Research On Application Of Self-organizing Map Neural Network In Information Processing

Artificial neural network is one of the most important research topics in artificial intelligence, and an important study method in the field of information processing. With highly parallel, distributed storage, fault tolerance, nonlinear structure, variability and other characteristics it becomes an important technology in the field of intelligent information processing. Self-orgnizing map(SOM) neural network is an unsupervised artificial neural network, it has many unique and effective characteristics, such as order preserving mapping, data compression, feature extraction, etc., all these characteristics make SOM be able to apply for classification and recognition in many fields. In this paper, on the basis of its characteristics, further construction and mining are carried out on ares of signal processing, medical field and weapons field. The detailed contents are as follows:(1)Directions of arrival(DOA) of arrary signal are estimated based on SOM’s character of keeping topological order. For the DOA estimation problem of sigle arrary signal, we build two 2-D array processing models for detection of DOA: uniform linear array and arbitrary array, and both of them are based on DDOA vectors. Take the relationship between DDOA vector and angle of arrival(AOA) as a mapping from DDOA to AOA, and then through analysis and deduction, we get the conclusion that the topologies of them are similar. Therefore, it is not nessesary to get the specific mapping expression between DDOA and AOA, and a two-level SOM neural network is enough to simulate the map, for the sake of that similar DDOA maps to similar AOA. Further, combined with the Lipschitz condition, the reliability of the system is verified and analyzed in detail. The network is trained by simulated signals of uniform distribution in advance, and then used to test by simulation signals without noise, simulation signals with Gaussian noise, and signals of lake experiment, furthermore, the results of test are Compared to MUSIC and RBF, we get this conclusion thar the two-level SOM network performs fast and accurate, which makes it can be considered to implemented in real-time.(2)The potential information genes of colon cancer location by using SOM’s character of automatic classification are discussed. Whether it is feasible to predict cancer by gene expression is a controversial issue. In this chapter, we establish the location prediction system based on the random selection of gene expression by using the character of classification of SOM neural network. Though choosing different amount of genes and different genes combination, together with the principle that low ration of misclassification means potential information genes, we get some potential information gene groups. Then these gene groups are taken for nonlinear principal component analysis, Analysis of intersection operation and cross validation, we get the conclusion that the potencial information gene groups of colon cancer location are not unique, and it is not feasible to classify the location of colon cancer by using gene in clinical.(3)We take optimized auto regression self-organizing(AR-SOM) neural network to fit the curve of bore pressure of gun and the maximum bore pressure for the function partition fitting character of AR-SOM. The gun pressure variation research has an important role in the study of gun. This chapter we use BP neural network and AR-SOM neural network sepretely for the prediction and fitting of measured data of a certain type of artillery gun pressure, the experimental results show that AR-SOM can better grasp the changes of bore pressure in each stage, and has good prediction effect.The main innovations of this paper are as follows:(1) We propose a DOA estimation method based on two-level SOM neural network, and verify the reliability of the method. The proposed method is based on the theory that DDOA and AOA have the same topological structure, which is suitable for both uniform linear array and arbitrary array. The network can be trained in advance; in application, AOA estimation is rapid, accurate, little influenced by noise and good robustness.(2) We established the SOM loop classification system, and analyzed the relationship between the gene and different kinds of colon cancer, and drew the conclusion that the gene can not be used to predict the position of colon cancer.(3) We applied optimized AR-SOM neural network to bore pressure curve fitting, improves the precision of curve fitting, and provides a powerful method for limited bore pressure from the grasp of the law.