符号序列的预训练HMM分类方法

隐马尔可夫模型(Hidden Markov Model,HMM)是一种双重随机概率模型,已广泛应用于序列数据建模.针对符号序列分类中距离度量定义的困难,提出一种符号序列的预训练HMM分类新方法.首先,定义一种基于HMM状态转移矩阵的序列距离新度量;其次,为得到不同序列在HMM隐状态共享条件下的状态转移矩阵,提出一种两阶段的预训练方法,先在所有序列上进行HMM预训练以学习所有序列共享的隐状态,再使用共享状态为每条序列进行训练得到各自的状态转移矩阵;最后用近邻分类器对符号序列进行基于距离的分类.在三个应用领域的真实序列上进行了实验,并与基于子序列、HMM变体模型等现有分类方法进行对比,结果表明,所提出的方法能使用较低的特征维度取得较理想的分类精度.

Pattern recognition and prediction study of rock burst based on neural network

Many monitoring measures were used in the production field for predicting rockburst.However, predicting rock burst according to complicated observation data is alwaysa pressing problem in this research field.Though the critical value method gets extensiveapplication in practice, it stresses only on the superficial change of data and overlooks alot of features of rock burst and useful information that is concealed and hidden in the observationtime series.Pattern recognition extracts the feature value of time domain, frequencydomain and wavelet domain in observation time series to form Multi-Feature vectors,using Euclidean distance measure as the separable criterion between the same typeand different type to compress and transform feature vectors.It applies neural network asa tool to recognize the danger of rock burst, and uses feature vectors being compressedto carry out training and studying.It is proved by test samples that predicting precisionshould be prior to such traditional predicting methods as pattern recognition and critical indicatormethod.