PIRSON’S METHOD FOR STRATIFICATION IN SEVERAL LAYERS EARTH USING NEURAL NETWORK José Francisco Rodrigues 1, Renato Crivellari Creppe 2, José Alfredo Covolan Ulson 3, Paulo José Amaral Serni 4, Rogério Andrade Flauzino 5 1 School of Engineering Bauru/UNESP, Bauru, Brasil, jfranc@feb.unesp.br Scholl of Engineering Bauru/UNESP, Bauru, Brasil, creppe@feb.unesp.br 3 Scholl of Engineering Bauru/UNESP, Bauru, Brasil, ulson@feb.unesp.br 4 Scholl of Engineering Bauru/UNESP, Bauru, Brasil, paulojas@feb.unesp.br 5 Scholl of Engineering Bauru/UNESP, Bauru, Brasil, flauzino@feb.unesp.br 2 Abstract: Systems based on artificial neural networks have high computational rates due to the use of a number of simple processing elements and the high degree of connectivity between these elements. In this paper, a neural approach is developed to aid in designs of electric grounding. More specifically, artificial neural networks are used for mapping of the ground in horizontal layers. Simulation results are presented to demonstrate the validity of the proposed approach. Keywords: Pirson’s method, soil stratification, neural networks. 1. INTRODUCTION The basic purpose of an electric grounding system is to provide safety for human beings as well as to maintain the integrity of equipments protected by such system. Specified the place for installation of the grounding system, it will be give beginning to project through the soil electric resistivity measures () for different spacings (a) or depths among the electrodes. The measurement methods are obtained from results provided by the analysis of Maxwell’s equations from electromagnetism, which are applied to the soil [1]. In the curve of resistivity in relation to the depth, obtained from experimental measurements, is based on whole methodology used in soil stratification, which allows the elaboration of the grounding project. Among the methods of soil stratification in several layers, which use in their calculations the resistivity curve in relation to the depth, are the following: Pirson’s method; Graphic method In this paper, the Pirson’s method is used as basis for the development of the work. This method has been chosen because it represents an extension of the well-known method titled as “two-layers method”. 2. INITIAL CONSIDERATIONS The organization of this paper is presented as follows. In item 3, the neural networks with general considerations and biological neuron is apresented. In item 4, the soil stratification process is formulated in its conventional form. In item 5, the Pirson’s method is introduced and used as basis for the development of this work. In item 6, the neural applications are developed to aid in the electric grounding project. Simulation and results analysis are presented in item 7 in order to validate the proposed application. Finally, in Section 8, the key issues raised in the paper are emphasized and conclusions drawn. 3. NEURAL NETWORKS 3.1. General considerations The purpose of neural computation is to discover the general principles upon which the solutions the human brain devises are based and to apply these principles to computational systems. ANN (Artificial Neural Network) researchers believe that the brain, constituted of neurons, builds its own information processing strategies based on what is commonly called “experience”. ANNs have been used in several applications, among them Pattern Recognition, Sign Processing, Speech Processing, Robotics, and Systems Optimization. The characteristics that make the ANN methodology interesting from the standpoint of problem solving are: a) The capacity to “learn” through experience. This “learning” process is believed to occur when the network manages to map generalized solutions for a class of problem. b) Efficient performance in processes which lack explicit knowledge about how to identify feasible solutions. c) It does not require the identification of possible mathematical models that describe the behavior of the analyzed system. d) High immunity to noise; in other words, the network’s performance does not collapse when information is incorrect or absent. 