Diagnostic Autonomous Control System for Power Plant Feed Water Process

Diagnostic Autonomous Control System for Power Plant Feed Water Process

Copyright © IFAC 12th Triennial World Congress, Sydney, Australia, 1993 DIAGNOSTIC AUTONOMOUS CONTROL SYSTEM FOR POWER PLANT FEED WATER PROCESS P. La...

1MB Sizes 0 Downloads 60 Views

Copyright © IFAC 12th Triennial World Congress, Sydney, Australia, 1993

DIAGNOSTIC AUTONOMOUS CONTROL SYSTEM FOR POWER PLANT FEED WATER PROCESS P. Lautala, Y. Majanne and J. Henttonen Control Engineering Laboratory, Tampere University o/Technology, p.a. Box 692, SF-33JOJ Tampere, Finland

Abstract. A scheme for a diagnostic auto no mous control strategy of a power plant feed water process is discussed. The contro l system is based on M odel fredi cti vc !::o ntro I strategy, MI'C, applying diagnostic information to optimise the control result and the avai labi lity of the process. In this paper condi ti o n monitoring and Jiognostics ore corried out by diagnostic algorithms based on models describing the general properties and characteristics of monitored compone nts as a function o f operating points and time. How ever , information from any kind of diagnosis system ca n be applied. The diagnostic information is used 10 update the co nstraints and contro l increment weights in the cost function under which the MP(: system calculates uptimal l~"ntrol signals . The proposed system is able to change the dynamics and the slructure of control system automaticaJly during the operat io n of the process, if a ny malfunctions have been detected. Key words. Model basc-u di"gnostic.s, Model predictive contro l. Power plant

performance control syste m capable of operating without expert intervention for long peri ods oftime (Morari and Zafiriou). The MPC concept has a long hi story but si nce its rediscovery in the lat e 1970's, its popularity in the process industry has increased stead ily. Mehra and others (1982) review a number of applications including a supe rheate r, a steam generator, a wind tunnel , an utility hoiler connccted to a distillation column and a glass furnac c.

I . INTRODUCTION One ofthe main goals of condition monitoring and fault diagnosis in industrial plants is to improve the availahility of critical suh processes, Diagnostic data is usually applied to reveal develop ing faults as soon as possihle to schedule the maintena nce operations to the most suitahle moment of time. Howeve r, the diagnostic information can he applied to improve the controllability and availability of a process automatically hy using intel ligent control system considering the' operating state of prm:,'ss components.

Diagnosis method s utili sing standard process measurements are hased on models de sc ribing the static and dynamic behaviour of the process. In model predicti ve control same type of models are nee ded. Thc diagnosis system and th e intelligent control system can he integrated as a sys te m applying same models and measurements monitoring the state of the process and optimising the operation of the process due to the current state of process components. Optimal control and diagnostic s can utilise each other in both directions. Diagnostic infomlation can be used to set constraints and weights of performance criteria to adapt the process control to meet th e restrictions coming from the operating state of the process . In o ptimising control the process can be utilised safe ly to ex tn: me limits if the robustness of the system is quaranh;ed by diagnostics. Conventional control must be all ways designed to compromise with all possible process situations.

Today the signal processing capacity of modem digital automation systems has made it technically and ewnomically possihlt: to apply condition monitoring and diagnostic ll:chnology with process industry and con ventionalenergy industry . However, so far the existing real time condition monitoring applications arc quite rare and they are mainly hased on vibration monitoring. Vibration monitoring can be applied to ddect only a limited type of faults. It requires special instrum entation and signal processing system and it cannot he integrated easily ;l' a part of the existing automation system, Despi te all these di sadvantages the method is popular, because it is the only ge:ncral method to apply so far. So, general condition monitoring and diagnosis methods based on normal process measurements are needed in order to get these systems more widely applied in process and energy industry .

