Font Size: a A A

Signal Classification Based On Hidden Markov Model

Posted on:2012-08-26Degree:MasterType:Thesis
Country:ChinaCandidate:Z H ZhaoFull Text:PDF
GTID:2178330338997754Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
Signal classification has always been a research focus. And how to classify signal accurately has been one of the objectives pursued by many researchers. Classification plays an important role in fault diagnosis and texture image retrieval .Used in fault diagnosis, classification can identify failure signal to determine the causes, thereby locating the faulty equipment; while in texture image retrieval, a good classification algorithm can provide fast and accurate image to meet customer requests. In recent years, the study about wavelet theory and hidden Markov models have been making advance by leaps and bounds. The wavelet-domain hidden Markov model is a combination of both, which has become a hot topic as well.Based on the study of wavelet-domain hidden Markov model, this article discussed the impact the choice of wavelet domain exerts on the classification of signal, as well as the effect that hidden Marko model has on signal classification.It's application in both one-dimensional signal and two-dimensional signal classification. Gear fault signal and texture image retrieval are respectively studied. The work of this thesis is as follows:1) In gear fault diagnosis. We have proposed a method based on wavelet-domain HMM model. In this paper, and by establishing a binary tree hidden Markov model structure, we use the coefficients obtained from wavelet transformation directly, to train the hidden Markov model parameters. Through compute the similarity between test samples and the model, fault signal categories are thus identified. Wavelet coefficients of signal can be a very good reflection of macro and micro information, which ,when use to train parameter model ,can achieve more accuracy; while the binary tree hidden Markov model can well describe the correlation of wavelet coefficients between father and son.Through simulation experiments, this approach has achieved good classification results.2) In the texture image retrieval, we proposed a Hidden Markov Tree algorithm basing on Non-aliasing Contourlet Transform and discussed through two aspects, i.e. feature extraction and similarity measurement. In terms of feature extraction, the paper discussed the advantages and disadvantages of Contourlet Transform and Non-aliasing Contourlet Transform, the scale and direction information on the well-functioning of Non-aliasing Contourlet Transform of HMT. As for similarity measurement, in the application based on HMM, the traditional KLD prevails in calculating the distance between the two models. But as the nature of the traditional KLD does not satisfy the triangle inequality, an improved model of KLD to measure the distance between two HMM is adapted consequently. By comparative experiment, it's concluded that different choices of wavelet domain can improve the texture retrieval rate, and improved KLD distance in the mean time can enhance the accuracy of similarity measurement, hence improve the retrieval rate.
Keywords/Search Tags:wavelet transform, Hidden Markov Model, Non-aliasing Contourlet Transform, KL distance
PDF Full Text Request
Related items