3.2. Biological neuron The origin of the ANN theory goes back to the mathematical and engineering models of biological neurons. The nerve cell, or neuron, was identified anatomically and described in considerable detail by Spanish neurologist Ramóny Cajal in the 19th century. The neuron has a cellular 1 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 675 Pirson’s Method for Stratication in Serveral Layers Earth Using Neural Network José Francisco Rodrigues, Renato Crivellari Creppe, José Alfredo Covolan Ulson, Paulo José Amaral Serni, Rogério Andrade Flauzino body, or soma, that contains within it most of its organelles. From the soma of each neuron emerge prolongations that are functionally divided into connections called dendrites and axons [2], as illustrated in Figure 1. The neuron is a highly specialized cell that receives electric impulses from its dendrites, processes them in the soma and finally transmits them through its axon (usually a single one) to the dendrites (or even to the some) of other neurons. The connection between an axon of a neuron and a dendrite of another neuron is called a synapse, as illustrated in Figure 2. The synapse is the basic functional unit for the construction of biological neural circuits and involves the position of the plasmatic membranes of two neurons so as to form a point junction (the size of a synaptic junction is smaller than a post-synaptic neuron). to the solution of this problem. The training of an ANN may involve the use of different learning mechanisms: supervised – in which the desired results are supplied entirely, and unsupervised – in which the ANN itself is capable of adjusting its own functioning. Most ANNs use the supervised learning mechanism, which can be considered as the network’s capacity to modify its performance based on a comparison between the response obtained and the desired response. ANNs are trained with experimental learning data. The quality of the data is important for the learning process, exerting a strong influence on network performance. Learning is a previous stage and consists of adjusting the weights and biases of the ANNs whose transfer functions and neural structures have been predefined. Once an ANN has been trained, it is used for the reproduction of output data corresponding to new input data. The ANN’s performance is evidenced in this stage. All ANNs are composed of mathematical elements called neurons, as clearly shown in Figure 3. Fig. 1. Simplified diagram of a neuron. Fig. 3. Diagram of a neuron. The neuron receives, as input, a signal (number), p, multiplied by a weight, w, and a linear term, the bias, b. This input is added and treated by a function, F, of the neuron, producing the signal, i.e., the number a, as output. Expressed in mathematical terms, one has: a F(w.p b) (1) The weight w and the bias b are adjustable parameters of the neuron and the ANN. Now, using a general model of the artificial neuron, as shown in Figure 4, one has: Fig. 2. Simplified diagram of a synaptic connection. An ANN is composed of a great number of processing elements, also called processing units, which are widely interconnected. Each of these links connects only two processing elements, usually in the same direction, and has a value that determines the degree of connectivity between them, called weight of the connection. Thus, the entire processing is carried out distributively among the network’s processing elements, sending its results to other units through the connections between them. For this reason, ANNs are known as distributed and parallel processing (DPP) systems. The manner in which the processing elements are interlinked is called topology or interconnection pattern. Modifications of the synaptic weights of an ANN do not represent a major problem from a mathematical standpoint, but they may be a serious problem in practice, for computational resources are limited and the processing time is proportional to the amount of resources required. The ANN training process should be capable of gradually modifying an initial interconnection pattern so as to adapt it x1, x2 ... xN : are the input signals. w1, w2 ...wN : are the weights. is the polarization associated to the neuron. u is the output of the linear combiner. g(.) is the neuron’s activation function. y is the neuron’s output signal, or its activation state. In this model, the input signals xi are weighted (multiplied) by the respective weights wi (synapses); if the value of wi is positive, the synapse will be excitatory; otherwise, the synapse will be inhibitory. The polarization value and the input signals xi, weighted by the neuron’s respective wi synapses, are then added and the value of u is applied to the activation function g(.) in order to limit the value of the neuron’s output signal y. In mathematical terms, one has: 2 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 676 N u wi .x i θ i 1 (2) y = g(u) (3) 4. SOIL STRATIFICATION PROCESS Considering