文献类型：学位论文

中文题名：智能多故障识别方法在过程监控中的应用研究

英文题名：Application Research on Intelligent Multi-faults Recognition Methods in Process Monitoring

作者：何菲[1];

机构：[1]华东理工大学;

导师：钱锋;华东理工大学

授予学位：硕士

语种：中文

中文关键词：故障识别;过程监控;集成SVM;LTSA;主元分析

外文关键词：fault recognition; process monitoring; combines SVM; LTSA; PCA

摘要：随着科技的发展,工业过程的复杂性日益提高,为了保证生产过程的安全性、可靠性,提高产品质量,减少生产成本,各种设计方案、控制策略、优化算法层出不穷。过程监控是个系统工程,需要紧密结合自动控制、人工智能、计算机科学、模式识别以及系统工程的相关学科知识,从而提供高控制系统的可靠性和安全性。本文主要针对流程工业中的过程监控问题,提出了几种智能多故障识别诊断方法。主要工作有如下几点。 主元分析方法通常与Hotelling T2和平方误差预测/(SPE/)统计量方法相结合,来解决线性、类别较少的故障检测和分类问题。本文首先运用三类鸢尾花数据作为实验样本,将选取某类作为基准样本,训练出主元模型,运用Hotelling T2和Q统计量方法对其余两类进行检测,从而判断是否是基准样本的同类。同时将其他两类样本投影到基准样本的得分空间,从二维主元分布图就可以明显看出每类样本的分布,从而达到识别的效果。本文第三章提出了一种PCA/_SVM的多故障分类方法,运用传统PCA方法对故障数据进行降维,并将所有故障数据投影到正常工况样本的PCA主元空间中,由正常数据样本计算出T2统计量的阈值,根据Hotelling T2统计原理,对所有故障数据进行检测,将检测到的故障样本通过SVM的多分类方法进行故障分类。通过TE过程仿真平台的实验表明,PCA/_SVM方法与PCA/_KNN、C/_SVM方法相比较,算法简单,容易实现,计算速度较快,同时可以达到较高的多分类准确率。此外,本文还采用了流形学习方法中的局部切空间整合算法/(LTSA/),将其作为多类故障数据的降维工具,并提出一种集成的支持向量机分类/(CSVM/)思想,在多分类过程中,对每个2分类器的核函数进行选择,择优选取分类效果较好的核函数,从而建立一个集成分类器,能够对新进的样本进行准确分类。将LTSA和集成SVM方法进行结合,运用LTSA对数据进行降维处理,然后把数据输入到SVM的不同核函数的二分类器中进行训练,并通过测试样本进行实验测试,实验表明,LTSA CSVM跟其他的支持向量机分类方法比较具有一定的优越性。
As the development of science and technology, industry process is becoming more and more complicated. In order to ensure process safety and reliability, decrease the production costs and improve products' quality, related researchers have developed many kinds of design program, control strategy and optimization algorithm. Process monitoring is a systematic engineering, it is integrated with automation control, artificial intelligent, computer science, pattern recognition and system engineering knowledge. This paper is focused on industry process monitoring, especially about multi-class faults recognition. Main research work is presented as follows: Principle component analysis /(PCA/) is usually combines with Hotelling T2 and SPE statistical methods to solve fault detection and classification problem with linear and little sample label. Fisher data was used in the second section. Train a PCA model with one class of the sample data; calculate the T2 and SPE control limits of the train samples, all the test samples' T2 and SPE value would be compared with control limits to decide whether it is abnormal data or not. Meanwhile, the other two class data was projected to the principle space. It can be obviously discovered from the 2-dimension distribution plot that different class is scattered in different space. In the third part of this paper, PCA/_SVM is presented as a multi-label recognition method. Traditional PCA is used as a dimension decrease tool, then all fault data is projected into the normal PCA space, detect the new coming data with Hotelling T2. All the detected data would be transport to the SVM classifier to judge which class the sample belong to. From the experiment based on the TE data, it can be concluded that PCA/_SVM method is simple, fast computation and higher classification accuracy. Manifold learning method is introduced in the last section. Local tangent space alignment algorithm /(LTSA/) which is one of the manifold learning method, imported as a dimension reduction method. A combined support vector machine /(CSVM/) classification is created in this part. During the multi-labels SVM classification process, each lvsl classifier will choose an appropriate kernel function Instead of just one kernel function for all the classifiers. After experiment test, the LTSA/_CSVM shows some advantages in the multi-label classification.