You are on page 1of 15

Robust Tuning Tuning of of PI and PID Controllers

USING DERIVATIVE ACTION DESPITE SENSOR NOISE


By BIRGITTA KRISTIANSSON and BENGT LENNARTSON

T
1066-033X/06/$20.002006IEEE

oday, proportional-integral (PI) and proportional-integral derivative (PID) type controllers are by far the most widely used regulators. Historical remnants show that proportional control (P type) was used by the ancient Greeks and Babylonians for water-level regulation more than 2,000 years ago [1], [2]. Integral action (I) was added in the beginning of the 19th century [3]. The first commercial controller with derivative action (D), the Fulscope, was announced by Taylor Instrument Companies in 1939. In the 1930s, academics began to study automatic control, and the design and tuning of PI and PID controllers has since been the focus of a great deal of research [4][6], resulting in numerous design strategies. In light of these developments, one may ask whether there is any further need for additional tuning rules. According to [5], such a demand still exists. Unfortunately, ideas presented by researchers over the years have reached end users only to a limited extent. Several studies (for example, [7][10]) have shown that most process controllers installed in industry are poorly tuned, resulting in lack-of-control costs. For instance, poor control of a regularly sized paper mill has been observed to cost millions of dollars [9]. Our hope is that the tuning rules presented in this article will help everyday users of PI and PID controllers avoid poorly tuned control loops. The strategy given in this article includes a lowpass filter together with derivative action. Provided with a suitably tuned filter, a PID controller can reject process disturbances without too much control action and sensitivity

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 55

IMAGESTATE

to sensor noise. Our rules are applicable to many kinds of plants, including stable plants with all poles on the negative real axis, as well as plants with integral action. The resulting closed-loop systems offer good midfrequency (MF) robustness combined with the best possible tradeoff between performance and control activity. The tuning rules introduced by this study demand moderate plant knowledge, while providing freedom for users to make tradeoffs between such factors as stability margin, output performance, and control cost. In almost every situation, derivative action with a lowpass filter improves the properties of the controlled system compared to PI control. We show that designing a PID controller with a filter is no more difficult than designing an ordinary PI controller. In many design methods, not all of the available degrees of freedom are utilized. For example, with no motivation, the ratio between integral and derivative time constants is often fixed to four [11][13], and the filter time constant is set to a tenth of the derivative time constant [14][16]. Experience in optimizing PID controllers for a variety of plants with all design parameters free shows that the best tradeoffs among performance, robustness, and control activity often result in complex controller zeros and a filter time constant that is larg-

er than that which is usually recommended [17]. A reformulation of the classical PID controller for design purposes is thus motivated. This article also investigates the performance sensitivity related to various controller parameters, in particular, the influence of the damping factor and the time constant of the controller zeros. When plants experience uncertain time delays or unmodeled high-frequency (HF) resonances, or when measurement noise is severe, the controller must have low gain in the HF range. This requirement can be met by introducing rolloff in the controller. A PI or, preferably, a PID controller equipped with additional lowpass filtering (F), then becomes an attractive alternative. The resulting PIDF controller with five tuning parameters can also be tuned according to the rules presented in this article. Aside from the Ziegler and Nichols tuning rules [11], the most commonly used tuning method is probably the approach based on internal model control (IMC) [18]. When IMC is used, the identified or derived plant model must be reduced to first or second order, respectively, before the PI or PID design rules can be applied. This reduction can be accomplished in various ways [19], [20], and the choice of reduction method greatly

Notation
a b Ci e(t ) G(s) Gm GMS HF IAE JHF Ju Juec Jv K K (s), Kx (s) ki , Ki Kp k L(s) Ld LF m MF MHF MS MT Ratio Ti / Td Ratio Td / Tf Arbitrary constants, i = 1, 2, . . . Control error Plant model transfer function Gain margin Generalized maximum sensitivity, stability margin High frequency Integrated absolute error High-frequency criterion Control activity criterion Economic limit for Ju Performance criterion Static gain Controller transfer function, x = PI, PID, PIF, or PIDF Controller integral gain Proportional controller gain Controller high-frequency gain Loop transfer function Delay time Low frequency Rolloff rate Midfrequency Mid to high frequency Infinity norm of the sensitivity function Infinity norm of the complementary sensitivity function m 150G , 180G r (t ) S (s) Su (s) Sv (s) T T (s) Td Teq Tf Ti T63 u (t ) v (t ) w (t ) y (t ) z f , 150 , i150 Reference signal Sensitivity function Control sensitivity function Disturbance sensitivity function Time constant Complementary sensitivity function Derivative time constant Equivalent time constant (T63 Ld ) Filter time constant Integral time constant The time for a step response to reach 63% of its steady-state value Control signal Process (load) disturbance signal Sensor noise Controlled output Plant model zero MS /MT Filter constant k /( ki ) Controller zero damping Controller pole damping Parameter characterizing the dynamics of a plant model Phase margin Vector of tuning parameters Frequency at which the phase lag of G(s) is 150 or 180 , respectively

56 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

influences the resulting design. A comparison between the main tuning rules presented in this article and other tuning methods, including different versions of IMC, is given in [21]. Since system properties are not independent, improvement of the control in one respect often implies deterioration in another. A fair method for comparing two controllers must, accordingly, ensure that all aspects not immediately compared are equally restricted during the comparison. Such an evaluation method, which was presented in [17], [22], and [23], is used in this article to compare tuning rules. This evaluation method is described in the next section.

v r + K(s) u + + G(s) y

+ -

FIGURE 1 Closed-loop SISO system. The plant G(s) is controlled by the controller K (s). The system has three input signals, namely, the reference signal r (t ), the process disturbance V (t ), and the sensor noise w (t ), as well as two output signals, namely, the controlled output y (t ) and the control signal u (t ).

EVALUATION OF CONTROLLERS
We present a systematic procedure for evaluating control structures and design parameters. Our procedure is based on four criteria representing low-frequency (LF), MF, mid- to high-frequency (MHF), and HF performance and robustness properties. By evaluating these criteria, the tradeoff between, for example, output performance and control activity can be investigated while passband robustness (stability margin) is kept constant. Another objective is to keep the MF, MHF, and HF criteria equal, or at least equally bounded, to evaluate the improvement in LF performance in an objective manner. The introduced evaluation criteria, which are based on H -norms, also facilitate comparisons of discrete-time and multi-input, multi-output controllers. For further details, see [21] and [23]. k i , s 0, s K( s) k , s , sm (1)

where ki is the integral gain, k is the HF gain, and m is the rolloff rate of the controller. The performance criteria, which are based on the sensitivity functions, are Performance criterion Jv = 1 G yv ( s) s = 1 Sv ( s) s ,

