基于NRST的转子故障数据集属性约简方法研究
发布时间：2018-11-28 20:18
【摘要】：如何从海量数据中,挖掘出有用信息,寻找出数据之间蕴含的反映机械设备运行状况规律,解决复杂诊断建模难的问题,实现对故障模式智能化识别,成为当前急切需要解决的问题。然而采集到的反映复杂机械系统运行状况的工业数据往往夹杂着大量噪音,具有较强的非线性和耦合性,严重影响了有效信息的获取,且目前单一的故障诊断模型无法有效的对复杂的机械系统做出全面的诊断。针对以上问题,本文开展了邻域粗糙集理论(Neighborhood Rough Set Theory,NRST)与其它数据驱动方法结合的转子故障模式识别方法的研究工作,重点对NRST的属性约简方法及NRST结合统计分析及机器学习的方法进行了探讨。本文的主要工作概况及取得的研究成果如下:(1)介绍NRST的定义及前向贪心属性约简方法,充分利用了NRST能直接处理连续数值型属性的优势,提出以转子工频倍频为条件属性以故障类型为决策属性构建NRST决策表进行特征提取的方法并且结合典型故障类别的频率谱特性分析了可行性。实验结果也同时证明该方法获取的特征属性更符合物理意义,避免了离散化过程中关键属性的丢失。(2)在NRST属性约简的基础上,提出了NRST结合费舍判别(FDA)对故障类别进行分类的方法,求出了判别函数和累积判别能力,探讨了二次降维和去冗余后对故障模式识别的影响,完成了数据从高维到低维的映射,实现了低维下的故障分类效果。实验结果证明该方法在特征属性少的情况下能够达到同样的识别正确率,从而可以节省存储空间提高运算效率。(3)为了寻求高效、准确的故障诊断方法,同时也为了探讨NRST属性约简对机器学习的影响,提出了NRST结合径向基神经网络(RBFNetwork)对故障类别辨识的方法,选用高斯标准函数作为径向基函数,采用自组织选取中心法确定基函数中心、宽度及连接权重。实验结果证明该方法明显缩短了建模时间,提高了识别准确率,值得推广。(4)为了解决知识的存储与发现难的问题,推动智能化诊断技术的发展,设计了基于WEKA数据挖掘平台连接My SQL数据库的故障识别系统。完成了故障知识的存储、数据流的展现及WEKA中调用My SQL语句对数据库的访问。
[Abstract]:How to mine the useful information from the massive data, find out the rule of reflecting the running condition of the mechanical equipment contained in the data, solve the difficult problem of complex diagnosis and modeling, and realize the intelligent recognition of the fault pattern. Become the current urgent need to solve the problem. However, the industrial data collected to reflect the running state of complex mechanical systems are often mixed with a large amount of noise, which has strong nonlinearity and coupling, which seriously affects the acquisition of effective information. At present, a single fault diagnosis model can not effectively make a comprehensive diagnosis of complex mechanical systems. In order to solve the above problems, the research work of rotor fault pattern recognition based on neighborhood rough set theory (Neighborhood Rough Set Theory,NRST) and other data-driven methods is carried out in this paper. The attribute reduction method of NRST and the method of NRST combined with statistical analysis and machine learning are discussed. The main work and results of this paper are as follows: (1) the definition of NRST and the method of forward greedy attribute reduction are introduced. The advantage that NRST can directly deal with continuous numerical attributes is fully utilized. This paper presents a method of constructing NRST decision table based on rotor power frequency doubling as conditional attribute and fault type as decision attribute, and analyzes the feasibility of constructing NRST decision table based on the frequency spectrum characteristics of typical fault categories. The experimental results also show that the feature attributes obtained by this method are more physical and avoid the loss of key attributes in the discretization process. (2) on the basis of NRST attribute reduction, In this paper, a method of classifying fault categories with NRST and Fisher discriminant (FDA) is proposed. The discriminant function and cumulative discriminant ability are obtained. The effects of quadratic reduction and redundancy removal on fault pattern recognition are discussed. The mapping of data from high dimension to low dimension is completed, and the effect of fault classification under low dimension is realized. The experimental results show that the method can achieve the same recognition accuracy in the case of less feature attributes, thus saving storage space to improve the computational efficiency. (3) in order to seek an efficient and accurate fault diagnosis method, At the same time, in order to discuss the effect of NRST attribute reduction on machine learning, a method of fault classification identification based on NRST combined with radial basis function neural network (RBFNetwork) is proposed. Gao Si standard function is selected as radial basis function. The center, width and connection weight of the basis function are determined by the self-organizing selection center method. The experimental results show that the method can obviously shorten the modeling time and improve the recognition accuracy. (4) in order to solve the difficult problem of knowledge storage and discovery, the development of intelligent diagnosis technology is promoted. A fault identification system based on WEKA data mining platform is designed to connect My SQL database. The storage of fault knowledge, the display of data stream and the access of database by calling My SQL in WEKA are completed.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TH133;TP18
