Study Of Multi-Model Soft Sensor Modeling Method And Application In Fermentation Process

Microorganism fermentation engineering is a complex biochemical reaction process which has high nonlinear, time-varying and relevant character. So if we want to study microbial fermentation process and increase productivity, the most important prerequisite is to get more state variables and master more responsive process information. However, in the actual fermentation process, due to technical and technological constraints, on-line measurement of a number of important biological parameters is difficult, and soft-sensor technique provides an effective way to solve this problem.In recent years, considering the characters such as high dimension of inputs, nonlinearity and strong correlation between inputs and outputs of complex systems, the common modeling method is a single model based on neural network and fuzzy logic. However, the model would cause deviation if we use a single function fit the relationship between the secondary variable and the outputs, and without considering the links between the data sets. When the samples are huge, only using a function to establish the model will cause large network architecture, and the training would cost a longer time.This project is sponsored by national hi-tech research development plan "Soft Sensor Technology Based on Fuzzy ANN Inverse in Biological Process and Its Optimal Control". For the defects of single model, researches about multi-model have been done in soft measurement area to find effective schemes to solve parameter measurement problems of fermentation process. At first, principal component analysis (PCA) is used to pre-process the sample data, which can effectively remove redundant information among variables, reduce the relevance and complexity of the model. Then, clustering algorithms are proposed to deal with the pro-processed data set. The divided subsets are trained to build a sub-model using back propagation neural network (BPNN). In the part of classification, two different modified clustering algorithms are proposed. One is particle swarm-based kernel fuzzy c-means clustering (PSKFCM) algorithm, the other is affinity propagation clustering (AP) algorithm. In order to avoid the defects of traditional clustering algorithms such as the dependence on data distribution, sensitivity to initial value and noise and being easy to be trapped in local minima, we present PSKFCM algorithm which combines the advantages of kernel fuzzy c-means clustering (KFCM) and particle swarm optimization (PSO) algorithm. And AP algorithm is presented to solve the commonly existed problems in original clustering algorithms, such as clustering number should be determined in advance and clustering accuracy depends on data distribution. These two clustering algorithms are applied to group the training data into overlapping clusters. Finally, probability-weighted method is applied to fuse each sub-model to achieve the output result of soft senor model. The proposed modeling method is applied to predict the biomass concentration online for an erythromycin fermentation process. Case studies show that the approach has better performance compared to the conventional single model method. The comparisons between a single model and multi-model based on probability weighted show that the soft measurement proposed in the paper, can approach the expected result effectively, and can also improve the prediction precision.