“Plant-Friendly” System Identification: A Challenge for the Process Industries

1. “Plant-Friendly” System Identification: A Challenge for the Process Industries Daniel E. RiveraControl Systems Engineering LaboratoryDepartment of Chemical and Materials EngineeringArizona State UniversityTempe, Arizona 85287-6006http://www.fulton.asu.edu/~cseldaniel.rivera@asu.edu 2004 CSChE Mtg.

2. Presentation Outline • What is “plant-friendliness” in identification testing? • Definition and Origins • Practical Motivation • Survey of some plant-friendly identification approaches • Friendliness criteria • Optimization-based formulations • Constrained multisine signals for highly interactive processes • Minimum crest factor • Uniformly distributed/geometric discrepancy • Identification Test Monitoring • Summary and Conclusions 2004 CSChE Mtg.

3. “Plant-Friendly” Identification Testing • The term originates from the chemical process control community; first used by Dupont control researchers and collaborators in the early 90’s (a 1993 ACC paper by Pearson, Ogunnaike and Doyle makes the first mention in print). • Is principally motivated by the desire for informative identification experiments while meeting the demands of industrial practice. • Broadly speaking, a plant-friendly test yields data leading to a suitable model within an acceptable time period, while keeping the variation in both input and output signals within user-defined constraints. 2004 CSChE Mtg.

4. “Plant-Friendly” Identification Testing (Continued) • The ideal plant-friendly identification test should: • be as short as possible, • not take actuators to limits, or exceed move size restrictions, • cause minimal disruption to the controlled variables (i.e., low variance, small deviations from setpoint). 2004 CSChE Mtg.

5. Motivation for a Fundamental Examination of Plant-Friendly ID • Plant operations desires plant-friendliness, but classical identification theory is “plant-hostile” 2004 CSChE Mtg.

6. Reducing Variance Effects Asymptotic Variance Expressions for independent open-loop estimation, per Ljung (1987, 1999) Reducing the number of estimated model parameters, increasing the length of the data set , and i ncreasing the power of the input signal all contribute to variance reduction in system identification 2004 CSChE Mtg.

7. Motivation for a Fundamental Examination of Plant-Friendly ID • Plant operations desires plant-friendliness, but classical identification theory is “plant-hostile” • Identification testing is an expensive proposition, and improper execution can endanger a project. • There is an absence of fundamentally based, systematic guidelines in the literature for problems of practical significance 2004 CSChE Mtg.

8. Process Testing Duration(as reported by Mitsubishi Chemical engineers, from guidelines presented by a major process control software vendor) Suggested Test Duration = (6...8)*(Estimated Settling Time Process)*(Number of Independent Variables) Example: Ethylene Fractionator: 6*6 (hrs)*17 = 612 (hrs) = 25.5 (days)8*6 (hrs)*17 = 816 (hrs) = 34 (days) 2004 CSChE Mtg.

9. Inputs Outputs Estimate for a large Air Separation Unit: 2 months at the plant 24 hrs/day! Incentives for “Fast” Identification Testing Per Kothare and Mandler, Air Products & Chemicals, (presented at the 2003 AIChE Annual Mtg.) 2004 CSChE Mtg.

10. Typical Costs of Step Testing(from Mathur and Conroy, “Multivariable Control without Plant Tests” 2002 AIChE Annual Mtg.) • Cut throughput, 5-10% for 6-8 weeks $ 50,000 • One off-grade excursion, 100% production loss $ 60,000 • Engineering (testing) 6-8 weeks, 24 hours/day $140,000 • Engineering (commissioning), 2 weeks, 24 hours/day $ 20,000 Total:$270,000 2004 CSChE Mtg.

11. Motivation for a Fundamental Examination of Plant-Friendly ID • Plant operations desires plant-friendliness, but classical identification theory is “plant-hostile” • Identification testing is an expensive proposition, and improper execution can endanger a project. • There is an absence of fundamentally based, systematic guidelines in the literature for problems of practical significance • Some now well established identification topics (e.g., classical optimal input design, control-relevant identification, closed-loop identification) are helpful but do not address all the issues. 2004 CSChE Mtg.

12. Optimal Input Signal Design • Classical formulations (summarized in Chpt. 13 of Ljung’s System Identification: Theory for the User) address minimizing the constrained variance of the input and/or output signals • The optimal experimental design depends on the (unknown) true system and noise characteristics • In practice, process control engineers tend to think more in terms of maintaining high/low limits, move size constraints, and minimizing test duration rather than constrained variance. 2004 CSChE Mtg.

