Research Of The Data Clustering Method And Its Application In Soft Sensor

As a non-supervised learning method, cluster analysis has been widely applied in the area of production as well as daily life with the development of computer science. In the theoretical research and practical application, many clustering algorithms have been put forward, but each of them has some advantages as well as disadvantages. There has no universal method for every situation so far.As a completely developed clustering algorithm which has been widely employed, the FCM also has many disadvantages such as the possibility to fall into the local minimum while iterating, the uncertain caused by the categories number determination, low speed of convergence, and so on. In this thesis, two disadvantages of the FCM are analyzed--one is the determination of the cluster categories number, and the other is the iterating speed. An improved algorithm is proposed in this paper, and the application of the new algorithm has proved its advantages in the soft sensor modeling. The following part shows an overview of the thesis:Firstly, the background of the clustering research methods and the theoretical basis of soft measurement technology, and the development and present research situations of cluster analysis are discussed.,. The discussion does not only cover four kinds of the realization methods and their improvements, but also pays attention to the key points and steps of clustering methods.Secondly, development of the fuzzy clustering and four kinds of the clustering realization are introduced, following which the establishment of the aim functions, fuzzy C mean algorithm, and the steps of FCM are introduced. The main problems of FCM algorithm are briefly analyzed as well.Thirdly, several fuzzy cluster validity indexes are discussed and process of cluster is shown. Then the improved algorithm based on the density function and S-FCM is put forward, and it is improved with regard to the number determination of the cluster initial categories and the convergence speed. The validity of the proposed algorithm is proved by the data and pictures in the experiment with Matlab.Fourthly, steps of the soft-sensing models building are discussed regarding the definition, application and math description of the soft sensor. Then processes about several common soft-sensing methods and their improvements are introduced.Finally, BP neural networks, RBF neural networks, and PLS are discussed in order to build a soft-sensing models. PCA is used in dimensionality reduction of high-dimensional data and dealing with massive datasets. At last, data of the collected oxygen is classified according to the proposed improved algorithm. Three categories cluster results are built respectively as the soft-sensing models by three categories data mentioned above.