Journal of Process Control 17 (2007) 489499 www.elsevier.

com/locate/jprocont

Automatic on-line estimation of backlash in control loops

Tore Ha gglund
*
Department of Automatic Control, Lund University, Box 118, SE-22100 Lund, Sweden Received 16 October 2006; received in revised form 11 January 2007; accepted 13 January 2007

Abstract This paper describes a new method for detection and estimation of backlash in control loops. The detection procedure is based on normal operating data. It is not assumed that the output from the backlash is measured. The procedure is automatic in the sense that no information has to be provided from the user to run the procedure. Since an estimate of the dead band caused by the backlash is provided by the procedure, the procedure gives all information needed to compensate for the backlash. 2007 Elsevier Ltd. All rights reserved.
Keywords: Backlash; Hysteresis; Dead band; Supervision; Detection; Monitoring; Diagnosis; Estimation; Compensation

1. Introduction Control valves are subject to wear. After some time in operation, this wear results in friction and backlash that deteriorates the control performance. Therefore, valves have been identied as the major source of problems at the loop level in process control [1,2]. Valves with a high level of static friction (stiction) results in stick-slip motion that causes the control loops to oscillate. As the amount of friction increases, so does the backlash in the linkage mechanism in the positioner and actuator of the valve. The backlash adds a time delay to the control loop which deteriorates the control. In [3], it is reported that a backlash of 10% increases the peak error at load disturbances with 50% and the integrated absolute error (IAE) with 100%. These gures are dependent on the magnitude of the load disturbances, but simulation studies presented later in this paper verify them. Since control loops in process control applications often are coupled to surrounding control loops, there is also a risk that the disturbances caused by backlash in one loop will propagate to other loops.

Tel.: +46 46 2228798; fax: +46 46 138118. E-mail address: tore@control.lth.se

When the stiction or backlash becomes large, the valve should, of course, be repaired or replaced. However, this can normally not be done without interrupting the process. For this reason, and other economical reasons, it is of interest to try to keep the valve running for as long time as possible. Stiction can be compensated for using the method presented in [4]. Backlash is easier to compensate for, since it is an invertible nonlinearity. This will be discussed further in Section 3 in this paper. Even though the problems caused by stiction and backlash are severe, they are often not discovered by operators in process control plants. The main reason is that the reduction of personnel has resulted in a situation where each operator simply has too many loops to supervise. For this reason, the research on procedures for automatic performance monitoring has been very active in the last decade. The industrial use of these procedures has also increased rapidly in recent years. Good surveys of performance monitoring procedures are given in [58]. There are several methods suggested for detecting control loops with stiction, e.g. [912]. However, no ecient procedure to detect backlash has been presented so far, but this paper provides such a procedure. In the next section, a description and an analysis of backlash is presented. Section 3 gives some methods to compensate for backlash. The main section of this paper

0959-1524/$ - see front matter 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.jprocont.2007.01.002

490

T. Ha gglund / Journal of Process Control 17 (2007) 489499

1

is Section 4, where the new backlash estimator is presented. Section 5 shows some simulation examples of the backlash estimation procedure. The procedure has also been tested in a paper mill. Results of these industrial tests are presented in Section 6. 2. Backlash

0.5 0 0.5

Fig. 1 shows a block diagram of a control loop with backlash. The controller C has setpoint ysp and process output y as inputs, and control signal u as output. The controller output u is not the input to process P, but it goes through a backlash that gives the true process input ub. Fig. 2 describes the function of the backlash, where the dead band caused by the backlash is denoted d. When the control signal u is reversed, the process input ub remains constant until u has passed the dead band d. The describing function YN of a backlash is !    s   1 p d d d d arcsin 1 2 ReY N a 1 ; p 2 a a a a   d d 2 ; ImY N a pa a 1

