【作者】 赵宇；

【导师】 褚健； 苏宏业； Biao Huang；

【作者基本信息】 浙江大学 ， 控制科学与工程， 2011， 博士

【摘要】 控制系统性能评估与监控技术(Control Performance Assessment & Monitoring, CPA&M)是流程工业过程控制领域自上世纪80年代末新兴的一项重要技术,旨在不影响系统正常运行的情况下,利用过程常规运行数据,通过建立基准控制器以及多种不同的性能指数,对控制系统控制运行性能的变化进行实时监控,给出性能评估结果并诊断引起系统控制性能下降的原因,进而提出相应的保持／改善系统控制性能的策略,目前已成为维护现代流程工业自动化系统保持平稳安全高效运行的主要技术。已有的控制回路性能评估与监控算法大多需要已知太多过程信息,包括系统时滞、对象甚至扰动模型。此外,大多性能评估算法基于最小方差控制(Minimum Variance Control MVC)基准,不能对先进控制策略(Model Predictive Control MPC)给出可靠的评估结果,对于非线性过程的控制性能评估技术尚未有太多的成果。针对上述存在的问题,本文结合当前控制系统性能评估与监控技术理论研究以及工业应用现状,在基于预报误差方法的框架下,开展了以下几个方面的研究工作：1.针对多变量闭环控制系统提出了一种基于拓展预报误差方法的控制性能监控算法。利用系统常规闭环运行数据建立滑动平均(Moving Average MA)时间序列模型,计算系统的预测误差,根据预测误差均方和定义了新的控制系统闭环潜能指数。在不同时滞条件下得到的闭环潜能曲线来反映控制系统的性能存在多大的提升余地。由于所定义闭环潜能指数同时考虑了对过程静态控制性能和动态控制性能的综合评估,可以对控制系统的跟踪性能和调节性能给出全面客观的评估结果。2.针对多变量MPC控制系统提出了一种基于多步预报误差方法的性能监控算法。由于MPC的目标函数当中存在预测时域,实现的是在预测时域内的多步最优控制策略,因此对MPC控制系统进行性能评估需要建立基于多步预报误差的基准。基于多步最优基准定义了新的闭环潜能指标,该指标同MPC的预测时域建立了联系,在固定预测时域的条件下,根据不同的过程时滞取值可以获取系统的潜在性能提升曲线,所得曲线可以反映MPC控制系统当前的控制性能趴离最优控制性能还有多大提升余地,可以为工程师进行MPC的维护和控制器参数整定提供有效的参考。3.针对多变量控制系统提出了一种基于广义最小方差(Generalized Minimum Variance GMV)控制基准的性能评估算法。针对最小方差控制基准存在的高增益、广带宽、需要频繁剧烈调节控制作用等缺点,在广义最小方差控制基准的目标函数当中引入了Quadratic Gaussian对控制项的约束,并且GMV中采用了动态权重矩阵,不仅可以根据变量的的重要性赋予不同的权重大小,还可以根据系统在不同频段的重要程度修改动态权重系数的大小,从而能够确定更加客观的评估基准对控制系统的控制性能进行评估。4.针对一类线性参数时变(Linear Parameter Varying LPV)过程提出了基于预报误差方法的建模和性能评估算法。文中首先将线性系统中广泛应用的预报误差建模方法拓展到线性参数时变过程的建模当中,对具有Box-Jenkins (BJ)结构的LPV模型进行了建模。在预报误差方法框架下,具有模型插值结构和参数插值结构的LPV模型参数都可以被精确估计。获取模型参数后,在LPV最优控制策略作用下可以得到LPV的方差下界,以此来作为控制性能评估基准,对其他控制器作用下的LPV过程进行性能评估。5.针对一类非线性过程提出了基于广义预报误差方法的性能评估算法。实际工业过程本质上均为非线性系统,对于某些复杂对象采用简单的线性系统不能够有效捕捉其动态特性,非线性模型能够提供更精确的过程特性描述能力和更准确的过程行为预测能力。文中利用正交最小二乘方法对一类在线性时不变扰动作用下的非线性多项式自回归(Nonlinear Polynomial AutoRegressive NPAR)过程进行了建模,通过计算得到的非线性预报误差定义了非线性过程潜能指数,该指数同时考虑了过程的静态和动态控制性能,能够对非线性过程的跟踪和调节控制性能给出综合客观的评估结果。更多还原

