[2]. Astrom, K. 4 Higghond, T. PID Controllers: Theory, Design, and Tuning; Instrument Society of America: Research Triangle Park, 1995. [8] Bialkowski, W. L. “Dream ve Realty: A View from Both Side ofthe Gap.” Pulp Paper Cen. 1999, 94,19. {4] Bristol, B, Ml. “Pattern Recognition: An Alternative to Parameter Adaptive PID Controller,” Automatica 1977, 15,197 (5) Cao, R.; Medvoy, TJ. “Evaluation of Pattern Recognition Adaptive PID Controller,” Audomatica 1990, 26, 797. [6] Ender, D. B. “Process Control Performance: Not as Good as You Think,” Control Eng, 1993, 40, 180 [7] Mang, C. Cs Lee, T. HL; Ho, T. MH. Adaptive Control Instrument So cioty of America: Reseatch Triangle Patk, 1993 [8] Hersh, M. A. Jobnson, M. A. “A Study of Advanced C in the Workplace,” Control Eng. Prac. 1997, 5, 771. {9} Loyben, W. L. "Derivation of Transfer Functions for Highly Nonlinear Distillation Columns,” Ind. Eng. Chem. Res. 1987, 26, 2490, [10] Layben, W. L.; Luyben, M.L. Bssentials of Process Control, MeGraw- Hill: New York 1997, (11) Luyhen, W. Lj Tyreus, B.D. Luyben, M. L. Plantwide Control, McGraw-Hill: New York, 1998. (22) MeMfillan, GK. Peeing and Control Loop Performance; Instrument Society of America: Research Triangle Park, 1994 [13] Morati, My Zafirion, E. Robast Process Control; Prentice Hall: Engle- ‘wood Clif, 1989. [14] Sehorg, D. Bj Edgar, . F.; Mellichamp, D. A. Process Dynamics and Controt, John Wiley & Sons: New York, 1989. M,; Seborg, D. B. “A New Method for On- Tuning,” AICKE J. 1982, 28, 434 [16] Ziegler, 3. G.; Nichols, N. B. "Optimum Settings for Automatic Con trollers,” Trans, ASME 1942, 12, 759, rol Systems ne Controller CHAPTER 2 RELAY FEEDBACK Astrém and Wgglund (1984) suggest the relay feedback test to generate tained oscillation as an alternative to the conventional continuous ¢y- cling technique. It is very effective in determining the ultimate gain and ul rate frequency. Luyben (1987) popularizes the relay feedback method and calls this method ATV" (aulotune variation). The acronym also stands for all terrain vehicle since ATV provides a useful tool for the rough and rocky road of system identification ‘Aa pointed out by Luyben, the motivation for using the relay feedback (ATV) hos grown out of a study of an industrial distillation column. The Gisillation column is an important unit in chemical provess industries It is rather difficult to obtain a linear transfer function model for highly nonlinear columns. Attempts have been made using step or pulse tests Unfortunately, the system results in extremely long time constant (e.g, 1 = 870hr; Luyben, 1987). Moreover, very large deviations occur in the Tinear model as the size or direction of the input is changed. Simulation studies also reveal that, sometimes, very small changes of magnitude (less ‘han 0.0156) have to be made to get an accurate linear model. This imme- diately rules out the use of this kind of input design in real plants because plant data are never known (o anywhere near this order of accuracy. Luy- ben shows that the simple relay feedback tests provide an effective way to determine linear models for such processes, It has become s standard practice in chemical process control as can be seen in recent textbooks in process control (Seborg et al., 1989; Luyben and Luyben, 1997) “The distinct advantages of the relay feedback are: 1. Ie identifies process information around the impartant frequency, the ultimate frequency (the frequency where the phase angle is —") 2, Its a closed-loop est therefore the process will not drift away from. the nominal aperaling poin 3. For processes with a long time constant, it is a more time-cfficient method than conventional step or pulse testing. The experimental time roughly equals to 2~ 4 times of the ultimate period, 2.1 Experimental Design Consider a relay feedback system where G(s) is the process transfer fune- tion, y is the controlled output, yset is the set point, ¢ is the error and w@) FIGURE 2.1, (A) Block diagram for a relay feedback system and (B) relay feed- Ibs test for n system with positive steady-state gain is the manipulated input (Figure 2.14) An on-off (ideal) relay is placed in the feedback loop. The Astrém- Haglund relay feedback system is based on the observation: when the output lags behind the input by —7 radians, the closed-loop system may oscillate with a period P,. Figure 2.1(8) illustrates how the relay feedback system works. A relay of magnitude h is inserted in the feedback loop. Initially, the input wis increased by A. As the output y starts to increase (after a time delay D), the relay switches to the opposite position, «= —h. Since the phase lag is —n, a limit eyele with a period Py results (Figure 2.1). The period of the limit cycle is the ultimate period. ‘Therefore, th ultimate frequency from this relay feedback experiment is 2n z (21) From the Fourier series expansion, the amplitude @ can be considered to” be the result of the primary harmonic of the relay output. Therefore, the ultimate gain can be approximated as (Ogata, 1970; Astrém and Hagelund, 1084) Ke# (22) 2. Relay Feedback 9 where h is the height of the relay and a is the amplitude of oscillation, ‘These two values can be used direetly to find controller settings. Notice that Bys. 2.1 and 2.2 give approximate value of w,, and Ky. A more accurate expression will be derived shortly. ‘The relay feedback test can be carried out manually (without any auto- tuner). The procedure requires the following steps. 1. Bring the system to steady-state 2. Make a small (eg, 5%) inctease in the manipulated input. The mag- nitude of change depends on the process sensitivities and allowa- ble deviations in the controlled output. ‘Typical values are betiveen 3 ~ 10%. 3. As soon as the output crosses the set point, the manipulated input is switched to the opposite position (e.g., 5% change feom the original value). 4. Repeat step 2 until sustained oscillation is observed (Figure 2.1). Read off ulin Bq. 22. period P, from the cycling and compute Ky from This procedure is relatively simple and efficient. Physically, it implies you ‘moves the manipulated input against the process. Consider a system with 1 positive steady-state gain (Figure 2.1). When you increase the input (as in step 1), the output y tends to increase also, As a change in the output, is observed, you switch the input to the opposite direction. It is meant to bring the output back down to the set point, However, as soon as the output comes down to the sct point, you switch the input fo the upper position, Consequently, e continuous cycling results but the amplitude of osila is under your control (by adjusting hk). More importantly, in most cases, you obtain the information you need for tuning of the controller Several characteristics ean be seen from the relay feedback test. Consider the mest common frst order plus Lime delay systeras Kye! os) Ts+1 (2.3) where K; is the steady-state gain, D is the time delay and r is the time constant, Figuse 2.2 indicates that if the normalized time delay , D/>, is less than 0.28, the ultimate period is smaller than the process time constant, In terms of plant test, that implies the relay feedback test is more time-eficient than the step test. The reason is that it takes almost 3.1 to reach 95% of the steady-state value in a step test and the time required for the relay feedback is also roughly equal to 3 Py (to establish a stable oscillation).10 ‘wey Ta | FIGURE 22. Pafr (top) and Gljas)/G(0) (bottom) as fanetion of the worma- lined time delay, D/=) Therefore, the relay feedback system is more time-eflicient th test for systems with: the step D/= <0.28 en Since the time delay cannot be Loo large (it often comes from the meas rement delay), the temperature and composition loops in process industries seem to fall into this eategory. In other words, Eq. 2.4 is fairly typical for many slow chemical processes, especially for units involved with somposi- tion changes. The bottom graph in Fig. 2.2 shows that the amplitude of oscillation decreases with D/r. This gives a good indication of the magni- tude of oscillation. 