【作者】 杨文安；

【导师】 廖文和；

【作者基本信息】 南京航空航天大学 ， 机械工程， 2012， 博士

【摘要】 产品的质量是现代企业增强市场竞争力、赖以生存和发展的基础。制造过程的质量控制与诊断是保证产品质量的重要环节。对制造过程质量进行控制，是实施过程质量连续改进的起点，而制造过程质量诊断则可为过程质量连续改进指明方向。通过制造过程质量诊断发现过程异常并采取纠正措施，可以使过程恢复并保持稳定受控状态。随着制造过程现代化和复杂程度的提高，对过程质量控制与诊断提出了更严更高的要求，单纯使用传统的统计过程控制与诊断技术并不能很好地满足这些要求。本文针对制造过程质量控制与诊断中存在的4个核心问题进行了研究，包括缺乏快速而经济的Shewhart控制图设计方法、缺乏高效统一的制造过程质量量化控制方法、缺乏实时准确的过程均值与方差控制图异常模式并行识别方法以及缺乏精准便捷的质量特性相关多工序制造过程质量诊断方法。本文的创造性研究成果主要有：（1）提出了一种基于过程历史波动知识的控制图统计经济设计方法针对传统Shewhart控制图设计中存在如下问题：(1)统计设计控制图的使用成本较高，(2)经济设计控制图的统计特性并不理想，(3)只注重在控参数而不考虑过程历史波动知识，提出了基于过程历史波动知识的控制图统计经济设计方法。实验结果表明，该方法不仅降低了控制图使用成本，而且保证了控制图的相关统计特性在需求值以内，并能更准确反映当前制造过程真实运行状态。为了求解控制图统计经济设计中的优化问题，分别提出了一种单目标和多目标粒子群优化算法。针对多目标优化问题通常有大量Pareto最优解，还结合聚类分析和伪权系数向量法设计了一个控制图多目标统计经济设计决策支持算法。在此基础上，成功地将上述方法应用于某企业机加工车间在轴承机械加工过程控制图单目标和多目标统计经济设计中。（2）提出了一种基于混合智能学习模型的制造过程质量量化控制方法针对传统Shewhart控制图实施中存在如下问题：(1)无法对过程状态作出定量评估，(2)过程因漂移而偏离正态分布时将会增加两种错误(虚发警报和漏发警报)的风险，提出了一种基于混合智能学习模型的制造过程质量控制方法。该混合智能学习模型由两个基于智能学习的序列模块组成：ModuleⅠ和ModuleⅡ。ModuleⅠ使用一个基于自组织特征映射神经网络的量化误差控制图来侦测过程异常并对过程异常的严重程度作出定量评估，ModuleⅡ使用一个基于离散粒子群优化算法的选择性神经网络集成(DPSOSEN-BPN)来辨识被ModuleⅠ中基于自组织特征映射神经网络的量化误差控制图侦测到的过程失控信号的异常源种类。实验结果表明，ModuleⅠ中基于自组织特征映射神经网络的量化误差控制图在侦测过程失控上的性能表现优于文献中常用的一些方法，ModuleⅡ中DPSOSEN-BPN具有较好的泛化学习能力，不但能在均值或方差异常单独出现时，具有快速而准确的辨识能力，也能在均值及方差异常同时出现时，取得较好的辨识绩效。在此基础上，成功地将混合智能学习模型应用于某企业喷涂车间白色面漆喷涂过程异常侦测和异常源种类辨识中。（3）提出了一种制造过程均值与方差控制图异常模式并行识别方法针对均值及方差控制图上出现的某一种模式通常是由不同的原因所引起，提出了一种基于选择性学习矢量神经网络集成的同时识别均值及方差控制图异常模式识别方法，此方法使用原始过程数据及统计特征值作为神经网络的训练样本。实验结果表明，该方法不但能在均值或方差控制图模式单独出现时，具有快速而准确的识别能力，也能在均值及方差控制图模式同时出现(亦即混合控制图模式)时，取得较好的识别绩效。实验亦显示结合了过程数据与统计特征可提升异常的控制图模式识别准确率。在此基础上，成功地将上述方法应用于某企业喷涂车间白色面漆喷涂过程均值及方差控制图异常模式识别中。（4）提出了一种质量特性相关多工序制造过程质量诊断方法针对传统的Shewhart分析方法在显示异常时，并不能告知是什么异常，发生在哪个或者哪些工序中，提出了质量特性相关的多工序制造过程质量诊断方法。实验结果表明，该方法不但可以诊断上道工序对下道工序的影响，分清上下两道工序的质量责任，还可以找出多工序制造过程中质量改进的关键工序。在此基础上，成功地将上述方法应用于某企业机加工车间滚柱零件机械加工过程质量诊断中。更多还原

