Intelligent PID Product Design

Size: px

Start display at page:

Download "Intelligent PID Product Design"

Reynard Blankenship
5 years ago
Views:

1 IFAC Conference on Adances in PID Control Intelligent PID Product Design Willy K. Wojsznis*, Terrence L. Bleins*, John Caldwell*, Peter W. Wojsznis*, Mark J. Nixon* *Emerson Process Management, 1301 Research Bld., Austin, TX USA (Tel: ; Abstract: This aer outlines intelligent PID design for DCS. The design includes a PID algorithm with dierse standard otions and algorithm extensions for wireless/eent-drien control and for surge control. The core of the PID intelligence is adatie rocess modeling based on model switching and arameter interolation. The deeloed rocess model is alied to loo tuning, adatie control, loo erformance ealuation and ale diagnostics. A user-friendly interface roides insight into a loo s current state and history eents. The interface also roides adice about how to imroe loo erformance. Keywords: Intelligent control, Adatie algorithms, Feedback control, Feedforward control, First order systems, Performance monitoring, Diagnostic rograms, User interfaces 1. INTRODUCTION The label intelligent is attached to many control roducts. Some roducts are named intelligent because they erform self-diagnostics or use more comlex algorithms. It seems the name intelligent is justified for many roducts when a collection of simle features imroe roduct functionality and make it easy to use. In this aer we adhere to the more demanding definition of intelligent. An intelligent control system has the ability to learn about rocesses, disturbances and oerating conditions Åstrӧm et al. (199). The key features for such an intelligent control system are adatie rocess modelling, adatie control and enhanced oerall functionality. Adatie control for industrial alications started from non-model based PID tuning, like the attern recognition technique Åstrӧm et al. (006a), where PID arameters are adjusted by an automated rocedure used by an exerienced rocess engineer. Another technique alied for auto-tuning in industry is the relay oscillation technique Åstrӧm et al. (006b). The controller tuning alies Ziegler- Nichols rules or its ariations and is characterized by its simlicity. A modification of this technique made it ossible to identify a first order lus time delay rocess model Bleins et al. (003). This modified technique has been alied to auto-tuning in the PID controller design described in this aer. The adatie rocess identification technique with model switching was found to be robust and reliable in arious alications - Morse et al. (1994), Narendra et al. (1997). Reducing the number of models and alying arameter interolation to imroe accuracy simlified the technique and made it suitable for the industrial imlementation Wojsznis et al. (00), (003). Seeral ariations of the model switching technique hae been comared in controlling simulated H Bӧling et al. (005), confirming the adantages of the model switching adatation. The modelling accuracy of the switching technique has been further imroed by running the the algorithm recursiely with the same data Wojsznis et al. (003). The enhanced algorithm is a main adatie technology for the design. The design imlements erformance monitoring based on the concet by Harris (1989), Desborough et al. (199). The concet has been extended on the tuning index, which indicates otential ariability reduction due to the imroed tuning. This tye of tuning index for PI controllers has been exlored by Ko et al. (1998). Vale diagnostics for control loos has been the focus of research by Horch (1999) and Choudhury et al. (004). They resent ractical results in detecting ale stiction and in roiding qualitatie assessment in a grahical form. In the discussed PID roduct, the rocess model obtained from the adatation is used for ale diagnostics. This noel aroach allows automatic identification of the ale dead band and resolution. The PID controller has an otion for wireless oeration with eent-drien mode of oeration for otimizing erformance and saing energy. The PID control design roides an automatically configured user interface for setting loo oeration and obsering the results of auto-tuning, adatie tuning and erformance ealuation. The aer is structured as follows section outlines the PID design, section 3 - adatie tuning and control, section 4 - loo erformance ealuation and fault detection. Section 5 resents ale diagnostics concets followed by conclusions and acknowledgments..1 PID Function Block. PID DESIGN OUTLINE The PID algorithms are imlemented in the fieldbus function block, which combines all of the necessary logic to erform analog inut channel rocessing, roortional-integralderiatie control with the otion for nonlinear control (including error-squared and notched gain), and analog outut channel rocessing.