2. CONDITION MONITORING AND FAULT DIAGNOSIS IN P< )WER PLANTS

Autonomous means having the power for self government. Autonomous control systems have the power and ahility for self governance in the perfomlancc of control functions. Conventional controllers can be consider"d to have only a low degree of autonomy because they can tolerate only restricted disturbances and variations in the process parameters. The control syst,'ms having a high degree of autonomy must he ahle to compensate faults without external intervention and th"y mu st provide high level adaptation to changes in the plant and environment. In addition to supervising and tuning the control algorithms, the autonomous controller must also provide' a high degree, of tolerance to faults. To enSUft' system rdiability, faults must first be detected , isolated and identified and suhs"'1u"ntly a new control law must be designed. (Antsaklis c/ al., 1988; Antsa kli s et aI., 1991)

Thc total efficiency and avai lability of the power plant process depends on the cfficic ncy and availability of individual sub proccsses and process components, Weari ng of pump s, fans and val ves, fouled or leaking heat exchangers as well as blockages in pipclinc.:s and valvcs impair the e fficiency and controllability of thc process remarkahly. The pumps and th e fan s are also very important due to the availability of a plant, because faults in these componc.:nts usually lead to process outages or at least remarkable reductions in production capacity. These malfunctions are quite difficult to reveal by operators' visual inspection because the symptoms arc usually de veloping slowly and the control loops tend to compensate the effects of the changes in the process values.

Model Predictive Control algorithms, MPC, arc, a family of control algorithms which use explicit and separately identifiahle models to optimize process behaviour. Control design methods based on the MPC concept have found wide: acceptance in industrial applications because of their ahility to yidd high

2.1.

Dia~nosis

Algorithms

The condition monitoring and fault diagnosis methods studied here are based on fault sclec tive diagnosis algorithms (Lautala ct a/., 19X9; Lautala cl al., 1991 ; Maja nn e ct al., 1991) . The

557

Figs. :1 and 4 illustrates the hehaviour o f effi ciency characteri stics as a fun cti on of time in different cases of faults. E xtensive fl ow ga p ca us(~ e.g. hy the weari ng of the impell er can be detected as a smooth impai rment of efficie ncy hoth as a function of o peratin g point a nd tim e. Cavitati on in a pump can he detected hy the qui ckl y changes in d fici e ncy in a certain o perating point range. Cavitati on speeds up the wearing o f the pump whi ch also can be see n from the he haviour of the e ffi ciency characteristics as a fun cti on of time.

al gorithm s calcul ate di ffe rent characteri stics and generate test values describing the o perating conditio n of the monitored compo nent s. Faults can be detected and isolated eithe r by the behavio ur of a sin gle test value or by co mbining informati on fro m few differe nt algorithm s. Thi s hierarchical structure leads to simple reaso ning and fau lt isolat ion processes. Di agnosis algorithm s for pumps, fans a nd valves are based on refere nce characteri sti cs and individual values calcul ated from on-line measurements. The characteri stics plot the co mponent effi ciency, capacit y, nomi nal kinematic rotating speed etc. as a functio n of the ope ratin g point. The individual points of c haracteri sti cs are stored in tabl es. The characteristi c tahles include the latest test values from each o perating point range and mean values calculated wi th differe nt averaging tim es. The developme nt speed of a fault is mo nitored by cal culating the test value c ha nging speed. By these mea ns it is poss ible to ohse rve the behaviour of the monito red co mpo nent bo th as a fun ctio n of o perating point and tim e. Most of the test values are calculated using static models. The analysis data is pre-processed by the "steady-state" ftlter (V illisuo) whi ch cut s off the dynamic transients he fo re the test values are calcul at!!d . The static nature of the algo rithm s make them suitah le fo r detec tin g slo wl y develo ping fau lt s. TESTSIUNAL

. ,,d

..............--;1

o ~.~.:s-..r'T'...,-:.---..-

' >f'I!'JIAl"J' OHr r · .

Fig. 3.

Behav iour (If the e fficiency characteri st ics of the worn pump.

Fig. 4.

Ilchaviuur of the efficie ncy characteristi C) of the cav itati ng pump_

DEVELOPING SPEED

A

Afel

i:

;

d

r;o

i'"iO

~

i700

"lOO

X""iQOo

A "00

XIoOOO

10203040 o .0 Oso . 00 Of'FJ{ATING POINT 1% )

~

'-;."'0"'2""0"'3"'0'-;."'0'-;,"' 0 ":'.o;h7,f0"'80"""9Q;;'-;-;:!. 00 OPERAnNG PO.NT 1"' 1

Fig. I. Tables for s.oring the characteristic tes t value and fa ult developing speed of the monitored component.