2.2. Direct Tuning Rules ‘The Ziegler-Nichols tuning is stil popular in control engineering practice. It ‘works reasonably well for come loops but tends to be too underdainped for many process control applications. Many modified versions of Z-N tuning have been proposed over past 50 years (Seborg et al., 1989; Luyben and Luyben, 1997 ; Hang et al., 1991). Frequency domain interpretation of the Z-N method is also given (Astrém and Hégeland, 1988; Zhung and 2. Relay Feodback 1 ‘Table 2-1 The original Ziegler-Nichole settings. mirollet Ke tr _7> P 7 PL K,/22 BA /l2 = PID Ky? Pal? ‘Table 2.2 Different versions of the Ziegler-Nichols settings for PID con- troller. Tale Ra some overshoot Ky/3Pa/2 Pa/® no overshoot __Ku/5__Pa/2_P./® Anthetton 1991). The original Z-N eettings ean be calculated directly from K, and P,. For P, PLand PID controllers, the settings are shown in Table au Although widely used, the Ziegler-Nichols settings may have the disad- vantage of a large overshoot for some loops especially for set point chan- ges. Thus, more conservative tunings are often preferable. Two detuning Sotlings aro often used (Seborg et al., 1989). One is the some overshoot rule and the other is the no overshoot rule. From the standpoint of jet fightor manouvering, the threo aottingn can be viewed aa the attack mode, cruise mode and landing mode, respectively (Luyben and Luyben, 1997). ‘Therefore, they serve as distinet handles to tighten or loosen the speed of response. Based on the integrator plus time delay system, Tyreus and Luy- bben (1992) also proposed a tuning rule which also utilizes the information of Ky and Py. Shen and Yu (1994) also proposed a modified tuning rule for multivariable systems. These settings are rather similar to the some overshoot rule. For a PI controller, the settings are shown in Table 2.3 [Notice that these of seltings tend to work well for first order systems with ‘Table 2.3 Additional versions of the Zieglor-Nichols settings for PI con- troller Rule Ke 7 TyreusLuyben K,/3.22 02. P, Shen-Yix Kyf3_ 2 Py2 ‘along time constant (ic., small D/r value, e.g., less than 0.2). This can bbe seen for the derivation of the Tyreus-Luyben tuning (For integrator pl time-delay process). Therefore, once we have the ultimate gain and alt ‘mate period, the tuning rule can be applied direetly. In many cases, that completes the controller tuning. 2.3. Approximate ‘Transfer Functions After the relay feedback experiment, the estimated ultimate gain (K.), ultimate frequency (@) can be used directly to ealeulate controller pa. rameters. Alternatively, it is possible to back-calculate the approximated process transfer functions. The other data useful in finding the transfor function are the time delay (D) and/or the steady-state gain (Ky) In theory, the steady-state gain can be obtained from plant data, One simple way to find Ky is to compare the input and output values at two different steady-states, That is Ky = Ay/Su (25) where Ay denotes the change in the controlled variable and Aw stands for the deviation in the manipulated input. However, precautions must be taken to make sure that the sizes of the changes in u are made small enough such that the gain in Eq, 2.5 truly represents the linearized gain. For highly nonlinear processes, these changes ate typically as small 2s 10° to 10~* percent of the full range (Luyben, 1987). Such small changes would only be feasible using # mathematical model. Trying to obtain reliable steady-state gains from plant data is usually impractical The time delay D in the transfer function can be easily read off from the initial part of the relay feedback test. It is simply the time it takes y to start responding to the change in u (Figure 2.1). Therefore, itis more likely that sve will have information on the time delay rather than the steady-state sain, ‘Now we are ready to find an approximate model. Typical transfer funeti ‘ons in process control are assumed and paramotere can be calculated. The transfer functions have the following forms: Model I (integrator plus time delay ) Gye? G(s) = Ree (2.8) Motel 1 (Gast onder pls time deley) Gls) = eet en TF 2 Relay Feedback 13 Model 2a (second order plus time delay) Kye?" OO) = STIR es) Model 2b (second order plus time delay with two unequal