【Abstract】 Control Performance Assessment/Monitoring (CPA&M) is an important technology developed since the 80’s of last century, which aims to monitor the system’s performance by using only the normal route operating data without influencing the system’s operation. Performance assessment results and root causes for performance degradation or deteriora-tion are supposed to report according to different controller benchmarks and performance indices, corresponding maintenance and improvement suggestions are expected to provide to control engineers. The technology is becoming more and more important for the main-tenance of industrial automatic with smooth, safety and high efficiency operation.Many of the current available CPA&M algorithms require considerable amount of process information, the time delay, process model or even the disturbance model, etc. Moreover, most algorithms are minimum variance control (MVC) based, which cannot provide reasonable performance assessment results for the model predictive control (MPC) systems. Currently, there are only a few results for the control performance assessment of nonlinear processes. For the above mentioned challenging problems, the thesis will focus on the following research topics in both theoretic and industrial application aspects:1. An extended prediction error method based algorithm is developed for the control performance monitoring of multivariable control systems. By using only the routine closed-loop operating data, moving average model of the system can be obtained, based on which the prediction error of the system can be calculated. Consequently a new closed loop potential index is developed. At different time delays the cor-responding potential indices can form a potential trajectory, which can reflect how much performance improvement potential the current control system can have. In-dividual potential index is also defined for each sub-loop, depending on which the performance of the sub-loops can be determined. The proposed algorithm can be used for the performance monitoring of control systems with different kinds of con-trollers (PID, MPC, etc.). Additionally, both the static performance and dynamic per-formance of the system are considered in the proposed closed-loop potential index, based on which a comprehensive tracking and regulating performance assessment result can be obtained.2. A multi-step prediction error based algorithm is presented for the control perfor-mance monitoring of MPC systems. The traditional prediction error methods are all based on the single-step optimal prediction, and essentially it is equal to the min-imum variance control. As for the objective function of MPC, multi-step optimal control action is supposed to be implemented within the prediction horizon, hence the multi-step prediction based benchmark should be more appropriate for the per-formance monitoring of MPC systems. The proposed algorithm first builds the re-lationship between single-step optimal prediction and multi-step optimal prediction; based on the latter a new closed-loop potential index is defined, which is prediction horizon related. With the fixed prediction horizon, a potential trajectory can be ob-tained via different time delays to reflect how much potential performance can be improved for the current control system in comparison with the optimal benchmark. Individual potential trajectory can indicate the performance status for each sub-loop. The above results can provide useful suggestions to control engineers and operators for the maintenance and controller tuning of MPC.3. A generalized minimum variance control benchmark based algorithm is proposed for the control performance assessment of multivariable control systems. To cope with the drawbacks of minimum variance control benchmark, such as high gain, wide bandwidth and unrealistically large control signal variations, etc., constraint on the control action is added to the objective function of generalized minimum variance (GMV) benchmark, which is similar to the objective function of Linear Quadratic Gaussian (LQG). However, in comparison with the constant weighting in LQG, dynamic weighting is used in GMV, based on which the weighting can be assigned not only depending on the importance of different variables but also depend on the importance of different bandwidths. Hence it can be served as a more realistic benchmark. The overall performance index and individual performance indices can both reflect the performance of the whole system and that of the sub-loops. The overall performance index can be obtained using only the closed-loop operating data, which avoids the requirement of process model and the estimation of the interactor matrix, and hence is convenient for the practical industrial application.4. A prediction error method based modeling and control performance assessment al-gorithm is developed for a type of linear parameter varying (LPV) process. Accurate process model can provide useful information to the application of control perfor-mance assessment. The widely used prediction error methods (PEMs) in the identi-fication of linear systems are extended to the modeling of LPV processes. By using the Box-Jenkins (BJ) model structure, both process model and disturbance model are considered. Under the PEMs framework, both the model interpolation based and parameter interpolation based LPV model structures can be effectively identified. With the obtained LPV model, optimal control strategy is applied to obtain the low bound variance, which can be served as the benchmark for the control performance assessment of the LPV process under different control strategies.5. A generalized prediction error method based algorithm is developed for the control performance monitoring of a class of nonlinear process. The practical industrial pro-cesses are essentially all nonlinear processes. For some complex processes the linear models cannot effectively capture the dynamic behaviors, while nonlinear models can provide more accurate process description capacity and more reliable process prediction ability. Orthogonal Least Squares (OLS) method is applied to the param-eter estimation of Nonlinear Polynomial AutoRegressive (NPAR) model with linear time invariant disturbance, based on which the prediction error of the nonlinear pro-cess can be estimated. Finally the nonlinear process potential index is defined con- (?)tering both the static and dynamic performance to achieve the purpose of control pedormance assessment of nonlinear processes.更多还原