Vous êtes sur la page 1sur 6

Preprints of the 8th IFAC Symposium on Nonlinear Control Systems University of Bologna, Italy, September 1-3, 2010

FrA01.4

A Generalized Predictive Force Controller for electropneumatic cylinders


Lot Chikh Philippe Poignet Franois Pierrot c Mical Michelin e

Fatronik France Tecnalia, Cap Omega Rond-point B. Franklin, 34960 Montpellier Cedex 2, France. (e-mail: lchikh@fatronik.com and mmichelin@fatronik.com) Laboratoire d Informatique, de Robotique et de Microlectronique de e Montpellier (LIRMM). 161 rue Ada, 34392 Montpellier Cedex 5, France. (e-mail: poignet@lirmm.fr and pierrot@lirmm.fr)

Abstract: Is it possible to synthesize an easy-to-implement predictive force controller for electropneumatic cylinders? In this paper, this problem is treated in details. As an electropneumatic cylinder is a highly nonlinear actuator, our strategy is based on the precise nonlinear modeling of the actuator and the application of a feedback linearization strategy. This enables to have an equivalent linearized model and therefore, to nd an explicit solution of the predictive optimization problem. The obtained controller is then an easy-to-implement one and the number of control parameters is very reduced: the weighting coecient and the prediction horizon. Experimental results prove the availability of this control approach and good performances in terms of capacity of tracking long duration static forces of high amplitudes. Keywords: Force control, pneumatic actuators modeling, feedback linearization 1. INTRODUCTION Force control and mechanical impedance are important in many robotics and automation tasks. For instance, force reecting master arms or haptic interfaces are used in many robotics medical devices Zarrad et al. (2007). Other possible applications are walking robots Verrelst et al. (2005), soft grippers, cooperative robots Shin et al. (2008) and smooth grinding. These applications need some actuators which fulll specic requirements: they should be able to generate long duration static force and provide a high force output per unit weight Richer and Hurmuzlu (September 2000). The usually chosen solution is electrical motors which are not always welladapted to these requirements. First, they have a limited power-to-weight ratio and this necessitates the use of some specic direct-drive motors which are expensive. Furthermore, electrical actuation limits considerably the range of possible applications as they cannot be used in rough and dusty environments. In this context, the inspection of other actuation type -typically pneumatic actuators- becomes primordial. Pneumatic actuators present many advantages that make them a good alternative for force control of robotics devices. Indeed, in addition to their high power-to-weight ratio (as they only use air under pressure), they are low cost actuators which enables their use at a very competitive and attractive cost. Another advantage is that they can easily be used in rough and dusty environments. In case of the pneumatic muscles, these actuators have a long durability and are maintenance free as they have at least 10 millions switching cycles. However, pneumatic actuators are not widely used in robotics systems contrary
Copyright by IFAC

to electrical ones. An important reason of that is that they are still very complex in their control as their model is highly nonlinear. In this paper, we focus on the force control problem of the electropneumatic cylinders. Many studies have been carried so far in literature. Only one study of GPC pressure control of a pneumatic chamber has been found Chaewieang et al. (2008). The model used by authors is linear and the nonlinearities of the pneumatic chamber and the valve are not taken into account. No studies of predictive force control of pneumatic actuators have been found in literature. In Ben-Dov and Salcudean (1995), authors have synthesized a lead-lag controller with a low pass lter for attenuating measurement noise. This control strategy enables tracking step signals of amplitude up to 15N while reaching the force sensor resolution of 0.22N . Sliding modes have been studied by Richer and Hurmuzlu (2000) and sine signals of amplitudes of 75N have been tracked. Results have shown promising results even if a phase dierence exists between the reference and the output signal. Finally in Ilchmann et al. (2006), an interesting approach consists on feedback linearization of the system and the application of a proportional output feedback controller with saturation. Authors have shown satisfactory control results with steady state errors less than 1N and good time responses for reference step changes of 5N . As far as we know, there is no application of predictive control on force tracking of electropneumatic actuators. This study brings two main contributions comparing to the state of the art of force control of pneumatic cylinders. First, the possibility of application of force predictive