lags) Kye? OO) = Gat ite ti) (2.9) models, the model I has two unknown parameters, models I and 2a have three unknown parameters and model 2b has four unknown parameters, Therefore, additiosial information such as D or Kis needed if the last three models are employed. As pointed out by Tyreus and Luyben (1902), the simplest. integrator-plus-time-delay model (model i) provides {good approximation for slow chemical processes, et, systems showing 2 small D/r value. Iti the model we recommend for slow processes, "The relay feedback experiment has the following steps: 1, Inecessary, the time delay D can be read off from the initial response (Figure 2.1) and the steady-state gain can be obtained from steady- state simulation 2. The ultimate gain (iy) and ultimate frequeney (sq) are computed (Bqs. 2.1 and 2.2) after the relay feedback experiment. 3. Different modet structures (Eqs. 2.6-9) are fitted to the data 2.3.1 SIMPLE APPROACH Once the model is selected, we can back-calculate the model parameters froin two equations describing the ultimate gain and ultimate frequency, Model J (Friman and Waller, 1994) (2.10) ay Notice that. no a priori procass knowledge is needed for this model. More- over, computation of Ky and D is quite straightforward, Model 1 (2.42) (2.13)“ For the model I, either D or Ay is needed to solve for the time constant. For example, ifthe time delay is read off from the relay test, we can compute + from Bq. 2.12. Then, ip can be found by solving Bq, 2.13. w= pane a = vii es) The equations describing model 2a are quite similar to those for model 1 Again, we need to know D or Kp before finding model parameter Model 8b a= wD ~ tan“ (eyr) — tan 1 Kp Ke f+ eu + Gand] Since we have four parameters in the model 2b, both Ky and D have to be known in order to solve for two time constants 7, and 7p. This is the most complex model structure in our models and itis often sufficient for procose contra) applications. Let us use a first order plus time delay system to illustrate the parameter estimation procedure. urs) (218) (2.47) Example 2.1 WB column (Wood and Berry, 1978) 128 Gs) = Te8e+T ‘This is the transfer function between the top composition (zp) and the reflux flow (R). From a relay feedback test, we obtain the following ultimate sgain and ultimate frequency: Ky = L-7L and oy = 1.615, Note trat these ‘wo values are only an approximation to the true values: Ay = 2.1 and ay = 1.608, Parameters can be calculated for different model structs Model F(no priori knowledge on KX, and D) 0.94¢-827" os Model 1 (assume Dis known, ie., D = 1) 13.26 OO) = Tate) 2 Relay Feedback 18 Model 2a (assume D is known) Le GO) = Gs9e4 a ‘Model 26 (assume D and Kp are known) 2 O54 10 000067) Gs) = Despite varying in model parameters, all these four models have the same ultimate gain and ultimate frequency. ‘That is, the models are correct around the ultimate frequency which is important for the controller de- sign. However, if we extrapolate the model to different frequencies, e.., © =0, the results can be completely misleading. For example, the steady- state gain of the model 2a is only 1.12 which is less than 10% of the true value. We have to be very cautious when using these models. . 2.3.2 IMPROVED ALGoRITIM In theory, if the model structure is correct. and the ultimate gain and ul timate frequency are correctly identified, we could have a very good ap- proximation of the transfer function. For example, ifthe Kw and dy in the previous example are close to the true values, then we will not have errors im the steady-state gains and time constant for model 1. Unfortunately, since Eqs. 2.1 and 2.2 only give approximations to the ultimate gain and ultimate frequency, the parameters derived from Eqs, 2.12 and 2.13 can deviate significantly from the true system parameters, This implies the ob- served ultimate period, Py, and the computed ultimate gain aze not the f value. In order to have a better approximation of the transfer function, funda- mental analysis of the relay feedback eystem is necoseary. First, one would like to know what the period of oscillation from the-relay feedback ex- periment