IFAC Conference on Adances in PID Control The PID function block suorts mode control, signal scaling and limiting, feedforward control, oerride tracking, alarm limit detection, and signal status

2 IFAC Conference on Adances in PID Control The PID function block suorts mode control, signal scaling and limiting, feedforward control, oerride tracking, alarm limit detection, and signal status roagation. To suort testing, a simulation mode is aailable. Two PID equation forms are suorted in the block: standard and series. Both forms include external reset and feedforward control. The PID block roides enhanced control for alications in the following areas: nonlinear (notch gain) control excetion reorting execution enhanced saturation recoery Figure 1 illustrates how the nonlinear tuning arameters are used to establish notched gain - KNL. Fig. 1. Notch and notch gain (KNL) arameters The PID roduct uses a ositie feedback network to create the reset contribution to the PID outut. For the standard mode of oeration, the time constant of the filter included in the ositie feedback network is the reset time. For the excetion reorting execution (wireless alication) the filter is restructured to roide integral action to match the rocess resonse in the elased time as in (1). The deriatie calculation is modified to comute a rate of change oer the elased time from the last udate (). Reset and rate action are only comuted when there is a new alue. T TReset FN = FN 1+ ( ON 1 FN 1) 1 e (1) where F - new filter outut N FN - filter outut for the last execution 1 ON - controller outut for the last execution 1 T - elased time since a new alue was communicated en en 1 OD = KD () T where e N - current error en 1 - last error O - controller deriatie term D T - elased time since a new alue was communicated Performance of the algorithm has been reorted in Kaltiokallio et al. (010). Due to the use of the ositie feedback network for the reset comonent, each of the elements in the forward ath must contain a filter comonent 1 Howeer, when a Ts+ 1. i rocess is oerating using a saturated condition, the PID is at its outut limits. Under these conditions the filter in the forward ath limits resonse to rocess conditions that require the PID to moe from its outut limits. The oint at which the ale will start to oen as the PV aroaches setoint deends on the PV rate of change and the magnitude of the error. A better resonse to major usets can be achieed by reducing the filtering that is alied in the forward ath calculation for this rocess saturation condition. Such reductions are ossible if the rocess measurement is relatiely noise free. Thus, there is merit in allowing the user to select the amount of filter alied in the forward ath when the PID outut is limited. The extended PID algorithms constitute a solid base for imlementing adanced functionality: auto-tuning, adatie tuning and control, loo erformance ealuation and fault detection. The PID function block oerates in the DCS controller ( the DCS is a DeltaV System) and uses all the standard DCS tools for grahical control design and debugging (Control Studio) and on-line oeration monitoring and adjusting (dynamos, facelates and detailed dislays).. PID alication The PID alication (Insight) is installed on the workstation and roides suort for adanced functionality. The entry alication screen roides a summary oeriew of the control loo s oeration and statistics on incorrect mode, limited control, large ariability, loo oscillation and incorrect tuning Figure. Further details for eery loo through arious tabs. The Tune tab roides a comlete interface for on-demand tuning, adatie tuning, adatie control, and models iewing for learning setu and for testing loo oeration in simulation Figure 3. The simulation allows the user to exlore loo erformance with new tuning arameters, adjust arameters for a desired erformance or to design the loo in a unique grahical way using the robustness lot in the gain-hase margin lane Figure 4. The robustness lot resents an area of accetable gain and hase margin for a secific rocess model. Fig.. Summary screen of the adanced PID alication.

IFAC Conference on Adances in PID Control Fig. 3. Auto-tuning and adatie tuning screen. tuning. The model quality is alidated by taking into account the most recent adatation and the last adatation.

Relay-based auto-tuning has been resented in many ublications Åstrӧm et al. (005b), Bleins et al., (003), therefore in this aer we focus on the adatie modeling and control.