1058

control in an easy-to-implement way on electropneumatic cylinders is inspected. Generalized Predictive control technique which is very intuitive, as it uses only two control parameters, enables time response reduction in an natural manner as any change is anticipated and depends on the prediction horizon. All the nonlinearities are taken into account and based on the nonlinear model, a feedback linearization strategy is adopted. The second contribution is in the results themselves as this control strategy brings an improvement in control performances especially for brutal force reference changes of high amplitudes. Reference step jumps of 300N are tested and give promising results in terms of small time response, reduced overshoot and reasonable steady state error which greatly depends on the pressure sensor resolution. This paper is organized as follows. Section II presents the versatile electropneumatic testbed used for the dierent tests and its nonlinear modeling. Section III, focuses on the control strategy. Feedback linearization equations are presented and the GPC background is introduced. Finally in section IV, the GPC force controller is implemented experimentally and control results are exposed in details. 2. EXPERIMENTAL PNEUMATIC TEST BED AND NONLINEAR MODELING The versatile electropneumatic setup is presented in Fig. 1. It includes two 5/3-way proportional valves 1 and three kinds of actuators. The rst actuator is a standard double acting cylinder 2 which is very widespread in industry. It is one of the cheapest pneumatic actuators and has been largely studied in literature (Sawodny and Hildebrandt (2002), Xiang and Wikander (2004), Ilchmann et al. (2006) and Brun (1999)). The second one is the rodless cylinder 3 which has the advantage of being symmetric, that is to say, the maximum force provided in one direction equals the one in the inverse direction. Another important advantage for the rodless cylinder is that it has generally a larger stroke than double acting one. Finally, Pneumatic Articial Muscles 4 (PAMs) are also used in the setup but not tested in this paper. More details about PAMs can be found in Daerden and Lefeber (2002) and Van Damme et al. (2008). Both of the two cylinders can be connected to a moving arm which may represent a robot arm. In the long run, our aim is to design ecient parallel robot which relie on pneumatic actuation with robust control: the moving arm of this test bed is very similar to arms used for various parallel robots such as Delta or Quattro robots. At its extremity, dierent masses can be attached for robustness tests of the controllers. An electrical motor is added at the rotational axis of the arm in order to experiment disturbance rejection. Three types of sensors are used on the setup; a high resolution incremental encoder, two pressure sensors and two force sensors. The pressure sensors have a pressure measurement range from 0 to 10 bars with an accuracy of 2% of the full scale measurement value. The real time prototyping environment is xPC TargetT M from Mathworks.
1 2 3 4

Fig. 1. Versatile electropneumatic setup for robotics application In the sequel, we present the nonlinear model of the pneumatic cylinders. Every electropneumatic positioning device includes an actuation element (the pneumatic cylinder), a command device (the valve), and a mechanical part and position, pressure and/or force sensors. A schematic representation of the electropneumatic system is given in Fig. 2. Supply pressure ps is supposed to be constant. p0

Fig. 2. Schematic representation of the experimental setup denotes the atmospheric pressure. It is supposed that any variation of the muscles volume or pressure behave between the ideal isothermal and adiabatic condition Hildebrandt et al. (2003) and can be described by the polytropic gas law: (1) p1 V1 = p2 V2 Where pi is the pressure in one of the two cylinders chambers (indexes 1 and 2 are related to two pressure states) and Vi is the volume of one chamber or the muscle, is the polytropic constant. The ideal gas equation describes the dependency of the gas mass: pV m= (2) rT where m is the gas mass inside the cylinder chamber or the muscle, T is the air temperature which is considered to be equal to the atmospheric temperature and r is the specic gas constant. Therefore, combining equations (1) and (2) leads to the pressure dynamics expression: dp dV = [rT qm (u, p) p s] (3) dt V (s) ds

MPYE-5-1/4-010-B from FESTO Double acting cylinder: FESTO DNC 32 320 PPV A Rodless cylinder: FESTO 532448 DGC 32300304132 ZR PAMs: FESTO MAS - 20-450N-AA-MC-K