the characteristics that usually present the soils, by its own geologic formation along the years, the modeling in stratified layers, that is, in horizontal layers, it has produced expressive results that are proved in practice. The mathematical modeling of the soil in two horizontal layers is generally made applying the foundations and theories of the electromagnetism. With the aid of measurements carried out by the Wenner’s Method [2], it is possible to found the soil resistivity referent to the first and second layers, as well as their respective depths. I Soil Plane A Fig. 4. Model of an artificial neuron. Based on equations (2) and (3), one finds that the activation function (g) simply processes the set of inputs received and transforms it into the activation state. The range of variation of the neuron’s output normally lies within the interval of [0,1] or [-1,1]. The most typical activation functions are: h First Layer Second Layer a) Step function: In this type of activation function, one has: 1, if u 0 g(u) 0, if u 0 Fig. 5. Soil into two layer scheme. In the Wenner’s method, an electric current flowing through the point A, in a soil of two layers as observed in Figure 5, generates potentials in the first layer that should satisfy the following equation: (4) b) Sigmoidal function: In this type of activation function, one has: 1 g(u) β u 1e where is the tangent at the point of inflection 2V = 0 From development of equation (7), for any point p into first layer of the soil and distancing r of the current source A, it has a potential Vp defined by: (5) Vp c) Hyperbolic tangent function: 1 e 1 e u u I1 1 Kn 2 2 r 2 2 n 1 r ( 2 nh ) (8) where, Vp is the potential of a point p (belonging to the first layer) in relation to the infinite; 1 is the first layer resistivity; h is the relative depth to the first layer; r is the distance from point p to the current source A; and K is the reflection coefficient defined by: In this type of activation function, one has: g (u ) (7) (6) 2 1 2 1 K 1 2 1 2 1 1 For this activation function, the neuron’s output can assume real negative and positive values in the domain of –1 and 1. (9) where 2 represents the second layer resistivity. 3 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 677 Pirson’s Method for Stratication in Serveral Layers Earth Using Neural Network José Francisco Rodrigues, Renato Crivellari Creppe, José Alfredo Covolan Ulson, Paulo José Amaral Serni, Rogério Andrade Flauzino From analysis of equation (9), it is verified that the variation of the reflection coefficient is limited to interval [1, 1]. Applying this formulation in the Wenner configuration, it is verified that the relative resistivity in relation to the superficial resistivity for a spacing a can be expressed by: (a) 1 4 [ 1 n 1 Kn 1 2n h a 2 Kn 4 2n h a 2 To stratify a soil in two layers, it is first necessary to draw in a graph the curve ( x a) obtained by Wenner’s method. In this curve, the value of the superficial resistivity is not determined, that is, the value of the soil resistivity ( 1) when h/a is zero it is ignored. Therefore, this value should be estimated according to some extrapolative numeric method. Therefore, it is recommended to make several measures through Wenner’s method for small distances. This is justified because the penetration of currents is made predominantly through the first layer. Observing the behavior of the curve ( x a) in figure 6, the sign K is determined by: If the curve is descending, the sign K is negative; If the curve is ascending, the sign K is positive. (10) ] As the variation of the reflection coefficient K is small, and it is limited between [-1, 1], can then to trace a family of curves (a)/1 in relation to h/a, for negative and positive values K, covering its entire variation zone. Soon after, it is chosen an arbitrary value of spacing a1, belonging to the group of measurements, and is calculated (a1)/1 or 1/(a1). From the corresponding theoretical curves, shown in Figure 6, it is obtained the values corresponding of K and h/a. With these values, a table is generated with values of K and h/a (multiplied by a1 previously chosen) that will serve for to draw a (K x h) graph. Another value of a is then chosen and the same procedure is repeated until the drawing another (K x h) graph, which should be done in the same (K x h) graph used for a1. The intersection point of both curves (K x h) will result in the real values of K and h1; consequently, through equation (9), the value (2) is obtained. (a1) 1 (a) Curve for