维修备件需求预测系统的设计及实现
|
摘要
产品备件,工业领域称为service parts,它是指生产企业针对自己已经销售的产品向终端客户提供后期维修保用的备品备件。随着市场竞争的加剧,维修售后服务成为企业的重要竞争能力之一并且伴随着额外的维修服务利润。然而由于产品故障的不确定性和专业性,备件需求一直以来都难以准确预测。因此针对维修备件需求预测的系统化解决方案有助于企业充分利用有限的成本和精力实现最佳的售后服务,从根本上提升企业的市场竞争能力。
本文基于一个真实的企业案例提出维修备件需求管理的最佳模式应该是企业对其产品维修相关的知识管理。由于缺乏知识管理,面对复杂的维修备件的需求诱因,现行大多数的商用软件主要是根据历史需求数据采用线性回归分析或移动平均等单纯基于历史需求数量变化的预测模型,其预测能力十分有限。因而本文详细研究了可靠性工程,统计学和信息工程等相关领域的知识,结合维修备件需求的特点,提出了一套完整的系统解决方案且成功构建了基于贝叶斯网络的预测模型。整个研究工作包括:
1.详细分析企业中对维修备件需求管理不同层次的需求,结合其现有的系统架构,设计了具有学习能力的维修备件需求预测系统的系统蓝图和系统架构,并实现原型系统的构建。对系统中关键的产品备件生命周期管理模块、预测备件需求模块和预测评估模块等三大功能模块做了具体的阐述。
2.在需求预测这个核心功能上,借鉴可靠性工程领域的经典理论,从产品的生命周期入手,结合“浴盆理论”,详细分析了维修备件发生故障的主要原因以及其诱因的复杂性,提出备件的需求预测可以根据备件的生命周期进行推测。本文中将维修备件的生命周期分为“早期故障期”、“偶然故障期”和“正常失效期”。“正常失效期”内,维修备件的发生故障的概率可以由其寿命函数分布确定。且根据不同料件的特性,其寿命函数分布可使用指数分布、Weibull分布、正态分布或者对数正态分布。而对于处于“早期故障期”和“偶然故障期”中的维修备件,其发生故障概率则采用如备件精密性,客户行业等关键因素预测。
3.在预测模型的构建和实现方面,根据维修备件需求变化与其诱因之间的因果关系选择贝叶斯网络作为其预测的主要工具,并在此基础上构建预测模型。最终模型的输入变量不仅包括了备件的生存期,还考虑了备件的磨损特性、精密特性、是否定制件、产品最终客户所在的行业、地区、是否拥有同类型设备和备件需求预测时期的季节等影响因素。文章详细说明了在预测模型的贝叶斯网络中,具体节点之间概率表的学习方法。并且针对模型中的计算关键点,提供了细致的解决方案:确定了适用于系统开发的备件寿命分布函数参数估计的算法和推导过程、备件寿命分布函数的自动选择算法、如何根据备件的生存周期合理推算其对应的生命周期、并使用朴素贝叶斯分类法对产品最终客户所在行业进行自动分类。
4.最后对原型企业进行数据采样,完成了对核心预测模型的准确性和可行性的验证。
本文提出的需求预测模型有效去除了传统维修备件预测方法中完全脱离备件特性和实际维修需求基础的弊端,在提高备件需求预测准确率的同时提供给企业一个完整的维修备件相关知识管理的系统解决方案原型。
Service parts, also known as spare parts, are the accessories for after market service. With the keen competition, after market service became one of the key competencies accompanying with additional service profit. However due to the uncertainty and very complicated situation of product failures, service part requirement is always hard to predict. Therefore a specific system solution for service parts requirement forecast will help enterprise to achieve best after market service with limited cost and resources, which will improve the enterprise’s competency ultimately.
