You are on page 1of 13

ARTICLE IN PRESS

Control Engineering Practice 16 (2008) 12311243 www.elsevier.com/locate/conengprac

Comparison of friction models applied to a control valve


Claudio Garcia
Laboratory of Automation and Control, Department of Telecommunications and Control Engineering, Polytechnic School of the University of Sa o Paulo, 05508-900 Sa o Paulo, Brazil Received 27 June 2006; accepted 24 January 2008 Available online 9 June 2008

Abstract Eight different models to represent the effect of friction in control valves are presented: four models based on physical principles and four empirical ones. The physical models, both static and dynamic, have the same structure. The models are implemented in Simulink/ Matlabs and compared, using different friction coefcients and input signals. Three of the models were able to reproduce the stick-slip phenomenon and passed all the tests, which were applied following ISA standards. r 2008 Elsevier Ltd. All rights reserved.
Keywords: Control valves; Stiction; Stick-slip; Static friction models; Dynamic friction models; Data-driven models

1. Introduction Performance assessment of control loops is an important research theme, and there are many tools to detect variability in control loops. These tools are employed to diagnose different causes of variability, such as friction in the control valve, oversized valves, improperly tuned controllers, disturbances coming from other control loops, and so on. Data extracted from real processes is usually used to test the performance assessment tools. An easier way to perform the initial tests of the performance assessment techniques might be to use simulators, in which the cause of variability is simulated. After these preliminary tests, the tool can be applied to diagnose real situations with data collected from existing plants. Control valves are the most common nal control elements in industry. One of the main factors that affect the behavior of the control loops is friction in control valves. Among the variability causes previously mentioned, the most difcult one to model is friction, and in particular static friction (stiction). The purpose of this paper is to implement and test different friction models applied to control valves. The idea is to

Tel.: +55 11 3091 5648; fax: +55 11 3091 5718.

E-mail address: clgarcia@lac.usp.br 0967-0661/$ - see front matter r 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.conengprac.2008.01.010

analyze the behavior of the models with the valve operating in open loop, simulating a valve installed in a bench. It is necessary to take into account that valve behavior changes signicantly as friction increases. Consider, for instance, an ideal frictionless pneumatic valve with a full stroke of 0100%. If this same valve is affected by friction, it will not move until a certain pressure is applied to its actuator. Besides, when a valve is affected by stiction, the behavior of the control loop presents variability, since the valve does not respond instantaneously to the control signal. What happens is that the signal that comes from the controller has to reach a value high enough to overcome the stiction and move the stem. When this occurs, the stem slips and the valve position normally goes to a point beyond the desired value, causing oscillations and variability in the control loop. Models based on physical principles as well as empirical or data-driven ones have been proposed to simulate valve friction. Physical models describe the friction phenomenon using balance of forces and Newtons second law of motion. The main disadvantage of these models is that they require knowledge of several parameters such as mass of the moving parts, spring coefcient, and various friction coefcients (viscous, Coulomb and static), which are not easily estimated (Garcia, 2007; Romano & Garcia, 2007, 2008). On the other hand, the data-driven models simplify

ARTICLE IN PRESS
1232 C. Garcia / Control Engineering Practice 16 (2008) 12311243

the simulation of a sticky valve and have been used to study valve stiction (He, Wang, Pottmann, & Qin, 2007). Many papers on friction modeling in control valves have been published in the last few years (Choudhury, Jain, & Shah, 2006; Choudhury, Thornhill, & Shah, 2004; Choudhury, Thornhill, & Shah, 2005; Eborn & Olsson, 1995; He et al., 2007; Jain, Choudhury, & Shah, 2006; Kano, Maruta, Kugemoto, & Shimizu, 2004; Kayihan & Doyle, 2000; Stenman, Gustafsson, & Forsman, 2003), but a full comparison of different models to describe the behavior of control valves affected by friction has not been presented. In Eborn and Olsson (1995) the authors compare some friction models, but the results are presented with the valve inserted in a control loop, in such a way that it is difcult to visualize how the isolated valve responds when submitted to different kinds of input signals. In He et al. (2007) the authors present one gure comparing some data-driven models, considering just the case when the valve is ideal, that is, with no friction. In this work, the simulated valves are modeled with three different levels of friction and are submitted to tests that are recommended in ISA standards for real control valves (ISA, 2000, 2006). This form of testing the models is a contribution of this work, since there is not any other related paper that performs tests in simulated valves according to international standards. The paper is organized as follows: in Section 2, the eight valve friction models applied to a pneumatic springdiaphragm sliding stem valve are presented. In Section 3, the applications of the valve friction models analyzed in this paper are listed. In Section 4, the tests applied to control valves according to ISA standards are presented. In Section 5, the characteristics of three valves with different friction coefcients are presented. In Section 6, the responses of the model simulations, with valves with different friction coefcients, applying the ISA recommended testing, are shown and an evaluation table is presented. Finally, in Section 7, the conclusions are drawn. 2. Control valve friction models As the main purpose of this paper is to compare friction models applied to a control valve, eight different models of friction in pneumatic sliding stem control valves are presented, starting from simple models, with just one parameter, and moving to more complex ones, with seven parameters: Classical (Olsson, 1996), Karnopp (Karnopp, louvry, Dupont, & 1985), Seven Parameters (Armstrong-He Canudas de Wit, 1994), Lugre (Canudas de Wit, Olsson, stro A m, & Lischinsky, 1995), Stenman (Stenman et al., 2003), Choudhury (Choudhury, Jain et al., 2006; Choudhury, Thornhill et al., 2004; Choudhury et al., 2005), Kano (Kano et al., 2004) and He (He et al., 2007). The rst four are physical models, the rst two (Classical and Karnopp) being static models and the next two (Seven Parameters and Lugre) dynamic ones. The last four are empirical models. Notice that the more recent models are all data driven.

