【作者】 刘银华；

【导师】 林忠钦； 金隼；

【作者基本信息】 上海交通大学 ， 车辆工程， 2013， 博士

【摘要】 轿车车身制造偏差是影响产品质量、制造成本和市场占有率等的重要因素之一，因此车身装配偏差的诊断与控制方法一直受到学术界和工业界高度重视。车身产品由数百个薄板零件在数十个工位上焊装而成，一方面装配过程中零件、夹具、焊接、操作等偏差源因素众多，对最终装配质量的影响关系极其复杂；另一方面车身批量生产中通常只对若干装配工位的有限测点进行抽样检测，无法观测到装配过程中各类偏差输入、传递与输出的完备信息。由于装配系统的复杂性和观测信息的不完备性，使得各类偏差源与车身装配质量之间的影响关系难以准确描述，故障诊断的溯源结果呈现出一定的不确定性。传统基于模式匹配的诊断方法主要用于解决完备信息下单个夹具定位失效模式引起的偏差诊断问题，而实际生产中偏差源诊断受到检测样本小、信息不完备的制约，特别是多偏差源的诊断仍然依赖工程经验，诊断结论准确性差、诊断验证周期长、偏差控制成本高。贝叶斯网络理论是不确定知识表达与诊断推理的有效方法之一。它一方面可以将车身装配偏差间复杂影响关系以不确定的概率模型形式进行表达；另一方面能够对工程经验、设计知识等先验信息和当前检测数据进行多源信息融合，从而实现不完备信息下装配偏差关系模型的学习更新，不断提高故障诊断准确性。与传统基于经典统计理论的方法相比，贝叶斯网络在检测样本小、信息不完备条件下的复杂系统故障诊断中具有显著优势。本文将贝叶斯网络理论应用到车身装配偏差诊断的不确定性问题，提出基于贝叶斯网络的装配偏差影响关系建模与诊断方法。针对小数据集下贝叶斯网络模型构建的难题，提出基于先验信息与检测数据等多源信息融合的模型学习方法,并给出了多偏差源诊断推理、偏差源可诊断性、观测节点优化等系统方法，最后结合工程应用案例验证了本文所提出方法的可靠性。本文主要研究内容及创新成果如下：(1)车身装配偏差诊断问题的贝叶斯网络表达方法针对车身装配过程的偏差源诊断问题，引入不确定性建模的贝叶斯网络方法，对车身装配偏差的影响关系进行表达。首先通过对装配偏差源与质量检测特征的提取，获得贝叶斯网络的输入输出节点；然后结合大样本下贝叶斯网络学习方法，给出了装配偏差影响关系的贝叶斯网络模型表达形式;最后结合车身装配过程小样本检测等特殊性问题，分析了上述贝叶斯网络模型在进行实际偏差源诊断时需解决的问题。(2)车身装配偏差贝叶斯网络的多源信息融合建模针对小数据集下复杂偏差影响关系的贝叶斯网络建模难题，提出基于多源信息融合的贝叶斯网络建模方法。首先结合工程经验、设计阶段装配偏差仿真结果等先验信息，提出从偏差敏感度矩阵到贝叶斯网络初始结构与先验参数的映射方法，建立初始贝叶斯网络模型；然后通过节点先验分布与检测数据的融合，实现网络结构关系与各节点条件概率的迭代更新。在此基础上，提出基于信息熵的测点组评价方法，结合偏差源的可诊断性分析和有效独立准则，实现了面向偏差源可诊断性的观测节点优化。(3)基于贝叶斯网络的多装配偏差源诊断方法研究针对车身多级装配过程多偏差源诊断的需求，提出小样本下证据变量值获取以及多偏差源诊断推理方法。首先研究小样本下基于贝叶斯估计的装配质量评价方法，获得偏差源概率推理中的证据变量值；然后通过贝叶斯网络的最大后验假设和最大后验概率问题的求解，实现车身装配过程多偏差源问题的故障诊断；最后结合对贝叶斯网络诊断结果的分类，分析了不完备质量检测以及噪声等因素对诊断准确率的影响，验证了贝叶斯网络诊断方法的有效性。(4)工程应用在上述研究基础上，结合某车型装配车间的实际情况，在侧围总成装配、前围下部总成装配等案例的偏差源诊断中开展应用研究，详细描述了从输入输出节点定义、贝叶斯网络模型构建、观测节点质量评价到后验概率推理的车身装配偏差诊断过程，验证了贝叶斯网络方法解决多偏差源诊断问题的有效性与模型在线迭代学习的可行性。本文面向汽车车身装配过程的偏差控制，研究基于贝叶斯网络的装配偏差建模方法，实现了小样本检测环境下车身装配多偏差源的诊断，不仅为汽车车身装配过程精度控制提供新的理论方法与技术指导，对飞机、列车等复杂产品的装配精度控制也具有借鉴意义。更多还原