1.5

2 2.5

3 3

2.5

1.5

0.5

0.5

Fig. 3. The negative inverse of the describing function of a backlash (solid line) and the Nyquist plots of the loop transfer functions obtained when the processes P1 = e0.2s/(s(1 + 0.8s)) (dashed line) and P2 = 1/(1 + s)4 (dashed-dotted line) are controlled with PID controllers.

ysp C

u
Backlash

ub

y P

Fig. 1. Block diagram of a control loop with backlash. The controller C has inputs setpoint ysp and process output y. The controller output u goes through the backlash and is modied to ub before it enters the input of process P.

ub

where a is the input amplitude and d is the backlash given in Fig. 2 [13]. The negative inverse of the describing function of the backlash is shown in Fig. 3. The gure shows also the Nyquist plots of two loop transfer functions obtained when an integrating and a stable process, respectively, are controlled by PID controllers. From this gure it can be concluded that backlash generates limit cycles when integrating processes are controlled by controllers with integral action. The gure also shows that backlash will normally not generate limit cycles when the process is stable, provided that the controller is well tuned. Since d is divided by a at every position where it appears in (1), the shape of the describing function is independent of d. This has an interesting consequence. It means that the magnitude d of the backlash will inuence the oscillation amplitude, but since the intersection with the Nyquist plot occurs at the same position, the oscillation period will remain the same independent of the magnitude of the backlash. Throughout this paper it is assumed that the controller C is a PID controller. This is mostly the case in practice. However, the results presented in this paper can quite easily be modied to other controllers having integral action. The PID controller used in the following examples has the structure:   Z 1 dy t ut K y f t ; y sp t y f t dt T d f Ti dt 2

Fig. 2. The output ub from a backlash with input u. The dead band of the backlash is d.

where u is the controller output, ysp is the setpoint, yf is the ltered process output, and the controller parameters are gain K, integral time Ti, and derivative time Td. The controller has setpoint weights equal to zero in both the proportional term and the derivative term. This is common

T. Ha gglund / Journal of Process Control 17 (2007) 489499

491

in industrial controllers [14]. The process output is ltered through a second-order low-pass lter Y f s 1 1 sT f 2 Y s; 3

This is fairly close to what was predicted by the describing function analysis. Example 2 (Control of a stable process with backlash). A process with transfer function P 2 s 1 1 s
4

where Y and Yf are the Laplace transforms of the process output and the ltered process output, respectively. A second-order lter is used to guarantee high-frequency roll o in the controller, and the lter-time constant is Tf = Td/5. If a PI controller is used, it is suggested to use the lter-time constant Tf = Ti/10 [14]. The following two examples illustrate the problems caused by backlash in the feedback loop. Example 1 (Control of an integrating process with backlash). An integrating process with transfer function P 1 s 1 e0:2s s1 0:8s 4

is controlled by a PID controller of the form (2) with parameters K 1: 2; T i 2: 2; T d 1: 2: The controller parameters are derived using the MIGO design method [14]. A backlash of 5% (d = 0.05) is introduced in the control loop. Fig. 3 shows the Nyquist plot of the loop transfer function and the negative inverse describing function of the backlash. The curves do not intersect, which indicates that no limit cycle will occur. Fig. 5 shows the results of the simulations, where a setpoint change is made at t = 0 and a load disturbance is applied at the process input at t = 100. Furthermore, noise with a standard deviation of 1% is added to the process output. The gure shows that even though there is no limit cycle, as in the previous example, there is a severe deterioration of the control caused by the backlash. Because of the noise, the control error will never settle. The control signal has to pass the dead band every time the process input is to be reversed. This means that there will be low-frequency disturbances of the process output. The describing function analysis and the examples illustrate the control problems that arise when backlash is introduced in the control loop. Control loops where integrating processes are controlled with controllers having integral action will go into a limit-cycle oscillation. These oscillations will be detected by oscillation detection procedures [9].

is controlled by a PID controller of the form (2) with parameters K 1: 9; T i 2: 4; T d 0:67:

The controller parameters are derived using the MIGO design method [14]. A backlash of 5% (d = 0.05) is introduced in the control loop. Fig. 3 shows the Nyquist plot of the loop transfer function and the negative inverse describing function of the backlash. The curves intersect, which indicates that a limit cycle will occur. The describing function analysis predicts a limit cycle with an amplitude in the process output of 4.4% and an oscillation period of 7.7 s. Fig. 4 shows the results of the simulations, where a setpoint change is made at t = 0 and a load disturbance is applied at the process input at t = 100. The gure shows that the control loop oscillates. The amplitude of the process output is 3.2% and the oscillation period is 5.7 s.

0.6 0.5 0.4 0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

0.6 0.5 0.4 0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

0.4 0.5 0.2 0.4 0 0.3 0.2 0.2 0.1 0.4 0 20 40 60 80 100 120 140 160 180 200 0 0 20 40 60 80 100 120 140 160 180 200

time (s)
Fig. 4. Control of the integrating process P1 with 5% backlash.

time (s)
Fig. 5. Control of the stable process P2 with 5% backlash.

492

T. Ha gglund / Journal of Process Control 17 (2007) 489499

150

IAE

emax
50 40

100 30 20 50 10 0 0

10

10

Fig. 6. The percentage increase of the IAE value (left) and the peak error emax (right) at load disturbances caused by backlash in Example 2 for values of the backlash up to d = 10%. The solid line corresponds to a load change of 10%, and the dashed line to a load change of 20%.

Except for extremely lag-dominant processes, the control loops where the process is stable will normally not go into limit-cycle oscillations. However, the control performance is deteriorated even in these cases. This is illustrated in Fig. 6. The gure shows how the IAE value and the peak error emax at load disturbances increase when backlash appears in the control loop presented in Example 2. Both the IAE value and emax increase as the backlash d increases, even though emax is very noise sensitive. The increase is dependent on the magnitude of the load disturbances. The results agree well with those given in [3], where it is claimed that 10% backlash results in 100% increase in the IAE value and 50% increase in the peak error. The backlash introduces a dead time in the control loop. The length of this dead time is dependent on several states and parameters. The dead time appears only when the control signal action is reversed. The dead time is the time it takes for the control signal to pass the dead band. A low integral gain K/Ti gives a long dead time. The dead time becomes short if the control error is large. It means that the dead time is shorter for large load disturbances than for shorter. This explains the results shown in Fig. 6. Stable loops with backlash are normally not detected by oscillation detection procedures, since the oscillation amplitude is quite small. A new detection procedure for these processes is presented in Section 4. 3. Backlash compensation When it is discovered that a control valve has got so much backlash that the control is deteriorated, the best action to take is, of course, to replace or repair the valve. The fact that the amount of backlash normally increases with time makes this even more important. However, to replace or repair a valve means normally that the production has to be stopped. For this reason, and for the economical reason that it is of interest to use a valve for as long time as possible, it is of interest to compensate for the backlash. A control valve will normally not move on its own or when the control signal is constant unless the actuator is undersized or the positioner is unstable [15]. Therefore,

the position of the control signal with respect to the backlash is given by the control signal and its history. This means that the backlash is an invertible nonlinearity. An obvious way to compensate for the backlash is to make the control signal jump through the backlash every time the control action is reversed. The compensation can be seen as a feedforward compensation u uFB uFF ; 6