Fault s in the pump motor and gear ho x can be separated from the fau lt s in the pump it sdfh y inspecting hoth the e ffi ciency and the nominal kinematic rotatin g speed . The fault s in the pum p itself effect hoth on the e ffi ciency a nd the nominal kinemati c rotation speed whil e th e motor ,md gearhox fault effect only on th e effi ciency.

The co nditi o n of pumps is monitored by cal culating effi ciency and no minal kinematic rotati o n speed characteri sti cs. Fig. 2 illustrates the use of characteri stic tahles. In the exampl e the effi ciency of the feed water pump is moni to red. If the effic iency is j ust calcul ated and shown as a functi o n o f time . it is very difficult to d raw any co ncl us io ns ahout the operating state of the co mponent. However. if the signal is examined as a fun cti on of tim e. operatin g point and refere nce c harac teri stics, it is much eas ie r to diagnose the state of the co mpo ne nt.

Figs. 5 and 6 sho w the ,~ ffe ct of the simul ated gearbo x and extensive now gap fault s to Ihe effi ciency a nd kinemati c rotati on speed charac t" ri sti cs of the pump.

PUMP nl-'(lENCY n.!ioO

0.",--- - - - -- -_____ ____ --,

054

0,"

O.S2

0 .6'.. UO
0,",

o ss

oso

0.4'

0.., O.Y-

u. ~

0. 42 0 . 411

~

- - ---- ~

-20--_·- - -i \-·---YO ...-

- -3.S

Pump Power IMW I

Fig. 5_ Efrc(.,.t of the simu lated gearho x and ex tensive now gap fau lts to thc cffi c i ~ n cy of the pump.

SAMPl£S

Nomi na..! Kme ma tu,: Kol.atlll~ S l'ccd ··,tr--~----~--~-------__,

,/

0<1010

jO

11)

10 o'uA~o

to

.

....,

100

..m""I " l

"""

P ump Rotating S pee d (l / rrun]

Fig. 2.

Moni.o ri ng the efficiency of a of pu mp w ith th e characteristics table.

Fi g. 6.

558

Effect of .he simu la.ed gear hox and extensive now gap f,u lts to the kim.'mati c rotatio n speed characteristics of th e pump_

The same test signals can be used also tll monitor and analyse the operatIng status of I·.D. and I.D. fans. Operation of the valves is monitored by using estimated capacity coeffiCIent Kv as a test signal. Blocked and eroded valves can be detected by comparing measured capacity coefficient characteristics with reference ~har~ctcristics. Stuck or backlashing valve IS detected by taking Into account the direction of the movement of the valve when characteristics are determined. Characteristics for opening and dosing valves are determined and compared with each other. Fig. R.

~ softw~c env~ronment V/O/ for creati ng. updati ng and anal ysIng t?e ~Iagnosls system based on these algorithms is developed by Flnrush powercompany Imatran VoimaOy, ComputerTechnology Laboratory of Finnish Resl~a rch Centre VTT and Control Engineering L~boratory ofTampere University of Technology. The system utilIses expert system technology for creating and updatIng the system. Numerical algorithms can be created and connected to the process nWlLsurements with graphic editor. Test results from different diagnosis algorithms an~ combined togetherformore detailed analysis. which is carried out with event trees. These trees arc also creat"d graphically with an event tree edItor. The VIOl system is presented more detailed in Majanne ct al..199l; Kurki ct al.. 1991.

The fact that in MPC the optimal process control inputs are calculated during every control cycle by taking into account the pn,vailing constraints. makes MPC very suitable companion for process diagnostics and condition monitoring. With the fl exible ?andling. of constraints MPC is possible to use the diagnostic InformatIOn to update the constraint vectors and by this mean effect on the dynamics and even on the structure of the control system . 4. CONTROL or POWER PLANT FEED WATER SYSTEM Conventionally feed water flow is controlled both by changing the speed of the pumps and by throttling the flow with the val ves. Throttling is dynamically fast but economically poor way to control the flow. The nonnal way to control the flow is to operate the process With almost open valves to minimise the pressure losses. Only a liltk control reserve is le ft for the valve to compensate the dynamically fast disturbances. The flow control is realised by controlling the flow by valves and pressure differe~ce over t.he valves hy pumps. The set point for the pressure difference IS usually related with the feed water flow.

