Title: Analysis of techniques for automatic detection and quantification of stiction in control loops
1Analysis of techniques for automatic
detectionand quantification of stiction in
control loops
- Henrik Manum
- student, NTNU
- (spring 2006 CPC-Lab (Pisa))
Made 23. of July, 2006
2Agenda
- About Trondheim and myself
- Introduction to stiction and its detection
- Yamashita stiction detection method
- Patterns found in sticky valves
- Quantification of stiction
- Conclusions
3About Trondheim
4About Trondheim
5About myself
- Professional experience
- Summer 2004 Norsk Hydro. Development of
flow-sheet solver for the fertilizer industry
(Yara). (YASIM) (Group with 1 professor, 1
PhD-engineer, 2 PhD students, and myself.) - Summer 2005 Statoil. Development of company-wide
PID tuning rules, and tuning of new LNG plant at
Melkøya. - Projects, NTNU
- Phase equilibria for sorption enhanced hydrogen
production (fall 2004, supervisor prof. De Chen) - Extension of the SIMC rules to oscillatory and
unstable processes. (fall 2005) - Thesis
- This presentation (spring 2006, University of
Pisa) - From September 2006
- PhD student with prof. Skogestad on the Norwegian
Research Council -funded project Near-optimal
operation of chemical plants using feedback.
6Agenda
- About Trondheim and myself
- Introduction to stiction and detection
- Yamashita stiction detection method
- Patterns found in sticky valves
- Quantification of stiction
- Conclusions
7Introduction to stiction
- MV(OP) plot. In this work we focus on flow loops
with incompressible fluids
1.) Valve at rest and subject to static
friction 2.) e(t) gt 0 3.) Integral action in
the controller changes its output 4.) Valve slips
and subject to dynamic friction.
8How to detect stiction
- Popular methods
- Horchs cross-correlation technique
9How to detect stiction
- Popular methods
- Horchs cross-correlation technique
10How to detect stiction
- Popular methods
- Higher-Order Statistics
11How to detect stiction
- Popular methods
- Curve-fitting / Relay Technique
Stiction
12Agenda
- About Trondheim and myself
- Introduction to stiction and its detection
- Yamashita stiction detection method
- Patterns found in sticky valves
- Quantification of stiction
- Conclusions
13How to detect stiction
- Pattern recognition techniques
- Possible to detect the typical movements using
symbolic represenations?
14How to detect stiction
- Pattern recognition techniques
- Neural networks
Neural network
15How to detect stiction
- Pattern recognition techniques
- Simpler Use differentials (Yamashita method)
16Yamashita method
17Yamashita method
(I,I,I,D,D,S,D,I,....,D)
18Yamashita method
sticky movements
Threshold 2/8 0.25
19Yamashita method
Threshold 2/8 0.25
20Yamashita method
21Yamashita method
- Application to simulated data
- Choudhury model used
22Yamashita method
- Application to simulated data
- Noise-free VERY GOOD
23Yamashita method
- Application to simulated data
- With noise Performance degraded
- Important parameters Sampling time, frequency
content of noise (method sensitive to
high-frequency noise) - Setting sampling time equal to dominant time
constant seems good. - For case of no stiction, rho_1 high, but rho_3
always below threshold (0.25) - For the case of sampling time equal to dominant
time constant and some filtering of the noise,
the method seems to work sufficiently good. - Good enough for plant data?
24Yamashita method
- Set-point changes Good as long as set-point
changes occur well within band-width for outer
loop (assuming linear changes from cascaded
loops) - Found with simulation on noise-free data with
setpoint changes (See next slide) - The band-width for the outer loop is (1/10)(1/?)