really represents. In other words, given a transfer function with known parameters, what is the expression for the period of oscillation ob- served from relay feedback experiment, /,, The following theorem (Astréim and Higglund, 1984) provides the answer THEOREM. Consider the relay feedback system with a transfer fune- lion G(s) and an ideal relay (Figure 2.1). Let HG(I,,2) be the pulse transfer function of G(s) with a sampling time of 7. If there is a peti- dic oscillation, then the period of oscillation P, is given by HG(P,/2,-1) =0 Astoem and Hagglund (1984) prove the theorem starting form the discrete- time state-space equations, The result, 17G(P,/2,~1) = 0, is obtained16 by finding the 2domain equivalent, The continuous-time response of an ideal relay (Figure 2.1) can be discretized at the point when the relay switches, The -tcansforms of the input and output are h/(z + 1) and 0, eespectively, Since this is 2 self-oscillation system, the propagation of the input is described by the gain HG(P,/2,—1). This equation can be used to find the period of eecllation for a known system. In identification, P,, is observed from the response and one is able to use this to back- calculate systern patameters, Unlike the contintous-time analysis based on the primary harmonic, the diserete-time expression gives a sound basis for finding the system parameters, since no assumption is made in the derivation. Based on the theorem, a belter relationship between dy (or ,) and the system parameters can he derived, For the transfer functions of inte- rest (models 1, 2a and 2b), the following results can be derived fom the modified z-transform (Chang et al., 1992) Model 1 s ne * a h@exw(DP)— 1) Model 2a e 2z[m-+ (m~ Dexr(-se)] 1) Sul exp(— el exrBEIL 4 xp(-Z5-)) — 3] whore m= Model £ ofesmeai, Fg 2.18-20 provide alternative expressions between the observed ultimate petiod , €f du, and system parameters. For example, Eq. 2.18 relates dy to D and r ina way differs substantially from the standard phase angle ‘equation (ie., Ba. 2.12) yD — tan" (Gyr) Again, we can derive a better expression for the amplitude ratio part at the ultimate frequeney, since the expression in Eq. 2.2 is based on the first harmonic of the Pouier series expansion. The square-wave response of w (Figure 2.1) consists of many frequency components Cen 2 Relay Peedback 17 Table 2.4 Equations for different model structures. Model dimple algorithm —hinproved algo model? Eqs 10H IT = none model 1 Eqs. 21213 Bgs. 218 13 Dor Ky model 2a Eqs. 21415 Bags. 219 15 Dor Ky model 2b Eqgs.216 & 17 __Bgs. 2.20 & 17 Dand Ky model 2 Bigs. 216k 17 Fo 2.2017 Dand kp ‘Therefore, it becomes obvious that the amplitude observed in the relay feedback response is contributed from multiple frequencies, w = 0, 33,5 etc. In theory, one can have a better estimate of the arnplitude ratio by employing mote terms. An iterative procedure is necessary if more than ‘one Lerin is employed (e.g, finding G(s) from the single-term solution and including the higher frequency information, w = 804, to find a new G(s), and the procedure is repeated until ((s) converges). However, experimental results show that the estimation of sysiem parameters can be improved substantially by improving the expression for period of ostllation alone as shown in the next section. Furthermore, for higher order systems, there is little incentive to improve the expression for the amplitude by including more terms since higher order harmonics (c.f. w = 3, o w = Sy) are attenuated by the process. If only one term is employed, the equations describing the amplitude ratio are exactly the same as Bys. 2.19, 2.15 and 217 2.3.2 PARAMETER ESTIMATION From the ongoing analysis, the procedure for the ev function has the following steps L. solect model structure 2, compute model parameters according to ‘Table 2.4 ‘Table 2.4 summarizes the information required and corresponding equ- ations to find approximate transfer function, Most of these equation sets can be solved sequentially. Notice that if the improved algorithm is used, better estimates of the ultimate gain and ultimate frequency can be