Passband robustness criterion

Evaluation Criteria
Consider the single-input, single-output (SISO) system in Figure 1, where the plant G( s) is controlled by the controller K( s). The closed-loop system has three inputs, the reference signal r( t), the process disturbance v( t), and the sensor noise w( t). Relevant outputs are the controlled output y( t), the control signal u( t), and the control error e( t) = r( t) y( t) . The loop transfer function is L( s) = G( s) K( s), while the sensitivity functions are Sensitivity function S( s) = 1 = Ger( s) , 1 + L( s) where ||.|| denotes the H -norm. The criterion Jv is a measure of the systems ability to handle LF load disturbances. This criterion is a frequency domain alternative to criteria based on an integral function of the error signal, for example, IAE = 0 |e( t)| dt [4], [24]. For servo problems, a more relevant criterion is obtained by replacing G yv ( s) with Ger( s). The passband robustness criterion given by generalized maximum sensitivity GMS is a combination of the constraints S = max |S( )| MS and T MT . The parameter = MS /MT transforms the restriction on ||T|| to a similar restriction on ||S|| . The maximum S of the sensitivity function is often used as a robustness measure [4], [25], [26] since S is equal to the inverse of the minimum distance from the loop transfer function to the critical point (1, 0) in the Nyquist plot. The components of GMS correspond to two circles in the complex plane given by ||S|| = MS and ||T|| = MT , from GMS = max( S( s) Control activity criterion Ju = Gur( s) HF criterion JHF = sm Gur( s)
,

T( s)

),

= Guw ( s)

= Su ( s)

= sm Guw ( s)

= sm Su ( s)

Complementary sensitivity function T( s) = L( s) = G yr( s) = G yw ( s), 1 + L( s)

Disturbance sensitivity function Sv ( s) = G( s) = G yv ( s) , 1 + L( s)

Control sensitivity function Su ( s) = K( s) = Gur( s) = Guw ( s) . 1 + L( s)

The controller K( s) can be either strictly proper or exactly proper. Including integral action in K( s) implies the asymptotic properties

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 57

which the Nyquist plot of L( ) is excluded to enforce a specified stability margin. In Figure 2, these circles are shown for the default values, specifically, MS = 1.7 and MT = 1.3. These values represent an empirical compromise. A larger value of MS implies a faster response with a larger overshoot in the reference step response and an undershoot in the disturbance step response; a smaller value implies a slower system with a larger IAE. The chosen value of MT guarantees a minimum phase margin of 45 . A criterion similar to GMS is formulated in [27]. The combined measure GMS replaces the two classical measures, phase margin PM and gain margin GM, which are commonly used to characterize MF robustness [28], [29]. The criterion Ju is related to the MHF region, which is close to and slightly above the closed-loop bandwidth. The control sensitivity function most often reaches its maximum in this range. In the HF region, two requirements are paramount, namely robustness toward model uncertainties and reduction of the effects of sensor noise. According to the small gain theorem [30], plants with significant model uncertainty can be rendered closed-loop stable by keeping the complementary sensitivity function T( s) small. Since T( s) = G( s) K( s)S( s) and, consequently, T( s)G( s)1 = Su ( s) = Gur( s), keeping Su ( s) small also keeps T( s) small. Since Su = Guw , this restriction also reduces the influence of sensor noise on the control signal. These facts motivate the HF criterion JHF = || sm Su ( s)|| . For high frequencies, |S( )| 1 , which implies |Su ( )| | K( s)| k m according to (1). Hence JHF , as well as Jv , is almost independent of the nominal plant model. When m = 0, which is valid for most PI and PID controllers without extra filters, JHF = Ju .

in a vector ) are modified to minimize the fourth criterion. For example, the LF performance is evaluated by solving the constrained optimization problem min Jv (),

GMS () C1 ,

Ju () C2 ,

JHF () C3 ,

(2)

where C1 , C2 , and C3 are constants. The default value of C1 is 1.7, while the values of C2 and C3 can vary. The last restriction in (2) is relevant only for strictly proper controllers. By this optimization procedure, completely different controllers can be compared under equal conditions. This type of evaluation method is discussed in [31]. Similar ideas with different criteria and constraints are presented in [32] and [33]. Alternatively, the loop-shaping strategy in [34] and [35] focuses on L( ), assuming complete knowledge of the frequency response of the plant. The expression optimal controller refers to a controller optimized according to (2), with all available parameters included in the tuning vector . We used the MATLAB Optimization Toolbox for the computations.

CONTROLLER AND PLANT MODELS


Before evaluating the closed-loop properties by the above criteria, we discuss the parameterization of PI and PID controllers as well as the dynamic characteristics of the plant models included in this investigation.

Parameterization of Controllers
There are alternative ways to formulate and parameterize the transfer function of a PID controller. We consider one-degreeof-freedom controllers, as shown in Figure 1. Our goal is to design PID controllers to reject process disturbances, as in most applications [15], [36]. In certain cases, such as in model predictive systems, where the PID controllers serve at lower levels in hierarchical structures with more sophisticated controllers at higher levels, or when good servo properties are demanded, the controller can be augmented by a filter in the feedforward path. The traditional PID controller KPID ( s) = Kp + K i/ s + K d s, with the parameters proportional gain Kp, integral gain K i, and derivative gain K d, has the drawback of being improper. To limit the HF gain, the PID controller is often augmented by a lowpass filter on the derivative part, and may then be formulated with a proportional gain Kp, integral time Ti, derivative time Td, and filter time constant T f as KPID ( s) = Kp 1 + 1 sTd . + sTi 1 + sT f (3)

Evaluation Procedure
An objective evaluation of different controllers is obtained when three of the four criteria are kept equal or at least bounded from above, while the tuning parameters (collected

2 MT 1 MS 0

1 |L( j )| 2 4 3 2 1 0

FIGURE 2 Definition of generalized maximum sensitivity GMS . When a system satisfies a given value of GMS , the Nyquist plot of its loop transfer function L( ) does not enter into the MS circle (default MS = 1.7) and the MT circle (default MT = 1.3).

This formulation has the advantage that a change in the gain margin requires adjustment of only Kp. With a = Ti/ Td and b = Td / T f, a double zero in the controller corresponds to a = 4 when b = , and a slightly higher value of a when b < (T f > 0). In [22] and [23], it is shown that a PID controller, optimized with the filter included and all parameters free, often has complex zeros. Consequently, we reformulate the PID controller with a first-order, lowpass filter as

58 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

Derivative action can be used without high sensitivity to sensor noise, thanks to inclusion of the lowpass filter in the design procedure.

