带时滞估计的软测量建模方法研究

发布时间：2018-09-03 07:13

【摘要】：化工过程对象呈现显著的非线性和时变性,为了对过程实施高效的监控策略,广泛以软测量技术的手段对反映产品质量指标的难测变量(即主导变量)进行推断估计。如今,随着过程工况复杂度的日益增加,工业界对于软测量技术的精度和可靠性方面的要求也相应提高。在实际过程中,主导变量的获取通常受到装置成本、仪表可靠性或技术瓶颈等方面的限制,存在很大的测量滞后性。尽管软测量建模领域的研究不断迈向自适应时代,时滞信息却往往不被考虑在建模过程中。为了进一步改善传统软测量建模方法的预测精度,本论文不仅针对工业过程的时变和非线性特征,同时还考虑了过程数据集中隐含的时滞信息,在现有的软测量技术研究成果的基础上,以时间差高斯过程回归算法为基础,对带时滞估计的自适应软测量建模方法进行了研究。全文的主要研究内容如下:1.针对建模数据时序匹配不一致和变量漂移的问题,提出一种基于模糊曲线分析(Fuzzy Curve Analysis,FCA)的时间差高斯过程回归(Time Difference Gaussian Process Regression,TDGPR)建模方法。该方法利用离线估计的时滞参数重新匹配建模样本时序,对于查询样本,采用TDGPR模型对主导变量进行在线预测。2.针对传统全局时间差(Time Difference,TD)模型的“老化”问题,基于选择性集成思想,提出一种局部时间差高斯过程回归(Local Time Difference Gaussian Process Regression,LTDGPR)的自适应建模方法。首先,对数据库中的时滞动态信息进行挖掘,并利用该信息对建模数据进行重构;然后,采取局部化策略对差分后的重构样本进行统计划分,得到LTDGPR模型集。对于查询样本,在线选择部分泛化能力强的LTDGPR子模型进行集成,估计出含一定时间差的主导变量动态偏移值;最后,基于TD模型思想对主导变量值进行实时预测。3.考虑到过程非线性和时滞呈现出的阶段性特征,提出一种基于局部时滞重构(Local Time-delay Reconstruction,LTR)的滑动窗时间差高斯过程回归(Moving Window Time Difference Gaussian Process Regression,MWTDGPR)建模方法。该方法以滑动窗和TD组合策略的方式逐步跟踪过程局部非线性突变和缓变特征,同步实现了局部样本集的时序校正和重构,为其它的时滞非线性时变化工过程建模问题提供了一种可行的框架。论文通过实际过程的数据仿真研究验证了上述方法的可行性和精度,仿真结果充分显示考虑时滞估计的自适应软测量建模对于化工过程的经济效益和安全平稳运行具有重大意义。
[Abstract]:Chemical process objects show significant nonlinear and time-varying characteristics. In order to implement an efficient monitoring strategy, the difficult variables (i.e. dominant variables) which reflect the product quality index are inferred and estimated by soft sensing technology. Nowadays, with the increasing complexity of the process, the requirements of the industry for the accuracy and reliability of soft sensing technology are also increased. In the actual process, the acquisition of dominant variables is usually limited by the cost of the device, the reliability of the instrument or the technical bottleneck, etc., so there is a great lag in measurement. Although the research in the field of soft sensor modeling is moving towards the era of adaptation, time-delay information is often not considered in the modeling process. In order to further improve the prediction accuracy of traditional soft sensor modeling methods, this paper not only aims at the time-varying and nonlinear characteristics of industrial processes, but also takes into account the implicit time-delay information in the process data set. Based on the existing research results of soft sensing technology and based on the time difference Gao Si process regression algorithm, the adaptive soft sensor modeling method with time delay estimation is studied. The main contents of this paper are as follows: 1. Aiming at the inconsistency of time series matching and variable drift of modeling data, a time difference Gao Si process regression (Time Difference Gaussian Process Regression,TDGPR (Time Difference Gaussian Process Regression,TDGPR) modeling method based on fuzzy curve analysis (Fuzzy Curve Analysis,FCA) is proposed. In this method, the time series of modeling samples is rematched by off-line estimation of time-delay parameters. For query samples, the TDGPR model is used to predict the dominant variables online. Aiming at the problem of "aging" of traditional global time difference (Time Difference,TD) model, an adaptive modeling method of local time-difference Gao Si process regression (Local Time Difference Gaussian Process Regression,LTDGPR) is proposed based on the idea of selective integration. Firstly, the time-delay dynamic information in the database is mined, and the modeling data is reconstructed using this information. Then, the LTDGPR model set is obtained by statistical partitioning of the reconstructed samples after the difference by using the localization strategy. For the query samples, the LTDGPR submodel with strong generalization ability is selected online to integrate to estimate the dynamic offset value of the dominant variable with a certain time difference. Finally, based on the idea of TD model, the dominant variable value is predicted in real time. 3. Considering the stage characteristics of process nonlinearity and time-delay, a sliding window time-difference Gao Si process regression (Moving Window Time Difference Gaussian Process Regression,MWTDGPR modeling method based on local time-delay reconstruction (Local Time-delay Reconstruction,LTR) is proposed. The method uses sliding window and TD combination strategy to track the local nonlinear mutation and slowly varying feature step by step, and synchronously realizes the timing correction and reconstruction of the local sample set. It provides a feasible framework for other nonlinear time-varying chemical process modeling problems with time delay. The feasibility and accuracy of the above method are verified by the data simulation of the actual process. The simulation results show that the adaptive soft-sensor modeling with time-delay estimation is of great significance for the economic benefit and safe and stable operation of chemical process.
【学位授予单位】：江南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TQ018

【参考文献】