enielated from the model For model 2b, if some information is not known, then a different procedure should be employed, For example, if Ky is not available, we can perform a second rolay foodback test (Li et al., 1991) or use a biased relay (Chapters 6 and 10) to find additional information. Nonetheless, equations in Table 24 ate generally applicable eegardless of the procedure18 109 tor FIGURE 2.3. Multiplicative erors of first order ps time delay system obtained from Bqs. 212, 12 and 18 2.3.4 EXAMPLES Several examples are used to illustrate the adventages of the improved algorithm. Consider a first-order system with time delay. Esample 2.2 First onder system with time delay 1.65 eT From a relay feedback experiment with A = 0.04, we have P, = 33.26 ‘and 0 = 0.26. 1f D and/or Ky are available, we can back-calculate +. The 1's calculated from Eqs. 2.12 and 2.13 are r = 16.3 and 16.09, respectively. ‘The improved algorithm (Eq, 2.18) gives a better estimate in r, + = 19.97, by improving the expression in the period of oscillation alone. ‘The result from Eq. 2 18s almost exact (the difference may have resulted from reading off a and Py from the response curve). Figure 2.8 shows the multiplicative ‘modeling errors, em = |(G(iw) ~ Gliw))/G(iw)], for the transfer funetion G estimated from Eqs. 2.12, 2.13 and 2.18, The results show that the error, ms i significantly less when + is ealeulated from Eq. 2.18 alone. . In the following examples, we assume Kp and D are known and the time constant r for Models 1 and 2a is obtained by taking the average G(s) = 2, Relay Feedback 1» Brrr (1) FIGURE 2, Percent errors in Ky and ty forthe frstorder lag over a range of time delay D. of the values calculated from the corresponding equations for the case of the simple algorithm. Next, the effects of time delay on the estimation of the ultimate gain and ultimate frequency are also investigated. In the original ATTV method, Kis calculated from Bq. 2.2 and Gy. is derived from Eq. 2.1. In the proposed method, Ky and oy are back-calculated from the estimated transfer function G(s). Again, this is shown in the following transfer function: Bsample 2.3 Variable time delay 1.656-?* 0547 ‘The percent errors in Ky and wy are compared for these two methods over a tange of time delay (D = 0.1 ~ 60). The reals (Figure 24) show that the errors in K, for the simple method are quite significant (5~20%), Furthermore, the error in wy is almost nil for the improved method. ‘Similar behavior can also be observed for a second-order lag with time- delay aystem. G(s) Example 24 Second onder system with two unequal lags2» : sme agit Erorin fe (8) FIGURE 2.5, Percent ettors in Ky and wy for two unequal first-order lng sith time delay systems over a range of delay time D. ars (avs FTAs Figure 2.5 shows that a better estimation of G(s) can be achieved over a fange of D (D < 6D). Again, impiovements can be made in finding the Correct Ky and wy by using a more accurate expression in the period of oscillation. 7 ‘Since the estimated transfer function is typically employed in the analysis and design of a feedback control system, the impact of the modeling errors in closed-loop performance is evaluated. A model-based controtfer, IMC {internal model control), is employed to analyze the performance. One of the advantage of the [MC is that we can specify the desired trajectory in the design. Figure 2 6 compares the set point responses of IMC when different models, G's, are employed in the design of the controllers. Consider the first-order plus time delay system. als) als) Ihe set point response of the control system, designed according to G(s) from the simple algorithm, tends to be mote sluggish than the desired trajectory (Figure 2.6). The proposed method improves this situation as shown in Figure 2.6. Despite the fact that a tighter response can be achieved 2, Relay Feedback — 2k & é os Desired festa a ‘hepa geri Teme FIGURE 2.5. Set point responses of IMC designed according to the estimated transfer functions G(s) (the clooed loop constant i 20 for the desired trajectory) by shortening the closed-loop time constant under