This dissertation was built up on a real business case, and it proposed that an ultimate model of enterprise service parts management is the knowledge management of after market service. Due to the lack of service parts knowledge management, majority commercial software often rely on purely historical data based forecast models, such as linear regression and moving average, etc. Its forecasting capability is quite limited when facing so complicated causes of service parts requirement. Hence this dissertation investigated throughout all related theories that may improve service parts requirement forecast in reliability engineering, statistics and information technologies. Then it introduced a full system solution of service parts requirement forecast and successfully built up the forecast model based on Bayesian Network. The key accomplishments include:
1. Through detail analysis of the different requirements in enterprise and considering the existed system architecture, the blue print and new system architecture design of the service parts requirement forecast system have been completed. It is a system with learning capability. And a prototype was finished as well. Its three key modules, "products and service parts life cycle management module", "service parts requirement forecast module" and "forecast assessment module”, are elaborated in this article.
2. For the core of requirement forecast, it believes that service parts requirement forecast has to be built up on part life cycle after detail investigation of service parts failure modes and causes combining with classic theories in reliability engineering and Bathtub curve. In this dissertation, the life cycle of service part has been divided into 3 periods "Initial failure period”,”Occasional failure period" and "Nature failure period". While part is in "Nature failure period", the probability of part failure could be calculated according to its life time distribution function. In the meanwhile, according to the attributes of the part, the life time distribution function could be exponential distribution, Weibull distribution, normal distribution or lognormal distribution. However if the part is in "initial failure period" or "occasional failure period", its failure probability may be forecasted via some key factors such as the part's precision attribution or the industry of its end customer.
3. In the phase of data modeling, Bayesian Network was chosen as main approach of forecast modeling because there are very clear cause and effect relationships in service parts requirement variation. The forecast model has been successfully built up and its input variants contain not only service parts life time, but also part wear out attribute, part’s precision attribute, customized part or not , industry of end customer, location of end customer , any products for backup on customer site, and seasonal factor of the forecast period. The article detail explained the probability table learning approach in the Bayesian Network of the forecast model. It gave detail solution for every key step of node calculation in the forecast model: firming the proper algorithms for parameter estimation of parts life time distribution function that fits system development mostly; explaining the auto selection algorithm of service parts life time distribution function; defining how to classify a part's life cycle through its life time; introducing Native Bayesian Classifier for industry classification of end customers.
4. Finally through data sampling from the original enterprise, the dissertation completed accuracy and feasibility verification of the core forecast model.
The forecast model proposed by this dissertation has eliminated the fatal weakness of traditional service parts requirement forecast models which generally ignore the consequence of part attributes and other repair requirements causes in service parts requirement forecast. In accompany with a satisfactory forecast result, this article provides enterprise a complete system prototype of service parts knowledge management.
本文基于一个真实的企业案例提出维修备件需求管理的最佳模式应该是企业对其产品维修相关的知识管理。由于缺乏知识管理,面对复杂的维修备件的需求诱因,现行大多数的商用软件主要是根据历史需求数据采用线性回归分析或移动平均等单纯基于历史需求数量变化的预测模型,其预测能力十分有限。因而本文详细研究了可靠性工程,统计学和信息工程等相关领域的知识,结合维修备件需求的特点,提出了一套完整的系统解决方案且成功构建了基于贝叶斯网络的预测模型。整个研究工作包括:
1.详细分析企业中对维修备件需求管理不同层次的需求,结合其现有的系统架构,设计了具有学习能力的维修备件需求预测系统的系统蓝图和系统架构,并实现原型系统的构建。对系统中关键的产品备件生命周期管理模块、预测备件需求模块和预测评估模块等三大功能模块做了具体的阐述。
2.在需求预测这个核心功能上,借鉴可靠性工程领域的经典理论,从产品的生命周期入手,结合“浴盆理论”,详细分析了维修备件发生故障的主要原因以及其诱因的复杂性,提出备件的需求预测可以根据备件的生命周期进行推测。本文中将维修备件的生命周期分为“早期故障期”、“偶然故障期”和“正常失效期”。“正常失效期”内,维修备件的发生故障的概率可以由其寿命函数分布确定。且根据不同料件的特性,其寿命函数分布可使用指数分布、Weibull分布、正态分布或者对数正态分布。而对于处于“早期故障期”和“偶然故障期”中的维修备件,其发生故障概率则采用如备件精密性,客户行业等关键因素预测。