where u is the controller output, uFB is the feedback term, e.g. the output from the PID controller (2), and uFF is the term compensating for the backlash. An ideal backlash compensation would be   d du uFF sign : 7 2 dt This compensation is not realizable in a noisy environment. A possible modication is to lter the control signal before taking the derivative. It gives the compensation   d duf uFF sign ; 8 2 dt where uf is the ltered control signal. Note that the gain of the compensator is changed from the true backlash d to a value d where d 6 d. The ltering of the control signal will introduce a delay in the detection of the sign changes of the control signal rate. This means that the control signal has already started its way through the backlash when the rate change is detected. Therefore, the compensation must be smaller than in the ideal case. There are other possibilities to perform the backlash compensation. In (8), the control signal u has been passed trough a low-pass lter to reduce the noise introduced in the controller by the process output y. Inside the controller, the measurement signal is fed through a high-pass lter because of the derivative term. So, the noise level is rst amplied and then reduced by the low-pass lter. A more direct way is to base the feedforward on the measurement signal directly. One approach that will be used in this paper is d uFF signe; 2 9

T. Ha gglund / Journal of Process Control 17 (2007) 489499

493

where the control error is e = ysp yf and yf is the ltered process output given by (3). When the control error e changes sign, so does the rate of the integral term in the controller. Therefore, the feedforward (9) can be seen as an approach where only the noise-insensitive integral part of the controller is considered, and the noise-sensitive proportional and derivative parts are excluded from the compensation. The backlash compensation will now be illustrated for the two examples in the previous section. Example 3 (Backlash compensation for an integrating process). Consider the control problem in Example 1. A backlash compensator of the form (8) is added to the controller. The ltered control signal is generated as U f s 1 1 sT d =52 U s: 10

0.6 0.5 0.4

0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

0.5 0.4 0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

time (s)
Fig. 8. Control of the process P2 with 5% backlash (d = 0.05) and a backlash compensator with d = d = 0.05.

This is a lter with relatively high bandwidth. On the other hand, the process output is noise free in this example. Because of this, the gain of the compensator was chosen equal to the backlash, i.e. d = d = 0.05, so the compensator coincides with the ideal compensator (7). The results of the simulations are given in Fig. 7. Comparing Figs. 4 and 7 it is obvious that the backlash compensator gives an almost ideal compensation in this noisefree case. Example 4 (Backlash compensation for a stable process). Consider the control problem in Example 2. A backlash compensator of the form (9) is added to the controller. In this example, the compensation is more complicated than in the previous example since the process output is corrupted with noise. Fig. 8 shows the result when a compensator with d = d = 0.05 is used. It is obvious that the gain of the compensator is too high, and that the compensator causes

0.6 0.5 0.4

0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

0.5 0.4

0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

time (s)
Fig. 9. Control of the process P2 with 5% backlash (d = 0.05) and a backlash compensator with d = 0.04.

0.6 0.5 0.4

0.3 0.2 0.1 0 0 20 40 60 80 100 120 140 160 180 200

the loop to oscillate. Reducing the compensator gain to d = 0.4 gives the results shown in Fig. 9. This compensator gives a process output that is almost unaffected by the backlash. The control signal has some high-frequency shifts at certain periods. This could have been avoided by adjusting the ltering of the process output. On the other hand, these variations do not cause any valve movements because of the backlash. Fig. 6 shows that a backlash of 5% gives an increased IAE value of about 45% when the load changes 20%. With the compensator, this increase is reduced to about 15%.

y u

0.4

0.2

0.2

4. Backlash estimation
0.4 0 20 40 60 80 100 120 140 160 180 200

time (s)
Fig. 7. Control of the integrating process P1 with 5% backlash (d = 0.05) and a backlash compensator with d = d = 0.05.