2.1. Force balance on the components of a pneumatic sliding stem valve The function of the valve actuator is to move the valve stem to modulate its opening. Pneumatic control valves are still the most used in the process industries, due to their low cost and simplicity. In order to model a sliding stem valve, it is assumed that the input variable is the signal that comes from the controller, converted to a pressure signal, and that the stem position is the output variable. In that way, the force balance equation is as follows (Choudhury et al., 2005; Kayihan & Doyle, 2000): F pressure F spring F friction F fluid F seat , mx (1)

where m is the mass of the valve moving parts (stem and plug); x is the stem position; F pressure S a P is the force applied by the actuator, Sa being the diaphragm area and P the air pressure; F spring k x is the spring force, k being the spring constant; F friction is the friction force; F fluid a DP is the force due to the uid pressure drop across the valve, with a the plug unbalanced area and DP the pressure drop; and F seat is the extra force required for the valve to be forced into the seat. Following Choudhury et al. (2005) and Kayihan and Doyle (2000), the contributions of F fluid and F seat are negligible in practical situations. F fluid is disregarded because it is two orders of magnitude smaller than the friction and spring forces, which means that the valve is modeled as if there was no uid in the line. F seat is not considered for simplicity. The main issue is how to model the friction force in Eq. (1). This will be done in the following sections through different friction models. 2.2. Static friction models According to Olsson (1996), friction models can be classied as static and dynamic. The classical friction models are static, which means that the friction is modeled as a static function of velocity. In the dynamic models there are time-varying parameters. This classication does not agree with what is normally dened as static or dynamic systems, but it has been kept in this work, to be in agreement with the related published papers. In the static models, three components are normally considered:

  

static friction or stiction; viscous friction and Coulomb friction.

The total friction force can be calculated as follows: h i 2 F friction v F c F s F c ev=vs sgnv F v v, (2) where Fc is the Coulomb friction coefcient, Fs is the static friction coefcient, v is the stem velocity, vs is the Stribeck velocity and Fv is the viscous friction coefcient.

ARTICLE IN PRESS
C. Garcia / Control Engineering Practice 16 (2008) 12311243 1233

Substituting Eq. (2) in (1), considering the simplications about F fluid and F seat , the force balance equation in the valve results in: Sa P k x mx  h i 2 F c F s F c ev=vs sgnv F v v . (3)

(Olsson, 1996): F friction x s0 x when sticking, ( F friction v; t F c F s g; t d 2 1 vt tl =vs when sliding. (6) 1 ) (5)

sgnv F v v Here 2.2.1. Classical model Some static friction models are summarized in Olsson stro (1996) and Olsson, A m, Canudas de Wit, Ga fvert, and Lischinsky (1998). The following is the classical static friction model: 8 > < F friction v if va0; if v 0 and jF e jpF s ; F friction F e (4) > : F sgnF if v 0 and jF j4F ; s e e s where F friction v can be calculated by Eq. (2) and F e F pressure F spring is the external applied force. The rst line of Eq. (4) refers to the situation in which the stem is moving. The second line represents the friction force when the stem is stopped. The third line represents the situation in which the stem is about to move, because the external force surpassed the static and Coulomb friction forces. A problem with this model is its behavior around null velocity, because in simulations the speed never reaches exactly v 0 and so the stem never stops completely. This problem was worked out as shown in the sequel. 2.2.2. Karnopp model Karnopp (1985) presented a model, aiming at solving the problems with null speed detection and avoiding the switching between the model equations that describe the sticking or sliding body. The model denes an interval around v 0, creating a dead zone for jvjoDV . Depending on jvjoDV or not, the friction force is a saturated version of the external force or a static function of velocity, as presented in Eq. (2). DV is a parameter to be dened and, in this work, after many simulations, it was chosen as DV 0:6vs 1:52 104 m=s. The criterion used to select it was to observe the simulations and to search the largest value for DV in which the stem effectively stopped moving when stuck. According to Karnopp (1985), the limit velocity DV has a negligible effect in the model response. 2.3. Dynamic friction models 2.3.1. Seven Parameter model In practice, the stiction force may vary with time while the moving body is stuck. To reproduce this effect, it is used a dynamic model in which the static friction coefcient increases during the time the body is stuck. louvry et al. This model was proposed in Armstrong-He (1994) and it is described by the following equations F s g; td F s;a F s;1 F s;a

td , td g

(7)

where Fs,a is the static friction coefcient at the end of the previous sliding period and td is the dwelling time. In the sliding friction presented in Eq. (6), the velocity value is affected by the frictional delay tl. This model indeed requires seven parameters: s0 is the pre-sliding stiffness, Fc is the Coulomb friction coefcient, F s;1 F s F c is the steadystate friction coefcient, Fv is the viscous friction coefcient, vs is the Stribeck velocity, tl is the frictional delay, and g determines the temporal behavior of the stiction. As it can be noticed in Eq. (5) and (6), this model is composed by two sub-models: one for sticking and one for sliding. The switching between the sub-models remains undetermined, which might demand an eighth parameter. Furthermore, the model states have to be appropriately initialized every time a switching occurs (Olsson, 1996; Olsson et al., 1998). To avoid this problem, it is proposed in this work the implementation of a model that is a fusion of the Karnopp and the Seven Parameter models. With this approach, sub-model (5) is not used, because it applies Karnopp proposal during the sticking phases. The submodel (6) is used inside the Karnopp model, to replace the equation that calculates the sliding friction force. 2.3.2. Lugre model A dynamic model that describes the friction effects more accurately, mainly for low velocities and during velocity reversals, was proposed in Canudas de Wit et al. (1995), Olsson (1996) and Olsson et al. (1998). The name Lugre comes from the fact that it originated from research done at the LUnd Institute of Technology in Sweden and the Laboratory of Automatic Control of GREnoble in France. The Lugre model is presented next (Olsson, 1996): dz jvj v z dt g v i 2 1h g v F c F s F c ev=vs s0 dz F v v. F friction s0 z s1 (8) dt Here z is the average deection of the bristles (as surfaces are very irregular at the microscopic level and they make contact at a number of asperities, this is represented as a contact through elastic bristles), v is the relative velocity between the two surfaces, g(v) is the positive function that species how the average deection depends on the relative