modeling errors, one has to realize that the value of » model-based controller is that one can foresee the closed-loop response. In other words, a good model always helps Generally, the proposed method improves the estimation in G(s) at the nominal condition (with perfect knosledge of Xp and D). The robustness with respect to etrors in the time delay is investigated. Since the improved method ealculates Ky and w, by finding the transfer function G(s) first, fullowed by solving the corresponding equations for them, it ie more eenci tive to the errors in the time delay than the original method. Let us take another first-order plus time delay system as an example, Ezample 2.5. Brror in the observed time delay 1.656 20s47 Figure 2.7 shows the estimate of Ky and wy for both methods when the pereent errors in time delay ranging from ~50% to 509%. Despite the fact that the errors in Ky and Ww, are less for the improved method over a reaconable range of errors in time delay, itis more sensitive to the error in D. Therefore, cate should be taken in reading off the time delay from the initial responses. 7 Ge)” SS Sa eS we Srrorin (8) 4 original ar method Eerorinpay FIGURE 27. Percent errors in Ky and wy for a fist order system over a range of variation in the time delay 2.4 Conclusions In this chapter, the relay feedback test is introduced and steps required to perform the experiment are also given. It can be carried ou! with or without a commercial autotuner, Once you have obtained the information con the ultimate frequency, the controller settings can be decided using the original or modified Ziegler-Nichols methods. You can also go a step further to find an appropriate transfer function for the process. This ean be useful for implementing model predictive control, dead time compensator (Smith predictor). Better approximation ean be achieved using the improved al- gorithm. Finding transfer fiametions using the biased relay plus hysteresis ‘was discussed by Wang et.al. (1997). The relay feedback works well for very difficult processes and, more importantly, the test is so simple that you should try it out on your systems. 2.5 References [1] Astrdm, K. 4. Hagglond, 7. “Aotomatic Toning of Simple Regulators ch Specifications on Phase and Amplitude Margins," Aatomatica 1984, 20, 645 2. Relay Feedback — 23 [2] Actrém, K. 3; Higelund, T. Automatic Tuning of PID Controllers, Instrumentation Society of America: Research Triangle Park, 1088. [3] Chang, B. C.;Shen, 8. H. Yu, C. G, “Derivation of Transfer Function from Relay Feedback Systems,” Ind. Eng. Chem. Res. 1992, 91, 856. (4) Friman, Mt; Waller, K. V. “Autotuning of Maltiloop Control Systems,” Ind. Eng. Chem. Res. 1994, 99, 1708. Hang, C. C.; Astiém, K. J.; Ho, W. K. ‘Refinements on the Ziegler- Nichols ‘Tuning Formula.” IEE Proc. Pi. D 1991, 138, 111 [8] Li, W; Rskinat, B.; Luyben, W. L. “An Improved Autotune Identifi- ‘ation Method,” Ind. Eng. Chem, Res, 1991, 30, 1530, [7] Luyben, W. L. “Derivation of Transfer Functions for Highly Nonlinear Distillation Columns,” Ind. Eng. Chem. Res. 1987, 26, 2490-2495. [8] Luyben, W. ea Luyben, M. 1, Bssentiale of Process Control; MeGraw- Hill: New York, 1997, (9] Ogata, K. Modern Control Engineering; Prentice-Hall: Englewood lifts, 1970, [10] Papestathopoulou, H..; Luyben, W. b. “Tuning Controllers on Distil- lation Columns with the Distillate Bottoms Structure,” Ind. Bng. Chem. Res. 1990, 29, 1858. [11] Seborg, D. E., Edgar, T. F.; Mellichamp, D. A. Process Dynamics and Control; Wiley: New York, 1989, [12] Shen, S. Hs Yu, C. C. “Use of Relay Peedback Test for Automatic Tuning of Multivariable Systems,” A/CKE J. 1994, 40, 627. (13) Tyres, Dz Luyben, W. L. “Tuning PI Controllers for Integra- tor/Deadtime Processes,” Ind. Eng. Chem. Res, 1992, 31, 2625. (14) Wang, Q. G3 Hang, C. ©.; Zou, B. “Low-Order Modeling from Relay Feedback,” Ind. Eng. Chem, Res, 1997, 36, 375 (15] Wood, R. K.; Berry, M. W. “Terminal Composition Control ofa Binary Distillation Column,” Chem, Eng. Sci, 1973, 28, 1707 {16] Zhuang, M; Atherton, D. P. "Automatic Tuning of Optimum PID Controllers,” JEE Proc. Pt. D 4993, 140, 216