Because of the reduction of personnel in process control plants, the times between the manual inspections of the control loops is often long. Therefore, it is of interest to

494

T. Ha gglund / Journal of Process Control 17 (2007) 489499

detect and estimate the amount of backlash automatically based normal operating data. Automatically means that no parameters or other guidance should have to be provided by the personnel. This feature is a prerequisite for the acceptance of the procedure in process control applications. Such a procedure is presented below. The procedure treats only stable process. As mentioned before, integrating processes with backlash will result in an oscillating control loop. These loops are detected by the procedure presented in [9]. Fig. 10 shows a part of the simulation given in Fig. 5. The process output y has been ltered through the lter (3). It means that the process output presented in Fig. 10 is the signal that enters the PID algorithm. The signals show the typical pattern obtained when a stable process is controlled by a controller having integral action and when there is backlash in the control loop. The process output remains a distance Dy from the setpoint while the control signal drifts through the dead band caused by the backlash. When the control signal has changed an amount Du, the process output is moved towards the setpoint. The time instances when the process output crosses the setpoint are marked in the gure. The time between these zero crossings is Dt = ti+1 ti. The change Du of the control signal is mainly caused by the integral part of the controller. It means that Z t i1 K K Du  jejdt Dy Dt; 11 T i ti Ti where Z Dy
ti

If the signals change slowly, the process dynamics can be neglected and the relation between the process output and the control signal is mainly determined by the static process gain Kp. The relation is Dy K p Dutrue ; 13

where Dutrue is the part of Du where the backlash is closed and the valve moves. It means that Du Dutrue d : 14

From Eqs. (11)(14) the following equation for estimating the backlash is obtained:   K Dy K 1 ^ Dt d Du Dutrue Dy Dt Dy : 15 Ti Kp Ti Kp The backlash estimator (15) assumes that the signals change slowly. A convenient way to check this is to see if Dt is long compared to the closed-loop time constant. In the examples presented later, estimation is only performed when Dt P 5Ti. The information required to determine the backlash online is the controller parameters K and Ti, and the static process gain Kp. Further, it is necessary to measure Dy from (12), i.e. to integrate the control error between zero crossings, and the time Dt between zero crossings. Note that it is not necessary to have access to the control signal u. It is a drawback that the process gain Kp is used in the algorithm, since this gain often is unknown. On the other ^ is quite insensitive to errors in the estihand, the estimate d mate of Kp. To see this, rewrite (15) to   ^ K Dt 1 Dy : d 16 T i KK p

ti1

jejdt=Dt;

12

see Fig. 10.

ti
0.4

ti+1

yf

0.39

t
0.38 35 40 45 50 55 60 65

0.4

0.35 35

40

45

50

55

60

65

time (t)
Fig. 10. Part of the simulation given in Fig. 5.

T. Ha gglund / Journal of Process Control 17 (2007) 489499

495

The rst term inside the brackets is always greater than 5, since estimation is only performed when Dt P 5Ti. For well-tuned controllers applied to processes that are not delay-dominated, the product KKp is normally larger than 0.5 [14]. Assuming these extreme values, i.e.   ^ K 5 2 Dy ; d 17 Kp a default value Kp = 1.5 will give estimation errors less than 20% as long as the true process gain is in the range 1 < Kp < 3.4. This is the case in most process control plants, since industrial controllers normally work with normalized signals. It is important that the noise does not cause zero-crossings. Therefore, the process output y is not only ltered by the second-order lter (3), but an additional second-order lter is applied before the signal is treated in the estimation procedure. In the examples presented in this paper, the time constant of this lter is Ti/2. On-line procedures like this must always have a security net [16,17]. The nal derivation of this security net must be developed during industrial eld tests. One element of this net has already been implemented. Load disturbances may obviously deteriorate the backlash estimation. To check that the process output has a form similar to the one in Fig. 10, estimation is only performed when emax < 2Dy, where emax is the absolute value of the largest control error in the interval [ti, ti+1]. A skeleton code describing the backlash estimator is given in Fig. 11. The estimation procedure can be used in several ways. First of all, it can be used as a detection procedure in a way similar to the oscillation detection procedure in [9]. In [9], the magnitudes of the IAE values between zero

crossings of the control error are studied, and it is concluded that an oscillation is present if the rate of large IAE values becomes high. In this paper, the control performance between zero crossings is also studied. Analogously to [9], it can be concluded that backlash is present in the loop if the rate of backlash detections gets high. If backlash is detected, and if the estimated backlash values are close, then one can also draw a conclusion about the amount of backlash. This is necessary if the goal is not only to detect backlash, but also to compensate for it. If there is stiction present in the control loop, the backlash estimated by the estimation procedure is the sum of the backlash and the dead band caused by stiction. This is probably a good feature, since the backlash compensator will then not only compensate for the backlash, but also for the stiction. The derivation of the backlash estimator is made assuming that a PID controller is used. However, it is straightforward to modify the method to other controllers having integral action.