13. Control-Relevant and Closed-Loop Identification • Important topics in system identification since the late 80’s • A control-relevant or closed-loop design may emphasize a narrower bandwidth than a traditional design, which may result in an experimental test of shorter duration • As before, a priori system knowledge is critical • Plant-friendliness still needs to be maintained, regardless. 2004 CSChE Mtg.

14. Guillaume (blue) vs. Schroeder-phased (red) multisine signals - control-relevant design example Power Spectral Density One Data Cycle Schroeder signal has 121% larger input span, 47% larger input move size, and 49% larger output span than the Guillaume signal. 2004 CSChE Mtg.

15. Closed-Loop Identification Issues From Ljung, "Identification in Closed Loop: Some Aspects on Direct and Indirect Approaches," invited paper for SYSID '97, Fukuoka, Japan. • "... the basic problem in closed-loop identification (is this): the purpose of feedback is to make the sensitivity function small, especially at frequencies with disturbances and poor system knowledge. Feedback will thus worsen the measured data's information about the system at these frequencies." • There are no difficulties, per se, with closed-loop data; simply that in practical use, the information content is less • One could make closed-loop experiments with good information contents (but poor control performance) 2004 CSChE Mtg.

16. Survey of Friendliness Approaches • Plant-friendliness criteria • Friendliness index • Crest Factor (and the Performance Index for Perturbation Signals) • Optimization-based problem formulations 2004 CSChE Mtg.

17. Friendliness Index(Doyle et al., 1999; Rengasamy et al. 2000; Parker et al., 2001) • Defined as part of a design procedure for inputs intended to identify Volterra series models • A constant sequence is 100% friendly, while one that changes at every sampling instant is 0% friendly 2004 CSChE Mtg.

18. Crest Factor The Crest Factor (CF) is defined as the ratio of (or Chebyshev) norm and the norm A low crest factor indicates that most elements in the input sequence are located near the minimum and maximum values of the sequence. • Seminal paper by Schroeder (1970) presents an analytical formula for determining phases in multisine signals that leads to near-optimal crest factors (for wide-band signals) • Work by Guillaume et al. (1991) provides a very efficient numerical technique for computing minimum crest factor multisine signals with arbitrary power spectral densities • The Performance Index for Perturbation Signals (PIPS, Godfrey, Barker, and Tucker, 1999) is an equivalent yet practical alternative. 2004 CSChE Mtg.

19. Crest Factor Signal Comparison Two signals with identical spectra and different crest factors can have markedly different “plant-friendliness” properties. The Performance Index for Perturbation Signals (PIPS) is a practical alternative (Godfrey, Barker, & Tucker, IEE Proc. Control Theory Appl.,1999): 2004 CSChE Mtg.

20. Optimization-Based Plant-Friendly Problem Formulations • Addressing control-relevance with constraints (Chikkula and Lee (1997); Cooley et al., (1998); Cooley and Lee (2001), Li and Georgakis (2002, 2003)) • Multiobjective approach involving friendliness index, constraints, and other criteria (Narasimhan et al. (2003, 2004)) • Minimizing crest factor with time-domain constraints for arbitrary signal spectra (Rivera et al. (2002, 2004), H. Lee et al. (2003a,b)) 2004 CSChE Mtg.

21. Multisine Input Signals A multisine input is a deterministic, periodic signal composed of a harmonically related sum of sinusoids, 2004 CSChE Mtg.

22. “Zippered” Power Spectrum 2004 CSChE Mtg.

23. Multisine Signal Design Guideline(H.Lee, D.Rivera, H. Mittelmann, SYSID 2003) 2004 CSChE Mtg.

24. Case Study: High-Purity Distillation High-Purity Distillation Column per Weischedel and McAvoy (1980) : a classical example of a highly interactive process system, and a challenging problem for system identification and control system design 2004 CSChE Mtg.

25. Output Gain Directionality (Stec and Zhu, ACC 1999) A plant-friendly identification test for a highly interactive system should be able to improve the gain-directionality of the output while meeting operating requirements 2004 CSChE Mtg.

26. Identifying Highly Interactive Systems (Stec and Zhu, 2001 ACC) The sequential cycles of correlated and uncorrelated signals provide a mechanism for generating a data set with good information content in both high and low gain directions (e.g., tested with a linear model) 2004 CSChE Mtg.

27. Modified Zippered Spectrum Correlated harmonics are now present! 2004 CSChE Mtg.

28. Design Guideline for Modified Zippered Harmonics(relies on an estimate of the steady-state gain) 2004 CSChE Mtg.

29. Problem Statement #1 2004 CSChE Mtg.

30. Problem Statement #2 This problem statement requires an a priori model to generate output predictions 2004 CSChE Mtg.