ARTICLE IN PRESS
1234 C. Garcia / Control Engineering Practice 16 (2008) 12311243

velocity of the surfaces in contact, s0 is the stiffness coefcient, F friction is the friction force, and s1 is the damping coefcient. The calculation of the friction force is inserted in the part of the Karnopp model which is in charge of calculating the
84
S

sliding friction force. Therefore, the structure of the Karnopp model is also applied to simulate the Lugre model. 2.4. Data-driven models A detailed physical model has many unknown parameters, so it is often difcult to estimate them. Besides, complex models are much slower to run in a computer. A data-driven model, on the other hand, is useful because it has only a few parameters to identify, and can be run faster. This is the reason why, in recent years, the models presented in the technical literature are all data-driven, as can be observed next. 2.4.1. Stenman model The basic concept behind this model is to try to imitate the jump that occurs in the stem position, when stiction is overcome. The Stenman friction model (Stenman et al., 2003) is parameterized by one parameter d and may be expressed by the following equation: ( xt 1 if jut xt 1jpd ; xt (9) ut otherwise;

72
Stem position x (%)

60
J

48
J

36 24
Slip jump J

12 0

Deadband

Stickband

10

20

30 40 50 60 70 Input pressure P (%)

80

90

100

Fig. 1. Example of valve signature showing the parameters J and S (Choudhury, Jain et al., 2006).

Input pressure x(k)

xss = xss y(k) = 0

no

x(k) > 0 yes

xss = xss y(k) = 100

no

x(k) < 100 y(k) = x(k) yes

yes no
S=0&J=0? v_new = [x(k) - x(k-1)]/t

no

sign(v_new) = 0

no

sign(v_new) = sign (v_old)

yes I=0 I=1 no


|x(k) - xss| > J

yes yes I = 1? no yes y(k) = y(k-1) I=0 no


|x(k) - xss| > S?

xss = x(k-1) y(k) = y(k-1)

yes
y(k) = x(k) sing (v_new) (S J ) / 2

Value sticks

Remain stuck

Value slips and moves

Value position y(k)


Fig. 2. Flowchart for the Choudhury model (Choudhury, Jain et al., 2006).

ARTICLE IN PRESS
C. Garcia / Control Engineering Practice 16 (2008) 12311243 1235

where x(t1) and x(t) correspond to past and present stem positions, respectively, u(t) is the present controller output and d is the valve stiction band. 2.4.2. Choudhury model Successive versions of the Choudhury model were presented in Choudhury, Thornhill et al. (2004), Choudhury et al. (2005) and Choudhury, Jain et al. (2006). It requires just two parameters (J and S), as shown in Fig. 1, which are extracted from signature experiments. S represents the amplitude of the input signal (pressure) during the time in which the stem is stuck (stickband+ deadband). J represents the size of the stem slip (slip jump). This model is described in the owchart (Choudhury, Jain et al., 2006) presented in Fig. 2. 2.4.3. Kano model The Choudhury model is not able to deal with both deterministic and stochastic signals (Kano et al., 2004). The
Stem position x (%)

84 72 60 J 48 J 36 24 Slip jump J 12 0 fD 10 20 fS 30 40 50 60 70 Input pressure P (%) J

80

90

100

Fig. 4. Relation between controller output and valve position (Kano et al., 2004).

Input pressure u(t)

cum_u = ur + [u(t) - u(t-1)]

yes

abs(cum_u) > fs?

no

uv(t) = u(t) - sign(cum_u fs) fD ur = sign(cum_u fs) fD

uv(t) = uv(t-1) ur = cum_u

Fig. 5. Flowchart for the He model (He et al., 2007).

Kano model is an extension that requires the same two parameters used in the Choudhury model. Its algorithm is presented in a owchart (Kano et al., 2004) shown in Fig. 3. 2.4.4. He model The He model (He et al., 2007) is less complex than the Kano model. It requires two parameters: fS (static friction band) and fD (kinetic friction band). J is dened as f S f D and S as f S f D , as shown in Fig. 4 (Kano et al., 2004). The central line in Fig. 4 presents the situation of an airto-open valve with no friction at all. As the friction increases, the static (fS) and the dynamic (fD) friction bands also increase. In Fig. 4 the controller output and the valve position are normalized and both vary in the range 0100%, so all the lines indicating the valve movement have a slope of 45%. The owchart of the He model can be seen in Fig. 5 (He et al., 2007). 3. Applications of the control valve friction models Friction is a common problem in spring-diaphragm type valves. Bialkowski (1993) reported that about 30% of the

Fig. 3. Flowchart for the Kano model (Kano et al., 2004).

