【作者】 胡益；

【导师】 侍洪波；

【作者基本信息】 华东理工大学 ， 控制科学与工程， 2014， 博士

【摘要】 随着现代工业朝着大规模、复杂化的方向发展,工业生产的过程监控和故障检测成为了工业系统关注的主要问题之一。通过对复杂生产过程的运行状态进行监测,及时发现过程干扰、故障以及其它异常工况,诊断出故障发生的原因,从而可以保证生产过程安全运行,提高产品质量和经济效益。近年来,随着计算机技术的发展和分布式控制系统(DCS)在工业过程中的广泛应用,大量的过程数据被采集并存储下来。因此,如何从海量数据中挖掘出隐藏的有用信息,使其服务于生产安全和产品质量控制,已经成为迫切需要解决的问题。在这种工程背景下,基于数据的多元统计方法受到了广泛关注,并被成功应用于过程建模、监控和控制领域。传统的多元统计监控方法以主成分分析(PCA)和偏最小二乘(PLS)为主,这些方法通过降维的策略提取高维、变量相关的过程数据的主要信息,从而对过程的运行状况进行分析。然而在实际应用中,传统的基于多元统计的方法存在一定的局限性,表现之一在于这些方法要求过程变量之间是线性相关的,然而流程工业中过程变量之间、过程变量与质量变量之间往往存在着非线性关系,用线性的方法对数据的非线性特性进行建模将难以得到精确的模型。本文主要将针对过程数据存在的非线性问题,将一种非线性的PLS方法一核PLS (Kernel PLS, KPLS)方法应用于工业过程的在线监控,并考虑到实际应用中存在的各种问题,对基本的KPLS方法进行改进,使其能在过程监控、故障诊断和质量预测中得到更好的应用。本文的主要研究工作包括以下几个方面：(1)针对过程监控数据的非线性特点,提出了一种基于核偏最小二乘(KPLS)的过程监控方法。该方法首先通过核函数将输入数据映射到高维特征空间,再在高维特征空间进行PLS运算。与线性PLS算法相比,KPLS算法在处理非线性问题时能够很好地建立起输入变量与输出变量之间的非线性关系,充分提取非线性系统的特征。通过构造相应的监控统计量,将KPLS方法应用于过程的在线监控,取得了比普通的PLS方法更好的监控效果。(2)针对非线性过程采集的数据存在离群值的问题,提出了一种基于球形化KPLS (Spherical KPLS, SKPLS)的鲁棒在线监控策略。该鲁棒策略通过将高维特征空间的特征向量投影到一个单位球球面得到新的特征向量,再对新的特征向量进行KPLS算法。通过球形化的处理,减少甚至消除了离群点对建模精度的影响。与普通的KPLS方法相比,提出的SKPLS方法具有鲁棒特性,能够处理含有离群点的建模数据。此外,SKPLS方法的性能要优越于其它线性的鲁棒算法,因为它能有效地捕捉变量之间的非线性关系。通过将具有鲁棒性的方法与基于核的非线性方法相结合,提出的新方法同时具有鲁棒特性和非线性特性,在应用于过程监控时能获得更精确的过程模型。(3)针对间歇过程的数据特点,将KPLS方法应用于间歇过程的数据处理,提出了一种基于多向KPLS (Multiway KPLS, MKPLS)的间歇过程监控方法。在提出的监控策略中,首先将正常间歇过程的三维数据按批次方向展开成两维的数据,再用KPLS方法建立起输入输出之间的非线性关系,最后构建相应的统计量对新的操作批次进行监控。由于KPLS方法具有很强的非线性特征提取能力,所以将提出的MKPLS方法应用于间歇过程中能获得比多向PLS (Multiway PLS, MPLS)方法更好的监控性能。(4)针对间歇过程存在的时变问题、多阶段问题以及阶段不等长问题,将“即时学习(Just-in-time-learning, JITL)"策略引入到KPLS方法中,提出了一种新的基于JITL-KPLS方法的间歇过程在线监控策略。在该策略中,用JITL方法来选取用于建模的最相似样本,然后基于相似样本建立KPLS模型用于对过程进行监控。基于局部邻域信息的JITL模型可以很好地捕捉过程的变化特性,因此适合于解决间歇过程的时变和多阶段问题。同时,KPLS方法还可以有效地处理复杂的非线性数据。此外,本文提出的监控策略应用于在线监控时不需要对未来的未知测量值进行预估。因此,提出的基于JITL-KPLS方法的监控策略在很大程度上改进了间歇过程监控性能。(5)针对KPLS方法在实际应用中存在的核矩阵计算复杂的问题和故障变量辨识的问题,提出了一种有效的非线性过程监控和故障辨识策略。从大量历史数据库中挑选出一定数量的特征样本,再基于该特征样本集建立KPLS模型,由于样本数量的减少,核矩阵的计算变得较为简单,从而缩短了KPLS方法的建模时间。同时,由于特征样本包含有足够的过程信息,所以建立的模型具有较高的精度,能获得很好的过程监控性能。为了对故障变量进行辨识,还提出了一种基于回归重构的方法,通过构造故障辨识指标可以有效地找出引发故障的重要变量,该方法提供了一种可供选择的非线性故障辨识方法。在对上述方法进行分析的同时,通过不同方法之间的比较,利用Tennessee Eastman过程和青霉素发酵过程等仿真应用进行验证,结果充分证明了本文所提出的新方法的有效性。最后,在对本文主要工作进行总结的基础上对未来的研究工作进行了展望。更多还原