3. MODEL PREDICTIVE CONTR( JL Malfunctions !ead always to constraints in process operation. Mo st applIcatIOns. where model predidive control MPC has been proved successful arc multivariahk and involve constraints. It is exactly these types of prohkms which motivated the development of the MPC control techniques (Mehra cl al.). Because most practical complex processes involve important control and state constraints which cannot he taken into account in most advanced control synth,~sis methods. the MPC method is an interesting candidate for ohtaining the solution of these type of control prohlems.

The schematic structure of the autonomous control system is presented in Fig. 9. In principal this type of solution has been proposed for autonomous actuating systems by Isermann , 1992. The system discussed in this paper is consisting of condition monitoring and fault diagnosis block and MPC block. The diagnostic system is monitoring the status of the process and sending diagnostic information to the MPC system. In MPC the estimated behaviour of the process is optimised during every control cycle under the constraints and the weights determined by using the diagnostic information. The constraints and the weights can be changed between control cycles. Also the process model used in MPC can be changed if the diagnostics has revealed that characteristics of some process components have changed. However. if the model is changed during the process operatIOn. the robustness of new control system must be verified before switching the models.

3.1. Theoretical Background (Morllli and Zafiriou) The name "Modd Predictiw Control" arises from the manner in which the control law is wmputed. At the prcSl:nt time k the behaVIOur of the process over the prediction horizon is considered: Using a modl'l the process response to changes in the marupulated vanablc is predicted. The moves of the manipulated vanables ~re selected such that the predkted response has ~ertam desl.rahle characteristics . Only the first computed change In the marupulated variahle is implemented. At time k+l the computation is repeated with the horizon moved by one time interval. An innovative feature in this system is considering both the process input and state constraints. In previous control algorithms the constraints were not considacd while optimising the process mputs. but the Signals of controller were limited afterwards . past

The hl oc k diogram of the model predictive controller (Li cl al.).

fUIIlft"

n prcdldiofl for opLiroalluplll set pOlnt

measur('lIrnl

prrcllctlOlI for Input ;-t!mOllrnl k-J

,

1",1'

Fig. 9.

The hlo c k. diagr;}1ll of pro posed intelligent co ntrol system for power

plant

Fig. 7.

fc~d

water process.

4.1 Simulation Experiment

The principle of model predictive control. A prediction hascd on i.~()lltrol signals al the moment k-I o ~ prediction corrt'l·{(,.'U with optimal conlrol signals

Backlash in process componets often cause oscillations in actuators shortening their operating life time remarkably. With the pro~osed control system vibration s can be eliminated by constraIning the speed of the actuators or by increasing the weights of control costs of faulty component. The operation of the control system was tested in cases. where the power plant feed water process was first operating nomlally and then a 100 rpm

In the syst~m discussc:d in this paper the optimis;uion problem is solved usmg L] solution of owrdetcnnined system of linear equations (Barrodale and Roberts , 197~; Barrodale and Roberts 1978). The prediction is calculated using open loop predictor

(Li

et al.).

559

5. CONCLUSION

backlash was added to the gear of the feed water pump controlling the speed. Figs. 10, 11 and 12 show the behaviour of the control variables with normal process operation, with faulty process and with faulty process and control parameters manipulated according to the diagnostic information.

i~LC '-~

il o

o

500

1000

1500

I

r 2 x>o

2000

3000

Tpe\sl

,-'~ S((I

1000

1500

2000

2SOO

f~Jt :\ ~~b-T o

A proposal for an autonomous control system is discussed. The system is consisted of the diagnostic modules and the control module. Both the diagnostic and the control strategies are based on models, which make it possible to integrate these systems together and to use the properties of both systems to complement each other. The control module of the system is realised by using model predictive control algorithm (MPC). The goal of the control is to update the control variables to provide the optimal control result in changing process circumstances_ Changes in process parameters can be taken into account with constraining the control variables and states and with changing the weights of the performance criteria of the optimisation. Even the model used in the control system can be changed automatically, but in that case the robustness of the new system must be checked carefully. Due to the robustness of the system, constraints are more recommended way 10 change the properties of the control system than switching the model.