5. Simulation example The detection and estimation procedure derived in the previous section will now be illustrated through simulations of the control loop presented in Example 2. Process P2 given by (5) is controlled with a PID controller tuned by the MIGO tuning rules. Measurement noise with a standard deviation of about 1% is added to the process output. The setpoint is changed at time t = 0 and a load disturbances at the process input is applied at time t = 100. Figs. 1215 show the results obtained for dierent amounts of backlash, varying from d = 1% to d = 10%.

Fig. 11. Skeleton code for the backlash estimator.

496
1

T. Ha gglund / Journal of Process Control 17 (2007) 489499

1

d = 1%
0.8 0.6 0.8

d = 5%
0.6 0.5 0.4 0.5 0.4 0.6

= d

= d

4.9

4.6

4.5

4.9

4.7

4.6

4.6

y
0.4 0.2 0 0.4 0.2 0 0 20 40 60 80 100 120 140 160 180 200 0

20

40

60

80

100

120

140

160

180

200

0.8

0.8

0.6

0.6

0.2

u
0 20 40 60 80 100 120 140 160 180 200

0.4

0.4

0.2

20

40

60

80

100

120

140

160

180

200

time (s)
Fig. 12. Backlash estimation applied to the control loop in Example 2 ^ are with the backlash d = 1%. The times of detection and the values of d indicated in the upper graph.

time (s)
Fig. 14. Backlash estimation applied to the control loop in Example 2 ^ are with the backlash d = 5%. The times of detection and the values of d indicated in the upper graph.

d = 3%
0.8 0.6

d = 10%
= d
0.8 2.8 2.6 2.6 2.5 2.6 2.6 2.5 0.6

= d

11.2

10

9.6

10.1

9.7

y
0.4 0.2 0 20 40 60 80 100 120 140 160 180 200 0 0

0.4 0.2 0

20

40

60

80

100

120

140

160

180

200

0.8 0.8 0.6 0.6 0.4

u
0.2 0 0 20 40 60 80 100 120 140 160 180 200

0.4

0.2

time (s)
Fig. 13. Backlash estimation applied to the control loop in Example 2 ^ are with the backlash d = 3%. The times of detection and the values of d indicated in the upper graph.

20

40

60

80

100

120

140

160

180

200

time (s)
Fig. 15. Backlash estimation applied to the control loop in Example 2 ^ are with the backlash d = 10%. The times of detection and the values of d indicated in the upper graph.

The times of detection are marked in the graphs, as well as ^. the values of the backlash estimates d The gures show that the deterioration of the control increases as the amount of backlash increases. During the simulation time of 200 s, between ve and seven detections are made in each case. The accuracy of the backlash estimates increases as the backlash increases. The estimated ^ is in most cases smaller than the true backlash backlash d d. This is a good property, since it is important that the backlash compensator does not use a gain that is too high. The simulation experiments show that the estimation procedure works well. Even for a backlash as small as d = 1%, where the eects in the process output are hardly noticeable because of the noise level, the procedure manages to detect the backlash and to provide fairly accurate estimates.