【Abstract】 The dimensional variation of auto bodies is one of the most important factors thataffect the product quality, manufacturing cost and the market shares, so the techniqueson the root cause diagnosis and process control are paid great attention to by bothresearchers and manufacturers. The auto body is assembled with hundreds ofcompliant sheet metal parts in dozens of assembly stations. In the assembly process,different variation sources such as incoming parts, tooling, welding etc, will affect theassembly variation. The relationship between variation sources and the assemblyvariation are extremely complicated. Besides, the measurement data about the processare always small samples in some limited stations, which only provide incompleteinformation for the process monitoring and diagnosis. The traditional fault diagnosismethods, such as pattern matching method, are different to diagnose the root cause ofthe assembly variations by considering the incomplete measurement data. In realassembly process, variation sources diagnosis is still mainly relying on theengineering experience.The Bayesian network theory is one of the most effective methods for knowledgerepresentation and uncertain reasoning. It can describe the complex relationshipbetween root causes and assembly deviation with an uncertain probability model. Onthe other hand, multi-source information, such as engineering experience, designknowledge and also the measurement data can be incorporated into the model to avoidthe shortcomings of small sample data. Compared with the traditional statisticmethods, Bayesian networks have a significant advantage for assembly variationdiagnosis under incomplete measurement data set.In this paper, Bayesian networks are used to solve the fault diagnosis problem inthe auto body assembly process. The Bayesian network model is used to illustrate thecausal relationships between the root cause nodes and the sensor nodes. A newmodeling method under small sample data set by considering multi-sourceinformation is proposed. Based on the diagnostic model, the methods on probabilityreasoning, diagnosability of the root nodes and optimal sensor placement aredeveloped. At last, the assembly diagnostic cases are studied. The results andcomparative analysis showed the proposed methods are effective and reliable.The main contents are shown as follows:(1) Description of the dimensional variation source diagnosis problem in theassembly process with Bayesian networks For the diagnosis problem in the assembly process, the Bayesian network theoryis introduced to describe the causal relationship of the assembly deviations. At first,the root nodes and sensors nodes of the diagnostic network are acquired by extractingthe key characteristics in the assembly process. Afterwards, the directed acyclic graphand conditional probability tables are learned based on the learning algorithms underlarge data sets. Furthermore, the probability reasoning methods based on thediagnostic Bayesian network and the variation sources diagnosis procedures arepresented. At last, the further studies concerning the small sample problem in the realmeasurement process are proposed.(2) The Bayesian network modeling of the causal relationship of theassembly variation by considering multi-source informationFor the Bayesian network modeling problem under the small sample data set, anew method based on multi-source information fusion is proposed. At first, theoriginal Bayesian network structure and conditional probabilities are acquired basedon the mapping of the variation simulation results which are obtained in the variationdesign stage. Furthermore, the diagnostic model is updated by incorporated with thenew measurement data set. The causal relationships are updated based on theconditional independence test method, and the corresponding parameters of thenetwork nodes are updated based on Bayesian estimation approach. At last, on thebasis on constructing the diagnostic information matrix, the effective independencemethod is used for optimal sensor placement in the auto body assembly process.(3) Multiple fault diagnosis in the auto body assembly process based onBayesian networksMultiple fault diagnosis in the multi-station assembly process is always different.Based on the above Bayesian network model, methods on the process monitoring forevidence variables, diagnostic reasoning and diagnostic capability analysis are studied.At first, under small sample data condition, a parameter estimation method based onBayesian approach is presented. The estimation results are used as evidence states ofthe sensor nodes for probability reasoning. Afterwards, the diagnostic methods andprocedures based on the Bayesian network are proposed by updating the posteriorprobability of the root nodes. At last, by classifying the diagnosis results, thesensitivity analysis by considering the number of evidence variables and measurementnoises is performed to evaluate the diagnostic performance of Bayesian networks.(4) Application to the assembly variation source diagnosisBased on the above theoretical research, the diagnostic method based on Bayesiannetworks is applied in the real assembly cases. The corresponding programmingmodule for assembly process monitoring is developed. The side frame and front areaassembly cases are studied for fixture fault diagnosis and optimal sensor placement. The results showed the Bayesian network method for multiple fault diagnosis in theauto body process is practicable and effective.This paper presented a framework of Bayesian network model and fault diagnosisin the auto body assembly process, which provides a practicable variation sourcesdiagnosis method under small sample data sets. The proposed methods in this papercan not only be used in dimensional quality improvement of auto bodies, but also canbe applied to aircraft, train bodies and other mechanical products for assembly qualityimprovement.更多还原