KPID ( s) = ki

1 + 2 s + ( s)2 . s(1 + s/)

(4)

drops slowly at increasing frequency). In these cases, can be modified as 150 = |G( 150 G )| , G(0)

The four tuning parameters are damping and time constant of the controller zeros, integral gain ki , and HF gain k = KPID (). The variable = k /( ki) in (4) is used for convenience. Moreover, in most cases the zeros can be fixed by simple and general rules, such as (9) given below. The user is then left with just two intuitively understood parameters, k (essentially control activity), which can be given a suitable value, and ki (performance), which can be increased until the overshoot (damping) becomes unacceptable. Translating from the parameters in (4) into the traditional parameters in (3) is given by Tf = ,

where 150 G represents the frequency at which the plant has a phase lag of 150 [21]. In this article, we use 150 together with 150 G and G(0) to characterize plant dynamics and formulate uncomplicated tuning rules. These plant characteristics can be found in practice by means of a relay experiment with hysteresis [12] and by investigating the step response. For plants with integral action, should be modified as i150 = 150 G |G( 150 G )| . lim0 |G( )|

(2 1), 2 2 1 , Tf = Td = Ti 2 1 K = ki Ti = ki (2 1). Ti = 2 T f = With > 1 and 1, the PID controller (4) can be reformulated as a lead/lag compensator, that is, a serial PID controller. With 0 < < 1, the PID controller has complex zeros. Finally, for = 1 and = 1, the PID controller becomes the PI controller KPI ( s) = ki 1 + s . s (5)

For details, see [17] and [37]. For most integral plants, i150 lies in the interval [0.6, 1]. For illustration, we reconsider the models used in [17], which are suggested as standard benchmark models for testing PID controllers according to groups 15 in [38]. Among these plants are stable, nonoscillating plants (such as those with time delay or nonminimum phase zeros), plants of high and low order (including plants with multiple poles), as well as the same plants with an additional pole at the origin. Table 1 illustrates examples of transfer functions for common plant models and the corresponding values of 150 or i150 .

TUNING PI AND PID CONTROLLERS FOR STABLE PLANTS


Examples show that control systems for stable, nonoscillating plants have common characteristics. This observation is the basis for formulating general tuning rules. Accordingly, the rules at the end of this section are intended for plants whose poles are all on the negative real axis.

Description of Plant Dynamics


To obtain useful tuning rules for PI and PID controllers, the requirements for plant knowledge must be moderate. The parameter is used in [37] to capture the dynamics of a controlled process. For stable plants, this parameter is defined as |G( 180 G )| , = G(0) where 180 G is the frequency at which the plant has a phase lag of 180 , which normally gives values of in the interval [0, 1]. For higher values of , the plant is more complex and harder to control. Note that is the inverse of the gain margin that a P controller gives with Kp = 1/ G(0). However, there are stable plants for which neither nor 180 G is defined, since the plant phase never reaches 180 , as well as plants with values of that are less than 0.1 (the phase

Properties of Optimally Tuned PID Controllers


The PID controller with a first-order, lowpass filter (4) has four parameters to tune, namely, the HF gain k , the zero parameters and , and the integral gain ki . Sometimes the HF gain k is replaced by = k /(ki ) . Our goal is to determine the tradeoff between output performance and control cost while guaranteeing a required stability margin.

HF Gain k
Figure 3 illustrates the tradeoff between improved performance (decreased Jv ) and the cost of higher control activity (increased Ju ) when the MF robustness GMS is fixed. For an

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 59

optimal PID controller, the Jv / Ju curve tends to flatten as Ju grows, as in Figure 3. In fact, for nonminimum phase plants, there is a theoretical lower limit for Jv . Although the user has full freedom to choose the level of control activity Ju , there is no advantage in increasing Ju above a certain level Juec , where the benefit of decreased Jv is insignificant [17]. More complex dynamics (higher values of ) imply lower Juec . The tuning rules are formulated to achieve control activities close to Juec . For the PID controller, Ju is typically equal to k , which means that the required level of Ju determines the HF gain k .

sluggish response to a reference step, while too-high damping implies a somewhat sluggish response to a disturbance step and an undershoot in the response to a reference step.

Zero Time Constant


The optimal zero time constant depends on the plant dynamics. In fact, after normalizing the time scale of the plant, 1/ increases almost linearly along with 150 . However, is often proportional to T63 , the length of time that a step response takes to reach 63% of its final value. Figure 5 shows that, for plant models with 150 0.09, a reasonable choice is = T63 /3 . This figure also illustrates that plants with firstorder dynamics behave otherwise and consequently must be treated separately. Furthermore, the optimal value of is largely independent J of u , with the exception of small values of Ju , as seen by comparing the constant value = T63 /3 0.8 with the optimal solution in Figure 6(a). Likewise, Figure 7 shows that the optimal value of increases slightly when Ju = 5. This value also marginally increases with decreasing demand on GMS .

Damping Factor
In all cases considered here, the optimal PID controller for a stable, nonoscillating plant has a pair of complex zeros. The damping factor has no clear dependence either on the plant dynamics, as measured by 150 , or on GMS or Ju . When Ju is chosen near Juec , the optimal value of normally ranges from 0.70 to 0.85. Similarly, as in Figure 4, = 0.75 is often a good choice. Damping that is too low, however, implies undershoot in the response to a disturbance step as in Figure 4(b) and a

TABLE 1 Typical plant models and their values of 150 , or, in the case of plant integral action, values of i150 . The parameter 150 , or alternatively i150 , which is usually between zero and one, is used to characterize the dynamics of a plant. The higher the value, the more difficult the plant is to control. This parameter can be used to formulate simple design rules.

G(s)
1 (1 + 0.2s )(1 + s ) 1 (1 + s )2 e 0 .1 1+s 1 0 .1 s (1 + s )2 e 0 .3 s 1+s e 0 .3 s (1 + s )2 1 (1 + s )3 1 (1 + s )(1 + 0.7s )(1 + 0.72 s )(1 + 0.73 s ) 1 0.25s (1 + s )3 1 (1 + s )4 e s (1 + s )2 e s (1 + s )3 e s (1 + s ) 1 (1 + s )8 1 2s (1 + s )4 e s 1 + 0 .1 s

150
0.038 0.067 0.088 0.141 0.228 0.249 0.266 0.352 0.359 0.395 0.488 0.527 0.528 0.646 0.907 0.973

G(s)
1 s (1 + s ) e 0:3s s (1 + s ) 1 s (1 + s )(1 + 0.2s ) 1 s (1 + s )2 1 0 .1 s s (1 + s )2 e 0 .3 s s (1 + s )2 1 s (1 + s )(1 + 0.7s )(1 + 0.72 s )(1 + 0.73 s ) 1 0.25s s (1 + s )3 1 s (1 + s )4 e s s (1 + s ) e s s (1 + s )3 e s s ( 1 + 0 .1 s )

i150
0.500 0.643 0.657 0.749 0.774 0.810 0.852 0.857 0.870 0.877 0.902 0.997

60 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

Easily understood and applied tuning rules can give close-to-optimal PI and PID controllers.

Figure 6(b) illustrates that high values of imply slowly decaying responses to step disturbances, while low values imply undershoot. Figure 7(a) illustrates a minimum value of Jv with respect to . When this minimum is not reached, a value of that is marginally higher than the optimal value is preferable to a lower value. Figure 7 also confirms that the economic limit of Ju for this plant ranges between 10 and 15, while the derivative action in the PID controller decreases Jv by a factor of two compared to PI control.