ARTICLE IN PRESS
1236 C. Garcia / Control Engineering Practice 16 (2008) 12311243

Actuator pressure P (%) and stem position x (%)

loops are oscillatory due to control valve problems. About the same numbers are obtained in another investigation (Ender, 1993). Quantifying the valve friction coefcients may help to reduce control loop variability. It is part of what is known as control performance monitoring/assessment, an important asset-management technology to maintain highly efcient operation performance of automation systems in production plants (Jelali, 2006; Thornhill & Horch, 2006, 2007). The control valve friction models described here have been applied to stiction detection and quantication in several papers. The Stenman model is used in Stenman et al. (2003) for stiction detection. In Kano et al. (2004) the authors propose an algorithm for valve stiction detection. In Rossi and Scali (2005) and Singhal and Salsbury (2005) the Choudhury model is used to analyze the proposed valve stiction detection method. In He et al. (2007) the authors consider the He model to detect stiction. Concerning quantication of control valve stiction, the papers Garcia (2007) and Romano and Garcia (2007, 2008) focus on techniques to estimate the parameters of the Karnopp model (m, k, Fv, Fc and Fs). The idea of Detection and quantication of control valve stiction is to detect if the control valve presents stiction and to nd out the friction parameters. In Srinivasan, Rengaswamy, Narasimhan, and Miller (2005) the goal is to nd the parameter d of the Stenman model. In Choudhury, Shah, Thornhill, and Shook (2006), and Choudhury, Shah, and Thornhill (2004) and Ha gglund (2007) the idea is to estimate the parameter S, whereas in Choudhury, Jain and Shah (2008), Choudhury, Jain et al. (2006), Jain et al. (2006) and Schoene and Qin (2005) the concern is with J and S. These parameters are related to the Choudhury and Kano models. Papers on Friction compensation for control valves are concerned with developing friction compensators to deal with the control valves that are affected by this problem. The authors Kayihan and Doyle (2000) use the Classical model to develop their compensator. The compensator in Srinivasan and Rengaswamy (2005, 2006, 2008) employs the Stenman model. 4. Tests recommended by ISA to be applied to control valves The models presented in Section 2 were implemented in Matlab/Simulinks and tested with different friction coefcients and input signals. In order to perform the simulated tests according to international standards related to control valves, the choice was to apply the tests that are dened in the ISA standards (ISA, 2000, 2006). The aspects of these documents that are relevant to this paper are discussed next. 4.1. Control valve diagnostic tests The purpose of the two tests recommended in ISA (2006) is to provide control valve users with a uniform means of

acquiring and reporting data used for diagnosing valve operability. All data must be acquired with the valve out of service, with no internal dynamic forces acting on the closure member. For both tests, data must be acquired at equally spaced intervals of input signal specied over the travel in the opening and closing directions. The recommended minimum number of data points in each direction is 101 for a full signal span. Each method is briey summarized as follows:

Dynamic test method: Starting with the user dened minimum signal specied, ramp the input signal (control signal to a positioner or pressure signal to an actuator) to the maximum signal specied, wait for the pause time and ramp down to the minimum signal. Then wait for the pause time again. The format of the input signals corresponds to a trapezoidal wave. During the travel, record the input signal and valve position. This test generates what are known as signature curves of the valve, which are produced by operating a valve through its signal range which, according to ISA (2006), are plotted with stem travel in the vertical axis and actuator pressure in the horizontal axis, with both scales ranging from 0% to 100%. Thus, the signature curves are derived by plotting the valve position versus the input signal in both directions, over the minimum to the maximum input signal specied for the application. The dynamic test response obtained in this paper in shown in Fig. 6 and the signature curves can be seen in Figs. 711. Ramp and pause test method: Starting at the user dened minimum input signal specied, ramp the input signal at a slow rate. Wait for the ramp and pause time specied. Repeat the procedure up to the maximum input signal specied. Record the input signal and valve position. Repeat the preceding process in the opposite direction.

100 90 80 70 60 50 40 30 20 10 0 0 2 4 6 8 10 12 14 16 18 20 Time t (s)


Input signal Classical model Karnopp model

Fig. 6. Dynamic test response with two input cycles applied to the rough valve with trapezoidal input.

ARTICLE IN PRESS
C. Garcia / Control Engineering Practice 16 (2008) 12311243
100 90 80

1237

100 90 80 Stem position x (%)


Vendor Nominal Rough

Stem position x (%)

70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 Actuator pressure P (%) 80 90 100

70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 80 90 100 Actuator pressure P (%)


Vendor valve Nominal valve Rough valve

Fig. 7. Signature curves of vendor, nominal and rough valves employing the Seven Parameter model.

Fig. 9. Signature curves of vendor, nominal and rough valves employing the Stenman model.

100 90 80 Stem position x (%)

100 90 80 Stem position x (%)


Vendor valve Nominal valve Rough valve

70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 Actuator pressure P (%) 80 90 100

70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 Actuator pressure P (%) 80 90 100


Vendor valve Nominal valve Rough valve

Fig. 8. Signature curves of vendor, nominal and rough valves employing the Karnopp and Lugre models.

Fig. 10. Signature curves of vendor, nominal and rough valves employing the Choudhury and Kano models.

In the sketch shown in ISA (2006) there are ve pause times, at 20%, 40%, 60%, 80% and 100% of the input signal. See this test result in Fig. 12. 4.2. Control valve response measurement from step inputs In ISA (2000), three alternative environments are dened to perform the tests: bench, laboratory and in-process testing. The bench test, which is performed without process uid, is the one carried out in this work. In Eq. (1) it is assumed that the force due to the uid pressure drop across the valve F fluid is null. This is a characteristic of bench testing. The input and output signals to be measured in bench testing are the pressure input signal and the stem position.

