Research Of Data-driven Strategies For Control System Performance Monitoring

Identification, control and performance monitoring are three main themes inSystem and Control. Identification is to find a mathematical model to describe thissystem from given measurement data. The task of control lies in designing acontroller for the system to meet certain requirements of control and index. And theaim of control performance monitoring is to monitor control performance changes ofthe system, analyze and diagnose the root cause of poor performance in real-time.Generally, the design of controller and control performance monitoring are based onsystem model. Thus, the identification of process model is the precondition andfoundation of control and performance monitoring.Nowadays, control loop performance monitoring and diagnose is one of themost active research field in process control. Model identification is the keyprocedure prior to designing the optimal predictive controller. Control performancemonitoring plays an important role in keeping the control system in good status.When the control and performance monitoring strategies based on model is used andthis process model changes over time, we need to re-identify the system to ensureperformance of the controller and the accuracy of performance monitoring. Inaddition, open-loop model identification method is not fit for identify theclosed-loop system in operation and closed-loop identification becomes an essentialsolution. Therefore, process model identification, controller design and controlperformance monitoring are closely linked in industry process control.Control performance monitoring based on model need process control, sodiscussion are made from two aspects, system identification and performancemonitoring, to achieve control performance monitoring strategies based on processmeasurement data in this paper. For open-loop system identification problem,non-negative garrote variable selection method is employed to develop a NNG-based(Non-Negative Garrote) open-loop system identification algorithm. Analysis and comparison are made between NNG method and OLS (Ordinary Least Square)algorithm on the accuracy of open-loop identification, time series variable selectionproblem in regression equation and the identification of system delay. Forclosed-loop system identification problem, a novel recursive closed-loop subspaceidentification method is proposed to obtain process model based on orthogonaldecomposition (ORT). The proposed two identification method have goodapplication signification for performance monitoring of control systems. For theperformance monitoring, firstly an online control performance monitoring strategy isprovided based on the minimum-variance-benchmark-related performanceassessment method and index proposed by Harris. Secondly, an online performancemonitoring algorithm is proposed on the basis of recursive updating of covariancematrix. Following are the main content of this paper:(1) History and current status of system identification and control systemperformance monitoring is introduced. Statistical summarization of the applicationof system identification and performance monitoring in industrial process are givenand the main existing problems in system identification and performance monitoringare analyzed.(2) NNG-based open-loop system identification algorithm. For open-loopsystem identification problem, NNG variable selection method is employed todevelop a NNG-based open-loop system identification algorithm. Analysis andcomparison are made between NNG method and OLS algorithm on the accuracy ofopen-loop identification, time series variable selection problem in regressionequation and the identification of system delay. By the simulations on SISO (SingleInput Single Output) system, MIMO (Multi-input Multi-output) system and apractical industry process, better identification performance of the proposed methodis proved, and information about the delay time can be obtained.(3) Recursive closed-loop subspace identification algorithm based onorthogonal decomposition. A Recursive closed-loop subspace identification algorithm based on orthogonal decomposition is proposed by combining two stageclosed-loop subspace identification method based on orthogonal decompositionwhich is raised by Katayama and Bona-Fide algorithm to update factorization whichis used in LQ decomposition. Parameters of time-varying system can be effectivelyidentified by this algorithm. It solves the computation load problem in recursiveupdating and meets the online identification requirement. The effective ness of thisalgorithm is proved by the simulation research.(4) Online control system performance monitoring algorithm based onminimum variance benchmark. According to the performance assessment methodand the performance index based on minimum variance benchmark proposed byHarris and the kernel algorithm for PLS (Partial Least Square), an online controlperformance monitoring strategy is provided. It needs only closed-loop output dataand system delay. By simulation on typical system, it is proved that this algorithmcan be used to assess performance changes of the control system in real time andguide the adjustment of controller parameters.(5) Online control system performance monitoring algorithm based oncovariance benchmark. Selecting ideal ‘golden’ period process output data asbenchmark without the information about the system, a covariance matrix updatingmethod is provided and then an online control system performance monitoringalgorithm based on covariance benchmark is proposed. The proposed method canmonitor performance change of the system in real time and detects abnormalperformance immediately. The effectiveness is proved by applying this algorithm inthe performance monitoring of model predictive control for a heavy oil fractionsystem.