6. Industrial tests The backlash estimation procedure has been tested on a ow control loop in a paper mill. The process section is a pipe where pulp is transported from a recycling-pulp tower to a tank. See the P&I diagram in Fig. 16. A PID controller (FIC) controls the pulp ow through a valve. The process output y is the pulp ow, measured in the range 0900 m3/h, and the controller output u is in the range 0100%. The setpoint is external, and is given by a controller (LIC) that controls the tank level downstream. This means that the ow controller is a slave controller in a cascade conguration. The pulp ow is driven by a pump which is controlled by a pressure controller (PIC). The ow and

T. Ha gglund / Journal of Process Control 17 (2007) 489499

497

LIC Pulp tower PIC PT FIC FT

LT Tank

Fig. 16. Process and instrumentation diagram of the pulp ow section.

Flow [0-900m3 / h]
450

440

430

420

410

50

100

150

200

250

Control signal [0-100%]

40 38

36

34

32

50

100

150

200

250

time (s)
Fig. 17. Manual test to check the amount of backlash in the valve. The backlash is estimated to d = 3%.

the pressure controllers interact. To reduce this interaction, the bandwidth of the ow loop, which normally is quite fast, has been reduced by introducing a low-pass lter with a time constant of 20 s in the loop. A manual test was performed to check the amount of backlash in the valve. The result is shown in Fig. 17. The ow controller output is rst increased to ensure that the gap is closed. Since the ow increases, the gap is closed when the control signal is at the nal value u = 39%. The controller output is then reversed and decreased in steps of 1%. The rst steps do not result in any ow decrease, indicating that the control signal is inside the dead band. However, the step made from the value u = 36% gives a ow decrease, showing that the gap is closed near this value of the control signal. The test shows that the backlash is around d = 3%. The ow controller is a PI controller with parameters K = 0.6 and Ti = 28 s. The signals used in the controller are normalized to the range [0,1]. The static process gain was estimated to Kp = 1.3. Fig. 18 shows the result of a recording made for about 4000 s. The loop is oscillating

because of the oscillating setpoint. The setpoint oscillations are probably generated by the ow variations caused by the backlash. Fig. 18 shows that backlash was detected ve times during the test, with backlash estimates ranging from 2.5% to 3.1%. These estimates are close to those obtained from the manual tests in Fig. 17. Setpoint variations may disturb the backlash estimator. If the test emax < 2Dy were not present, fteen detections would have been obtained during the test, and especially the last two major setpoint changes would have given backlash estimates that are far too large. To get rid of the disturbances caused by the external setpoint, experiments with a xed internal setpoint were also performed. Fig. 19 shows the results of such a test. Comparing Figs. 18 and 19, it is obvious that the oscillations caused by the external setpoint have disappeared. Fig. 19 shows that there are some low-frequency disturbances present. They are probably caused by interaction from the pressure control loop. These disturbances are more dicult to handle than the high-frequency noise in the simulation experiments in the previous sections. Three backlash detections were made during the experiments, with the estimates 1.6%, 2.4%, and 2.4%, respectively. These values are slightly lower than those obtained in Fig. 18. This is expected, since the setpoint variations in the previous example amplies the eect of the backlash. Furthermore, already the derivation of the method and the simulation examples have shown that the backlash estimates are expected to be conservative. To summarise, the industrial tests show that the backlash estimation procedure works also in an industrial environment with dicult low-frequency disturbances. To obtain a robust procedure that is automatic in the sense that no user interaction is needed, more industrial eld tests have to be performed, and it is likely that the supervisory net must be extended.

7. Conclusions Stiction and backlash in control valves are the major problem at the loop level in process control plants. There are two aspects of the problem. First of all, the nonlinear-

498

T. Ha gglund / Journal of Process Control 17 (2007) 489499

0.5

Normalized flow
2.7 3.1 2.9 2.9

0.48 d = 2.5
0.46 0.44 0.42 0.4 0.38 0 500

1000

1500

2000

2500

3000

3500

4000

Normalized control signal

0.32 0.3 0.28 0.26 0.24 0 500 1000 1500 2000 2500 3000 3500 4000

time (s)
Fig. 18. Backlash estimation applied to the pulp ow control loop. The upper graph shows the external setpoint (noisy signal) and the process output ^ are indicated in the upper graph. (ow). The lower graph shows the control signal. The estimated values of d