The three valve tests recommended in ISA (2000) can be summarized as follows:

Baseline test: Used to evaluate measurement noise, the presence of limit cycles, and the baseline response time T86, that corresponds to the interval of time between initiation of an input signal step change and the moment at which the response reaches 86.5% of its full steadystate value. In the example given in ISA (2000), two steps up are applied in the input signal, from 50% to 52% and from 52% to 54%; and two steps down, from 54% to 52% and from 52% to 50%. Each step lasts 1 min. This test result is presented in Fig. 13. Small step test: Used to determine dead band and resolution. The dead band is the range through which an

ARTICLE IN PRESS
1238
100 90 80 Stem position x (%) 70 60 50 40 30 20 10 0 0 10 20 30 40 50 60 70 Actuator pressure P (%) 80 90 100
Vendor valve Nominal valve Rough valve

C. Garcia / Control Engineering Practice 16 (2008) 12311243

input signal may be varied, with reversal of direction, without initiating an observable change in output signal. In ISA (2000) it is dened as a percentage of input span. The resolution is the smallest step increment of input signal in one direction for which movement of the output is observed, expressed as percentage of input span. In ISA (2000) the test given as example was performed with incremental steps of 0.1%, starting at 50% of the input signal and reaching 50.6% and coming back to 50%, each step lasting 30 s. The whole cycle must be repeated twice. It is necessary to assure that there is at least one step in addition to the step causing initial movement. This test result is presented in Fig. 14. Response time test: Consists of an increasing sequence of step sizes. Start the test with a step size of 0.1% and end with 10%, assuming the following values: 0.1%, 0.2%, 0.5%, 1%, 2%, 5% and 10% of input signal span.

Fig. 11. Signature curves of vendor, nominal and rough valves employing the He model.
Actuator pressure P (%) and stem position x (%) 100 90 80 70 60 50 40 30 20 10 0 0 4 8 12 16 20 24 Time t (s) 28 32 36 40
Input signal Karnopp model Choudhury model

Actuator pressure P (%) and stem position x (%)

51.6 51.4 51.2 51 50.8 50.6 50.4 50.2 50 0 6 12 18 24

Input signal Karnopp model

30 36 42 Time t (s)

48

54

60

66

72

Fig. 12. Response of the Karnopp and Choudhury models to the ramp and pause test method employing nominal valve.
Actuator pressure P (%) and stem position x (%)
Actuator pressure P (%) and stem position x (%) 90 86 82 78 74 70 66 62 58 54 50 0 4 8 12 16 20 24 Time t (s) 28 32 36 40
Input signal Karnopp model

Fig. 14. Result of small step test with vendor valve.

100 90 80 70 60 50 40 30 20 10 0 0 3 6 9 12 15 18 21 24 27 30 33 36 Time t (s)


Input signal Karnopp model

Fig. 13. Results of baseline test with nominal valve.

Fig. 15. Results of response time test applied to the nominal valve.

ARTICLE IN PRESS
C. Garcia / Control Engineering Practice 16 (2008) 12311243 1239

Larger steps sizes such as 20% and 50% may be used if desired. In the example given in ISA (2000), each step lasts 1 min. In Fig. 15 this test result is presented. 5. Characteristics of the simulated tested valves It was selected a sliding stem globe valve, with diaphragm actuator, provided with a spring that moves the actuator stem in a direction opposite to the movement caused by the diaphragm pressure. Its nominal size is 4 in. and its full stem stroke is 4 in. (0.1016 m). So as to simulate the behavior of a control valve, it is necessary to dene the model parameters. The mass m and the friction force depend on the valve. The parameter values were extracted from Fitzgerald (1995) and Kayihan and Doyle (2000). In this last paper, the authors have dened three valves with different friction coefcients:

The parameters J and S for the Choudhury, Kano and He models are: Valve S J Vendor 0.0167 0 Nominal 0.5894 0.0550 Rough 0.9224 0.0860

S (width of the signal pressure during sticking) is very simple to measure, because the width of the signal pressure during sticking corresponds to the two horizontal lines in Fig. 1. The measurement of J (slip jump size) is not as simple, because it is not so easy to determine where it ends. The values shown above were calculated in order to generate similar gures to the Karnopp and Lugre models, since they had generated an expected signature curve. 6. Simulation of the friction models In this paper, the eight models are compared through the ve tests described in Section 4. The valve is supposed to operate in open loop in a bench, eliminating any test related to observing limit cycling. 6.1. Search for an adequate integration step time As it has already been mentioned, one of the great problems to simulate friction is related to the model behavior close to null velocities. If the integration step of the numeric integration method is not sufciently small, the stem velocity can cross the zero velocity and simply reverse, without stopping. Therefore, to make the stem stop, it is necessary for the integration step to be sufciently small. Fig. 1 shows the presumed signature of a valve with stiction. It is used as the basis of comparison for dening the adequate integration step time for the static and dynamic models. When the stem is stuck, the effect is named deadband+stickband, and when it is slipping, it is called slip jump. The slip jump corresponds to the momentary slipping suffered by the stem when the externally applied force overcomes the stiction force. At some instants, the resistance is much reduced, causing a large acceleration and peaks in the stem velocity. A good model of the rough and nominal valves, which present

  

vendor: it is the valve with smaller friction; nominal: it is the valve with mean friction; rough: it is the valve with higher friction.