31. Other Problem Formulations • Minimize worst-case of both input and output crest factors • Incorporate controller equations in the optimization problem for signal design under closed-loop conditions • Examine alternative criteria (e.g., geometric discrepancy via Weyl’s Theorem) in lieu of crest factor. 2004 CSChE Mtg.

32. Constrained Solution Approach Some aspects of our numerical solution approach: • The problem is formulated in the modeling language AMPL, which provides exact, automatic differentiation up to second derivatives. • A direct min-max solution is used where the nonsmoothness in the problem is transferred to the constraints. • The trust region, interior point method developed by Nocedal and co-workers (Byrd, R., M.E. Hribar, and J. Nocedal. “An interior point method for large scale nonlinear programming.” SIAM J. Optim., 1999) is applied. 2004 CSChE Mtg.

33. Standard & Modified Zippered Spectrum Design Standard Zippered Spectrum Modified Zippered Spectrum 2004 CSChE Mtg.

34. State-space Analysis Output State-Space Input State-Space +(blue): min CF(y) signal with a modified zippered spectrum and a priori ARX model *(red) : min CF(u) signal with a standard zippered spectrum 2004 CSChE Mtg.

35. min CF signal design: time-domain min CF(u) signal withStandard Zippered Spectrum min CF(y) signal with ARX model and Modified Zippered Spectrum SNR = [-0.04, -1.12] dB SNR = [-5.0, -5.0] dB 2004 CSChE Mtg.

36. Min CF Signals with ARX Model Output Predictions: Time-Domain Comparison 2004 CSChE Mtg.

37. Closed-loop Performance Comparison, MPC Setpoint Tracking: Models obtained from noisy data Model Predictive Control (MPC) optimizes the predicted future values of the plant output based on previous and future information MPC Tuning Parameters: Prediction Horizon : 100 Move Horizon : 25 Output Weighting: [1 1] Input Weighting : [0.2 0.2] 2004 CSChE Mtg.

38. ARX Model Prediction vs. Plant Data + (blue) : Model Prediction * (red) : Weischedel-McAvoy Distillation Simulation 2004 CSChE Mtg.

39. NARX Model Estimation We rely on a NARX model to predict the system outputs during optimization (Sriniwas et al., 1995) 2004 CSChE Mtg.

40. ARX vs. NARX Model Predictions NARX Model ARX Model + (blue) : Model Prediction * (red) : Weischedel-McAvoy Distillation Simulation 2004 CSChE Mtg.

41. Min CF Signal Comparisons - ARX and NARX (Time-Domain) 2004 CSChE Mtg.

42. Model-on-Demand Estimation (Stenman, 1999) • A modern data-centric approach developed at Linkoping University • Identification signals geared for MoD estimation should consider the geometrical distribution of data over the state-space. current operating point 2004 CSChE Mtg.

43. Weyl Criterion 2004 CSChE Mtg.

44. min Crest Factor vs Weyl-based Signals: Output State-Space Modified Zippered, min CF (y) Signal Modified Zippered, Weyl-based signal 2004 CSChE Mtg.

45. min Crest Factor vs Weyl-based Signals - Power Spectra Modified Zippered, min CF (y) Signal Modified Zippered, Weyl-based signals All harmonic coefficients are selected by the optimizer in the Weyl-based problem formulation 2004 CSChE Mtg.

46. min Crest Factor vs Weyl-based Signals - Time-domain Comparisons 2004 CSChE Mtg.

47. Some Pertinent Questions • How does one build process knowledge relevant to system identification in a systematic and (nearly) automatic way, with little user intervention and without demanding significant computational time and effort? • How is process knowledge systematically acquired in the course of identification testing, for purposes of improving the identification test? 2004 CSChE Mtg.

48. Identification Test Monitoring • Relies on the use of periodic, deterministic inputs (such as multisines or pseudo-random signals) to define a natural window for analysis, • Relies on concepts from signal processing, robust control, and optimization to develop measures that systematically acquire and apply process knowledge, and use this knowledge to refine the design parameters of the identification test 2004 CSChE Mtg.

49. Identification Test Monitoring Scheme Apply and analyze data from periodic test signals, cycle-by-cycle, to improving system knowledge and refine the experimental design in a control-relevant manner. 2004 CSChE Mtg.

50. Statistical Plant Set Estimation(Bayard, 1993) 2004 CSChE Mtg.