3 IFAC Conference on Adances in PID Control Fig. 3. Auto-tuning and adatie tuning screen. tuning. The model quality is alidated by taking into account the most recent adatation and the last adatation. A high quality model and the exected control erformance with the recommended tuning are used as criteria for switching to adatie control. Relay-based auto-tuning has been resented in many ublications Åstrӧm et al. (005b), Bleins et al., (003), therefore in this aer we focus on the adatie modeling and control. The adatie model identification uses switching strategy with arameter interolation and model re-centering. This aroach makes it ossible to dramatically reduce the number of models used for adatation. The salient features of the technology include: shorter adatation time, comlete rocess model identification, and reduction in required rocess excitation. The structure of the PID with model arameter interolation is shown in Figure 5. Figure 4. Loo simulation screen with gain-hase margin robustness lot. 3. ADAPTIVE MODELING, TUNING AND CONTROL Three aroaches to loo tuning are imlemented. On-Demand Tuning - Uses an on-demand test of the rocess to automatically roide tuning recommendations. Ondemand tuning is aailable for PID and Fuzzy Logic Control (FLC) blocks. Tuning recommendations are aailable ondemand by initiating automatic testing of the rocess. When testing is requested using on-demand tuning, the PID or FLC block's actual mode switches to Local Oerride (LO). Once in LO mode, the oeration of the loo's rimary control algorithm is susended and the controller resident relay control adjusts the control block outut (OUT). The original algorithm has been enhanced for identifying rocess dead time and time constant in addition to the ultimate gain and ultimate eriod. The controller settings are then comuted using modified Ziegler-Nichols, Lambda, or Internal Model Control tuning rules. Adatie Tuning - Uses normal oerator changes in setoint or outut to identify rocess models and roide tuning recommendations. Adatie Control and Model Scheduling - Includes all of the Adatie Tuning caabilities lus the ability to create models in u to 5 regions and to automatically change control loo Fig. 5. Adatie Feedback Control Design. The rocess identification is based on the ealuation of multile sets of models. Each model consists of three arameters (=3, gain, dead time, lag). Assigning n alues for eery arameter, the model set has M = n 3 models. The Adatie Control oerates in the following way: The suerisor detects changes in the rocess outut (i.e. the controlled ariable (y-pv)), the setoint or the maniulated rocess inut (u-out). If any changes exceed a minimum leel, model ealuation starts. This inoles: model initialization and adjustment of the model's outut with the current rocess outut model incremental udate based on the changes of the maniulated inut to the rocess comuting for eery model i squared error e( ) ( ( ) ˆ ( )) i k = yk yi k for each scan k, yk ( ) - the rocess outut at the time k yˆ i( k) - i-th model outut at the time k The squared error is assigned for eery arameter alue of the model i if the arameter alue is used in the ealuated model. A zero is assigned to any arameter alue that is not art of the ealuated model. Next, the model i+1 is ealuated. Calculated squared error is assigned for eery arameter

IFAC Conference on Adances in PID Control alue and reiously assigned squared error for eery arameter alue is udated. The model s ealuation continues until all models are ealuated.

4 IFAC Conference on Adances in PID Control alue and reiously assigned squared error for eery arameter alue is udated. The model s ealuation continues until all models are ealuated. As a result of the ealuation, eery arameter alue is assigned a sum of squared errors from all models in which this secific arameter alue has been used. In the one scan k therefore, eery arameter tye ν with alue has assigned squared error SE ( k). =1,,,P =1,,,V M ( ) γ i( ) i= 1 SE k = e k (3) M the number of models γ = 1 if arameter with alue is used in the model, otherwise γ = 0. Model ealuation is reeated in the scan k+1 and the sum of the squared errors for eery arameter alue is added to the sum of the aroriate arameter alue accumulated in the reious scans. The adatation cycle continues through a declared number of scans (1 to N), or until there is enough excitation on the inuts. As a result of this rocedure, eery arameter tye with alue has an assigned sum of squared errors SSE oer a eriod of ealuation: N SSE = SE( k) (4) k = 1 At the end of adatation cycle, the inerses of SSE are calculated: F 1 = (5) SSE An adatie arameter alue a for the arameter is calculated as a weighted aerage of all alues of this arameter ( ) = 1,,..., V. 1 V a = (1) f + () f ( V) f V 1 f = F F (6) 1 Calculated arameters define a new model set with center arameter alues a and the range of arameter changes defined as ± % a, =1,,,P. Within that range, two arameters at a minimum should be defined. As soon as a model has been udated, controller tuning takes lace based on udated a model arameters. In the PID roduct a first order lus dead time rocess model is used (in a discrete form). yk ( ) = a yk ( 1) + b uk ( 1 L) (7) h h a = e τ b= k 1 e = k( 1 a) (8) h = loo scan eriod L rocess dead time, L=hk. u - rocess inut incremental change since adatation started y - model outut incremental change τ = model time constant With three alues for eery arameter of the model, there are 3 3 = 7 models for ealuation. Imroed conergence and reduction in the number of models is achieed by the following amendments to the basic algorithm: Performing arameter adatation sequentially, one arameter at a time. In this way, the number of model combinations for the first order lus dead time model has been reduced to 3*3=9. Performing adatation for only two arameter alues with minimal errors. Using the original data set and erforming adatation iteratiely by running the algorithm seeral times. In a sequential rocedure in which one arameter is udated oer a calculation cycle, the udating is erformed in the sequence rocess gain, dead time and time constant Figure 6. Fig.6. Sequence of arameters adatation and interolation Figure 7 illustrates model adatation, showing how rocess outut is aligned with three models oututs. Fig. 7. Adatie models alidation lots After model adatation comletes, controller redesign begins using the first order lus dead time rocess model. Any tuning rules, tyically Lambda or IMC, can be alied. If there are infrequent changes in the maniulated inut, external excitations can be automatically injected into the maniulated inut in Automatic modes. PID outut ulse of the amlitude 3-7% and duration of seeral scans is normally enough for a rocess model identification Figure 8.