The vendor valve has only the Coulomb friction component and corresponds to a new valve with Teon stem packing. The nominal valve presents stiction and may correspond to a new valve but with graphite packing operating at ambient temperature. Finally, the rough valve corresponds to a rather worn out one. The features of each valve are shown in Table 1. According to Eborn and Olsson (1995), it is assumed that tl 0:02 s, g 100 s and F s;1 F s F c for the Seven Parameters model. In the simulations, it is assumed for this model that the time in which the stem was stuck at the start of the simulation was very long, so that the initial stiction has reached its maximum value F s;1 F s F c . In the Lugre model, the following parameters were used (Eborn & Olsson, 1995): s0 108 N=m and s1 9000 N s=m. To estimate parameter d in the Stenman model, a search was performed, aiming at best reproducing the expected signature curve of Fig. 1. The values found are: Valve d Vendor 0.010 Nominal 0.4666 Rough 0.334

Table 1 Features of the simulated valves Parameter Sa, area of the actuator diaphragm (m2) m, mass of the moving parts of the valve (kg) k, spring constant (N/m) Fc, Coulomb friction coefcient (N) Fs, stiction coefcient (N) Fv, viscous friction coefcient (N s/m) vs, Stribeck velocity (m/s) Vendor 0.06452 1.361 52538 44.48 44.48 612.9 2.54 104 Nominal 0.06452 1.361 52538 1423 1707.7 612.9 2.54 104 Rough 0.06452 1.361 52538 2224 2668.8 1226 2.54 104

ARTICLE IN PRESS
1240 C. Garcia / Control Engineering Practice 16 (2008) 12311243

stiction, should be able to reproduce the stick-slip phenomenon. The signature curve shown in Fig. 1 was obtained through a dynamic test method, applied as described in Section 4. To nd an adequate integration step for the static and dynamic models, simulations were performed employing integration steps of 103, 104, 105 and 106 s. The best results were obtained with 106 s. However, with such a small step, the running time of the models is very long. Among the numeric methods with xed and variable step tested to minimize running time without affecting the model accuracy, Eulers xed step method resulted in lower running time and responses similar to the others. The Stenman, Choudhury, Kano and He models are not rst principle ones, so there is no concern about the behavior of the velocity close to zero. The Euler method with time step of 103 s was employed for them. 6.2. Dynamic test method and signature curves In the papers (Choudhury et al., 2005; Kayihan & Doyle, 2000) about friction in control valves, the authors employed the valve signature curve, with a sinusoidal actuator pressure input, in order to show the behavior of the model in open loop. In this paper, following the ISA standard recommendations, the input signal to be used is trapezoidal or triangular and not sinusoidal. To perform this test, simulations were carried out, varying the input signal that represents the pressure on the actuator in the range of 012 psi (082,737 Pa) and observing the behavior of the stem position. The ramp time is 4 s and the pause time is null for the triangular signal or 1 s for the trapezoidal one. In the simulations, two complete opening and closing cycles are applied. In order to reproduce the signature curve of Fig. 1, it is necessary to have at least two complete cycles of the input signal, a fact which is not stated explicitly in the ISA standards. Only the data relative to the second cycle, corresponding to the situation when the valve reaches its normal operation cycle, is used. In the rst cycle the valve stem starts at position 0% (completely open), as shown in Fig. 6, which depicts the response of the Classical and Karnopp models, applied to the rough valve with trapezoidal input. It can be noticed that, after the rst input cycle, the stem motion is reduced to a small range of its full stroke. As stated in item 2.2.1, the main problem with the Classical model is its behavior around null velocity, in such a way that the valve is never sticky. As it is fundamental for the model to reproduce the stick-slip phenomenon, the Classical model is discarded. The signature curves presented next employ a trapezoidal input. Fig. 7 shows the response of the Seven Parameter model for the three valves with different friction coefcients. Because the time during which the trapezoidal input stops is only 1 s, the Seven Parameter model responses in Fig. 7 has a negligible slip jump with the parameters used. In order to make the stiction force appear in this model, it

would be necessary for the trapezoidal input to have a longer pause. Only then would it be possible to observe the slip jump. The pause time of the input signals to be used here is not large, so the Seven Parameter model is also discarded. The signature curves of the Karnopp and Lugre models are shown in Fig. 8, in which the stick-slip phenomenon appears for both models, as shown in Fig. 1. The response of the Lugre model with the assigned parameters is quite similar to the Karnopp one. Nevertheless, if the parameters are changed, the response of the model becomes different. For instance, if the values p (Olsson, 1996) s0 105 N/m and s1 2 s0 N s=m are considered, the Lugre model behavior is very different. The signature curves generated by the Stenman model are shown in Fig. 9 and they are completely different from what is expected. Therefore, this model is discarded. The signature curves derived from the Choudhury and Kano models are shown in Fig. 10. It indicates that the responses of both models are identical. Comparing Figs. 8 and 10, it can be noticed that the Karnopp, Lugre, Choudhury and Kano models present very similar signature curves for the three valves. The responses of the He model when submitted to a trapezoidal input are shown in Fig. 11 and they do not comply with what is expected, so it is discarded. Although not being exactly equal, the signature curves of all the models for the vendor valve are similar, which means that all of them are able to represent the behavior of valves with low friction coefcients. Nevertheless, when the friction coefcients increase, the differences appear. As the behavior of the Karnopp, Lugre, Choudhury and Kano models are the ones that responded as expected, they continue being analyzed in the next sections. 6.3. Ramp and pause test method The ramp and pause method was applied with ramp time of 1 s, pause time of 1 s and pauses in 20%, 40%, 60%, 80% and 100%. The test is performed twice, although this is not required in ISA (2006). The responses of the Karnopp, Lugre, Choudhury and Kano models for the vendor valve, not shown here, are quite similar, following closely the input signal. The responses for the nominal valve employing the Karnopp and Choudhury models are shown in Fig. 12. The Lugre and Kano models presented responses quite similar to the Karnopp one, so they are not shown in Fig. 12. Fig. 12 reveals that the Karnopp model responded similarly, keeping a certain distance from the input signal, due to the relatively high friction coefcients in this valve, whereas the Choudhury model did not move at all. Based on that, this last model is discarded. The behavior of the models for the rough valve are similar to the one for the nominal valve, that is, the Karnopp, Lugre and Kano models responded quite similarly, whereas the Choudhury model did not move.

