You are on page 1of 18

Int. J. Elect. Enging. Educ., Vol. 36, pp. 204221. Manchester U.P., 1999.

Printed in Great Britain

A SET OF CONTROL EXPERIMENTS ON AN OVERHEAD CRANE PROTOTYPE J. A. MENDEZ, L . ACOSTA, S. T ORRES, L . MORENO, G. N. MARICHAL and M. SIGUT Department of Applied Physics, University of L a L aguna, T enerife, Spain. E-mail: jamp@cyc.ds.ull.es
ABSTRACT A set of real-time control experiments applied to a non-linear system is presented. The system is an overhead crane and the experiments are applied on a low-cost scale model of it. This paper tries to compare the performance of both classical and intelligent control techniques applied to this plant. KEYWORDS digital control; intelligent control; non-linear systems; overhead crane

1 INTRODUCTION In digital control and intelligent control courses, in addition to a good theoretical formation, it is important to train the students experimentally. So the laboratory must be an essential component in verifying many of the theoretical concepts seen in the lectures. Moreover, the students deal with the problems that arise from implementation of the control algorithms in real systems: application of A/D and D/A converters, control actions in real time, system constraints on commands to apply, etc. The objective of the experiments described in this paper is to show the diculties involved in the study of non-linear systems. Because of the theoretical complexity associated with the control of these kinds of systems1, experiments with real plants are required to complement the topics studied during the course. Moreover, the students can compare the results obtained with the classical control strategy and with intelligent control ones. To implement the dierent strategies, a non-linear system with four degrees of freedom is considered. The plant, very typical in the process industry, is an overhead crane2. Its task is to transport containers between two points in a plane. The main diculty when controlling this plant is related to the oscillations that the load mass suers along the trajectory. These oscillations are especially undesirable at the arrival point. The laboratory is provided with a low-cost scale model of an overhead crane. In this prototype the students can check the proposed algorithms. Firstly, the students must validate the proposed model of the real plant. To do that, they perform a set of open-loop experiments to compare the practical results with the simulated ones, verifying the model parameters. Then, they
204

205

implement a pole placement controller. To improve the performance of this strategy, an intelligent subsystem is added to the controller. It consists of a neural network (NN) that provides adequate values for the feedback gains in order to optimise the system response. This operation is performed on-line. Finally, the design of a controller based on fuzzy logic is proposed. In view of the growing success of these intelligent strategies3, applied in many industrial elds, it is important for the students to understand their methodology. This does not present any diculties since the fuzzy controllers build the control rules according to linguistic reasoning. In the paper, the main design phases for the fuzzy controller are described, and application to the prototype is performed. The scale model of the crane has been used in post-graduate courses where more complex experiments have been tried (predictive control algorithms, robust control, etc.). This paper presents some experiments for undergraduate courses about digital control and intelligent control. So, the two rst experiments are proposed for a digital control course. The students have already taken a preliminary course on linear systems: time and frequency response, stability, controller design, etc. In the digital control course, some topics like digital ltering, discrete-time stochastic control systems, linear discrete-time optimal control, etc. are studied. In addition, some experiments with other plants have been carried out4,5. The last two experiments are proposed for an intelligent control course. In this, the students learn topics like neurocontrol, fuzzy logic and real-time expert systems. 2 PLANT DESCRIPTION In Fig. 1 the plant used in this work is shown. As can be observed, it is a scale model of an overhead crane. It has the following sub-systems:
$

Crab: A vehicle that transports the load mass. It is powered by two d.c. motors joined to the driving wheels. The input voltage of each d.c. motor

FIG. 1

General view of the crane prototype.

206

can vary between 8.5 V and +8.5 V. The voltage applied to a motor is proportional to the angular velocity reached by the motor. The maximum speed is about 60 rpm for an 8.5 V input. Load mass: A container joined to the crab by a metallic rope, assuming that it has a negligible mass. Metal bridge of about 1.2 m length where the crab can move horizontally.

The values of the crab mass, the length of the rope and the mass of the container used in the experiments have the following values: Crab mass: 0.3 kg Load mass: 0.15 kg Length of the rope: 0.5 m The scale model is complemented with an A/D D/A converter board, which sends the information to the control unit and the commands to the crab. The board is located on a PC, with the necessary control software. 3 PLANT MODEL A simplied model of this overhead crane plant is shown in Fig. 2. The task of the crane is to bring containers with dierent masses from an initial point to a nal one. There are four degrees of freedom, the position of the crab (x), the linear velocity (x), the rope angle (h) and its angular velocity (h ). It is assumed that the crab has a mass M, the length of the rope is l and the load mass is m. A simplication is made assuming that the rope has a negligible mass. The input of the plant, u, represents the horizontal force applied to the crab. The application of a non-zero force on the plant produces undesirable oscillations in the load mass. The control objective is to make the crab arrive at the nal position with h and h equal to zero. Dening the state variables x =x, x =x, x =h and x =h, the dynamics 1 2 3 4 of the plant can be described6 by the following two coupled non-linear second-

FIG. 2 Schematic view of the plant.

207

order equations ml(x cos x x2 sin x )+(M+m)x =u 3 3 4 3 1 (1) x cos x +lx =g sin x 1 3 3 3 where g is the gravity acceleration. The rst experiment proposed to the students is to validate the model with the real plant. This step is necessary for the later realisation of any control experiment on the plant. They can do it by applying step voltage inputs and ramp voltage inputs to get the open loop response of the real plant. The results are compared with the simulated results obtained with the masses and lengths of the real plant specications.

Experiment 1: Model validation To perform the validation of the model (1) for this prototype, the students are told to take into account the practical aspects. Firstly, the transfer function that relates the input voltage of the plant V (in volts) with the output of the controller u (force applied in Kg m/s2) has to be considered. This transfer function is: V (s)= U(s) , bs b=0.015 (Kg m/s V) (2)

Other eects to consider are due to non-linearities that are not represented in the model (1), e.g. command constraints, mechanical friction, dead zones, etc. Anyway, from among these non-linearities only the dead zone and timedelay have a signicant contribution in the system. To validate the model, dierent open-loop experiments have been carried out. A voltage ramp input has been considered to study the response of the system. In Fig. 3 the theoretical (dashed line) and the measured (continuous line) evolution of the x and x variables for an input ramp with slope equal 1 2 to 3 V is shown. In the simulation a dead zone from 0 to 1.8 V has also been included. As can be observed, the theoretical model ts the real plant response relatively well. In view of this, the model (1) can be seen as a proper representation of the plant dynamics. After model validation, the real-time implementation of the dierent controllers in the prototype has been performed. 4 CONTROLLING THE PLANT The experiments presented to the students during the course are described in this section. These experiments try to complement their theoretical background. The experiments are divided into two subsets: the rst subset related to a classical control strategy and the second subset related to the design of intelligent controllers. The objective of the proposed control strategies is to move the crab from an initial point to a desired one, arriving with no oscillations in the load mass. Using state variables, the control objective is to take the crane from the initial point x =x , x =0, x =0, x =0 to the set point x =x , 1 1i 2 3 4 1 1ref x =0, x =0 and x =0. 2 3 4

208

FIG. 3 Open loop response using a voltage ramp input with slope 3 V. T heoretical response (dashed line) and measured response (continuous line).

For this plant, a simple PID controller is not able to achieve the control objective because of the multivariable specication given by the fact of controlling the crab and the load. For this reason, other control strategies are designed. The rst of these consists of a pole placement controller. This seems to be the easiest and most direct way to control the system. Obviously, a linearisation of the model is necessary in order to apply this method. The main drawback when the controller is implemented is that it can present reduced performance due to the presence of non-linear components: dead zone, command constraints, time delay, etc. As an alternative algorithm a state feedback controller is presented in which the gains are adjusted automatically using NN. The last experiment is the design of a controller based on fuzzy logic. Once the dynamics of this system are known, e.g. by observing the behaviour of the system under the previous controllers, the students are ready to dene the fuzzy variables and their subsets, and also the inference rules required to achieve the control objective.

Experiment 2: Pole placement controller The design of a pole placement controller is proposed to the students. The rst step in applying this method is the linearisation of the model. It is obtained assuming that the angles are small and ignoring the products of angular

209

variables. In this way the model (1) is simplied to x =x , 1 2 mg 1 x = x + F, 2 M 3 M x =x 3 4 x = 4 (m+M)g 1 x F 3 lM lM (3)

where x is the position of the crab, x its linear velocity, x the angle of the 1 2 3 load, x its angular velocity and u is the command applied to the crane. The 4 general control law applied is given by the following equation: u(k)=k e (k)k x (k)k x (k)k x (k) 1 1 2 2 3 3 4 4 being (4)

e (k)=x x (k) (5) 1 1ref 1 where x is the reference value for the position x , while the reference value 1ref 1 for the linear velocity x , angle x and angular speed x is zero. 2 3 4 The values for the gains k are adjusted to get the desired closed-loop i response in the system. So, the procedure is, rst, to choose the desired closedloop poles, and then to calculate the gains. As is well known, this is accomplished by means of Ackermans formula7: K=[0 0 ... 1]M1 P(A) (6) c M being the controllability matrix of the system, A the matrix of the linearised c system and P the characteristical polynomial of the closed-loop system. The students in this experiment have to propose the closed-loop poles for the plant to perform satisfactorily. Several guidelines are provided to the students: the poles at the origin have to be moved out of it; and the imaginary poles have to present a damping factor between 0.4 and 0.7. Prior to implementation, the students have to verify the performance of the controller by means of simulation. In Fig. 5 a simulation trial is presented. The poles are located at 0.5, 178 and 3.733.81i. As can be observed, the complex poles are selected to have a damping factor of 0.7. The evolution observed in this experiment is satisfactory. However, the real-time implementation looses ecacy. This occurs mainly because the corresponding feedback gains are too large and the non-linear eects previously mentioned have more inuence on the process. To get better results it is necessary to reduce the damping factor of the complex poles. In Fig. 6 a real-time experiment is shown where the complex poles are selected to have a damping factor of 0.14. It has been carried out with a sampling time T =0.03 s. The value of T is chosen according to the following criterion: sampling periods of 10 to 30 times the bandwidth of the closed-loop system7. In Fig. 4 the Bode diagram for the linearised model of the system is presented. The value of the bandwidth is roughly 6 rad/s. Thirty times the bandwidth gives a value for T of about 0.03 s. It can be observed that the crane attains the set point smoothly but with overshooting. Simultaneously, the variable x starts oscillating until the crane 3

210

FIG. 4

Bode magnitude plot for the linearised transfer functions corresponding to the position and angle.

FIG. 5

Simulated trial for the pole placement controller. Poles at 0.5, 178, 3.733.81i. T he damping factor of the complex poles is 0.70. T he corresponding gains are k =19.98, k =44.13, k =99.50 and 1 2 3 k =2.91. Set point at 0.5 m. 4

211

FIG. 6 Real-time experiment for the pole placement controller. Poles at 0.5, 178, 0.765.28i. T he damping factor for the complex poles is reduced to 0.14. T he corresponding gains are k =19.36, k =39.79, k =18.30 and 1 2 3 k =3.54. Set point at 0.5 m. 4 is near to the nal position. At this point the mass oscillation decreases and tends towards the equilibrium point x =0, x =0. The settling time of the 3 4 closed loop system is about 10 s. The main drawback of this control strategy is related to the choice of closedloop poles. If accurate values for the poles are not encountered, the feedback gains may take inappropriate values and the performance of the closed loop system would decrease. In the following experiment an intelligent control strategy is introduced. The aim is to solve the problem of the controller in experiment 2 to adjust the controller gains. The neural network works on-line to give their appropriate values. Experiment 3: State feedback controller using neural networks Articial neural networks are mathematical algorithms designed using the biological principles of brain activities. Basically, a NN performs mapping between input space and an output space. The objective is to train the network in such a way that, when an input is applied, the network produces the desired response. Learning is accomplished by adjusting the interconnecting weights in response to training data sets. In this work the NN used are the most common family of feed-forward networks: the multilayer perceptron (MLP)8. The MLP is trained with the backpropagation algorithm to adjust the weights of the net. This well-known algorithm uses a gradient descent technique. The aim of this experiment is to propose to the students an eective method for adjusting adaptively the parameters in a controller. To do that, a NN is used to provide the feedback gains to the controller by using information on the system state. The NN proposed acts as an implicit self-tuner, providing the

212

parameters directly to the controller. In fact, there is no estimation and control design stage. The NN at each instant simply looks for the optimal parameters to minimise a cost function using information on the system state. This control scheme can be shown in Fig. 7 (Ref. [6]). As can be observed, a state feedback controller similar to (4) is applied on the x and x state 2 3 variables and on the x error variable. Thus, the applied command is 1 u(k)=[k k k ][e (k) x (k) x (k)]T (7) 1 2 3 1 2 3 So there are three parameters to tune: the gains k , k and k corresponding 1 2 3 to e , x and x respectively. To improve the convergence of the networks the 1 2 3 tuning scheme proposed was split into two self-tuning subsystems. One NN (NN1) was used to tune the parameters related to the crab variables, k and 1 k , and another NN (NN2) to provide the parameter related to the load mass 2 variable, k . The inputs to the net NN1 are the x error and the x derivative 3 1 1 and for NN2 the x error and the x derivative. This structure based on two 3 3 NN is expected to work better than a one-NN structure because for each NN, fewer parameters are aecting the most performance. So it is less dicult for each NN to nd the optimal values for the gains. Obviously, during an experiment the set-point x keeps constant until the 1ref crane completes the task. This also ensures the convergence of the neural networks. The training of the network is done on-line. At each stage of the process the NNs are trained to nd the values for k , k and k that minimise 1 2 3 a pre-specied cost function. This function is generally chosen to be quadratic. It is proposed to be like the following: J(k)=p (k)[x x (k)]2+p (k)x2 (k)+p (k)x2 (k)+p (k)x2 (k) (8) 1 1ref 1 2 2 3 3 4 4 where p (k) are time-dependent functions that weigh each variable in the cost i

FIG. 7

Structure of the neural network feedback controller for the crane.

213

function. Depending on the position of the crab, the p (k) variables will take i the appropriate value. Then the NN must change its weights to minimise this cost function. The backpropagation states that at each stage the weights are updated using this expression: J(k) w (k+1)=w (k)m L,j,i L,j,i w L,j,i where m is the learning rate. The computation of the gradient is done according to J(k) J(k) k j = w k w L,j,i j L,j,i where (9)

(10)

y(k) u(k) J(k) =e(k) , (11) k u k j j k are the feedback gains and y is the output of each NN; x in the case of j 1 NN1 and x in the case of NN2. The factor y/u(k) is obtained immediately 3 if the model of the plant is known. Avoiding any assumptions about the plant model, this term can be obtained by a dierence approximation: y(k) y(k)y(k1) # . u u(k1)u(k2) (12)

The factor u/k (k) is calculated from the expression (7) for the controller: j u(k) u(k) u(k) =e (k); =x (k); =x (k). (13) 1 2 3 k k k 1 2 3 Starting from the instant k, the algorithm can be summarised as follows: Step 1) READ x (k), x (k), x (k) and x (k) 1 2 3 4 FIX the measuring period T . Step 2) (Networks training) NN1 training: u(k) u(k) e (k); x (k) 1 2 k k 1 2 u(k) u(k) ; Backpropagation , ; e (k), e (k); p (k)| 1 1 i i=1,...,4 k k 1 2 UPDAT E k and k 1 2 NN2 training;

u(k) x (k) 3 k 3

214

Backpropagation UPDAT E k

u(k) ; ; x (k), x (k); p (k)| 3 4 i i=1,...,4 k 3

3 Step 3) COMPUT E command u(k)=(k k k )[e (k) x (k) x (k)]T 1 2 3 1 2 3 Step 4) APPLY u(k) and WAIT until t=(k+1)T Step 5) kk+1 Step 6) GOT O Step 1 This algorithm nds the optimal values for the gains of the controller. Then for this xed-parameter plant, the gains of the adaptive controller will converge to a set of values that minimises the performance index. However, if the plant dynamics varies then the gains supplied by the neural networks change adaptively to a new set of values. It is also worth pointing out that from a theoretical point of view and due to the incompleteness of the on-line training data, problems could be expected to appear in the experiments. However, the experiments carried out did not reveal this kind of problem. In Figs. 8 and 9 the results obtained in one of the experiments are presented. This experiment was performed after several training trials with the crane. The cost parameters chosen are p (k)=20, p (k)=0.2, p (k)=0.1 and p (k)=0. The 1 2 3 4 obtained results seem entirely satisfactory since both control objectives in the load angle and in the cart position are attained. The command applied to the plant in this experiment is shown in Fig. 8. Note that the command values never violate the command constraint. As an alternative to the state feedback controller, a controller based on fuzzy

FIG. 8 Real time implementation on the crane prototype. Evolution of the x 1 and x variable. 3

215

FIG. 9 Command evolution in the experiment shown in Fig. 8. logic is proposed. A very simple controller based on linguistics rules is presented. The rules are obtained based on the knowledge of a human expert. This reduces the complexity of the control algorithms. A comparison between these results and the preceding ones is proposed. Experiment 4: Control based on fuzzy logic Fuzzy logic theory allows the use of human reasoning in the control of systems. Since this reasoning is often qualitative, the resultant actions are vague or ambiguous, in contrast to the precise mathematics used by computers. Fuzzy logic solves this problem by considering that a variable can have a membership degree to a fuzzy set. This motivates the denition of fuzzy sets that cover the full range of the variable. Then, each value of the variable has a membership degree to the fuzzy sets. The reasoning expert rules will be applied onto these fuzzy sets (inference). Due to the non-existence of fuzzy actuators or sensors (that is, with their inputs or outputs being fuzzy), it is necessary to add two steps into the control scheme: fuzzication and defuzzication. Therefore, there are three dierent parts in a fuzzy control scheme: Fuzzication: The precise values of the inputs must be transformed into their correspondent fuzzy variables with the corresponding membership degree to each fuzzy number. Inference: Direct application of the rules onto the inputs. As the result the outputs are obtained with the correspondent membership degree. Defuzzication: The imprecise values of the outputs must be transformed into precise ones. The typical scheme of a fuzzy control system is shown in Fig. 10 (Ref. 9). The fuzzy numbers of each input or output variable and the rules that dene the control action must be specied. These sets and rules are proposed to be dened intuitively by the students, once they have studied the behaviour of the system in the previous experiments. For this purpose, a visual fuzzy software

216

FIG. 10 Scheme of a typical fuzzy controller. is used. This program allows an easy denition of the fuzzy sets and the inference rules. These fuzzy sets and the results of the control action are graphically represented. Mamdanis method10 is used in its discrete form, computing the output of the controller from the inputs read by the A/D converter. One of the possible choices of fuzzy variables and the inference rules, which oers acceptable results, is given below. The chosen input variables to the inference motor are the error in the position (e =x x ) and the angle of 1 1ref 1 the load (h). The output variable is the command applied to the crane (u). Every fuzzy variable has ve sets dened in their universe of discourse:
$ $ $ $ $

PL PS ZE NS NL

Positive Large Positive Small Almost Zero Negative Small Negative Large

The membership functions are represented in Fig. 11. Then, the controller will have two inputs and one output. Once the students completely understand the dynamics of the system, they have to propose a set of rules like the following R : if (e is A and h is 1 1 1 R : if (e is A and h is 2 1 2 ... B ) then (u is C ) 1 1 B ) then (u is C ) 2 2

R : if (e is A and h is B ) then (u is C ) n 1 n n n where {A } are fuzzy numbers of the rst input variable e ; {B } i i=1,...,n 1 i i=1,...,n of the second one h, and {C } the fuzzy numbers of the output variable i i=1,...,n u. Let e and h be two crisp input values; C the fuzzy number inferred from 1o o i e and h applying the rule R ; and C the fuzzy number given by the aggre1o o i gation of each C . Following Mamdanis method, the t-norm that operates on i the fuzzy inputs is the min operator (9). Then, the membership degree of the

217

FIG. 11

Membership functions of the dened fuzzy variables.

value consequence of the rule R , can be expressed generally as10: i m (e , h, u)=m (e )9m (h)9m (u) (14) C 1 Ai 1 Bi Ci i If the e and h values are specied, then the membership degree to the 1o o inferred fuzzy number C is: i m (u)=m (e )9m (h )9m (u) (15) C Ai 1o Bi o Ci i On the other hand, the s-norm that operates on the C fuzzy results is the i max operator (8) m (u)=m (u)8m (u)8 ... 8m (u) C C C C 1 2 n And the Centre-of-gravity method is employed for defuzzication. (16)

m (u) du C W The inference module proposed has 9 rules dened according to the following considerations: There is no action on the angle until the position is near the set point. The oscillations of the load are corrected by aligning the crab vertically with the load.

u = W o

um (u) du C

(17)

218

Once the angle is corrected, the crab is taken to the set point with a small command. With these guidelines the resultant rule table is the following: (e =NL) then (u=PL) 1 (e =NS or e =ZE or e =PS) and (h=NL) then (u=NL) 1 1 1 (e =NS or e =ZE or e =PS) and (h=NS) then (u=NS) 1 1 1 (e =NS) and (h=ZE) then (u=PS) 1 (e =ZE) and (h=ZE) then (u=ZE) 1 (e =PS) and (h=ZE) then (u=NS) 1 (e =NS or e =ZE or e =PS) and (h=PS) then (u=PS) 1 1 1 (e =NS or e =ZE or e =PS) and (h=PL) then (u=PL) 1 1 1 (e =PL) then (u=NL) 1 The algorithm used for the implementation of the controller can be summarised as follows: If If If If If If If If If Step 1) DEFINIT ION of the fuzzy subsets for the variables and inference rules. FIX the sample period T . Step 2) READ the inputs e (k), h(k). 1 Step 3) FUZZIFICAT ION of the input variables. Step 4) APPLY the inference motor rules to get the output variable. Step 5) DEFUZZIFICAT ION of the output variable. Step 6) APPLY u(k) and WAIT until t=(k+1)T . Step 7) kk+1. Step 8) GOT O Step 2. In Fig. 12 the results obtained in one of the experiments is shown. Starting from any initial error, the set point is attained smoothly and with zero oset. There is no overshooting and the settling time is about 6 s. With a thinner choice of the triangular sets around the nal target values, the system is unstable. This is because small variations in the variables will produce large variations in the fuzzied variables. Thus, the resultant command will change too quickly causing instability. An advantage of this kind of controller is its robustness to changes on the system parameters (mass of the crab, mass of the load or length of the rope). Moreover, choosing the rules and the subsets intuitively, once the performance of the system in the previous experiments is known, we get a good response from the controller. The comparison between these results and the previous ones is presented. It can be seen that the oscillations of the load mass are reduced in a shorter time than the preceding experiments. This is because the controller rst tries to correct the oscillations of the load once the crab is near the set point. Due to this, the variations on the applied command are higher than in the preceding experiments. In experiment 2 the command takes the maximal value to attain the crab to the desired position. Near to this point, it decreases slowly sup-

219

FIG. 12

Evolution of the position and the angle of the crane applying the fuzzy controller.

pressing the oscillations of the load. The command evolution in experiment 3 is very soft, but the position of the crab is attained in a longer time. With the fuzzy controller, the command evolves with sharper changes, but the angle is corrected earlier. Due to the value of the command applied, the amplitude of the oscillations varies for experiments with dierent controllers. Thus, high values of angles are obtained in experiment 2, while in experiments 3 and 4 these values decrease to roughly half. It is worth pointing out that with the fuzzy controller, if there were oscillations of the load later, the system retrieves the set point eciently and quickly. 5 CONCLUSIONS In this work a set of control experiments are proposed in order to complement the theoretical formation of the student in digital control and intelligent control courses. They have the opportunity to work on a real system and to learn the diculties that arise from this. The experiments are carried out on a non-linear plant. It is a scale model of an overhead crane. The students can compare the results obtained with both classical and intelligent strategies. For a classical control scheme, a pole placement controller is implemented. The main diculty in a successful implementation of this controller is that the feedback gains obtained may produce limited performance in the real-time implementation. To cope with this problem, a NN is used in order to tune these values automatically. The response obtained was completely

220

satisfactory. Another alternative to this scheme is a fuzzy controller. The knowledge of the students about the system can be expressed by means of rules in the inference process of a fuzzy controller. The main advantage of this is that no model of the plant is necessary for its denition. This fact is interesting because it is not necessary to re-dene the controller in view of changes in the parameters of the system. A comparison of the results obtained with these techniques shows that the introduction of intelligent algorithms improves the performance of the system. Moreover, this reduces the problems presented in classical techniques about the tuning of the controller parameters. The results obtained in these experiments show that the scale model presented in a useful and didactical module for the students. It is a simple plant, but with sucient complexity to make suitable the implementation of both classical and intelligent control strategies. Moreover, the students can observe the response of the system in the plant using the dierent controllers. These experiments are estimated to be carried out in about 15 hours during the courses. The necessary software for the control is provided to the students. They then can dedicate the time to studying in depth the relevant aspects of the controllers and the problems of implementation in real systems. REFERENCES
[1] Atherton, D. P., Nonlinear Control Engineering, Van Nostrand Reinhold, Student Edition (1992) [2] Cheng, C. and Chen, C., Controller design for an overhead crane system with uncertainty, Control Eng. Practice, 4, No. 5, pp. 645653 (1996) [3] Munkata, T. and Jani, Y., Fuzzy systems: An overview, Commun. ACM, 37, pp. 6976 (1994) [4] Moreno, L., Acosta, L., Hamilton, A., Pineiro, J. D., Merino, J. J., Sanchez, J. L. and Aguilar, R. M., Experiments on a d.c. motor-based system digital control course, IJEEE, 32, pp. 163178 (1995) [5] Moreno, L., Acosta, L., Mendez, J. A., Hamilton, A., Pineiro, J. D. and Sanchez, J. L., Stochastic optimal controllers for a d.c. servo motor: applicability and performance, Control Eng. Practice, 4, No. 6, pp. 757764 (1996) [6] Mendez, J. A., Acosta, L., Moreno, L., Hamilton, A, and Marichal, G. N., Design of a neural network-based self-tuning controller for an overhead crane, IEEE Conf. Control Appl., Trieste, Italy, pp. 168171 (1998) [7] Astrom, K. J. and Wittenmark, B., Computer-Controlled Systems, Prentice Hall, Information and System Sciences Ser. (1997) [8] Hush, D. R. and Horne, B., Progress in supervised neural networks, IEEE Signal Process. Mag., pp. 1214 (1993) [9] Lee, C. C., Fuzzy logic in control systems: fuzzy logic controller part I and II. IEEE T rans. Systems, Man Cybernet., 20 (1990) [10] Pedrycz, W., Fuzzy Control and Fuzzy Systems, Wiley (1989)

ABSTRACTS FRENCH, GERMAN, SPANISH Un ensemble dexperiences de controle sur un prototype de pont roulant ` ` Cet article presente un ensemble dexperiences de controle en temps reel appliquees a un systeme non-lineaire. Ce systeme est un pont roulant et les experiences sont appliquees a un modele reduit ` ` ` bon marche de celui-ci. Cet article essaie de comparer les performances des techniques de controle ` classique et intelligente appliquees a ce type de processus.

221 Ein Satz Steuerexperimente auf einem Prototypen eines Laufkrans Ein auf ein nichtlineares System angewandter Satz Echtzeit Steuerexperimente wird vorgestellt. Das System ist ein Laufkran, und die Experimente werden auf einen preisgunstigen mastablichen Modell desselben angewandt. Dieser Bericht versucht, die Leistung sowohl klassischer als auch intelligenter Steuertechniken, die auf diese Anlage angewandt werden, zu vergleichen. Experimentos de control sobre un prototipo de puente de grua Este artculo presenta un conjunto de experimentos de control en tiempo real aplicados a un sistema no lineal. El sistema es un puente de grua. Los experimentos se llevan a cabo en un prototipo de bajo costo. Estrategias clasicas e inteligentes de control han sido aplicadas en la planta.

You might also like