[Abstract]:How to mine the useful information from the massive data, find out the rule of reflecting the running condition of the mechanical equipment contained in the data, solve the difficult problem of complex diagnosis and modeling, and realize the intelligent recognition of the fault pattern. Become the current urgent need to solve the problem. However, the industrial data collected to reflect the running state of complex mechanical systems are often mixed with a large amount of noise, which has strong nonlinearity and coupling, which seriously affects the acquisition of effective information. At present, a single fault diagnosis model can not effectively make a comprehensive diagnosis of complex mechanical systems. In order to solve the above problems, the research work of rotor fault pattern recognition based on neighborhood rough set theory (Neighborhood Rough Set Theory,NRST) and other data-driven methods is carried out in this paper. The attribute reduction method of NRST and the method of NRST combined with statistical analysis and machine learning are discussed. The main work and results of this paper are as follows: (1) the definition of NRST and the method of forward greedy attribute reduction are introduced. The advantage that NRST can directly deal with continuous numerical attributes is fully utilized. This paper presents a method of constructing NRST decision table based on rotor power frequency doubling as conditional attribute and fault type as decision attribute, and analyzes the feasibility of constructing NRST decision table based on the frequency spectrum characteristics of typical fault categories. The experimental results also show that the feature attributes obtained by this method are more physical and avoid the loss of key attributes in the discretization process. (2) on the basis of NRST attribute reduction, In this paper, a method of classifying fault categories with NRST and Fisher discriminant (FDA) is proposed. The discriminant function and cumulative discriminant ability are obtained. The effects of quadratic reduction and redundancy removal on fault pattern recognition are discussed. The mapping of data from high dimension to low dimension is completed, and the effect of fault classification under low dimension is realized. The experimental results show that the method can achieve the same recognition accuracy in the case of less feature attributes, thus saving storage space to improve the computational efficiency. (3) in order to seek an efficient and accurate fault diagnosis method, At the same time, in order to discuss the effect of NRST attribute reduction on machine learning, a method of fault classification identification based on NRST combined with radial basis function neural network (RBFNetwork) is proposed. Gao Si standard function is selected as radial basis function. The center, width and connection weight of the basis function are determined by the self-organizing selection center method. The experimental results show that the method can obviously shorten the modeling time and improve the recognition accuracy. (4) in order to solve the difficult problem of knowledge storage and discovery, the development of intelligent diagnosis technology is promoted. A fault identification system based on WEKA data mining platform is designed to connect My SQL database. The storage of fault knowledge, the display of data stream and the access of database by calling My SQL in WEKA are completed.
【学位授予单位】：兰州理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TH133;TP18
【参考文献】
相关期刊论文 前10条
1 马宪民;张兴;张永强;;基于支持向量机与粗糙集的隔爆电动机故障诊断[J];工矿自动化;2017年02期
2 李辉;王毅;杨晓萍;贾嵘;罗兴，
本文编号：2364147
本文链接：https://www.wllwen.com/kejilunwen/zidonghuakongzhilunwen/2364147.html