ARTICLE IN PRESS
C. Garcia / Control Engineering Practice 16 (2008) 12311243 1241

In its rst movement in Fig. 12, the valve has to overcome the fs value, which corresponds to f s J S =2 32:22%. Therefore, the valve moves when pressure P reaches 32.22%. As the valve movement goes on in the same direction, without a valve reversion, the second valve movement occurs when the pressure overcomes J 5:50%, that is, it has to increase 5.50% above the pressure value when the valve stuck. In Fig. 12 the valve stuck when the pressure reached its second pause, that is, 40%, so the next pressure value for the valve to slip is 45.50%. When there is a valve direction reversal, the input signal has to change an amount of S 58:94% for the valve to move. Analyzing Fig. 12, it is possible to see that the pressure had to change from 100% to 41.08% for the valve to move, which is equivalent to S. 6.4. Baseline tests The valve considered in the simulations has no dynamics associated with the actuator, that is, it responds instantly as the pressure signal varies, so there is no meaning in estimating T86, and this test is applied simply to compare the behavior of the models. The responses of the three models are very similar for the three valves. Fig. 13 presents the results of the baseline test applied to the nominal valve. It is presented just the response of the Karnopp model, since the Lugre and Kano model responses are practically identical. The input pressure has to increase an amount of f S 100 S J =2 32:22% in order to give the rst jump. Thus, as its initial value is 50%, the valve moves as soon as the input signal jumps from 82% to 84%. For the second jump, the input pressure has to overcome the J value (5.50%), so it moves again when the input pressure jumps

from 88% to 90%. In order to move in the closing direction, the input signal would have to overcome the S value (58.94%). However, the input signal only varies 40%, which is not enough to move the valve again, so the valve does not return to its initial position.

6.5. Small step tests This test is carried out on the vendor and nominal valves, since the rough valve did not move with the input pressure varying from 50% to 100%. The results of the three models are quite similar. Fig. 14 shows the results of the vendor valve, just presenting the response of the Karnopp model. To fulll what is expected of the small step test, it might be enough to increase the input signal up to 51%, corresponding to the point where a second step is applied after the rst movement of the valve. Nevertheless, the input pressure was extended up to 61.7%, to allow measurement of the dead band. The input pressure has to increase an amount of f S 100 S J =2 0:835% in order to give the rst jump. Its initial value is 50%, so the valve moves as soon as it reaches 50.9%. For the second jump, the input pressure has to overcome the J value (0%), so in the next step it moves again. In order to move in the closing direction, the input signal has to overcome the S value (1.67%). Therefore, when the input signal reversion reaches 1.7%, the valve moves in the opposite direction. It keeps oscillating around 50.85% and does not return to the initial position. The valve output moved when the input changed from 50.9% to 51%, meaning that the resolution is less than or equal to 0.1%. The measured dead band is between 1.6% and 1.7%. Its true value is S 1:67%.

Table 2 Summary of the performed tests Test type Dynamic test method (signature curve) Valve type Vendor Nominal Rough Vendor Nominal Rough Vendor Nominal Rough Vendor Nominal Rough Vendor Nominal Rough Classical Y N N Y N N Y N N N N N Y N N Karnopp Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Seven Parameter Y N N Y N N Y N Y Y Y Y Y Y Y Lugre Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Stenman Y N N Y N N N N N N Y Y Y N N Choudhury Y Y Y Y N N N N Y N Y Y N N Y Kano Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y He Y N N Y N N Y Y Y Y Y Y Y Y N

Ramp and pause test method

Baseline tests

Small step tests

Response time test

Y, response as expected; N, response not as expected.

