基于PLS的关键性能指标相关的故障检测方法研究

发布时间：2018-03-24 15:14

本文选题：故障检测　切入点：关键性能指标　出处：《渤海大学》2017年硕士论文

【摘要】：随着对工业生产系统中鲁棒性需求的不断增加,工业系统中的基于数据的过程监控技术已成为学术界的热门课题。这种基于数据的监控技术能够快速准确地检测出工业过程中所存在的故障,并加以处理来保证系统的稳定性。然而,研究报告和工业反馈显示并非所有过程故障都会影响产品质量。如若忽略某些不必要的报警,可以减少检修时间和人力,提高生产效率。基于此,一种新型的故障检测方法应运而生,即关键性能指标(Key Performance Indicators,KPI)相关的故障检测技术。该技术可以有效地区分工业过程中的故障是否影响产品的KPI。但该方法无法很好地解决工业系统所采集的数据中存在离群点的问题。本文拟对多元统计分析技术中偏最小二乘法(Partial Least Squares,PLS)进行改进,并与KPI故障检测技术相融合。对线性和非线性静态系统,进行KPI相关的鲁棒故障检测方法具体研究。主要研究内容如下:(1)在线性过程中,针对现有的PLS对离群点的敏感以及故障判决逻辑复杂等情况,本论文将引进一种鲁棒的PLS算法降低建模数据中离群点对PLS模型的影响。之后,通过奇异值分解(Singular Value Decomposition,SVD)对过程变量分解为KPI相关和KPI无关两个正交子空间。从而,提出一种鲁棒的KPI预测和线性故障检测方法,该方法能够区分所发生的故障是否影响KPI。采用Tennessee Eastman(TE)化工过程,对两种算法进行KPI预测和KPI故障诊断。与PLS相比,该方法具有更强的鲁棒性和更优的故障检测性能。(2)对于非线性过程中测量数据含有离群值的问题,本论文在前人的基础上,提出一种改进的球形化KPLS(Modified Spherical KPLS,MSKPLS)在线检测方法。该方法利用球形化策略将核处理后得到的高维特征变量投影到一个单位球面上。而后,对球形化处理后的数据进行KPLS算法计算,并利用建模得到的系数矩阵对所得到的变量空间进行SVD分解,得到KPI相关和KPI无关两个正交子空间。从而可以对故障进行有效的检测。该策略不仅可以降低离群点对建模精度的影响,还可以解决非线性、KPI检测两大关键问题。通过数值例子和TE化工系统对已有算法和新方法进行仿真对比,结果显示新方法在KPI故障检测上有更优的性能。
[Abstract]:With the increasing demand for robustness in industrial production systems, The data-based process monitoring technology in industrial systems has become a hot topic in academic circles. This data-based monitoring technology can quickly and accurately detect the faults in industrial processes. However, research reports and industrial feedback show that not all process failures affect the quality of the product. Ignoring unnecessary alarms can reduce overhaul time and manpower. To improve production efficiency. Based on this, a new fault detection method emerged as the times require, That is, the key Performance indicators / KPI (key Performance indicators / KPI) related fault detection technology. This technique can effectively distinguish whether the failure in the industrial process affects the KPIs of the product. But this method can not solve the problem in the data collected by the industrial system. The problem of outliers. In this paper, the partial Least Squares (PLSs) in the multivariate statistical analysis technique is improved. And fusion with KPI fault detection technology. For linear and nonlinear static systems, the robust fault detection methods related to KPI are studied in detail. The main research contents are as follows: 1) in the linear process, In view of the sensitivity of PLS to outliers and the complexity of fault decision logic, this paper introduces a robust PLS algorithm to reduce the influence of outliers on PLS model. Singular Value decomposition (SVD) is used to decompose process variables into KPI correlation and KPI independent orthogonal subspaces. A robust KPI prediction and linear fault detection method is proposed. The method can distinguish whether the fault occurred affects the KPI. The chemical process using Tennessee Eastman Tet is used to predict the KPI and diagnose the KPI fault of the two algorithms. Compared with PLS, the proposed method can be used to predict KPI and diagnose KPI faults. This method has stronger robustness and better fault detection performance. An improved on-line detection method for spherical KPLS(Modified Spherical KPLS- MSKPLS is proposed, in which the high-dimensional feature variables obtained by kernel processing are projected onto a unit sphere by using the sphericity strategy. Then, the KPLS algorithm is used to calculate the spherically processed data. The coefficient matrix is used to decompose the variable space by SVD. Two orthogonal subspaces of KPI correlation and KPI independence are obtained, which can effectively detect faults. This strategy can not only reduce the influence of outliers on modeling accuracy. Two key problems of nonlinear KPI detection can also be solved. The simulation results of existing algorithms and new methods by numerical examples and te chemical systems show that the new method has better performance in KPI fault detection.
【学位授予单位】：渤海大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP277;O212.4

【相似文献】