Filter Factor
A higher filter factor usually means increased control activity Ju , as shown by the optimal curve in Figure 6(a). The optimal value of , when Ju Juec , varies according to 150 , although for many common plants the optimal value of is significantly lower than the recommended b = 10 corresponding to 20 [21]. For low-order and minimum-phase plants, the optimal value of corresponding to Juec may be as low as = 2 (see Figure 8). The performance Jv for such plants can be improved by using a higher value of but at the cost of high control activity Ju . Recall that a PID controller with = 1 and = 1 constitutes a PI controller, which shows the limitation on control activity in the PI case.

tionship for different values of . Figure 4 shows the best possible Jv / Ju combination with fixed controller time ratios a = Ti/ Td = 4 as suggested in [11] and [12] and b = Td / T f = 10, as recommended in [14] and [16]. With two parameter values specified, only two parameters remain to be optimized. Fixed a = 4 produces minor deterioration compared to optimal values, which for most stable plants is approximately a = 2.5. However, the often used value b = 10 is, in most cases, too high by a factor of two to three. As Figure 4 shows, the output performance Jv can be significantly improved with constrained control activity Ju and required MF robustness GMS by optimizing all accessible parameters, including the filter time constant.

4 Jv

12

J u 16

Integral Gain ki
Since ki 1/ Jv , the optimal integral gain ki increases with increasing Ju (see Figures 4 and 6). However, the optimal value of Jv is higher for more complex plants, which implies that ki decreases when 150 2 Jv increases. The optimal value of a=4, b=10 ki also decreases when a larger 1.5 stability margin (lower GMS ) is demanded. When the con=0.5 1 troller parameters , , and 1.0 are given by tuning rules, the 0.5 integral gain ki is also the para0.75 meter best suited for manual 0 tuning of the desired tradeoff 5 10 between rise time (quickness) and damping (robustness) of (a) the step response.
FIGURE 3 A typical relationship between Jv and Ju for fixed GMS . The curve has a tendency to flatten as Ju increases.

0.5 y 0.3 a=4, b=10 =1.5 1.25 1.0 0.75

0.1 0 15 Ju 20 0.1 0.5 0 5

10 (b)

15

Traditional Tuning, a = 4, b = 10
It is desirable to view a comparison of the optimal Jv / Ju rela-

FIGURE 4 (a) The dependence of the Jv /Ju tradeoff on the damping of the PID controller zeros. When and ki are optimized, the Jv /Ju relation is not greatly influenced by , as long as remains around 0.8. (b) The corresponding responses to disturbance steps when Ju = 10. The curves marked with a = 4, b = 10 represent tuning with a and b fixed, while Kp and Ti in (3) are optimized. The plant model is G(s) = e0.3s /(1 + s)2 .

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 61

Tuning Rules for PID Controllers


Based on the above discussion, we now present three sets of tuning rules for PID controllers. These rules are intended for stable, nonoscillating plants.

Tuning rules for stable nonoscillating plants of at least the second order 1 = 150 G (0.44 + 0.86150 ), 1 2 , 25 , min 3 + k = G(0) 150 G 150 G 0.45 ki = 0.1 . G(0) 150 + 0.07 = 0.75, An alternative to the expression for k is = 2 + 14150 . (7)

Tuning Rules Based on 150


When a relay experiment involving hysteresis [12] is used to obtain the required plant information, the controlled self-oscillating frequency is not 180 G but rather is somewhat lower, typically around 150 G . The first tuning method, which requires knowledge of 150 and 150 G , is suitable for plants with poles (at least two) on the negative real axis. Recommended parameter settings for a PID controller (4) are

(6)

0.4

/T63

0.2

0.5

150

FIGURE 5 The dependence of / T63 on 150 . The correspondence between rules (8) and (9) (solid) and optimal outcomes (+ ) for plants of order two or higher is good. The circles (o) show optimal outcomes from first-order plants with small time delays, which significantly differ from the outcomes for plants of higher order and thus must be treated separately.

The HF gain k is chosen to keep Ju as close to Juec as possible. Figures 5 and 8 illustrate these rules as well as optimal outcomes for a set of plant models with values of 150 between 0.04 and 0.91. When all four controller parameters are tuned according to (6), GMS is in the range of 1.651.85 for most of the plants tested, generally near 1.7. The resulting value of Jv differs from optimal values at corresponding values of Ju and GMS , in most cases by less than 5%. The only exceptions are plants with extremely long delay times Ld relative to time constants Teq (Ld / Teq > 4) leading to 150 > 0.95, as well as plants with stable zeros close to the imaginary axis 1 ). Precise comparisons between the optimal solu(|z| < T63 tion and the tuning rule (6) require that GMS be kept at exactly the same level, which is impossible with a simple tuning rule for ki . The parameter Ld is an equivalent delay time, defined as the time it takes for the step response to reach 5% of its final value, while Teq = T63 Ld. A good rule of thumb is that 150 0.5Ld / Teq (see Figure 9). For plants with 150 < 0.05, k tends to be unnecessarily high, a consequence of the tendency

2 Jv 1.5 y

0.6

=0.4
0.4 0.6 0.8

= 0.6
0.7 T63 /3 1.0

0.2

1.0 1.2

0.5 =3

0 5 Optimal

8 15

10 20 0 5 10 (b) 15 20

10 (a)

Ju

FIGURE 6 Optimal Jv /Ju relations and load disturbance step responses for different PID controller-zero time constants . The difference in performance between optimal (varying with Ju ) and constant = T63 /3 0.8 is marginal except for small values of Ju . High values of yield slowly decaying disturbance step responses, while low values imply undershoot. In (a) and (b), the plant is G(s) = e0.3s /(1 + s)2 with = 0.75. In (b), Ju = 10. Obviously, a value of that is too high is a better choice than a value that is too low.

62 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

of the optimal Jv 1/ ki to be extremely small. Therefore, the maximum value k = 25 is used in (6). The rules (6) are recommended for autotuning [4], as well as for finding appropriate initial values for optimization. The placement of the controller zeros is not as crucial as the tradeoff between ki and k in attending the specified value of GMS . Thus, one option is to use (6) to select only and . Thereafter, a suitable control activity k is chosen (Ju k ), with k according to (6) as a first choice. Finally, the integral gain ki is manually tuned until the desired tradeoff between rise time (output performance) and damping of the step response (robustness) is achieved. Both good performance and low control activity (small k ) can be obtained. The low control activity is a consequence of including the lowpass filter in the controller design.

= T63 (2.65150 + 0.11).

(8)

Given a step response, the controller zeros of a PID controller can be chosen as Extremely simple tuning rules for PID controller zeros = 0.35T63 T63 /3, = 0.75. (9)

Tuning Rules Based on T63


An alternative tuning procedure is based on the step response. Figure 5 shows that, except for plants with 150 < 0.09 , approximately one-third of T63 serves as an appropriate value for . For plant models with 150 < 0.09, the optimal value of decreases as

The remaining parameters k and ki can be used to tune the tradeoff between output performance and control activity. Start with between five and ten, noting that the control activity is increased by using a larger value of and decreased by using a smaller one. A higher value of is recommended, according to Figure 8, for plants with more complicated dynamics (higher 150 ) (see Table 1). Optimal values of as well as of are almost independent of . Finally, increase ki until acceptable overshoot (damping) is achieved in the closed-loop step response.