3.在预测模型的构建和实现方面,根据维修备件需求变化与其诱因之间的因果关系选择贝叶斯网络作为其预测的主要工具,并在此基础上构建预测模型。最终模型的输入变量不仅包括了备件的生存期,还考虑了备件的磨损特性、精密特性、是否定制件、产品最终客户所在的行业、地区、是否拥有同类型设备和备件需求预测时期的季节等影响因素。文章详细说明了在预测模型的贝叶斯网络中,具体节点之间概率表的学习方法。并且针对模型中的计算关键点,提供了细致的解决方案:确定了适用于系统开发的备件寿命分布函数参数估计的算法和推导过程、备件寿命分布函数的自动选择算法、如何根据备件的生存周期合理推算其对应的生命周期、并使用朴素贝叶斯分类法对产品最终客户所在行业进行自动分类。
4.最后对原型企业进行数据采样,完成了对核心预测模型的准确性和可行性的验证。
本文提出的需求预测模型有效去除了传统维修备件预测方法中完全脱离备件特性和实际维修需求基础的弊端,在提高备件需求预测准确率的同时提供给企业一个完整的维修备件相关知识管理的系统解决方案原型。
Service parts, also known as spare parts, are the accessories for after market service. With the keen competition, after market service became one of the key competencies accompanying with additional service profit. However due to the uncertainty and very complicated situation of product failures, service part requirement is always hard to predict. Therefore a specific system solution for service parts requirement forecast will help enterprise to achieve best after market service with limited cost and resources, which will improve the enterprise’s competency ultimately.
This dissertation was built up on a real business case, and it proposed that an ultimate model of enterprise service parts management is the knowledge management of after market service. Due to the lack of service parts knowledge management, majority commercial software often rely on purely historical data based forecast models, such as linear regression and moving average, etc. Its forecasting capability is quite limited when facing so complicated causes of service parts requirement. Hence this dissertation investigated throughout all related theories that may improve service parts requirement forecast in reliability engineering, statistics and information technologies. Then it introduced a full system solution of service parts requirement forecast and successfully built up the forecast model based on Bayesian Network. The key accomplishments include:
1. Through detail analysis of the different requirements in enterprise and considering the existed system architecture, the blue print and new system architecture design of the service parts requirement forecast system have been completed. It is a system with learning capability. And a prototype was finished as well. Its three key modules, "products and service parts life cycle management module", "service parts requirement forecast module" and "forecast assessment module”, are elaborated in this article.
2. For the core of requirement forecast, it believes that service parts requirement forecast has to be built up on part life cycle after detail investigation of service parts failure modes and causes combining with classic theories in reliability engineering and Bathtub curve. In this dissertation, the life cycle of service part has been divided into 3 periods "Initial failure period”,”Occasional failure period" and "Nature failure period". While part is in "Nature failure period", the probability of part failure could be calculated according to its life time distribution function. In the meanwhile, according to the attributes of the part, the life time distribution function could be exponential distribution, Weibull distribution, normal distribution or lognormal distribution. However if the part is in "initial failure period" or "occasional failure period", its failure probability may be forecasted via some key factors such as the part's precision attribution or the industry of its end customer.
3. In the phase of data modeling, Bayesian Network was chosen as main approach of forecast modeling because there are very clear cause and effect relationships in service parts requirement variation. The forecast model has been successfully built up and its input variants contain not only service parts life time, but also part wear out attribute, part’s precision attribute, customized part or not , industry of end customer, location of end customer , any products for backup on customer site, and seasonal factor of the forecast period. The article detail explained the probability table learning approach in the Bayesian Network of the forecast model. It gave detail solution for every key step of node calculation in the forecast model: firming the proper algorithms for parameter estimation of parts life time distribution function that fits system development mostly; explaining the auto selection algorithm of service parts life time distribution function; defining how to classify a part's life cycle through its life time; introducing Native Bayesian Classifier for industry classification of end customers.
