The Application Of Hybrid Modeling Based On Neural Network In Fermentation Process

Microbial fermentation is a complex, uncertain, multi-leveled and unstructured system. With the rapid development of fermentation industry, it requests high quality in the area of on-line measurement and processing control. Especially, the technology of on-line measurement for biological parameters in fermentation process plays a key role for improving the fermentation industry.However, there is no appropriate solution of on-line measurement for biological parameters because of the limitations of bio-sensing technology. A lot of biological parameters can't be measured on-line, but still be measued by the off-line measurement. It provides a new method of the on-line measurement for biological parameters in fermentation process with the rapid development of soft-sensor technology. But because of the particularities of the fermentation process and the limitations of soft-sensor technology, the method based on the soft-sensor technology needs to be further analysed and studyed before it is used in actual fermentation process. For the on-line measurement of biological parameters, the principium-model which is simplified by the actual process can show the change of parameters, but difficult to ensure the accuracy of on-line measurement; while black box model which is modeled for the sample datas can fit the parameters accurately, but difficult to improve the generalization ability for the dependency of the sample datas.By analyzing the fermentation process and the soft-sensor technology, the paper proposes a hybrid soft-sensor modeling method based on data processing of BP neural network and principium-model. The simulation in MATLAB and experiments in pharmacy factory prove that the model based on the hybrid soft-sensor modeling method makes the biological parameters more accurate and has strong generalization ability.In this paper, the advantage of the hybrid modeling is that it improves the accuracy by dealing with the original datas, while enhances the generalization ability by using the principium-model to reduce the dependence of the sample datas.