【Abstract】 Product quality is the foundation of modern enterprises’ survival, development and increase of itscompetitiveness in market competition. Quality control and diagnosis of the manufacturing processplays a more important role in ensuring product quality. Quality control of the manufacturing process isthe starting point in the continuous quality improvement cycle, and while quality diagnosis of themanufacturing process provides direction for continuous quality improvement. By means of qualitydiagnosis, it can substantially help to identify the process abnormalities, further facilitate correctiveaction, and eventually bring the process back in control. With the advance of the modernization andcomplexity of manufacturing process, more stringent and higher requirements are generated for processcontrol and quality diagnosis of manufacturing process. It is very difficult to fulfill these requirementsonly using the traditional technologies of statistical process control and diagnose.In this dissertation, four core problems standing in the way of quality control and diagnosis of themanufacturing process are investigated, which includes a lack of statistical and economic designmethods for Shewhart control charts that are effective and cost-efficient, a lack of effective anduniversal method for intelligent quality control of the manufacturing process, a lack of real-time andaccurate methods for simultaneous recognizing of abnormal process mean and variance control chartpatterns, and a lack of precise and convenient quality diagnosis method for the multistagemanufacturing process with dependent quality characteristic.The main contributions of this dissertation are summarized as follows:(1) A statistical and economic design method for Shewhart control charts is proposed.To address the issue in the design of Shewhart control charts that the statistical design of a controlchart requires high cost inputs, while the economic design provides disappointing statistical properties,and both of them only consider in-control parameters but without taking into account any historicalknowledge related to the process shifts, this research constructs a statistical-economic model todetermine the optimal parameters using previous process shifts, which can be extracted from the fieldoperation. Experimental results show that the proposed method can not only reduce the cost, but alsoinsure statistical properties within the required intervals, meanwhile correctly reflect the actual processconditions. In order to solve the optimization problem in the statistical-economic design of controlcharts, this study proposes particle swarm optimization-based single and multiple objectiveoptimization algorithms. Since the Pareto-optimal set can be extremely large in multi-objective problems, a multi-objective decision support method that combines clustering analysis andPseudo-weight coefficient vector approach is designed for multi-objective economic-statistical designof control charts. Besides, the aforementioned method is successfully applied in the single andmultiple-objective statistical-economic design of control charts for diameters of the bearings.(2) A hybrid intelligent learning model-based method for quantitative quality control ofmanufacturing process is proposed.To address the issue in the implementation of Shewhart control charts that control charts fail toperform quantitative assessment of process behavior and tend to encounter the increasing risk of typeⅠand typeⅡerrors when the process deviates away from the normal distribution or exhibits some degreeof autocorrelation, a hybrid intelligent learning model is developed. The proposed hybrid intelligentlearning model is comprised of two intelligent learning-based sequential Modules: ModuleⅠandModuleⅡ. ModuleⅠemploys a self-organizing map (SOM) based quantization error control chart todetect process mean and/or variance changes and to provide a quantitative assessment of the severity ofthe process abnormalities. ModuleⅡemploys a discrete particle swarm optimization based selectiveback-propagation neural network (DPSOSEN-BPN) to identify the category of signals (i.e., meanabnormality, variance abnormality or mean or variance abnormalities) that are judged as out-of-controlsignals by quantization error control chart of ModuleⅠ. Experimental results show that the developedself-organizing map (SOM) based quantization error control chart of ModuleⅠis better than thecommonly used approaches in literature in detecting process mean and/or variance changes and yetable to make a quantitative assessment of the severity of the process abnormalities, the developedDPSOSEN-BPN of ModuleⅡ has better generalization ability and can effectively recognize not onlysingle mean or variance abnormality but also mixed abnormalities in which mean and varianceabnormality exist concurrently. Besides, the proposed hybrid intelligent learning model is successfullyapplied in detecting process changes and identifying the categories of process abnormality for a whitemillbase dispersion process.(3) A method for the simultaneous recognition of process mean and variance control chart patternsis proposedTo address the issue that process mean and variance control charts are usually implementedtogether and that these two charts are not independent of each other and while a specific pattern of themean and variance charts can be associated separately with different problems, this study proposed aselective neural network ensemble based model for the simultaneous recognition of both mean andvariance control chart patterns. The numerical results indicate that the proposed model can effectivelyrecognize not only single mean or variance control chart patterns but also mixed control chart patterns in which mean and variance control chart patterns exist concurrently. Empirical comparisons also showthat both direct data and selected statistical features extracted from the process that are employed as theinputs of neural network yield better performance than previous works using raw data or statisticalfeatures only. Besides, the proposed method is successfully applied in abnormal control chart patternrecognition for a white millbase dispersion process.(4) A quality diagnosis method for the multistage manufacturing process with dependent qualitycharacteristic is proposedTo address the issue that the Shewhart control chart can show the indication of abnormality, but itcan not tell us what and where the abnormality is, this study proposes a quality control and diagnosismethods for the multistage manufacturing process with dependent quality characteristic. Experimentalresults show that the proposed method can not only diagnose the effect of preceding stage on the nextstage, discriminate the quality responsibility between the preceding and next stages, but also identifywhich stages are key ones for improving the quality of the finished product of a production line.Besides, the proposed method is successfully applied in quality diagnosis for the mechanical processingof boiler roller with multiple dependent process stages.更多还原