.soo

1000

1500

2000

3000

flexibility of the MPC and diagnostic information about the operating status of the process makes the suggested system more versatile compared with the conventional adaptive control. In the proposed system the availability of the whole process can be optimised instead of the performance of a single control loop_ This possibility makes MPC very promising method to handle processes under different circumstances and to increase the autonomy of control systems.

t j 2SOO

JOOO

Tm~lsJ

Fig. 10. Control variables during the normal operation of the process.

f:~~

RER:RENCES Antsaklis, PJ .. Passi no, K.M., Wang, S.1.(l991): An Introduction to AutonomOl<' Control Systems. IEEE Control Systems, Vol. 11, Number4,lune 1991. Antsaklis. P.1., Passino. K.M .. Wang, S.1.(1988): Autonomou.< Control Systems: Architecture and fundamental issues. Proc. 1988 American Control Conference, Atlanta. GA, lune 15-17. 1988, pp 602-607 .

I~OL---~'~OO~--~'OOO~--~"~OO~--~2~OOO~--~2'~OO~--~~~ Tme[s)

~~~ If~ o

o

500

.500

1000

1000

1500 Time Is,

1500

2000

2500

2000

Harrodale, 1.. Robcrts, F.D.K.( 1973): An ImproyedAlgorithmfor Discrete LJ Linear Approximmion. SIAM J. Numer. Anal., 10 (1973), pp. 839-848. Harrodale, 1., Roberts. F.D. K.( 1978): An Efficient A19orithm for Discrete LJ Linear Approximation with Lint'ar Constraints. SIAM 1. Numer. Anal.. 15 (1978), pp. 603-ld I. Dimmic, J.G .. ('obb, J.M.(1986): Ultrasonic Leak detection Cuts Valye Maintenance Costs. Power Engineering , August 1986, Vol. 90 pp 35-38.

3000

2!OO

Isermann, R., Raab, U.(1992); Intelligent Actutors -Ways to Autonomous Actuating Systems. l'reprints SISICA'92, Symposium on IntelligentComponenL' ond Instrument, for Control Applications. Malaga. Spain, May 20-22 , 1992 Kurki, M .. Majanne, Y. , Ruokonen, 1'.(1991): Real Time System for_Power Plant Diagnosis. 'Ihird Symposium on Expert Systems ApplicatIOn to Power Systems. April 1-5, 1991, Tokyo & Kobe. Japan, pp. 160-164.

3000

Tlmef'l

Fig. 11. Control variables during the faulty operation of the process. Lautala, P., Viilisuo, M.(J 989): A lIierarchical Expert System for Failure Diagno.,is in Power Plants. Pre-prints of the International Symposium on Power Systems and Power Plant Control, IFAC, Seoul. Kore
i~Lh_:~_~ J:

Lautal", P. , Antila. H.. Vilkko, M.(1991): A lIierarchical Model Based fault Diagnosis System "-or a Pea/Power Plant. Pre-prints of the IFACIIMACS Symposium on Fault Detection, Supervision and Safety for Technical Proc· esses -Safeprocess '91. Vol. 2. 10 -13. 9. 1991, Baden-Baden, Germany, pp 259-264

160

= -1

140O!-----;;:,.,.;;-----;;IOOO=-----c,::.SOO:;:----;;2000=-----:2~""':;:-----;;!~ limelsl

......; : E .l .~ = E FoodWMNFlOW



Li, S. , Lim , K.Y .. Fisher. DJ;.: A Statl! Space fonnulation for Model Predictive Control. AlChE Journal, Vol. 35. pp 241-249. Maj.nne, Y., Ruokonen, T., Kurki, M., Ala-Siuru. 1'.(1991): Hierarchical On-line Diagnosis Syst,'mfor Power Plants. l're-prints of the IF ACIIMACS Symposium 011 Fault Detectio n. Supervis io n and Safety for Technical Processes -Safeproces, '91. Vol. 2. 10 -13. 9. 1991, Baden-Iladen, Germany. pp H7 -n.

160 140

o

SOO

1000

1500

- ·· 2OC()

-

'· ----

~

JOOO

Time[~l

I ~~o_:_o_""2j

Mehra, R.K. , Rouhani, k., Eterno, j .. Richalet, J., Rault,A.(l982): Model Algoritmn;(' Control: }(ev;('w fUIlJ Rt'n'nl Dew-topmenl. Eng. Foundation Conference on Chcmic
o

SOO

1000

1500

2000

BOO

3000

Valisuo, M.(l99U): Compress ion Method for Measurement Data. lmatran Voima Oy, Project rep ort (in Finnish), Nov. 1990, Vantaa, Finland.

Time(s)

Fig. 12. Control variables during the faulty operation of the process with control parameters manipulated accordi ng to diagnostic information.

560