Research On Fuzzy C-means Clustering Algorithm And Its Application On Feed Property Identification Of Ethylene Cracking

Clustering is one of the core technology as data mining, and also is a way to analysis data to find useful information. In ethylene cracking process, the changes of feed have many kinds,and due to its expensive feed analyzer, little industrial site equips with it, so online recognition of oil property is important to achieve cracking online optimization.Fuzzy c-means clustering (FCM) method is based on objective function, it takes a certain weight to belong to each class, has the advantages of fast convergence speed, simple algorithm and can deal with the characteristics of large-scale data, which attract the scholars’s attention, and has become a main research direction in the field of data mining. However, the traditional FCM method is still unable to overcome such as the initial center of sensitive, easy to fall into local optimum, the unity of clustering results. This article improve performance of algorithm from the initial center selection, membership setting and other aspects of the analysis. The main research work and innovations are as follows;1.As clustering performance is sensitive to the initial value of cluster center, article combines with the effectiveness indicators, and utilizes indicators weighted to select better clustering center next iteration as the initial center computing. The algorithm makes full use of the prior information within historical information and optimize the initial center selection process, improve the convergence speed and the accuracy.2.As the traditional fuzzy C-means algorithm is based on the membership of the strike Euclidean distance, the algorithm contains only the mean center, bringing the unity of clustering results. To take full advantage of effective information of cracking feed, this paper proposes a fuzzy membership set method based on hybrid probabilistic model, namely through the establishment of Gaussian mixture model to achieve describing the probability distribution of clustering sample’s affiliation, and use EM algorithm to estimate the model parameter’s pole maximum likelihood. The algorithm can not only consider mean center of the sample, but also effectively use sample covariance and the weight coefficient information for mode discrimination.3.As oil feedstock has a lot of characteristic attributes, and each attribute influence the yield of product, a stepwise regression algorithm was used to grab the main attributes which influences the product yield most. At the same time, the main attributes were used in the improved clustering algorithm on oil feedstock,which is verified the validity and consistency with yield analysis.