基于支持向量机的自动加工过程质量控制方法研究

发布时间：2018-04-05 05:07

本文选题：制造过程质量控制　切入点：统计过程控制　出处：《重庆大学》2013年博士论文

【摘要】：制造过程质量控制是全面质量管理（Total Quality Control，TQC）的核心内容，是确保产品质量的重要措施。在现代生产条件下，加工自动化程度和复杂程度不断提高，单纯依靠统计过程控制（Statistical Process Control，SPC）方法已不能很好满足制造过程质量控制的要求。将人工智能与机器学习技术与SPC相结合实现制造过程质量的智能化控制，成为新的发展方向。近年出现的支持向量机（SupportVector Machine，SVM）技术，与人工神经网络（Artificial Neural Networks，ANN）等较早出现的机器学习方法相比，具有多方面优势。将SVM用于实现制造过程质量控制智能化成为目前该领域研究的热点。本文结合国家重大专项课题“齿轮加工自动生产线网络化现场管理与智能监控系统”（编号：2011ZX04001-041-06）需要，对基于SVM的自动加工过程质量智能化控制方法进行了系统研究，主要内容如下： 1）研究了一种将控制图模式的原始特征、形状特征及统计特征进行融合构建分类特征集，并以SVM作为分类工具的控制图模式识别方法。在该方法的特征融合中，分别采用了KNNC和ReliefF两种特征约简算法来消除特征合集中的冗余和无关特征分量；其中，约简算法的控制参数与SVM的超参数一起通过基于粒子群优化算法（Particle Swarm Optimization，PSO）和交叉验证（Cross Validation）实现的智能寻优模型进行同步优化，使融合特征集更利于模式分类，且SVM的泛化能力较好。仿真实验证实所提方法能获得更好的识别精度。 2）研究了基于优化多核支持向量回归（Support Vector Regression，SVR）的过程异常参数估计方法，提出了以三种基本核函数进行凸组合构建SVR的核函数。其中，各基本核的关键参数、组合权重系数以及SVR的惩罚系数通过PSO和交叉验证实现的智能寻优模型进行同步优化，以提高SVR的估计性能。将累积和（CUSUM）控制图与该多核SVR结合建立了过程异常监测模型，用于监控过程均值阶跃异常和估计均值阶跃异常幅度。仿真实验证实该优化多核SVR估计更准确。 3）研究了基于优化有向无环图SVM（Directed Acyclic Graph SVM，DAGSVM）的多变量过程异常源识别方法。针对多变量过程存在多种异常形式时模式类型数多、一般多分类SVM难以兼顾识别精度和效率的问题，以具有效率优势的DAGSVM作为研究对象，提出两种优化方法以改善其识别精度。一是根据核空间内的类型样本平均差异测度对DAGSVM拓扑结构进行启发式优化，在一定程度上缓解了其固有的分类错误累积问题，以提高识别精度；二是基于集成学习理论，以DAGSVM作为个体，通过二进制粒子群算法（Binary PSO，BPSO）构建选择性集成分类器，以获得更好的泛化能力，提高识别精度。通过仿真实验证明了所提方法的有效性。 4）研究了基于支持向量机混合模型的多变量过程异常在线监控与诊断方法。针对多变量过程控制中需要同时解决异常探测、异常源识别和异常幅度量化的要求，提出一种由SVM和SVR组成的混合模型（HSVMM），对两变量过程中的均值阶跃异常进行在线监控与诊断。其中，SVM对过程均值状态进行模式识别，以探测过程均值异常并对异常源进行识别；SVR被用作估计工具对均值偏移幅度进行量化。仿真实验结果表明，所提模型在探测小幅度异常和识别异常源上，与传统多元控制图和ANN模型相比具有优越性；并能较为准确估计异常幅值。基于以上研究成果，结合课题需要，设计开发了齿轮自动生产线过程质量智能化控制系统。系统基于.NET平台构建，，并集成MATLAB计算环境实现SPC控制图和支持向量机的高效运算，以满足加工过程质量的在线监控和诊断要求；系统结构设计为C/S与B/S结合模式，以满足围绕自动加工生产线过程质量控制的多方面需求。系统已在某齿轮加工企业实施应用，取得了初步应用效果。
[Abstract]:The quality control of manufacturing process is the core content of Total Quality Control ( TQC ) . It is an important measure to ensure the quality of products . By combining artificial intelligence with machine learning technology and SPC , it can not meet the requirement of quality control of manufacturing process .

1 ) The original characteristics , shape characteristics and statistical characteristics of the control chart pattern are fused to construct a classification feature set , and the SVM is used as the control chart pattern recognition method of the classification tool . In the feature fusion of the method , two feature reduction algorithms of KNNC and ReliefF are adopted to eliminate the redundancy and irrelevant feature components in the feature set .
In this paper , the control parameters of the reduction algorithm are synchronized with the super parameters of the SVM by the intelligent optimization model realized by Particle Swarm Optimization ( PSO ) and Cross Validation , so that the fusion feature set is more beneficial to the pattern classification , and the generalization ability of the SVM is better . The simulation experiment proves that the proposed method can obtain better recognition accuracy .

2 ) In this paper , we study the method of parameter estimation of process anomaly based on optimized multi - core support vector regression , and put forward the kernel function of combining three basic kernel functions . The key parameters , combined weight coefficient and penalty coefficient of each basic kernel are optimized by PSO and cross - validation to improve the estimation performance .

A multi - variable process anomaly source identification method based on optimized directed acyclic graph SVM ( DAGSVM ) is studied .
Second , based on the theory of integrated learning , using DAGSVM as an individual , a selective ensemble classifier is constructed by binary particle swarm optimization ( BPSO ) , so as to obtain better generalization ability and improve recognition accuracy . The validity of the proposed method is proved through simulation experiments .

4 ) On - line monitoring and diagnosis of multivariable process based on support vector machine hybrid model is studied . A hybrid model ( HSVMM ) , which is composed of SVM and support vector , is proposed to solve the problem of abnormal detection , abnormal source identification and abnormal amplitude quantization in multivariable process control .
The simulation results show that the proposed model is superior to traditional multivariate control graph and ANN model when detecting small amplitude anomalies and identifying abnormal sources .
and the abnormal amplitude can be estimated more accurately .

Based on the above research results , the intelligent control system for the process quality of gear automatic production line is designed and developed based on . NET platform , and the efficient operation of SPC control chart and support vector machine is realized by integrating MATLAB computing environment to meet the requirement of online monitoring and diagnosis of machining quality .
The system structure is designed as C / S and B / S combined mode , so as to meet the demand of quality control around automatic processing production line .

【学位授予单位】：重庆大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TH164;TP18

【相似文献】