【作者】 周雄；

【导师】 唐一科；

【作者基本信息】 重庆大学 ， 机械设计及理论， 2008， 博士

【摘要】 论文从研究传统的旋转机械故障的特征参数识别两类状态的能力出发,提出了特征参数灵敏度评价标准,并对其进行了实验验证;在传统的特征参数基础上,文中提出利用遗传算法对特征参数进行自组织生成新的高灵敏度的特征参数,并采用遗传算法对由所有特征参数组成的特征集合进行选择,去除了冗余的特征参数;最后建立了基于神经网络的逐次诊断模型,降低了诊断系统输入量的维数,并通过实验验证了整个诊断系统的有效性。该诊断系统结合了特征灵敏度的评价、特征提取、特征选择和逐次神经网络模型,有一定创新性,且易于实现,具有良好的工程应用价值。现将论文主要的研究成果归纳如下:①提出了用于评价特征参数识别故障的灵敏度的概念。并假设特征参数概率密度函数是正态分布情况下,从理论上推导出了灵敏度的计算公式。最后通过实验验证了灵敏度与特征参数区分两种状态的识别率是成正比关系的,灵敏度高的特征参数就具有高的识别率。②传统特征参数不能很好的区分两种状态,文中采用遗传算法重新提取新的高灵敏度的特征参数;该方法是对传统的特征参数进行再组织生成新的特征参数,文中用树形图来表示特征参数的公式,便于利用遗传算法进行交叉和变异;用特征参数灵敏度作为遗传算法的适应度函数。最后实验验证了新的特征参数的识别率高,证明了该方法的有效性。③文中将传统特征参数和由遗传算法提取的新的特征参数共17个组成特征参数集合。提出了以类内和类间距离为适应度函数,基于遗传算法的特征选择策略,选择出最有效的几个特征参数来降维,同时达到提高识别两类状态识别率的目的。该方法能够充分利用遗传算法的隐并行性,有效地剔除原始特征集中冗余特征参数。并利用实验验证了将优化后的特征集合用于神经网络分类器训练,能够提高故障的识别精度。④构建了逐次故障诊断神经网络模型,并从遗传算法特征选择后的特征参数集中选出灵敏度较大的三个特征参数作为逐次诊断神经网络的输入,降低了输入量的维数。更多还原

【Abstract】 The ability of rotating machinery’s symptom parameters (SPS) in distinguishing two states is studied in this paper, the evaluation standard of SPS is put forward. Based on traditional SPS, a way of using genetic algorithm (GA) to create new and high SPS by self-organization traditional SPS is proposed, the feature set composed of all s SPS is selected by GA , and redundant symptoms are removed by this way. The sequential diagnosis model based on neural network (NN) is build up, the dimension of input is decreased. The practical example of condition diagnosis is shown to verify the efficiency of the method proposed in this paper. The diagnosis system combined evaluation of SPS sensitivity, feature extraction, feature selection and sequential diagnosis model. This paper has innovation and engineering value. The main results and conclusions is presented as follows:①The sensitivity of SPS is proposed to evaluate the ability of distinguishing two states. And the formula sensitivity is derived theoretically. We verifies that parameters sensitivity is proportional to the distinction rate (DR). SPS with high sensitivity have best DR.②The sensitivity of traditional SPS is not high enough. We extract new and high sensitivity SPS with GA. SPS are self-organization and new SPS are created in this way. Arborescence is expressed to SPS formula and then intercrossed and variated by GA. Sensitivity of SPS is used as fitness.③In this paper, the feature set is composed of traditional SPS and new SPS created by GA. We use inner and outer distance as fitness function and select most effective SPS by GA. It can decrease the feature set’s dimension and increase the precision of SPS distinguishing two states. Redundant SPS are removed.④The sequential diagnosis model is build. Three SPS than have high sensitivity are selected and used as the inputs of sequential diagnosis neural network. Dimension of inputs is decreased.更多还原