【Abstract】 The modern industrial processes have become more and more large-scale and complex, so the process safety and product quality are two important issues in industries. Also, process monitoring and fault detection have attracted more attention. In the past several years, because of the development of computer technique and the wide utilization of the distributed control system (DCS), large amounts of data have been collected and stored. Therefore, it is necessary and desirable to solve the problems about how to extract useful information from the large amounts of data and how to utilize the obtained information for process safety and product quality control. In this context, data-based multivariate statistical methods have become more popular and have been successfully applied in process modeling, monitoring and control.Traditional multivariate statistical process monitoring (MSPM) methods mainly contain principal component analysis (PCA) and partial least squares (PLS). However, the traditional multivariate statistical-based method has several limitations. One of these limitations is that they require the process variables to be linearly correlated, but nonlinear relationships among different process variables are very common in the process industry, as well as between the process variables and the quality variables. This paper will focus on the nonlinear problem in the process monitoring and employ a nonlinear PLS method, called kernel PLS (KPLS), for online monitoring of industrial processes. The main contributions of this dissertation are summarized as follows:(1) To handle the nonlinear problem for process monitoring, a new technique based on kernel partial least squares (KPLS) is developed. KPLS is to first map the input space into a high-dimensional feature space via a nonlinear kernel function and then to use the standard PLS in that feature space. Compared to linear PLS, KPLS can effectively capture the nonlinear relationship between the input variables and output variables. For process monitoring, two statistics of KPLS method are constructed. KPLS can obtain precise process model and show superior process monitoring performance compared to linear PLS.(2) In order to eliminate the effect of outliers in the modeling data, a robust online monitoring approach is developed and presented for nonlinear process monitoring, which is based on spherical kernel partial least squares (SKPLS). Through projecting the feature vectors onto a unit sphere, we get new feature vectors, and then an ordinary KPLS is applied onto these new feature vectors. Due to the sphering, the influence of outliers is reduced heavily. Compared with ordinary KPLS, the proposed method incorporates a robustness feature for coping with the contaminated modeling data. By integrating a robust method into a kernel based method, the proposed method has both robust and nonlinear characteristic and can be applied to obtain more accurate models for process monitoring purposes.(3) For the batch process, a process monitoring method is developed based on multiway kernel partial least squares (MKPLS). Three-way batch data of normal batch process is unfolded batch-wise, and then kernel partial least squares (KPLS) is applied to capture the nonlinear relationship between the latent structures and predictive variables. Because of its good ability to descript the nonlinear characteristic, MKPLS can detect faults or disturbance more accurately and rapidly than multiway PLS (MPLS), especially for batch process having nonlinear characteristics.(4) An integrated framework consisting of just-in-time-learning (JITL) method and kernel partial least squares (KPLS) is described for the monitoring of the performance of batch processes. The training data set for modeling is defined by using JITL, and KPLS is employed to build the model for process monitoring purpose. The JITL model based on local neighborhoods of similar samples is very accurate and sensitive because it can well track the change of the process. Meanwhile, KPLS is a very efficient technique for tacking complex nonlinear data sets. Furthermore, the proposed monitoring strategy does not require the estimation of future values of the process variables during online application. As a result, the proposed scheme greatly improves the monitoring performance of batch processes.(5) An efficient KPLS-based fault diagnosis framework is proposed to address the problems of the complex kernel matrix calculation and the kernel method based fault identification. A small subset of feature samples is chosen from the large dataset and the KPLS model is built base on the selected feature samples. Since the small subset contains enough information of the process data, the precise model can be obtain and the calculation burden can be reduced significantly. Besides, the proposed fault identification approach can identify the fault source efficiently.Several monitoring methods are compared with the proposed methods, and the TE process and fed-batch penicillin fermentation process are applied to illustrate the efficient of the proposed methods. Finally, some conclusions and future research directions are discussed.更多还原