Research On Improved K-means And Self-organizing Map Neural Networks And Their Applications

Clustering analysis has many broad applications in pattern recognition,image processing,machine learning and statistics.K-means algorithm is one of the popular clustering methods in clustering analysis.With the rapid development of artificial neural network,the applications of artificial neural network become more and more widely.Self organizing maps(SOM)algorithm in neural network is often used for pattern recognition.In this paper,we study the k-means algorithm and SOM algorithm,and apply them on face recognition,defect detection and intrusion detection.The detailed contents are as follows:(1)In recent years,the global k-means algorithm is proposed,which dynamically adds one cluster center through a deterministic global search procedure from suitable initial positions,and employs k-means algorithm to minimize the sum of the intracluster variances.However,at sometimes the global k-means algorithm results in singleton clusters and the initial positions are bad.This can leads to the poor local optimist.In this paper,we modify the global k-means algorithm to eliminate the singleton clusters at first,and then we apply MinMax k-means clustering error method to global k-means algorithm in order to overcoming the effect of bad initialization,finally we propose the global MinMax k-means algorithm.The proposed clustering method is tested on some popular data sets and is compared with the k-means algorithm,the global k-means algorithm and the MinMax k-means algorithm.The experiment results show that our proposed algorithm outperforms other algorithms mentioned in the previous literatures.(2)The MinMax k-means algorithm is widely used to tackle the effect of bad initialization by minimizing the maximum intraclustering errors.The exponent parameter and memory parameter are involved in the executive process.Since different parameters result in different clustering errors,it is crucial to choose appropriate parameters in order to mitigate the bad effects.In the original algorithm,a practical framework extends the MinMax k-means to automatically adapt the exponent parameters to the data set.Generally,the program can reach the lowest intraclustering errors if the maximum exponent parameter has been set.In fact,this is not always correct.In this paper,we modify the MinMax k-means algorithm by particle swarm optimization to determine the proper parameter values which can subject the algorithm to reach the lowest clustering errors.The proposed clustering method is implemented on some favorite data sets in several different initial situations and is compared to the k-means algorithm and the original MinMax k-means algorithm.The experimental results indicate that our proposed algorithm automatically reaches the lowest clustering errors.(3)SOM neural networks always get more classifications than what we want,in order to tackle this defaults,we propose two algorithms: improved SOM neural networks based on hierarchical algorithm and on k-means algorithm,respectively.We apply the two proposed algorithm in face orientation recognition and intrusion detection.Our results show that the proposed algorithms have better clustering performance compared with learning vector quantization,fuzzy c-means and fuzzy clustering based generalized regression neural network algorithm.(4)Sparse Subspace Clustering constructs a sparse similarity graph by using the coefficient of sparse representation to subspace clustering.Based on sparse representation techniques,the algorithm gets the sparse coefficient by using l1-minimization and gets clusters' data by spectral clustering algorithm,respectively.As we known,the spectral clustering algorithm depends on k-means algorithm which is sensitive to the choice of initial starting conditions and it needs iterations for clustering.In order to avoid the drawbacks of k-means algorithm,we propose two modified Sparse Subspace Clustering algorithm,which are not affected by the centers and their iterations.In one of the methods,we get the clusters through comparing the positions of nonzero elements in the sparse adjacent matrix of similarity graph with the eigenvector.And in the second method,we use modified SOM algorithm instead of k-means algorithm and the experimental results show our proposed algorithm on face recognition outperforms the initial Sparse Subspace Clustering.(5)The classification for metal surface defect plays an important role in laser ultrasonic testing.The surface wave is classified by using three kinds of neural networks,i.e,self organizing map competition neural network,learning vector quantization neural network and probabilistic neural network.Several experiments on different input situations for the three kinds of neural networks are discussed.Experimental results indicate that the three kinds of neural networks have good performances in classification.