Research On Knowledge Discovery In Batch Process

With the swift development of the information technology, all trades and fields are all producing a large amount of data constantly, traditional database technology and method are hard to find the potential and useful knowledge, relations and rules exist in the data, so a phenomenon "data is affluent but knowledge is poor" is as a result. For resolving this problem, Knowledge discovery in Data (KDD) and Data-mining Technology are presented to find the information resources. Batch process is a kind of typical production process which producing batch products with ordered operating sequence, especially used in fine chemistry industry, food beverage and biological medicine trades. Compared with other processes, batch process has some characteristics, such as producing in batches according to recipes, non-stable state, resource-sharing, etc. Therefore batch process data have some characteristic such as multi-dimensions, strong-relations, non-linear and periodicity, etc. that against information express and deal withThis paper takes brewery saccharification process as the background, adopt KDD and Data-mining technology to find valuable knowledge from process data. Using the Fuzzy Cluster method, Hybrid Fuzzy Neural Network, Association rules mining methods, etc. find and excavate the recipes, periodic fouling, and operation strategy rule in the batch process.In general, the research work in this thesis could be summarized as follows:1.A quantitative analytical method for recipes in batch process was presented. Recipe as a variable with affluent and synthetic information has not quantitative analytical method before. This paper gave a recipe distance measurement method and recipe class distance measurement method based on hybrid variables measurement.2.According to distribution characteristic of recipes, a recipe fuzzy cluster algorithm based on kernel-function was presented. Firstly one recipe kernel-function was defined to represent recipe class, through minimizing all the distance of recipe samples to recipe class kernel, recipe samples were classed. The class number was gave out and each recipe was gave membership degrees belong to each classes. A membership degree matrix was presented in the end.3.A Hybrid Fuzzy Neural Network modeling method was presented. On the basis of CCT fuzzy neural network and fuzzy cluster method, this method can deal with discrete variables input. The fuzzy layer of the Hybrid FNN include two kinds of neurons, one is Gaussian fuzzification neuron which used to give the continuous input an fuzzy membership value, another is a presented fuzzy cluster neuron which also used to give the discrete input an fuzzy membership value. The number of fuzzy cluster neuron is equal to the fuzzy cluster number, and the neuron output is one element of the membership matrix.4.Use Hybrid fuzzy neural network modeling method and recipe fuzzy cluster method, a robust method to recipe-changing was presented. A multi-continuous parameters and one-recipe to one-output fuzzy neural network was designed to model the fouling in batch process.5. According to batch heat exchanger, washing and CIP (Cleaning in Place) operation cause its fouling and object gain changing periodically and thus conventional fixed parameter controller is hard to keep good performance, a novel recipe-FNN-based predictive method was developed for the periodical fouling measurement. Reversible fouling and irreversible fouling modeled the periodical fouling. Two MISO (Multi Input Single Output) four-layer recipe hybrid fuzzy neural networks were trained to learning the short-term reversible fouling growing trend and long-term irreversible fouling growing trend respectively. The combination of two network outputs provides the prediction for overall fouling. The results of experiments show the method has a better performance on a wort evaporator fouling prediction than the experiential formula. Using the predicted fouling values, a compensation factor for time-varying gain is applied succes...