ARTICLE IN PRESS
1242 C. Garcia / Control Engineering Practice 16 (2008) 12311243 Choudhury, M. A. A. S., Shah, S. L., & Thornhill, N. F. (2004). Detection and quantication of control valve stiction. In Proceedings of the 7th DYCOPS, Boston, USA. Choudhury, M. A. A. S., Shah, S. L., Thornhill, N. F., & Shook, D. S. (2006). Automatic detection and quantication of stiction in control valves. Control Engineering Practice, 14(12), 13951412. Choudhury, M. A. A. S., Thornhill, N. F., & Shah, S. L. (2004). A datadriven model for valve stiction. In Proceedings of the 5th ADCHEM International symposium on advanced control of chemical processes, Hong Kong, China (pp. 261266). Choudhury, M. A. A. S., Thornhill, N. F., & Shah, S. L. (2005). Modeling valve stiction. Control Engineering Practice, 13(5), 641658. Eborn, J., & Olsson, H. (1995). Modeling and simulation of an industrial control loop with friction. In Proceedings of the 4th IEE conference on control applications, Albany, USA (pp. 316322). Ender, D. B. (1993). Process control performance: Not as good as you think. Control Engineering, 40(10), 180190. Fitzgerald, B. (1995). Control valves for the chemical process industries. New York, USA: McGraw-Hill. Garcia, C. (2007). Parameter estimation of friction model for control valves. In Proceedings of the 8th IFAC DYCOPS, Cancu n, Mexico (Vol. 2, pp. 273278). Ha gglund, T. (2007). Automatic on-line estimation of backlash in control loops. Journal of Process Control, 17(6), 489499. He, Q. P., Wang, J., Pottmann, M., & Qin, S. J. (2007). A curve tting method for detecting valve stiction in oscillating control loops. Industrial and Engineering Chemistry Research, 46(13), 45494560. ISA. (2000). Standard ANSI/ISA-75.25.01-2000. Test procedures for control valve response measurement from step inputs. ISA. (2006). Standard ANSI/ISA-75.26.01-2006. Control valve diagnostic data acquisition and reporting. Jain, M., Choudhury, M. A. A. S., & Shah, S. L. (2006). Quantication of valve stiction. In Proceedings of the 6th IFAC ADCHEM, Gramado, Brazil (pp. 11571162). Jelali, M. (2006). An overview of control performance assessment technology and industrial applications. Control Engineering Practice, 14(5), 441466. Kano, M., Maruta, H., Kugemoto, H., & Shimizu, K. (2004). Practical model and detection algorithm for valve stiction. In Proceedings of the 7th IFAC DYCOPS, Boston, USA, CD-ROM. Karnopp, D. (1985). Computer simulation of stick-slip friction in mechanical dynamic systems. Transactions of the ASMEJournal of Dynamic Systems, Measurement and Control, 107(1), 100103. Kayihan, A., & Doyle, F. J., III (2000). Friction compensation for a process control valve. Control Engineering Practice, 8(7), 799812. Olsson, H. (1996). Control systems with friction. Ph.D. thesis. Department of Automatic Control, Lund Institute of Technology. stro Olsson, H., A m, K. J., Canudas de Wit, C., Ga fvert, M., & Lischinsky, P. (1998). Friction models and friction compensation. European Journal of Control, 3(4), 176195. Romano, R. A., & Garcia, C. (2007). Comparison between two friction model parameter estimation methods applied to control valves. In Proceedings of the 8th IFAC DYCOPS, Cancu n, Mexico (Vol. 2, pp. 303308). Romano, R. A., & Garcia, C. (2008). Karnopp friction model identication for a real control valve. In Proceedings of the 17th IFAC World CongressIFAC08, Seoul, Korea, accepted for publication. Rossi, M., & Scali, C. (2005). A comparison of techniques for automatic detection of stiction: simulation and application to industrial data. Journal of Process Control, 15(5), 505514. Schoene, C. B., & Qin, S. J. (2005). Blind identication for the detection and estimation of valve stiction. In Proceedings of the 2005 AIChE annual meeting, Cincinnati, USA (paper 520g). Singhal, A., & Salsbury, T. I. (2005). A simple method for detecting valve stiction in oscillating control loops. Journal of Process Control, 15(4), 371382. Srinivasan, R., & Rengaswamy, R. (2005). Stiction compensation in process control loops: A framework for integrating stiction measure

6.6. Response time test The responses of the three models for the three valves are very similar. In Fig. 15, just the response of this test applied to the nominal valve is presented. In it, the three models only responded to the input excitation when its size was 50% and their outputs did not return to 50%, even when the input did. 6.7. Summary of the testing results A summary of the testing results employing trapezoidal input is presented in Table 2. 7. Conclusions In this work, eight friction models applied to control valves were implemented and compared. They were tested with different input signals and with valves with different friction coefcients. ISA standards related to real control valve testing were also presented. All the bench tests recommended in these standards were applied. The tests were performed with the valve operating in open loop. The Karnopp, Lugre and Kano models were able to represent the expected behavior of the valves, mainly the stick-slip phenomenon. It was noticed that the friction values of the rough valve are so high that in some tests the stem did not even move. In order to work with more realistic friction coefcient values, it is proposed, as future work, to deal with model parameters estimated from real valves in operational conditions (Romano & Garcia, 2008), with different friction statuses, instead of the vendor, nominal and rough valves used here. Also, the intention is to test the three selected models operating in closed loop, with plants with different dynamics, in order to evaluate their behavior in this condition. Acknowledgment The author thanks the support provided by Cenpes/ Petrobras. References
louvry, B., Dupont, P., & Canudas de Wit, C. (1994). A Armstrong-He survey of models, analysis tools and compensation methods for the control of machines with friction. Automatica, 30(7), 10831138. Bialkowski, W. L. (1993). Dreams versus realityA view from both sides of the gap. Pulp & Paper Canada, 94(11), 1927. stro Canudas de Wit, C., Olsson, H., A m, K. J., & Lischinsky, P. (1995). A new model for control of systems with friction. IEEE Transactions on Automatic Control, 40(3), 419425. Choudhury, M. A. A. S., Jain, M., & Shah, S. L. (2006). Detection and quantication of valve stiction. In Proceedings of the 2006 American control conference, Minneapolis, USA, 20972106. Choudhury, M. A. A. S., Jain, M., & Shah, S. L. (2008). Stiction denition, modelling, detection and quantication. Journal of Process Control, 18(34), 232243.

ARTICLE IN PRESS
C. Garcia / Control Engineering Practice 16 (2008) 12311243 and compensation. Industrial and Engineering Chemistry Research, 44(24), 91649174. Srinivasan, R., & Rengaswamy, R. (2006). Techniques for stiction diagnosis and compensation in process control loops. In Proceedings of the 2006 American control conference, Minneapolis, USA (pp. 21072112). Srinivasan, R., & Rengaswamy, R. (2008). Approaches for efcient stiction compensation in process control valves. Computers and Chemical Engineering, 32(12), 218229. Srinivasan, R., Rengaswamy, R., Narasimhan, S., & Miller, R. (2005). Control loop performance assessment 2: Hammerstein model ap1243 proach for stiction diagnosis. Industrial and Engineering Chemistry Research, 44(17), 67196728. Stenman, A., Gustafsson, F., & Forsman, K. (2003). A segmentationbased method for detection of stiction in control valves. International Journal of Adaptive Control and Signal Processing, 17(79), 625634. Thornhill, N. F., & Horch, A. (2006). Advances and new directions in plant-wide controller performance assessment. In Proceedings of the 6th IFAC ADCHEM, Gramado, Brazil (Vol. 1, pp. 2936). Thornhill, N. F., & Horch, A. (2007). Advances and new directions in plant-wide disturbance detection and diagnosis. Control Engineering Practice, 15(10), 11961206.

You might also like