Copyright by IFAC

1059

where u represents the input voltage of the valve, s is the position of the piston in case of the cylinders. qm (u, p) represents the mass ow rate 5 . Dynamic eects of the underlying position controller for the valve-slide stroke are neglected. These hypothesis justify why the mass ow is a function of the input voltage and the pressure in the actuator chamber. The relation between volume and displacement is given by the following equations: Vi (s) = Vi (0) Ai s where i refers to one of the two l chambers and Vi (0) = Ai 2 represent the initial volume. l is the length of the cylinders. Ai represents the piston section of each chamber. For the rodless cylinder, sections are symmetric (i.e. A1 = A2 ). The valve model has been approximated -after identication- by the following expression (see Belgharbi et al. (1999) for details): qm (u, p) = (p) + (p)u (4) th where and denes 5 -order polynomials with respect to p. For force control experiments, it is mandatory to determine the force/pressure dependency. This dependency is linear and given by (for standard cylinder): Fstandard = A1 p1 A2 p2 (5) For the rodless cylinder, the force is simpler as we have piston area symmetry Frodless = A(p1 p2 ) (6) 3. CONTROL OF THE ELECTROPNEUMATIC CYLINDERS The approach handled here can be justied by dierential geometry concepts Isidori (1995) and dierential atness theory Fliess et al. (1995). Once the linearized system obtained, the GPC controller background is introduced and it application is detailed. The GPC based pressure/force control scheme presented in this section is represented on gure 3. 3.1 Feedback linearization equations We present in this section the feedback linearization equations for the electropneumatic cylinders. The approach handled here can be justied by dierential geometry concepts Isidori (1995) and dierential atness theory Fliess et al. (1995). Indeed, it can be easily proved that the valve input voltage satises the criteria for dierential atness when pressure is taken as an output. In the same manner and using dierential geometry concepts Isidori (1995), it can be proved that the system is completely linearizable. Equations of pressure dynamics in each chamber are given by: dp1 rT dV1 rT = (p1 ) p1 s+ (p1 )u dt V1 (s) V1 (s) ds V1 (s) (7) dp2 rT dV2 rT = (p2 ) p2 s (p2 )u dt V2 (s) V2 (s) ds V2 (s) In these equations, we have considered that u = u1 = u2 which means that the chambers rely on the same pressure source and use a control signal of an opposite sign. If we do the assumption that the valve is symmetric, this means that only one valve can be used for our control scheme
5 dm dt

which is in fact an important advantage of this approach. This is what is represented in gure 2. By replacing the volume variation by its value (ie, dV1 = A = dV2 ) we ds ds get: dp1 rT A rT = (p1 ) p1 s + (p1 )u dt V1 (s) V1 (s) V1 (s) (8) dp2 rT A rT = (p2 ) + p2 s (p2 )u dt V2 (s) V2 (s) V2 (s) Pressure dierence is given by the following expression: d rT rT A A (p1 p2 ) = [ (p1 ) (p2 ) p1 s p2 s] dt V1 V2 V1 V2
f1 (s,s,p1 ,p2 )

rT rT +[ (p1 ) + (p2 )] u V1 V2
f2 (s,s,p1 ,p2 )

(9) Therefore, the control input which linearizes the input/output behavior of the system is given by: rT rT u=[ (p1 ) + (p2 )]1 [uaux V1 V2 (10) rT A A rT (p1 ) (p2 ) p1 s p2 s)] ( V1 V2 V1 V2 Therefore, when we are interested by dierence pressure as an output of the system, this leads to the rst integrator linearized system: p1 p2 = uaux (11) Note that the expression that has to be inverted never equals zero in the physical functioning conditions (pressures and angle variations). 3.2 GPC outer position Controller Generalized Predictive Control (GPC) is one of the control techniques that belong to the Model Predictive Control (MPC) family. The various MPC algorithms only dier amongst themselves in the model used to represent the process and the noises and cost function to be minimized Camacho and Bordons (2004). Major ideas appearing in the MPC control family are given in Camacho and Bordons (2004): (1) use of an explicit model to predict the process output at future time instants (horizon), (2) calculation of a control sequence minimizing an objective function; and (3) a receding strategy, which involves the application of the rst control signal of the sequence calculated at each step. MPC advantages are detailed in Maciejowski (2002) and Camacho and Bordons (2004). The MPC strategy is represented in Fig. 4 After a state of the art study, only one experimental study of GPC applications on pneumatic cylinder has been found Song and Liu (2006). Actually, two main reasons may justify this. First, one important aspect is the need of an appropriate process model. Indeed, the design algorithm is based on prior knowledge of the model. That is why nonlinear models are developed in this study. The second point concerns the fact that MPC should need high computation amount when no analytical solution exists for the optimization problem. In our case, the feedback linearization inner loop provides a good alternative as the system is transformed into a linear SISO one. An

