1. © ABB Corporate Research Insert image here Control Performance Monitoring Alf Isaksson, Alexander Horch ABB Corporate Research PROST Seminar 22 January 2002

2. © ABB Coprorate Research - 2 - Goal: detect and diagnose malfunctioning control loops

3. © ABB Coprorate Research - 3 - Bad control manifests itself as oscillation or too high variance

4. © ABB Coprorate Research - 4 - Methods needed to detect oscillations diagnose oscillations determine of variance is too large Since there are hundreds of loops methods should be automatic

5. © ABB Coprorate Research - 5 - Oscillation detection Hägglund (1995). Consider areas between zero crossings (count if large enough). Stattin and Forsman (1998). Based on same idea, easier to use. Seborg and Miao (1999). Damping ratio of auto- correlation function.

6. © ABB Coprorate Research - 6 - Stattin index: Compare areas between zero crossings

7. © ABB Coprorate Research - 7 - Oscillation index 0.880.25 Controller re-tuned 0 = no oscillation, 1 = perfect osc.

8. © ABB Coprorate Research - 8 - Oscillation index trend plot days index Valve IP converter replaced

9. © ABB Coprorate Research - 9 - Major advantage: correlation analysis 00.20.40.60.81 0 0.2 0.4 0.6 0.8 1 oscillation loop 2 oscillation loop 1 Conclusion: The loops interact. One of them is likely to cause both oscillations

10. © ABB Coprorate Research - 10 - Potential causes are... F FC static friction cycling load tight tuning

11. © ABB Coprorate Research - 11 - If the cause is stiction... process output control signal cross-correlation

12. © ABB Coprorate Research - 12 - If the cause is NOT stiction... process output control signal cross correlation

13. © ABB Coprorate Research - 13 - Stiction diagnosis New method by Horch (1999) which utilizes that when stiction in valve, process variable and control signal have odd cross-correlation when ”not stiction” the signals are such that the cross-correlation is even (due to negative feedback)

14. © ABB Coprorate Research - 14 - Example: two coupled loops F FC Q QC water pulp Stiction O.K.

15. © ABB Coprorate Research - 15 - Example cont’d data concentration loopflow loop cross-corr. Diagnosis:stictionno stiction

16. © ABB Coprorate Research - 16 - Important assumptions Cross- correlation method O.K. 4O4O Oscillation detectedSelf-regulating process Integral action No compressible media

17. © ABB Coprorate Research - 17 - Example II: integrating plant no stiction two different level control loops no stiction stiction

18. © ABB Coprorate Research - 18 - CCF-method useless for integrating plants! Integration destroys the specific correlation in the stiction case. CCF is even, no matter if stiction or not. Re-calculation (differentiation) does not solve the problem level control loop

19. © ABB Coprorate Research - 19 - Idea! Look for discontinuities in the data!... ‘Second derivative is infinite’

20. © ABB Coprorate Research - 20 - 1.) Differentiate the process output! stiction Y dy dt d2y dt2 no stiction

21. © ABB Coprorate Research - 21 - 3a.) Histogram (ideally) no stiction d2y dt2 stiction d2y dt2

22. © ABB Coprorate Research - 22 - 3b.) Histogram (noise & filter) no stiction d2y dt2 stiction d2y dt2

23. © ABB Coprorate Research - 23 - Level control with stiction MSE: 0.97 2.01 d2yd2y dt 2 y(t) stiction

24. © ABB Coprorate Research - 24 - Level control without stiction MSE: 1.17 0.46 y(t) d2yd2y dt 2 no stiction

25. © ABB Coprorate Research - 25 - Use Camel method also for self-regulating processes! stictionno stiction Y dy dt d2y dt2 Y Y’

26. © ABB Coprorate Research - 26 - Flow control with stiction MSE: 1.19 4.20 y(t) dy dt stiction

27. © ABB Coprorate Research - 27 - Flow control without stiction MSE: 4.33 3.02 y(t) dy dt no stiction

28. © ABB Coprorate Research - 28 - Detect too large variance (too large 2-sigma) Is this good or bad? Basic problem: 2σ2σ -2σ

29. © ABB Coprorate Research - 29 - Performance index Possible to calculate denominator from normal operating data given knowledge of process time delay (deadtime). Proposed by Harris (1989). Modification presented in Horch and Isaksson (1999) Introduce a control performance measure: Current variance Theoretically opt variance Ip =

30. © ABB Coprorate Research - 30 - Before: After: Modified Index: 2.11 1.07

31. © ABB Coprorate Research - 31 - Commercial tools / suppliers... LoopMD KCL-CoPA ABB LoopAnalyst PROTUNER™

32. © ABB Coprorate Research - 32 - LATTS – Loop Auditing and Tuning Tool Suite Process model identification PID controller tuning Loop auditing Part of ABB Industrial IT concept and uses the new Aspect Integrator Platform (AIP). Consists of three Aspects:

33. © ABB Coprorate Research - 33 - Process Model Identification Aspect

34. © ABB Coprorate Research - 34 - PID Controller Tuning Aspect

35. © ABB Coprorate Research - 35 - Auditing Aspect Computes 21 different quantities/indices. For example: Control error standard deviation Oscillation index Stiction diagnosis (correlation) Stiction diagnosis (histogram) Modified Harris index

36. © ABB Coprorate Research - 36 - Auditing Aspect cont’d Combines these indices to test a number of hypotheses, such as Acceptable performance Possible valve problem Sluggish tuning The result is summarized in a report, either as a text file or in Internet Explorer

37. © ABB Coprorate Research - 37 - Auditing -- Index trend plots

38. © ABB Coprorate Research - 38 - Auditing -- Report

39. © ABB Coprorate Research - 39 - Conclusions New ABB Product LATTS under Beta testing right now. Product release approximately June 2002. Methods exist for non-invasive Oscillation detection Stiction diagnosis Minimum variance benchmark

40. © ABB Coprorate Research - 40 - Future work (industrial as well as academic) detection and diagnosis of mill-wide oscillations distinction of linearly and non-linearly caused oscillations performance assessment based on full process model (event-triggered estimation) application of multivariable performance index performance monitoring of MPC loops

41. © ABB Coprorate Research - 41 - abb