Jv 1.2 PI y

0.5 PI 0.3 Ju = 5 J=5 10 10 15 0.1 15

0.8

0.4 0.7

0.9

1.1 (a)

1.3

1.5

5 (b)

10

15

FIGURE 7 Optimal performance Jv = f ( ) and load disturbance step responses for different control activities Ju . The value of that minimizes Jv is largely independent of Ju except for small values. The economic limit of Ju for the plant G(s) = e0.3s /(1 + s)2 ranges from 10 to 15. The derivative action in the PID controller decreases Jv (Ju Juec ) by a factor of two compared to optimal PI control ( = 1, = 1).

30

15

kiG(0)/150G

k G (0)

20

10 5

2.5

10

0.5 (a)

150

0.5 (b)

150

0.5 (c)

150

FIGURE 8 Normalized PID controller parameters ki , k , and . The suggested tuning rules (6) and (7) for PID controllers based on 150 (solid) produce parameters with close-to-optimal outcomes (+) for most plants with all poles on the negative real axis.

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 63

A PID controller can reject process disturbances without too much control action and sensitivity to sensor noise.

For plants with 150 > 0.09 (Ld > 0.18Teq), including plants with highly nonminimum phase behavior, the results of this tuning procedure are remarkably close to optimal. The only problems arise from 1) plants with extremely long time delays and correspondingly high values of ( > 0.95), 2) plants with stable zeros close to the imaginary axis, and 3) plants with first-order dynamics and moderate time delay (see below). In Figure 10, responses are given to reference and disturbance steps, when a PID controller for G( s) = e0.3 s/(1 + s)2 is tuned using (9) and = 6.4. The integral gain ki should be manually tuned to GMS = 1.7. These step responses turn out

to be nearly optimal. Figure 10 confirms that nearly optimal behavior is obtained by the tuning rule (6).

First-Order Plants
A special group of plant models constitutes those with one time constant and moderate time delay, represented by the transfer function G( s) = K e sL d . 1 + sT (10)

First-order plants with extremely short time delays have simple dynamics and thus are not considered. PID tuning rules for first-order plants with time delay

=
150

1 , 0.68 + 0.84Ld / T 0.25(Ld / T)2

=T 0.06 + 0.68
0.5

Ld Ld 2 , Ld / T [0.05, 2], 0.12 T T 1 1 T T = min 5 + 2 , 25 , ki = 1.1 0.1 , K Ld KT Ld Ld / T [0.25, 2]. (11)

0 0

Ld /Teq

In Figure 11, step responses show that results based on (11) almost coincide with the optimal solution given by (2).

PI Controllers
FIGURE 9 Approximating the tuning paramter 150 . A good rule of thumb for computing 150 is 150 0.5 Ld / Teq .

In a PI controller, = = 1. Only two parameters need to be tuned. Unlike the PID case, an optimal PI controller yields a

0.3 y 1 KLT63 Optimal 0.5 0.1 KL 150 y KLT63 Optimal KL 150

0 0

5 (a)

10

0.1 0

5 (b)

10

FIGURE 10 Reference (a) and disturbance (b) step responses. The closed-loop behavior of a controller tuned in accordance with either the tuning rules (6) (KL 150) or (9) (KLT63) is remarkably similar to the optimal response. The plant considered here is G(s) = e0.3s /(1 + s)2 .

64 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

minimum in the Jv / Ju plot for most plant models (see [21], [22], and Figure 12). The goal of the PI tuning rules (12) and (13) is to reach that minimum, provided that the demand on MF robustness is fulfilled. Suitable rules for tuning PI controllers are PI tuning rules for plants of at least the second order 1 2 , or = 0.8(T63 Ld ), =150 G 0.06 + 1.6150 0.06150 150 [0.05, 0.9], 150 G 0.075 (12) 0.2 + , 150 [0.1, 0.9]. ki = G(0) 150 + 0.05 PI tuning rules for first-order plants with time delay

y 1.2 1 0.8 0.6 0.4 0.2 0 0

Tuning Rules Optimal

10

y 0.5 0.4 0.3 0.2 0.1 0 0.1 0

Optimal Tuning Rules

10

FIGURE 11 Reference and disturbance step responses for a firstorder plant with moderate time delay. These plots show that (11) yields closed-loop behavior close to the optimal solution. The plant considered here is G(s) = e0.6s /(1 + s) .

=0.5T63 = 0.5(T + Ld ), 0.57 1 ki = , Ld / T [0.2, 1]. TK Ld / T 0.055

(13)

Jv 2

PID: a = 4, b = 10

PI Versus PID Controllers


Rules (6)(13), especially (9), which is based on a step response and manual final tuning of the integral gain ki , show that PID controllers are no more difficult to design than PI controllers. The difference is that PID solutions provide additional freedom in selecting appropriate control activities Ju k (implicitly provided by ). PI control, where = 1 and = 1, means that the optimal control activity Ju is determined by the minimum in the Jv / Ju curve [see Figure 12(a)]. Often, the value of Ju can be somewhat increased by introducing derivative action with a value of typically in the interval [3, 10]. Performance is then improved significantly, without achieving excessive sensitivity to sensor noise and HF model uncertainties. Because of the concern for excessive sensitivity to sensor noise in the control signal, derivative action is often omitted in industrial applications. We thus optimize the tradeoff between noise sensitivity, the ability to compensate load disturbances, and MF robustness by fixing the controller zeros according to rules (6), (9), or (11) and then sequentially tuning or k and finally ki in the PID controller. As far as we know, this freedom for users is not available in any other tuning method. The IMC strategy [18], widely used in chemical process control, leaves only one parameter to tradeoff robustness, control costs, and performance qualities. To yield PI or PID controllers, this method demands first- or second-order plant models, respectively. The fundamental characteristics of a PID controller designed by IMC are two real zeros, chosen to cancel the plant poles, and a resulting closed-loop system with the characteristic polynomial ( s + 1/)2 or ( s + 1/)( s + z) , where s = z is a plant nonminimum phase zero, if present. The parameter is used for setting the desired response time. Comparisons with IMC are presented in [21].
1

Opt PI = f(T63)

= f(150) 0 0 5 (a) 10

Opt PID 15 Ju

0.5 PI KL y 0.3 PI Optimal PID: a = 4, b = 10 PID KLT63 PID Optimal 0.1 PID KL150

0.1 0

5 (b)

10

FIGURE 12 Comparisons of PI and PID controllers. For both types of controllers, tuning rules (6), (9), and (12) yield nearly optimal results. The price for performance Jv , compared to an optimal PI controller, is modest in terms of control activity Ju , when a PID controller is tuned by (6) or (9); it is high when a = 4 and b = 10 are fixed. Results of rules (6) (PID KL 150) and (12) (PI KL) are marked (o) in (a). In (b), Ju Juec for the PID controllers. The plant here is G(s) = e0.3s /(1 + s)2 .