K Varying Negatively 5. PIRSON’S METHOD h a (a Pirson’s method can be seen as an extension of the twolayer method. When dividing the curve ( x a), in ascending and descending spaces, is evidenced that the soil of two layers can be analyzed as a sequence of soil curves equivalent to the two layers. Considering the first space as a soil of two layers, it is obtained the values 1, 2 and h1. When we analyze the second space, firstly determine an equivalent resistivity, seen by the third layer. Thus, it is obtained the resistivity 3 and the depth of the equivalent layer. This methodology is followed for the computation of the resistivity of other layers. In Pirson’s method, the procedure for stratification is similar to that of two layers. Firstly, it is necessary to draw in a graph the curve obtained by the Wenner’s method. Again, in this curve the value of the superficial resistivty is not identified, that is, the value of soil resistivity when h/a is zero is indeterminate. Immediately, the curve ( x a) is divided in ascending and descending spaces. These spaces are identified from the inflection points, that is, where the curve concavity changes the sign. The inflection points can be obtained through the following equation: (a1) 1 (b) Curve for K Varying Positively h a (a Fig. 6 Descending and ascending curves ( x a). The curves drawn for K varying in the negative zone are represented in figure 6(a) and the curves for K varying in the positive zone are in figure 6(b). With base on the family of theoretical curves, figures 6(a) and 6(b), it is possible to establish a method that makes the matching of the curve, measured by Wenner, with a certain characteristic curve. This private curve is characterized by values that define the soil stratification process [1]. d2 da 2 0 (11) The values 1, 2 and h1 are obtained in the same way as presented in the two-layers method. 4 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 678 (12) To contour these imperfections, a new methodology, based on artificial neural networks, is developed to estimate the values that were previously obtained through approximations. The proposed application here for the soil stratification process in electric grounding is composed by four main phases. These phases are defined by: ĥ 2 is the estimated depth in second layer. at is the relative spacing to point of transition in second Phase (I): Obtaining of resistivity measurements in relation to the spacing. Phase (II): Estimation of resistivity values in the first layer through the ANN-1 (Artificial Neural Network-1). Phase (III): Identification of inflection points through the application of a numerical method. Phase (IV): Estimation of depths values h/a through the ANN-2. Considering the second space of the curve ( x a), the equivalent resistivity should be calculated seen from third layer. So, it is estimated the depth of the second layer ĥ2 by using the Lancaster-Jones’ method [1], that is: 2 hˆ2 d1 dˆ2 at 3 where: d1 = h1 is the thickness in first layer. d̂ 2 is the estimated thickness in second layer. space, which is obtained when d2 da 2 0. In Phase (I), the resistivity measures () in relation to spacing (a) are obtained (from experiments) by the Wenner’s method. Consequently, it is calculated the equivalent medium resistivity ̂ 12 , seen from third layer, using the Hummel’s method [3], that is the weighted harmonic average between first and second layer, that is: ˆ 12 d dˆ2 1 d1 dˆ2 1 2 1 (n-1) Z-1 1(n-2) (13) ANN-1 (TDNN) 1(n-p+1) For second space in the curve, the process is similar to that of two layers, presented in the previous section, 1(n) Z-1 1 (n-p) considering now ̂12 as the first layer resistivity. Thus, it is obtained the new estimated values of ̂3 and ĥ 2 . Fig. 7 ANN-1 architecture. The process is repeated for other spaces of the curve and the stratification by Pirson’s method is finished In Phase (II), the resistivity value (1) is obtained through artificial neural networks of type perceptron with time delay (Time-Delay Neural Network–TDNN) [4] as illustrated in figure 7. Based on the resistivity values obtained initially in Phase (I), it is determined other intermediary values that are used as input vectors to training of the TDNN. After the training process, the network is capable to estimate the resistivity value (1), that implies in an estimation of a forward step. The prediction order (p) assumed in the simulations was 5. In Phase (III), the identification of inflection points that define the ascending and descending spaces of the curves were automatically obtained by applying the Newton's method [5] used to find roots of functions. In Phase (IV), the estimation of the depth values h/a are obtained through a multilayer perceptron (MLP) network [5] with only a hidden layer (figure 8). The input variables used in the training of the network were the reflection coefficient K and the value (a1)/1 or 1/(a1). The number of neurons used in the hidden layer was 10, and the amount of vectors belonging to training set was around 1000 vectors. 