【作者】 刘锐；

【导师】 张宁；

【作者基本信息】 北京交通大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 随着高速铁路技术的日益发展,铁路经营线路得到不断拓展,高铁列车转向架的组装效率问题日益突出,而且,伴随着工业4.0时代的到来,智能化工业装配线逐渐兴起,企业车间的自动化程度越来越高,如何保证一条自动化装配线稳态高效的运转,成为了企业日益关注的重大问题。面对如此庞大的自动化装配生产线,一旦其中的某个关键部件出现问题,将会影响整条装配线的运转,带来严重的经济损失。因此,放弃传统费时费力的定检修和事后维修,采用一种根据设备健康状态而进行对应维修的视情维修策略,既可保证装配线高效运转,又可以极大地节约人力物力,提高经济效益。状态修的核心技术在于利用实际的在线监测数据对设备进行有效的健康状态评估,这也是PHM(Prognostics and Health Management)技术的重要组成部分。本文在对PHM常用的健康评估方法及电机传动系统在线监测数据的特性研究的基础上,确定利用无监督式学习中的模糊聚类算法(FCM)进行健康状态评估的研究,并对传统FCM算法的缺陷进行相应的优化,建立了有效的评估模型,以输出隶属概率的形式将数据分为了正常、警报、严重三类,实现了对电机传动系统的健康状态的评估,完成了机器学习算法与工业生产的重要结合。具体的研究有如下几点:(1)传统FCM算法初始值选取具有随机性,易导致陷入局部极小,本文利用遗传算法优化FCM,提出了 GA-FCM算法。根据FCM的准则函数创建适应度函数,并且构造自适应交叉、变异算子,可以动态选取最优初始聚类中心。经通用数据集验证,GA-FCM算法可以降低初始类中心的目标函数值,从而减少收敛次数,保证聚类的稳定性,避免陷入局部极小。(2)实际的监测数据中存有大量的数据异常点与噪声点,严重影响FCM的聚类划分,本文采用一种基于距离加权的方式,重写FCM的准则函数,将最终的聚类中心定格在区域密度大的地方。本文将FCM算法、遗传算法、加权策略三者结合,提出了 WG-FCM算法,这样既可以选取最优初始中心,又可以有效避免噪声点对聚类划分的影响。(3)构建基于WG-FCM算法的健康状态评估模型。本文利用实际的电机传动系统数据进行验证,结果表明WG-FCM算法可以获得比GA-FCM更优的聚类划分,聚类结果更加合理。本文将最终的研究成果融合至实际的监控平台之中,实现了电机传动系统的健康状态评估。更多还原

【Abstract】 With the rapid development of high-speed railway technology,the railway line has been continuously expanded,the assembly efficiency of high-speed rail train bogies has become increasingly prominent.Moreover,along with the advent of the industrial 4.0 era,the intelligent industrial assembly line is gradually emerging,and the degree of automation of enterprise workshops is getting higher and higher.How to ensure the steady and efficient operation of an automated assembly line has become a major issue that enterprises are increasingly paying attention to.Faced with such a large automated assembly line,once a key component fails,it will affect the operation of the entire assembly line and bring serious economic losses.Therefore,abandoning the traditional time-consuming maintenance strategy such as fixed maintenance and after-repair,adopt a maintenance strategy based on the health status of equipment,not only to ensure efficient assembly line operation,but also can greatly save manpower and resources,improve economic efficiency.The core technology of state repair lies in using the actual online monitoring data to conduct an effective health assessment of the equipment,which is also an important part of PHM(Prognostics and Health Management)technology.On the basis of the study of the classic PHM algorithm and the real-time data characteristics of the motor drive system,in this paper,the fuzzy clustering algorithm(FCM)in unsupervised learning was used to study the health status assessment.Aiming at the defects of the traditional FCM algorithm,the corresponding optimization is made and an effective evaluation model is established.Using probability of membership form,all data are divided into three categories:normal,warning,serious.This paper has realized the evaluation of the health state of the motor drive system,and completed the important combination of the machine learning algorithm and the industrial production.The specific research has the following points:(1)Due to the random selection of the initial value of traditional FCM algorithm,which can lead to local minimum,in this paper,the genetic algorithm is used to optimize FCM.Use the FCM criterion function to create a fitness function,and construct adaptive crossover and mutation operator.So that the optimal initial cluster center can be dynamically selected to avoid falling into local minimum.Through common data set validation,GA-FCM algorithm can reduce the initial class center of the objective function value,thus reducing the number of convergence,to ensure the stability of clustering and avoid falling into a local minimum.(2)There are a lot of outliers and noise points in the actual engineering data,which severely affect the clustering of FCM.This paper uses a method based on distance weighting to rewrite FCM’s criterion function,make the final clustering centers locate in the area where the density is high.This paper combines FCM algorithm,genetic algorithm and weighting strategy,called WG-FCM algorithm,which can not only select the optimal initial center,but also effectively avoid the impact of noise points on cluster classification.(3)A health state assessment model based on WG-FCM algorithm is constructed.In this paper,the actual motor drive system data is used to verify the results,it show that WG-FCM algorithm can get better clustering partition than GA-FCM,and the clustering results are more reasonable.The research results are integrated into the actual monitoring platform to complete the motor drive system health assessment.更多还原