IFAC Conference on Adances in PID Control Fig. 8. An examle of minimal excitation on the rocess inut (uer lot) and outut required for the rocess model identification. 4.

5 IFAC Conference on Adances in PID Control Fig. 8. An examle of minimal excitation on the rocess inut (uer lot) and outut required for the rocess model identification. 4. PERFORMANCE MONITORING AND FAULT DETECTION The ability to insect control and measurement loos quickly is of rimary imortance. The PID design roides adanced rocess monitoring that allows under-erforming loos and malfunctioning field deices to be identified instantly. In roiding this adanced caability, it takes full adantage of the fieldbus block architecture. A Bad, Uncertain, or Limited measurement, downstream limitations in control, and incorrect mode of oeration are automatically determined based on block mode and the status of block arameters. The user interface roides indices that quantify loo utilization, measurements with abnormal status, limitations in control action, rocess ariability, and aailability of recommended tuning. Other arameters used within the block are standard deiation, ariability and tuning index. The existence of an abnormal condition is determined within the controller based on function block arameters in the controller (such as measurement status, back calculation inut status, standard deiation, tuning index, and mode). The state of these monitored conditions is automatically reorted to the DCS serer on an excetion basis. Loo erformance is indicated by the ariability and tuning indexes. Variability index is a measure of the quality of control, while the tuning index indicates the otential for imroing control erformance by udating controller tuning according to the identified rocess model and the selected or default tuning rule. To suort the erformance calculations erformed in the function blocks, caability and al standard deiations are calculated eery scan ( ca, ). The al standard deiation and the caability standard deiation is calculated using a moing time window as in Shunta (1995): ca ( K) = ca ( K 1) + f ( MR ( 1) 1.18 ca K ) (9) K MR= 1 yk ( ) yk ( 1) ( K 1) - aerage moing range K the current window, K-1 the reious window The ariability index (VI) in % is comuted then as: ( ) mc + s VI = ( + s) (10) where s is the sensitiity factor that makes calculations stable. The default alue is 0.1% of the ariable scale. is the actual measured standard deiation. mc is the standard deiation that can be achieed with minimum ariance feedback control defined as: ca mc = (11) The tuning index is based on the ariability difference estimate for the current controller tuning (PID1) and the desired controller tuning (PID). res = PID 1 PID (1) The tuning index TI indicates the otential for imroing control erformance and is defined as the ratio of the otential PID ariability reduction to the actual PID ariability: res TI = (13) PID1 5. VALVE DIAGNOSTICS It is commonly recognized that the undesirable behaiour of control ales is the biggest single contributor to oor control loo erformance and the destabilization of rocess oeration. It is normal ractice to erform loo secial testing in manual mode and use the test results for calculating ale resolution and dead band Gerry et al. (001). There are seeral effectie techniques for comlete ale diagnostics when internal ale arameters, including ositioner ressure are aailable Grumstru et al. (001). The loo diagnostics are extremely imortant for the adatie loos. Loo adatation may bring instability caused not by the failed adatation but by the sticky ale or measurement failure. On the other hand, manual or semi-manual diagnostic rocedures as cited aboe are not adequate for the adatie loo working with no suerision. The on-line techniques by Horch (1999) and Choudhury et al. (004) roide qualitatie ale ealuation. For the PID roduct, ale diagnostics techniques has been deeloed that do not assume access to any ale arameters. Aailability of the ale stem osition imroes the diagnostics, though this arameter is otional. The aroach uses the rocess model gain K and is the best suited for the adatie control loos or automatically tuned loos where rocess gain is known. The ale dead band and resolution is reflected by a characteristic loo oscillation Figure 9. After calculating oscillation amlitudes on the controller inut A() and outut A(out) ale hysteresis is defined directly as: h = A( out) (14) The ale resolution r defines effectie ale stem moement during ale induced loo oscillations, which alied to the rocess with gain K roduce oscillation on the rocess outut with the amlitude A() Aml( PV ) = Kr (15)

IFAC Conference on Adances in PID Control r = Aml( PV ) (16) K Dead band is calculated as b= h r (17) Figure 9. A loo oscillation caused by the ale dead band % and the resolution 1%, noise leel is 0.

