To overcome the difficulty in constructing the above “assumed inherent sensor” inversion in an analytic manner, a static ANN is used to approximate the nonlinear function so that the ANN-inversion dynamic soft-sensing model or the desired soft-sensor is finally completed. This makes the proposed ANN-inversion soft-sensor stricter in construction principle and more credible in practical use than most proposed soft-sensors. The soft-sensor consisting of a static ANN and a set of differentiators has been put into use of estimating such crucial biochemical variables as mycelia concentration, sugar concentration and chemical potency in erythromycin fermentation process. The field results show that the soft-sensing values approximately coincide with the offline analyzing ones sampled from the production process.