基于ART2神经网络的聚类方法及其应用研究机械电子工程专业论文.docxVIP

哈尔滨工业大学工学硕士学位论文 哈尔滨工业大学工学硕士学位论文 II II Abstract Clustering analyse is one of the ways to know things. We can know the essential characteristics of things more clearly by clustering analyse. There are several means of clustering analyse now, such as hierarchical clustering method, partitioning clustering method and neural network clustering method. Neural network is a new clustering way which has been studied extensively and developed rapidly recent years. Some kinds of neural network such as self- organization mapping(SOM) and adaptive resonance theory(ART) network are all can be used as clustering methods. In this paper, ART2 neural network is chosen. Comparing to some kinds of neural network, ART2 neural network will not destory patterns already stored by system when learning new patterns. So it can quickly recognize, stored patterns and cluster patterns. It also can harmonize stability and flexibility of network by modifing vigilance parameter. Because traditional ART2 network only calculate the weight by current sample, this may cause pattern drifting problem. And the normalization makes the magnitude of samples lost. The clustering capability of traditional ART2 network is not very good if magnitude is very important information to the samples. These problems limit the use of ART2 in clustering. In order to solve pattern drifting and magnitude missing problems, a better weight modification algorithm is used and a new magnitude comparison method is added to the ART2. And we use Eular distance as a re-clustering method to improve the clustering performance. At last, perform experiments using bearing status samples. Discuss important parameters of ART2 and find out the rule of choosing the value. The experimental results show that the method proposed in this paper is better than three other ART2 clustering methods. Keywords clustering, neural network, ART2 III III 目录 摘要 I HYPERLINK \l _bookmark0 Abstract II HYPERLINK \l _bookmark1 第 1 章 绪论 1 HYPERLINK \l _