状态加权合成的CGHMM训练算法

为了克服训练数据不足的问题,提出了一种新的方法——基于状态加权合成的连续高斯混合密度隐马尔可夫模型(continuous Gaussian mixture hidden Markov model,CGHMM)训练算法.首先对每一个待合并模型中的每个状态都选定一个权值,当对多个训练样本进行CGHMM参数重估时,每一次迭代过程都分别对每一个训练样本获取CGHMM参数,再使用仅仅取决于状态数的权值加以合并.最后,将此新算法应用于轴承故障诊断,并与经典CGHMM算法进行了比较.实验结果表明,新算法的诊断精度更高,输出概率更好,获得了更优的训练模板.

A training algorithm of CGHMM based on state-weighted synthesis

In order to overcome the shortcoming of deficiency in training data,this paper proposed a training algorithm of continuous Gaussian mixture hidden Markov model(CGHMM) based on state-weighted synthesis.Firstly,weights are calculated for model merging,choosing one weight for every state of every model to be merged.When CGHMM parameters are reestimated based on multi-training data,for each iteration of the procedure,CGHMM parameters based on sample training data are obtained by computations,and then merged based on weights only depending on state number of HMM.Finally,the new algorithm is used in bearing fault diagnosis,and is compared with the algorithm of classic CGHMM.Experiment shows that the diagnosis accuracy of this new algorithm is higher and the output probability is better than the classic algorithm.Thus the new algorithm has more superior training model.