for well-tuned cascades. (? is effective delay
for inner loop)
25Yamashita method
26Yamashita method
- Application to plant data
- 167 industrial flow loops studied
- 24 of 55 loops same report Yam and PCU
- PCU Tool with the 3 methods mentioned earlier
implemented (cross-correlation, bi-coherence and
relay). - 8 more loops reported by Yam
- 7 of 8 loops sticky by bi-coherence method
- Last loop was sticky other weeks
- Conclusion
- Works good
- Reports stiction
- in about 50
- of the cases
27Yamashita method
- Application to plant data
- Alteration of sampling time
- Seems like increasing the sampling-time is not
too dangerous. should
be OK. - The original was 10 seconds
28Yamashita method
- Application to plant data
- Observation window
- OK to
- reduce
- obs. window
- to for example
- 720 samples
29Yamashita method
- Application to plant data
- Conclusions
- Detects stiction in about 50 of the cases for
which the advanced package reports stiction - Identifies the loops with clear stiction patterns
- Noise level less than worst case in simulations
30Agenda
- About Trondheim and myself
- Introduction to stiction
- Yamashita stiction detection method
- Patterns found in sticky valves
- Quantification of stiction
- Conclusions
31Patterns and explanations
- Some other patterns were found. For example
- Possible to find physical explanation?
32Patterns and explanations
- Reverse action ( negative valve gain) ?
- In this case no, because of wrong direction in
the plot
33Patterns and explanations
- Closer look at control equation (PI)
jump from below.
34Patterns and explanations
- The valve can (theoretically) also jump to the
left! - This can be a possible explanation for the
pattern showed in the example.
35Patterns and explanations
- Measurements out of phase
- 4 time-units 40 seconds. Unlikely in this case!
36Patterns and explanations
- Another (and maybe most likely) for why the Yam
method failed for the example
Strong increase followed by weaker in OP (want
differential gt 1)
37Patterns and explanations
- Conclusions
- More insight into control action on sticky valves
achieved. The Yamashita method can easily be
extended to cover to cover other known patterns.
The theoretical considerations in this chapter
needs to be checked with real valves.
38Agenda
- About Trondheim and myself
- Introduction to stiction
- Yamashita stiction detection method
- Patterns found in sticky valves
- Quantification of stiction
- Conclusions
39Quantification
- Some work already done at the lab with a method
developed and implemented in the PCU. - As with stiction detection methods, it could be
nice with more methods. - Necessary, as the
- detection methods
- dont report
- amount of stiction
40Quantification
- Basis Bi-coherence method.
FFT-filtering by setting all unwanted
coefficients to zero and then take the
inverse transform to get filtered data
41Quantification
- Filtering using FFT
- often problematic
Here lower limit too high
42Quantification
- Filtering using FFT
- Conclusion
- Need steady data (best with little SP-changes)
- Few examples of suitable data in our plant data
- Using default filter limits did not work good
- Still needs tuning
- Before industrial implementation quite a lot of
work needs to be conducted
43Quantification
- Chose to move on to ellipsis fitting...
- 3 different methods
- Simple centered and unrotated ellipse
- General conic with two different constraint
specifications (more details in the next slides)
44Quantification
- Simple unrotated ellipse
- equation for ellipse in the
- Set of observations - least squares
45Quantification
- General conic
- Easier set c -1 and solve by least squares
directly
46Quantification
- Results (ellipsis fitting)
- Very often the optimization problem found
strange solutions (often imaginary axes)
47Quantification
- Discussion (ellipsis)
- Does guarantee an ellipsis? (Probably
not) (See report for derivation) - Setting seems more promising
- Obviously still work to do here!
- Answer questions given above
- Consider other techniques, such as clustering
techniques
48Quantification
- Conclusions
- The work did not give industrial-ready results
- I got more insight into time-domain -gt frequency
domain filtering (FFT-filtering)
49Agenda
- About Trondheim and myself
- Introduction to stiction
- Yamashita stiction detection method
- Patterns found in sticky valves
- Quantification of stiction
- Conclusions
50Conclusions
- Yamashita method proved to work good on
industrial data. Findings submitted to ANIPLA
2006 as a conference paper. - Hopefully the thesis gives more insight into
patterns in sticky valves in MV(OP) plots. - Introductory work to filtering and ellipsis
fitting for quantification conducted.
51References
- See thesis for complete bibliography
- Thesis should be available from Sigurd Skogestads
homepage, www.nt.ntnu.no/users/skoge - Diploma students -gt 2006 -gt manum
- Contains more details about everything and also
description about software developed.