4. Finally through data sampling from the original enterprise, the dissertation completed accuracy and feasibility verification of the core forecast model.
The forecast model proposed by this dissertation has eliminated the fatal weakness of traditional service parts requirement forecast models which generally ignore the consequence of part attributes and other repair requirements causes in service parts requirement forecast. In accompany with a satisfactory forecast result, this article provides enterprise a complete system prototype of service parts knowledge management.
引文
[1]Aberdeen Group,The Service Parts Management Solution Selection Report,September 2005,1-4.
[2]张连富主编,高等学校教材-物流学,人民交通出版社,2005年01月第1版,74.
[3]虞益诚主编,电子商务概论,中国铁道出版社, 2006年6月,217.
[4]孙大鹏,关于物流中心选址的研究,大众科技,2006年第8期(总第94期),223.
[5] IBM, Logistics optimization tools, IBM Research, Aug 2004, 9, http://www.research.ibm.com/trl/projects/optsim/logiopt/index_e.htm.
[6]崔南方,丁留明,基于零件寿命函数的慢速流动备件库存模型,系统工程,第24卷第9期(总第153期),2006-9,24.
[7] Thomas R. Willemain*,Charles N. Smart,Henry F. Schwarz,A new approach to forecasting intermittent demand for service parts inventories,Science Direct,International Journal of Forecasting 20 (2004),375– 387.
[8]Dick Moss, Highly Accelerated Stress Testing,Agilent Technologies,April 2004,8-15.
[9]曲立,库存管理理论与应用,经济科学出版社,2006-8,53-56,56-60,71.
[10]韩家炜(加拿大),Micheline Kamber(加拿大),范明等译,数据挖掘:概念与技术,机械工业出版社,2001,9-12 ,152-162.
[11] Pete Chapman , Julian Clinton , Randy Kerber ,Thomas Khabaza , Thomas Reinartz, Colin Shearer, Rüdiger Wirth, CRISP-DM 1.0, The CRISP-DM consortium, Aug 2000, 13, (http://www.crisp-dm.org/Process/index.htm).
[12]郭亮,用CRISP-DM模型来规范企业数据中心建设,上海烟草(集团)公司信息中心,2006,2.
[13]杨斌,万胜春,数据挖掘在大型医疗设备故障诊断中的应用研究,医疗设备信息,第22卷第10期,2007年10月,42.
[14]茆诗松,贝叶斯统计,中国统计出版社,1999,1-2.
[15]杨晓楠,基于贝叶斯统计推理的结构损伤识别方法研究,同济大学,土木工程学院,2007年12月, 129.
[16]李俭川,贝叶斯网络故障诊断与维修决策方法及应用研究,机械电子工程,国防科学技术大学,2002,8-15,24-30.
[17]王军,周伟达,贝叶斯网络的研究与进展,电子科技,1999/8/5,7.
[18]董立岩,贝叶斯网络应用基础研究,博士论文,吉林大学,通信与信息系统,2007-4,25.
[19]J.F.Lawless (加拿大),寿命数据中的统计模型与方法,中国统计出版社,1998年03月第1版,11-30.
[20]顾瑛编著,可靠性工程数学,电子工业出版社,2004年08月,98-107.
[21]郝静如主编,计算机辅助工程,航空工业出版社,2000年01月第1版,2,170.
[22]茆诗松,罗朝斌,无失效数据的统计分析[J],数理统计与应用概率,1989,4(4),489-506.
[23]雷刚,Weibull分布寿命数据的参数估计,华中科技大学,概率论与数理统计硕士学位论文,2006年4月,6-14.
[24]彭鸿霖,韦伯分布可靠度评估技术,中国可靠性网,2000/12,3-5.
[25]程依明,Weibull分布n>25时的BLUE系数,应用概率统计,第八卷,第四期,1992年11月,420.
[26]彭鸿霖,指数分布可靠度评估技术,中国可靠性网,2000/12,3-4.
[27] Tom M. Mitchell,曾华军等译,机器学习,机械工业出版社,2003-1,127.