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 65

Jv 40

a = 4, b = 10 = 0.6

1.25 20 1.5

0.75 0 2 7

1.0 12 Ju

FIGURE 13 Optimal Jv /Ju relationship for various damping values of the controller zeros when the plant has integral action. The optimal value of for a plant with a pole at the origin is about 1.0. A higher value of is a better choice than a lower value. The plant is G(s) = e0.3s /(s(1 + s)2 ) .

Figure 12 shows results from applying tuning rules (6), (9), and (12). It can be seen that the rules for the zeros in the PID case can be used over a wide range of values of Ju . Note that the two rules for based on 150 or, alternatively, on T63, produce almost identical results in the PID case and the PI case. Also observe that tuning all accessible parameters by the 150 rules (6) provides results that are close to the optimal results. Figure 12 also shows the high cost in terms of control activity Ju for the improved performance ( Jv ) achieved by a PID controller compared to the performance offered by an optimal PI controller when a and b are held constant to the traditional values 4 and 10. In contrast, the cost is moderate when the PID controller is optimized with the filter included. In many commercial PID controllers, the filter time constant is fixed (often b = 10) and cannot be modified. This situation often motivates the exclusion of derivative action in practice.

TUNING PID CONTROLLERS FOR PLANTS WITH INTEGRAL ACTION


When the plant has integral action, alternative tuning rules for controllers are needed. However, formulating tuning rules in this case is more challenging than for stable plants. In particular, the necessary normalizations are more complicated, while a starting point near the optimal solution, which makes the optimization routine (2) converge, is more critical.

40 Jv 30 = 3.0

20

5.0 4.0 3.5 40 Optimal 15 Ju 20

Controller Zeros
For plants with integral action, the optimal value of is often close to one (corresponding to a double zero), but tends to grow with increasing complexity in terms of i150 . A value of slightly above the optimal point is a better choice than a lower value. Figure 13 illustrates the relationship Jv = f ( Ju ) with various values of as well as with fixed a = 4 and b = 10. For values of Ju larger than 6.5, the required GMS cannot be reached. The zero time constant is crucial. If is more than marginally lower than the optimal value, Jv increases drastically and the requirement on GMS cannot be met, as shown in Figure 14. Except for the lowest values of Ju , the optimal value of is almost independent of Ju . Figure 14(b) also illustrates that the improvement in output performance Jv is significant when derivative action is included, compared to the case in which a PI controller is used. The control activity in the latter case is Ju = 0.45.

10 = 20 0 5 30 10 (a)

80 Jv 60 PI

40 Ju = 5 10 20 15 0 0

HF Gain k , Filter Factor , and Integral Gain ki


10 (b) 15 20

FIGURE 14 (a) Optimal Jv /Ju relationship for various PID zero time constants for a plant with integral action. The value of is crucial. (b) When this value is more than marginally lower than the optimal value, Jv increases drastically and the requirements on GMS cannot be met. With a suitable value of , the PID controller offers significant improvement of the output performance compared to a PI controller. The plant is G(s) = e0.3s /(s(1 + s)2 ) .

As in the stable case, k is equal to Ju for the closed-loop system. In general, to meet the desired stability margins, more derivative action is required for plants with integral action than for stable plants. The filter factor increases rapidly with increasing Ju [see Figure 14(a)]. Without adverse consequences, may for most plants (except when i150 < 0.7) be fixed to 20, a value corresponding to the well-known value b = 10. In tuning formulas, we also observe that the integral gain must be normalized twice with respect to a time or frequency parameter, since both ki and the plant integral gain have units of inverse time.

66 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

When the plant has integral action, suitable tuning rules for a PID controller are PID tuning rules for plants with integral action and at least two time constants i150 0.7, = 20, ki = = 1 , 3.5 3i150

1 =0.6150 G , i150 <0.7,

k = 20

150 G , K

2 150 G 6.49 i150 , e K

Figure 14(b) shows that the optimal performance offered by a PI controller corresponds to Jv = 74.0, while a PID controller with Ju = 15 can offer Jv = 8.9, almost a tenth of that offered by the PI controller. Figure 15 illustrates a typical difference between PI and PID controllers in terms of process disturbance step responses, which implies that a controller with derivative action is recommended for plants with integral action, except for plants with large time delays (Ld / T > 10).

1 , 3.5 3i150 (14)

ADDITIONAL FILTER ACTION


To achieve a strictly proper controller, the PI controller can be augmented by a lowpass filter, or the first-order filter in the PID controller can be replaced by a higher order filter. This modification is advantageous in cases of significant sensor noise or unmodeled HF resonances [40]. The interesting observation is that a lowpass filter can be added to both PI and PID controllers with only marginal effect on LF performance [41]. When a first-order lowpass filter is included, the PI controller becomes a PIF controller with transfer function KPIF = ki 1 + s . s(1 + s/)

1 =0.68150 G ,

2 ki = 150 G e4.46.5i150 . K

The normalization factor K = lim0 |G( )|. For plants with integral action, only one time constant, and time delay of the form G( s) = Ke sL d , s(1 + sT)

more appropriate results are obtained by the rules PID tuning rules for plants with integral action, one time constant and time delay 50 >Ld / T 0.25, =20, = 1.7 eL d/(1.3T), = 1.7Ld + 1.3T, 1 2e4.5L d/ T Ld < T, ki = KT2 1 Ld T, ki = (0.06 + 0.004Ld / T), KL2 d 1 0 Ld / T < 0.25, k = 20 , = 1 0.35e4L d/ T , KT 1 5e8L d/ T . (15) = 2.4Ld + 0.9T, ki = KT2 When the plant has integral action, a PI controller, which introduces additional integral action, is usually inappropriate.

When the first-order filter in the PID controller (4) is replaced by a second-order filter, the resulting PIDF controller has the form KPIDF ( s) = ki 1 + 2 s + ( s)2
s 1 + 2 f s+

(16)

This formulation enables complex poles, with damping ratio f and undamped frequency / , as well as complex zeros. Figure 16 shows an example of a plant controlled by three exactly proper and four strictly proper controllers. All controllers are optimized with GMS = 1.7. The PI/PIF controllers have Ju values corresponding to an optimal PI controller without a filter. For one pair of PID/PIDF controllers, Ju = 10.8, close to

y 3 2 1

1.5 PI Optimal y 1 PI Optimal

y 2 PI Optimal

PID Optimal PID Optimal 0.5 PID Optimal Tuning Rules Tuning Rules 1 Tuning Rules

0 0

20 (a)

40

0 0

20 (b)

40

0 0

20 (c)

40

FIGURE 15 Load disturbance step responses for three plants with integral action controlled by optimal PI and PID controllers and by PID controllers tuned by the rules (14) and (15). A controller with derivative action is recommended for all plants with a pole at the origin. Plant G(s) = e0.3s /(s(1 + s)2 ) with rule (14) is used in (a), plant G = e0.1s /(s(1 + s)) with (15) in (b), and G = es /(s(1 + 0.1s)) with rule (15) in (c).

