Hybrid Model Based Soft-Sensor Techniques Research And Applications In Fermentation Process

Soft-sensor technology is a kind of valid method to solve complex measuring tasks. Especially in bioengineering, the measuring tasks have the characteristics of non-linear, time-variant, high-dimension, unstructured model, ignorance of principle mechanism, and scarcity of experimental data, etc. This makes it a very important job to restudy and extend the soft-sensor techniques.Currently, there are no clear definitions for soft-sensor technology, and it is considered to solve the primary problems of modeling and identifying. One method called "white-box" modeling just gather the most important process mathematical relations, another method called "black-box" modeling only approximates the process outside behavior. The mixed modeling method attempts to take advantage of the above two methods, but the selection of secondary variables and model structure is haphazard. The research of hybrid modeling method, which could make the most of prior knowledge and experimental data, has great value in theoretics and engineering.In this dissertation, the research focuses on three aspects, includes modeling structure, learning machine and filtering technique.Firstly, the soft-sensor concept is discussed from the view of generalized information theory, and a novel hybrid model structure is proposed in this paper. The hybrid model is a combination of non-linear algebraic equation group and differential or difference equation group, and possesses the ability to integrate equations in inhomogeneous expression form. Three essential problems are studied in chapter two. The first is the unique existence condition of model root, and some relative concepts such as the amount of knowledge utility, etc. The second is the relationship between measuring accuracy and model reliability, which is analyzed from the view of error theory by introducing model uncertainty concept. The third is the expression capability of some kinds of traditional soft-sensor models.Support vector machine method is a fine realization of statistical learning theory, some kinds of statistical modeling methods based on the support vector machine technique are researched. After theoretic analyzing support vector machine technique, the equipollence of support vector classification and regression problems is proofed, and a uniformed expression of support vector machine problem is gotten. The noise stained case is studied emphatically, and the experiment shows that a better result will be obtained by assigning weight factor to fit noise distribute characteristic. A multi-multiplier minimal optimization algorithm for norm-1 problems is deduced, and a kind of improved Gilbert geometric algorithm for norm-2 problems is introduced, which is applied in regression problems successfully.After the research of filtering techniques, the filter based on hybrid soft-sensor model is presented. By considering the filter robustness on modal, a kind of robust Kalman filter is studied. The proof shows that the innovation covariance matrix contains the information of model perturbation, and the one-step linear prediction of non-linear model could be more precise after compensation. The strong tracking filter proposed in current reference is a moderate implementation. A novel algorithm, unscented transformation robust Kalman filter, which use the unscented transformation to calculate the non-linear transform expectation and covariance is developed. Further, applying the improved Kalman filter to the hybrid model, a complete soft-sensor system is presented.At the experimentation part, the soft-sensor system implemented in Matlab program is put into the practice of biomass estimation in fermentation process. There are three main objectives. The first is to approve the better regression result based on weighted support vector machine. The second is to approve the robustness of unscented transformation robust Kalman filter on initial value and model perturbation, and to validate the new filter provided better performance. The third is to approve the validity of the presented hybrid model under the condition of uncertainty in prior knowledge or imprecision in experiment data.This dissertation proposes a theoretically perfect and practically convenience hybrid soft-sensor model and the method to construct the soft-sensor system based on it. This system avoids the adverse effects, and can improve the measuring accuracy, precision and reliability. It greatly improves the validity of the utility of the prior knowledge and experiment of the target system and primary variables, and can be applied effectively in complex measuring tasks.