[28]James O.Berger,统计决策论及贝叶斯分析,中国统计出版社,1998年5月,123.
[29][美]尼尔森(Nilsson,N.J.)著郑扣根等译,人工智能,机械工业出版社,2000年09月第1版,215-216.
[2]张连富主编,高等学校教材-物流学,人民交通出版社,2005年01月第1版,74.
[3]虞益诚主编,电子商务概论,中国铁道出版社, 2006年6月,217.
[4]孙大鹏,关于物流中心选址的研究,大众科技,2006年第8期(总第94期),223.
[5] IBM, Logistics optimization tools, IBM Research, Aug 2004, 9, http://www.research.ibm.com/trl/projects/optsim/logiopt/index_e.htm.
[6]崔南方,丁留明,基于零件寿命函数的慢速流动备件库存模型,系统工程,第24卷第9期(总第153期),2006-9,24.
[7] Thomas R. Willemain*,Charles N. Smart,Henry F. Schwarz,A new approach to forecasting intermittent demand for service parts inventories,Science Direct,International Journal of Forecasting 20 (2004),375– 387.
[8]Dick Moss, Highly Accelerated Stress Testing,Agilent Technologies,April 2004,8-15.
[9]曲立,库存管理理论与应用,经济科学出版社,2006-8,53-56,56-60,71.
[10]韩家炜(加拿大),Micheline Kamber(加拿大),范明等译,数据挖掘:概念与技术,机械工业出版社,2001,9-12 ,152-162.
[11] Pete Chapman , Julian Clinton , Randy Kerber ,Thomas Khabaza , Thomas Reinartz, Colin Shearer, Rüdiger Wirth, CRISP-DM 1.0, The CRISP-DM consortium, Aug 2000, 13, (http://www.crisp-dm.org/Process/index.htm).
[12]郭亮,用CRISP-DM模型来规范企业数据中心建设,上海烟草(集团)公司信息中心,2006,2.
[13]杨斌,万胜春,数据挖掘在大型医疗设备故障诊断中的应用研究,医疗设备信息,第22卷第10期,2007年10月,42.
[14]茆诗松,贝叶斯统计,中国统计出版社,1999,1-2.
[15]杨晓楠,基于贝叶斯统计推理的结构损伤识别方法研究,同济大学,土木工程学院,2007年12月, 129.
[16]李俭川,贝叶斯网络故障诊断与维修决策方法及应用研究,机械电子工程,国防科学技术大学,2002,8-15,24-30.
[17]王军,周伟达,贝叶斯网络的研究与进展,电子科技,1999/8/5,7.
[18]董立岩,贝叶斯网络应用基础研究,博士论文,吉林大学,通信与信息系统,2007-4,25.
[19]J.F.Lawless (加拿大),寿命数据中的统计模型与方法,中国统计出版社,1998年03月第1版,11-30.
[20]顾瑛编著,可靠性工程数学,电子工业出版社,2004年08月,98-107.
[21]郝静如主编,计算机辅助工程,航空工业出版社,2000年01月第1版,2,170.
[22]茆诗松,罗朝斌,无失效数据的统计分析[J],数理统计与应用概率,1989,4(4),489-506.
[23]雷刚,Weibull分布寿命数据的参数估计,华中科技大学,概率论与数理统计硕士学位论文,2006年4月,6-14.
[24]彭鸿霖,韦伯分布可靠度评估技术,中国可靠性网,2000/12,3-5.
[25]程依明,Weibull分布n>25时的BLUE系数,应用概率统计,第八卷,第四期,1992年11月,420.
[26]彭鸿霖,指数分布可靠度评估技术,中国可靠性网,2000/12,3-4.
[27] Tom M. Mitchell,曾华军等译,机器学习,机械工业出版社,2003-1,127.
[28]James O.Berger,统计决策论及贝叶斯分析,中国统计出版社,1998年5月,123.
[29][美]尼尔森(Nilsson,N.J.)著郑扣根等译,人工智能,机械工业出版社,2000年09月第1版,215-216.