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 67

10 PID Opt |K|

Finally, adjust ki to the desired damping of a closedloop step response. The simple design method for the PIDF controller can be compared to the H loop-shaping strategy described in [42]. The main idea in H loop-shaping is to augment the plant with a weight function W and modify W until a desired open loop shape is obtained for the augmented plant G = WG. The resulting controller is then combined with the weight function to obtain the final controller. When a PID filter is used for the weight function, the results of the PIDF controller and the H controller are almost equal. For details, see [41].

PIDF 1 PIF 0.2 101 PI Opt

101 (a)

103

0.5 y 0.4 0.3 0.2 0.1 0 0.1 0 2 4 6 (b) 8 10 t 12 PIF PI Opt PIDF PID Opt

CONCLUSIONS
This article presents some easily understood and applied methods for close-to-optimal tuning PI and PID controllers. By optimal we mean good MF robustness and the best possible tradeoff between output performance and control activity. For plants with all poles on the negative real axis, a simple step response can provide adequate plant knowledge. For plants with integral action, an impulse response or a relay experiment can be used. In all PID cases, the controller zeros can be fixed, the control activity can be varied by the filter factor, and, finally, the integral gain can be adjusted to the required damping of a step response for the closed-loop system. We maintain that, with this strategy, tuning a PID controller is as easy as tuning a PI controller, the difference being that the PID solution gives additional freedom for selecting slightly higher control activity, which significantly improves the output performance. When the situation demands HF rolloff of the controller, the PI or PID controller can be augmented by an additional lowpass filter. In these cases, the inclusion of derivative action is recommended. Four of the five controller parameters are then easily found, and the remaining gain can be manually tuned to obtain a desired tradeoff between output performance and damping (MF robustness). The major advantage of these tuning rules, compared to the previous rules, is that derivative action can be used without introducing high sensitivity to sensor noise in the control signal, a feature due to the inclusion of the low-pass filter in the design procedure. This approach can also be extended to include parameterized (larger) uncertainties using Horowitz bounds or -analysis [43].

FIGURE 16 PI and PID controllers with additional filter action. An extra lowpass filter in a PI or a PID controller can reduce the HF gain, and hence the sensitivity to sensor noise, with little deterioration of the LF performance. This property is illustrated by the controller gain |K | and the disturbance step responses for three exactly proper controllers and four strictly proper controllers, where k is limited to 20 and 80, respectively. The plant here is G(s) = e0.3s /(1 + s)2 .

Juec ; for the other pair, Ju = 5.6. In both cases, f = 0.4. When the HF gain k for the PIDF as well as the PIF controller is reduced to 20, which corresponds to very low bandwidth for the PIDF controller, then the performance Jv for the PIDF controller is deteriorated by 48% compared to the case of a first-order filter. Nevertheless, Jv is still only 53% of the value of Jv for the corresponding PIF controller. When the HF gain is k = 80, Jv for PIDF is increased by only 5.7% due to the filter, which is 40% of Jv for the corresponding PIF controller (with the same k = 80). These numbers and corresponding responses in Figure 16 show that derivative action is valuable for the output performance when significant rolloff is required in the controller. For more details about strictly proper PID controllers, see [23]. A set of simple tuning rules for PIDF controllers for stable plants can be formulated as Tuning rules for PIDF controllers = 10, = 0.8, = T63 /3, f = 0.4. (17)

ACKNOWLEDGMENTS
Prof. Karl-Johan strm and Prof. Sigurd Skogestad have both provided valuable comments on this manuscript, for which the authors are most appreciative.

AUTHOR INFORMATION
Birgitta Kristiansson (Birgitta.Kristiansson@chalmers.se) received the M.Sc. degree in engineering physics in 1965 and the Ph.D. in automatic control in 2003 from Chalmers University of Technology, Gteborg, Sweden. Her first major assignment was with the Swedish Council for Research Financing.

If more or less control activity is preferable, adjust .

68 IEEE CONTROL SYSTEMS MAGAZINE

FEBRUARY 2006

Later, she was responsible for technical audiology at the Sahlgrenska University Hospital in Gteborg, after which she taught control engineering and related subjects at the undergraduate and graduate levels. She is currently a senior lecturer in control engineering at Chalmers, and her research interests range from process and robust control to the design of PID controllers. She can be contacted at the Department of Signals and Systems, Campus Lindholmen, P.O. Box 8873, SE-402 72, Gteborg, Sweden. Bengt Lennartson received the M.Sc. degree in engineering physics in 1979 and the Ph.D. in automatic control in 1986 from Chalmers University of Technology, Gteborg, Sweden. In 1989, he became an associate professor at the Control Engineering Laboratory at Chalmers. Since 1999, he has been professor and chair of Automation in the Department of Signals and Systems. From 19982004, he was vice dean of the School of Mechanical Engineering; he is now dean of Education at Chalmers. His principle areas of research include robust, sampled-data, and PID control, in addition to discrete event and hybrid systems for manufacturing applications. He has been a member of IFACs Technical Committee on Manufacturing and an associate editor of Automatica. He is the supervisor for six Ph.D. candidates, and he is the coauthor of two books and over 100 peer-reviewed international papers.

REFERENCES
[1] R.C. Dorf and R.H. Bishop, Modern Control Systems, 9th ed. Englewood Cliffs, NJ: Prentice-Hall, 2001. [2] O. Mayr, The Origins of Feedback Control. Cambridge, MA: MIT Press, 1970. [3] B. Hick, Regulators or guvernors for steam engines etc, Tech. Rep., British Patent 8613, 1840. [4] K.J. strm and T. Hgglund, PID Controllers: Theory, Design and Tuning. Research Triangle Park, NC: Instrum. Soc. Amer., 1995. [5] M. Lelic and Z. Gajic, A reference guide to PID controllers in the nineties, in Proc. PID00. IFAC Workshop of Digital Control. Past, Present and Future of PID Control, Terrassa, Spain, Apr. 2000, pp. 7382,. [6] A. ODwyer, Handbook of PI and PID Controller Tuning Rules. London, UK: Imperial College Press, 2003. [7] D.B. Ender, Process control performance: Not as good as you think, Contr. Eng., vol. 40, no. 10, pp. 180190, 1993. [8] W.L. Bialkowski, Dreams versus reality: A view from both sides of the gap, Pulp and Paper, vol. 94, no. 11, 1993. [9] S. Rnnbck, Enkt om reglerproblem inom massa- och pappersindustrin, in Preprints fr Reglermte 98, Lund, Sweden, 1998, pp. 248252. [10] D. Toijer, Bttre trimning hgt p reglerteknikerns nskelista, Automation, vol. 27, no. 1, pp. 2528, 1999. [11] J.G. Ziegler and N.B. Nichols, Optimum settings for automatic controllers, Trans. ASME, vol. 64, no. 11, pp. 759768, 1942. [12] K.J. strm and T. Hgglund, A frequency domain method for automatic tuning of simple feedback loops, in Proc. 23rd IEEE Conf. Decision and Control, Las Vegas NV, Dec. 1984, pp. 299304. [13] A.A. Voda and I.D. Landau, A method for auto-calibration of PID controllers, Automatica, vol. 31, no. 1, pp. 4153, 1995. [14] K.J. strm and T. Hgglund, New tuning methods for PID controllers, in Proc. 3rd European Control Conf., Rome, Italy, Sep. 1995, pp. 24562462. [15] F.G. Shinskey, Process Control Systems, Application, Design and Tuning. New York: McGraw-Hill, 1996. [16] R.R. Pecharroman and F.L. Pagola, Control design for PID controllers auto-tuning based on improved identification, in Proc. PID00. IFAC Workshop of Digital Control. Past, Present and Future of PID Control, Terrassa, Spain, Apr. 2000, pp. 8994. [17] B. Kristiansson and B. Lennartson, Robust and optimal tuning of PI and PID controllers, IEE Proc. Contr. Theory Applicat., vol. 149, no. 1, pp. 1725, 2002. [18] D.E. Rivera, M. Morari, and S. Skogestad, Internal model control. 4. PID controller design, Ind. Eng. Chem. Process. Des. Dev. , vol. 25, pp. 252265, 1986.