= qm (u, p)

Copyright by IFAC

1060

Future force reference

Force Force /pressure conversion GPC predictive controller I/O linearizing block Valve Cylinder

GPC parameters Difference pressure loop

Fig. 3. General block diagram of the Force/pressure dierence predictive control strategy for the cylinders. The force control reference is specied and then transformed into a pressure dierence reference signal. minimum and maximum cost horizons. Nu is the control horizon. (j) and (j) are weighting sequences and w(t+j) is the future reference trajectory. The minimization of cost function leads to a future control sequence u(t), u(t + 1), ... where the output y(t + j) is close to w(t + j). Therefore, in order to optimize cost function, the best optimal prediction of y(t + j) (for N1 j N2 ) has to be determined. This needs the introduction of the following Diophantine equation: 1 = Ej (z 1 )A(z 1 ) + z j Fj (z 1 ) (14) with A = A(z 1 ) and polynomials Ej and Fj are uniquely dened with degrees j 1 and na respectively. is dened as 1 z 1 By multiplying (12) by Ej (z 1 )z j and considering (14): y(t + j) =Fj (z 1 )y(t) + Ej (z 1 )B(z 1 ) u(t + j d 1) + Ej (z 1 )e(t + j) (15) Since the noise terms in (15) are all in the future (this is because degree of polynomial Ej (z 1 ) = j 1), the best prediction of y(t + j) is: y (t + j|t) = Gj (z 1 ) u(t + j d 1) + Fj (z 1 )y(t) (16) where Gj (z 1 ) = Ej (z 1 )B(z 1 ) Polynomials Ej and Fj can merely be obtained recursively (demonstration can be found for instance in Camacho and Bordons (2004)). In the future, it will be referred only to N = N2 = Nu as the prediction horizon. N1 is chosen equal to 0. Lets consider the following set of j ahead optimal predictions: y (t + d + 1|t) = Gd+1 u(t) + Fd+1 y(t) . . (17) . y (t + d + N |t) = Gd+N u(t + N 1) + Fd+N y(t) It can be written in the following compact form: y = G u + F(z 1 )y(t) + G (z 1 ) u(t 1) (18) where terms y, u, G and F can be dened in Camacho (1993).

Fig. 4. The MPC strategy explicit solution exists. Furthermore, our system is a good application for fast MPC as the dimension of the state space vector is small. The GPC algorithm is based on a CARIMA model. Most Siso plants can be described by a CARIMA model after linearization. It is given by: A(z 1 )y(t) = z d B(z 1 )u(t 1) + C(z 1 )e(t) (12) where u(t) and y(t) are respectively the control and output sequences of the plant and e(t) is a zero-mean white noise. A, B and C are polynomials of the backward shift operator z 1 . They are given by: A(z 1 ) = 1 + a1 z 1 + a2 z 2 + ... + ana z na , B(z 1 ) = b0 + b1 z 1 + b2 z 2 + ... + bnb z nb and C(z 1 ) = 1+c1 z 1 +c2 z 2 +...+cnc z nc . For simplicity, it is assumed that C 1 equals 1. The GPC algorithm consists in applying a control sequence that minimizes a multistage cost function of the form:
N2

J(N1 , N2 , Nu ) =
j=N1

(j)[(t + j|t) w(t + j)]2 y


N2

(13) (j)[ u(t + j 1)]


2

+
j=1

where y (t + j|t) is an optimum j step ahead prediction of the system output on data up to time t. N1 and N2 are the
Copyright by IFAC

1061

Equation (18) can be rewritten in this form: y = Gu + f (19) Where f refers to the last two terms in Eq. (18) which only depend on the past. Now, it is possible to rewrite (13) as: J = (Gu + f w)T (Gu + f w) + uT u (20) where w = [w(t + d + 1), w(t + d + 1) . . . w(t + d + N )]T . Equation (20) can be written as: 1 J = uT Hu + bT u + f0 (21) 2 with H = 2(GT G + I), bT = 2(f w)T G and f0 = (f w)T (f w). Therefore, the minimum of J can simply be found by making the gradient of J equal to zero, which leads to: u = H1 b = (GT G + I)1 GT (w f ) (22) Since the control signal that is actually sent to the process is the rst element of vector u (receding strategy), it is given by: u(t) = K(w f ) (23) T where K represents the rst element of matrix (G G + I)1 GT . From equation (23), we can do many interesting remarks. The control signal depends on future errors - and not past ones as in conventional controllers. The control action is proportional to the future errors (with a factor K). Hence, if there is no error (i.e, wf = 0) , there will be no control move.

terms of static error and time response. Steady state error equals 5mbar which is the pressure sensor resolution. Time response equals to 0.035s. The control signals during all tests presented here are smooth and never reach the saturation value of the valve. There is no overshoot observed in the test. Another interesting test concerns dynamic pressure reference trajectories. A sine reference signal is tested on gure 6. Results are very satisfactory in terms of good tracking
Pressure sine tracking [bar]
4 3 2 39 40 41 42 43 44 45

Control signal [volt]

0.5 0

-0.5 39 40 41 42 43 44 45

Pressure error [bar]

0.05 0

-0.05 39 40 41

Time [s]

42

43

44

45

Fig. 6. Sine tracking of a pressure dierence signal using GPC of the reference signal. The maximum error equals 0.069 bar which is a very low value. According to the force control scheme presented in 3, and basing on the pressure dierence control loop, a force control test is handled. As one can see on gure 7, results are
Control input Force response [N] [Volt]
100 0 -100 -200 -300 0 2 1 0 -1 0 2 4 6 8 10

4. EXPERIMENTAL RESULTS In order to illustrate the eectiveness of our control strategy, some typical control tests have been performed and are presented in this section. As the force control strategy is based on pressure dierence control of the setup, the rst series of tests concerns pressure dierence control in order to evaluate the performance of the GPC controller. On gure 5, one can see control results for a step pressure signal of amplitude of 2bar. For this test, the linearized system takes the following
Step response [bar]
2 1.5 1 0.5 1 2 3 4 5 6

10

Force error [N]

-200 -400 0 2 4

Time [s]

10

Control signal [Volt]

2 0 -2 -4 1 2 3 4 5 6

Fig. 7. Step force response implementing the control scheme of gure 3 very good. The time response equals 0.171s and overshoot equals 8% which is very reasonable as the amplitude of the step change equals 421N (the initial force value is 171N and the desired value equals 250N ). Steady state error equals 4N and this is a normal value as the pressure sensor resolution is not very high (5mbar). It should be interesting to test this approach with more precise pressure sensors to improve the force control static errors. Finally in gure 8, a pulse signal of amplitude of +/150N and period of 4s is tested. Results clearly show that it is possible to have low static errors for high amplitudes and brutal changes reference signals. The oscillations which appear in the control signal can be removed by acting on the parameter of the GPC controller.

Error signal [bar]

1.5 1 0.5 0 1 2 3 4 5 6

Time [s]

Fig. 5. Step response for pressure dierence signal using GPC CARIMA form: B(z 1 ) = 0.001z 1 and A(z 1 ) = 1z 1 . Sampling time equals 1ms. The weighting sequence has been chosen equal to 0.09 and is xed to 1. The prediction horizon equals 100. Experimental results are very good in
Copyright by IFAC

1062

Tracking results
200

-200 5 10 15 20

4 2 0 -2 -4 6 8 10 12 14 16 18

200 0

-200 5 10

Time (s)

15

20

Fig. 8. Pulse response of the rodless cylinder 5. CONCLUSION In order to achieve good force control performances in terms of precision and robustness, a Generalized Predictive Controller is synthesized and experimentally implemented on a pneumatic cylinder. The nonlinear model of the actuator is developed and feedback linearization is applied in order to linearize the system and enable the existence of an explicit solution to the predictive optimisation problem. Dierent experimental results have been performed and show the reliability of the GPC approach as the pressure sensor resolution is reached with the dierent tests. This guarantees the best solutions according to the particular experimental devices which are used in this setup. This study proves that predictive control is a good, easyto-implement and reliable strategy for force control of electropneumatic cylinders. ACKNOWLEDGEMENTS The authors would like to thank the Fatronik Tecnalia Foundation which supported the experimental setup. REFERENCES Belgharbi, M., Sesmat, S., Scavarda, S., and Thomasset, D. (1999). Analytical model of the ow stage of a pneumatic servodistributor for simulation and nonlinear control. In The Sixth Scandinavian International Conference on Fluid Power. Tampere-Finlande. Pages 847-860. Ben-Dov, D. and Salcudean, S.E. (1995). A forcecontrolled pneumatic actuator. IEEE Transactions on Robotics and Automation, 11(6), 906911. doi: 10.1109/70.478438. Brun, X. (1999). Commandes linaires et non linaires en e e lectropneumatique. Mthodologie et applications. Ph.D. e e thesis, INSA de Lyon. Camacho, A. and Bordons, C. (2004). Model Predicitve Control. Second Edition. Springer, London. Camacho, E. (1993). Constrained generalized predictive control. 38(2), 327332. doi:10.1109/9.250485. Chaewieang, P., Sirisantisamrit, K., and Thepmanee, T. (2008). Pressure control of pneumatic-pressure-load system using generalized predictive controller. In Mechatronics and Automation, 2008. ICMA 2008. IEEE International Conference on, 788791. Takamatsu,. doi: 10.1109/ICMA.2008.4798857.
Copyright by IFAC

Daerden, D. and Lefeber, D. (2002). Pneumatic articial muscles, actuators for robotics and automation. European Journal of Mechanical and Environmental Engineering, 47(1), pages 10 21. Fliess, M., Lvine, J., Martin, p., and Rouchon, P. (1995). e Flatness and defect of non-linear systems: introductory theory and examples. International Journal of Control, 61, 13271361. Hildebrandt, A., Sawodny, O., and Neumann, R. (2003). A cascaded tracking control concept for pneumatic muscle actuators. In European Control Conference 2003 (ECC03), Cambridge UK, (CD-Rom). Ilchmann, A., Sawodny, O., and Trenn, S. (2006). Pneumatic cylinders: Modelling and feedback and feedback force-control. International Journal of Control, 79, Issue 6. June 2006, pages 650 661. Isidori, A. (1995). Nonlinear Control Systems. Third edition. Communications and Control Engineering Series. Springer-Verlag, Berlin. Maciejowski, J.M. (2002). Predicitve Control with Constraints. Prentice Hall. Richer, E. and Hurmuzlu, Y. (2000). A high performance pneumatic force actuator system part2 nonlinear controller design. Journal of Dynamic Systems, Measurement, and Control, 122, pp. 426434. Richer, E. and Hurmuzlu, Y. (September 2000). A high performance pneumatic force actuator system part 1 nonlinear mathematical model 2001. Journal of Dynamic Systems, Measurement, and Control, Volume 122 , Issue 3, pp. 426434. Sawodny, O. and Hildebrandt, A. (2002). Aspects of the control of dierential pneumatic cylinders. In In. Shimemura and M. Fujita, editors, Proc. of GermanJapanese Seminar, pages 247-256. Noto Hanto. Shin, D., Sardellitti, I., and Khatib, O. (2008). A hybrid actuation approach for human-friendly robot design. In Proc. of the 2008 IEEE International Conference on Robotics and Automation, Pasadena, CA. Song, Q. and Liu, F. (2006). Improved control of a pneumatic actuator pulsed with pwm. In Proceedings of the 2nd IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications. Van Damme, M., Vanderborght, B., Van Ham, R., Verrelst, B., Daerden, F., and Lefeber, D. (2008). Sliding mode control of a 2-dof planar pneumatic manipulator. International Applied Mechanics, 44, 11911199. Verrelst, B., Vanderborght, V., J, V., Van Ham, R., J., N., and Lefeber, D. (2005). Control architecture for the pneumatically actuated dynamic walking biped lucy. Mechatronics, 15, 703729. Xiang, F. and Wikander, J. (2004). Block-oriented approximate feedback linearization for control of pneumatic actuator system. Control Engineering Practice, 12, 387 399. Zarrad, W., Poignet, P., Cortesao, R., and Company, O. (2007). Towards teleoperated needle insertion with haptic feedback controller. In Intelligent Robots and Systems, 2007. IROS 2007. IEEE/RSJ International Conference on, 12541259. San Diego, CA,. doi: 10.1109/IROS.2007.4399085.

Force error [N]

Control input [Volt]

Force [N]

1063

Vous aimerez peut-être aussi