0.43

Normalized flow

0.42

= d

1.6

2.4

2.4

0.41

0.4

0.39

500

1000

1500

2000

2500

3000

Normalized control signal

0.28

0.27

0.26

0.25 0 500 1000 1500 2000 2500 3000

time (s)
Fig. 19. Backlash estimation applied to the pulp ow control loop. The upper graph shows the constant internal setpoint and the process output (ow). ^ are indicated in the upper graph. The lower graph shows the control signal. The estimated values of d

T. Ha gglund / Journal of Process Control 17 (2007) 489499

499

ities deteriorate the control performance. Furthermore, the loops facing these problems often remain undiscovered by the personnel in process control plants. Procedures for stiction detection and for stiction compensation have been available for over a decade, and they are used in many industrial plants today. Compensation for backlash is simple, but these procedures are seldom used in process control plants. The major reason for this is that no backlash detection and backlash on-line estimation procedure have been presented. This paper presents an on-line procedure for detection and estimation of backlash in control loops. It is given by Eq. (15) and some further details are summarized in Fig. 11. The method is automatic in the sense that no information has to be provided from the user. The only information needed except for the signals are the controller parameters. The eectiveness of the method has been demonstrated through simulations and industrial eld tests. The method is patent pending. References
[1] D.B. Ender, Process control performance: not as good as you think, Control Engineering 40 (10) (1993) 180190. [2] W.L. Bialkowski, Dream vs. reality a view from both sides of the gap, Pulp and Paper Canada 11 (1994) 1927. [3] G. McMillan, Valve response control, improvement, Encyclopedia of Chemical Processing and Design 61 (1997) 99113. [4] T. Ha gglund, A friction compensator for pneumatic control valves, Journal of Process Control 12 (2002) 897904.

[5] S. Qin, Control performance monitoring a review and assessment, Computers and Chemical Engineering 23 (1998) 173186. [6] T.J. Harris, C.T. Seppala, L.D. Desborough, A review of performance monitoring and assessment techniques for univariate and multivariate control systems, Journal of Process Control 9 (1) (1999) 117. [7] M.A. Paulonis, J.W. Cox, A practical approach for large-scale controller performance assessment, diagnosis, and improvement, Journal of Process Control 13 (2003) 155168. [8] M. Jelali, An overview of control performance assessment technology and industrial applications, Control Engineering Practice 14 (5) (2006) 441466. [9] T. Ha gglund, A control-loop performance monitor, Control Engineering Practice 3 (1995) 15431551. [10] A. Horch, A simple method for detection of stiction in control valves, Control Engineering Practice 7 (10) (1999) 12211231. [11] N.F. Thornhill, T. Ha gglund, Detection and diagnosis of oscillation in control loops, Control Engineering Practice 5 (1997) 1343 1354. [12] A. Singhal, T. Salsbury, A simple method for detecting valve stiction in oscillating control loops, Journal of Process Control 15 (4) (2005) 371382. [13] M. Vidyasagar, Nonlinear systems analysis, second ed., Prentice Hall, 1993. stro [14] K.J. A m, T. Ha gglund, Advanced PID Control. ISA The Instrumentation, Systems, and Automation Society, Research Triangle Park, NC 27709, 2005. [15] T.L. Blevins, G.K. McMillan, W.K. Wojsznis, M.W. Brown, Advanced Control Unleashed. ISA, Research Triangle Park, NC, 2003. [16] T. Ha gglund, Industrial implementation of on-line performance monitoring tools, Control Engineering Practice 13 (2005) 13831390. stro [17] T. Ha gglund, K.J. A m, Supervision of adaptive control algorithms, Automatica 36 (2000) 11711180.