[19] A.J. Isaksson and S.F. Graebe, Analytical PID parameter expressions for higher order systems, Automatica, vol. 35, no. 6, pp. 11211130, 1999. [20] S. Skogestad, Simple analytic rules for model reduction and PID controller tuning, J. Process Contr., vol. 13, no. 4, pp. 291309, 2003. [21] B. Kristiansson and B. Lennartson, Evaluation and simple tuning of PID controllers with high frequency robustness, J. Process Contr., vol. 16, no. 2, pp. 91102, 2003. [22] B. Lennartson and B. Kristiansson, Pass band and high frequency robustness for PID control, in Proc. 36th Conf. Decision and Control, San Diego, CA, Dec. 1997, vol. 3, pp. 26662671. [23] B. Kristiansson and B. Lennartson, Optimal PID controllers including rolloff and Smith predictor structure, in Proc. 14th World Congress of IFAC, Beijing, China, July 1999., vol. F, pp. 297302. [24] W.K. Ho, K.W. Lim, C.C. Hang, and L.Y. Ni, Getting more phase margin and performance out of PID controllers, Automatica, vol. 35, no. 9, pp. 15791585, 1999. [25] J. Langer and I.D. Landau, Combined pole placement/sensitivity function shaping method using convex optimization criteria, Automatica, vol. 35, no. 6, pp. 11111120, 1999. [26] H. Panagopoulos, K.J. strm, and T. Hgglund, Design of PID controllers based on constrained optimisation, IEEE Proc. Contr. Theory Applicat., vol. 149, no. 1, pp. 3240, 2002. [27] T.S. Schei, Automatic tuning of PID controllers based on transfer function estimation, Automatica, vol. 30, no. 12, pp. 19831989, 1994. [28] Q.G. Wang, C.C. Hang, Y. Yang, and J.B. He, Quantitative robust stability analysis and PID controller design, IEE Proc. Contr. Theory Applicat., vol. 149, no. 1, pp. 37, 2002. [29] J. Crowe and M.A. Johnson, Towards autonomous PI control satisfying classical robustness specifications, Proc. Inst. Elec. Eng.., vol. 149, no. 1, pp. 2631, 2002. [30] G.F. Franklin, J.D. Powell, and A. Emami-Naeini, Feedback Control of Dynamic Systems, 3rd ed. Reading, MA: Addison-Wesley, 1994. [31] R. Vilanova, I. Serra, and C. Pedret, A general approach to PID optimal design: Achievable designs and generation of stabilizing controllers, in Proc. PID00. IFAC Workshop Digital Control. Past, Present and Future of PID Control, Terrassa, Spain, Apr. 2000, pp. 444449. [32] C.A. Neto and M. Embiruco, Tuning of PID controllers: An optimizationbased method, in Proc. PID00. IFAC Workshop of Digital Control. Past, Present and Future of PID Control, Terrassa, Spain, Apr. 2000, pp. 415420. [33] G.P. Liu and S. Daley, Optimal-tuning PID controller design in the frequency domain with application to a rotary hydraulic system, Control Eng. Pract., vol. 7, no. 7, pp. 821830, 1999. [34] E. Grassi and K. Tsakalis, PID controller tuning by frequency loopshaping: Application to diffusion furnace temperature control, IEEE Trans. Contr. Syst. Technol., vol. 8, no. 5, pp. 842847, 2000. [35] E. Grassi, K. Tsakalis, S. Dash, W. MacArthur, S.V. Gaikwad, and G. Stein, Integrating system identification and PID controller tuning by frequency loop-shaping, IEEE Trans. Contr. Syst. Technol., vol. 9, no. 2, pp. 285294, 2001. [36] K.J. strm and T. Hgglund, The future of PID control, in Proc. PID00. IFAC Workshop of Digital Control. Past, Present and Future of PID Control, Terrassa, Spain, Apr. 2000, pp. 1930. [37] C.C. Hang, K.J. strm, and W.K. Ho, Refinements of the ZieglerNichols tuning formula, Proc. Inst. Elec. Eng., pt. D, vol. 138, no. 2, pp. 111118, 1991. [38] K.J. strm and T. Hgglund, Benchmark systems for PID control, in Proc. PID00. IFAC Work-shop of Digital Control. Past, Present and Future of PID Control, Terrassa, Spain, Apr. 2000, pp. 181182. [39] K.J. strm, H. Panagopoulos, and T. Hgglund, Design of PI controllers based on non-convex optimization, Automatica , vol. 34, no. 5, pp. 585601, 1998. [40] J.R. Corrado, W.M. Haddad, and D.S. Bernstein, H -optimal synthesis of controllers with relative degree two, Automatica, vol. 35, no. 6, pp. 11691173, 1999. [41] B. Kristiansson, B. Lennartson, and C.-M. Fransson, From PI to H control in a unified framework, in Proc. CDC00, Sydney, Australia, Dec. 2000, pp. 27402745. [42] D. MacFarlane and K. Glover, A loop shaping design procedure using H synthesis, IEEE Trans. Automat. Contr., vol. 37, no. 6, pp. 759769, 1992. [43] C.M. Fransson, B. Lennartson, T. Wik, K. Holmstrm, M.A. Saunders, and P.O. Gutman, Global controller optimization using Horowitz bounds, in Proc. 15th World Congress of IFAC , Barcelona, Spain, July 2002, pp. 24202425.

FEBRUARY 2006

IEEE CONTROL SYSTEMS MAGAZINE 69

You might also like