6 IFAC Conference on Adances in PID Control r = Aml( PV ) (16) K Dead band is calculated as b= h r (17) Figure 9. A loo oscillation caused by the ale dead band % and the resolution 1%, noise leel is 0.5%. The technique can be refined by using the back calculation signal from the ale. Other than imerfect rocess gain sources of error include difficulty with searation of the loo oscillation caused by the ale from the oscillation caused by oor tuning, external disturbances or set oint changes. The oscillation shae factor is alied for identifying oscillations caused by the ale. Some of the lab test results are in the table 1. Table 1 Vale stickness (resolution) and backlash diagnostics: real alue /diagnosed alue Dead band % Resol ution % 1/1 0.4/ 0.4.5/.49.1/ / / / / 0.1 /.1./.3 / /.5/.1.5/.49 1/.99 3/.3 1/.9 1/.96 The technique will be included into roduct after full field test trial. 6. CONCLUSIONS The intelligent PID design described here has been widely acceted by the users. The main factors that contributed to the accetance are the robust rocess model identification technique and a user interface that roides full insight into control loo oeration, control erformance, faults and tuning recommendations. Acknowledgments Our communication with rofessors Karl Åstrӧm, Dale Seborg and Thomas Edgar greatly imroed the roduct concets and design. The final shae of the roduct and its quality is the result of contributions from many control software deeloers including Dennis Steenson, John Gudaz, Peter Wojsznis, Mike Ott, Yan Zhang and Ron Ottenbacher. REFERENCES Åstrӧm, K.J. and McAoy, Th. J. (199). Intelligent control. J. Proc. Cont., Vol., No Åstrӧm, K.J. and Hägglund, T. (006a) Adanced PID Control, ISA, USA. Åstrӧm, K.J. and Hägglund, T. (006b) Adanced PID Control, ISA, USA. Bleins, T.L., McMillan, G.K., Wojsznis, W. K. and Brown, M. W. (003) Adanced Control Unleashed, ISA, USA. Bӧling, J.M., Seborg, D.E. and Hesanha, J. P. (005) Multimodel control of a simulated H neutralization rocess, IFAC World Congress, Prague, Czech Reublic. Choudhury, M., Shah, S.L. and Thornhill, N.F. (004) Diagnosis of oor control-loo erformance using higher-order statistics, Automatica, 40, Desborough, L. and Harris, T.J. (199) Performance assessment measures for uniariate feedback control, Can. J. Chem. Eng., 70, Gerry, J., Ruel, M. (001) How to measure and combat ale stiction on line, ISA Technical Conference, Houston, USA. Grumstru, B.F., Junk, K.W., Snowberger, J.L. and Taylor, E.R. (001) A method and aaratus for deterministically obtaining measurements, US Patent 6,19,31. Harris, T.J. (1989) Assessment of control loo erformance, Can. J. Chem. Eng., 67(10), Horch, A. (1999) A simle method for detecting of stiction in control ales, Control Engineering Practice, 7, Kaltiokallio, O., Eriksson, L.M. and Bocca, M. (010) On the erformance of the PIDPLUS controller in wireless control systems, 18 th Mediterranean Conference on Control and Automation, Marrakech, Morroco. Ko, B.S. and Edgar, T.F. (1998) Assessment of achieable PI control erformance for linear rocesses with dead time, ACC, 7, USA. Morse, A.S., Pait, F.M. and Weller, S.R. (1994) Logic-based switching strategies for self-adjusting control, 33 rd IEEE Conference on Decision and Control, Worksho no. 5, Lake Buena Vista, USA. Narendra, K. S. and Balakrishnan, J. (1997) Adatie Control Using Multile Models, IEEE Transactions on Automatic Control, Vol. 4, No., Shunta, J. (1995) Achieing World Class Manufacturing Through Process Control,. 8-33, Prentice Hall, USA. Wojsznis, W. K. and Bleins, T. L. (00) Ealuating PID adatie techniques for industrial imlementation, Proc. of ACC, Anchorage, Alaska, USA, Wojsznis, W. K., Bleins, T. L. and Wojsznis P.W. (003) Adatie feedback/feedforward PID controller, ISA Technical Conference, P137, Houston, USA.

EXPERIMENT 6 CLOSED-LOOP TEMPERATURE CONTROL OF AN ELECTRICAL HEATER

YEDITEPE UNIVERSITY ENGINEERING & ARCHITECTURE FACULTY INDUSTRIAL ELECTRONICS LABORATORY EE 432 INDUSTRIAL ELECTRONICS EXPERIMENT 6 CLOSED-LOOP TEMPERATURE CONTROL OF AN ELECTRICAL HEATER Introduction: