Research On Missing Data And Application In Soft-Sensing

In the field of each investigation and research, because of various factors, we often have to face with missing data. It not only affects the follow-up work of researchers, but also impacts the estimation accuracy. The approach in this field has always been a hot topic. Our purpose is to make full use of the existing materials scientifically and reduce the negative effects of the missing data. It has been one of the most difficulty and hottest topics in all scientific experiments. With the background of missing data in industrial process, this thesis introduces a method to solve the problem, that is:impute the missing data and get the complete data set, then analyze them to continue the follow-up work. The main researches are described as follows:(1) In the deep understanding of the industrial data characteristic, the causes and the classification of missing data in this field are put forward.(2) Aiming at the different classification of missing data, the common imputation methods are put forward, which includes two categories:single imputation and multiple imputation. The corresponding examples are also given. Simulation results show that we can get the satisfying result by the proper method.(3) With the background of penicillin fermentation process, this thesis proves that the data imputation methods are effective for the soft-sensing. Firstly based on the original training data set, data set with missing data is established artificially and arbitrarily. Then the proper imputation methods are selected to deal with the missing data. Finally, the predicted models for bacteria concentration, substrate concentration and production concentration are established by the three mentioned types of modeling data sets.Simulation results show that when the modeling data set contains missing data, the precision of soft-sensing will be improved obviously by using the proper imputation method.