6. NERUAL APPLICATION In this paper, a neural application is developed to aid the processes of soil stratification. The necessity of reduction in the approximation stages is due to fact of minimizing the imperfections caused in relation to the experimental interpretations involved with the process. These procedures types that are presented in the project of an electric grounding, they generally do not follow an explicit methodology for their estimation processes, but they involve the intuition and the experience of whom executes them. Therefore, some stages of stratification become empiric and the results obtained by any presented methods will not depend on only the efficien Stages as the extrapolation of the curve ( x a) in order to obtain the resistivity value in the first layer are frequently simplified from extension of the curve to the intersection with the ordinates axis. Another process of existent approximation is the standard use of curves, which implies, again, in readings of values directly in graphs. 5 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 679 Pirson’s Method for Stratication in Serveral Layers Earth Using Neural Network José Francisco Rodrigues, Renato Crivellari Creppe, José Alfredo Covolan Ulson, Paulo José Amaral Serni, Rogério Andrade Flauzino Estimating Value 1 – Phase II K ANN-2 (MLP) (a1)/1 Through ANN-1, it is estimated the resitivity value (1 ) in the first layer. This value is obtained through calculation of the resistivity in the spacing point {a = 0}. After the training of TDNN, the value 1 obtained through estimation of a forward step is given by 1 = 8575 .m. h/a 1/(a1) Identification of Inflection Points – Phase III Using the numeric method for identification of inflection points, it was obtained only one inflection point in the curve of figure 9, whose value is given by at = 8.0174 m. Fig. 8. ANN-2 architecture. After computation of these four phases, the soil stratification process can be concluded with the application of the equations defined in item 4. Estimate Values h/a - Phase IV Through ANN-2, it was obtained the values shown in table 2 taking into account the values of a1 and its respective resistivity values. 7. SIMULATION AND RESULTS ANALYSIS To elucidate the proposed application, the results obtained with the simulation of the four phases defined in the previous section are presented. Table 2. Estimated values (first spacing). a1 = 1m; 1/(a1) = 0.7183 Obtaining the Curve ( x a) – Phase I K Through Wenner’s method, it is obtained a group of spacing measures and soil resistivity as presented in table 1. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Table 1. Measurements data Spacing a (m) 1 2 4 8 16 32 Resistivity (.m) 11938 15770 17341 11058 5026 3820 x 10 h(m) K 0.2161 0.4502 0.5943 0.7074 0.8074 0.8894 0.9705 1.0465 1.1246 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 h (m) a2 0.1101 0.2671 0.3942 0.4922 0.5703 0.6443 0.7174 0.7834 h(m) 0.2201 0.5343 0.7884 0.9845 1.1406 1.2886 1.4347 1.5668 2 0 4 1.6 h 1.4 0.2161 0.4502 0.5943 0.7074 0.8074 0.8894 0.9765 1.0465 1.1246 a2 = 2m; 1/(a2) = 0.5475 Drawing the curves K x h (Figure 10) and applying a interpolative method, both them intercepted in the following point: h1 = d1 = 0.6266 m ; K = 0.4439 Figure 9 illustrates the graph of resistivity () in relation to spacing (a). 1.8 h (m) a1 1.2 -2 -4 1 -6 -1 0.8 -0.5 0.5 1 Fig. 10 Curves h x K (first spacing). 0.4 0.2 0 k 0.6 0 5 10 15 20 25 30 Using equation (9), the value 2 is calculated by: 35 a 2 = 22265 .m Fig. 9. Resistivity ( 1 ) in relation to spacing (a). 6 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 680 20 Considering the second space of the curve ( x a), it should be estimated the depth of the second layer. Applying equation (12), we have: -20 2 hˆ2 d1 dˆ2 at 3 ˆ ˆ h2 0.6266 d 2 2 8.0174 5.3449 m 3 dˆ 4.7183 m -40 -60 -80 2 -100 The equivalent medium resistivity can be calculated by the Hummel’s equation (13), that is: -120 -1 1 K 1 K 1 (0.7439) 3 18440 2708 Ω.m 1 (0.7439) 3 ˆ 12 Therefore, the obtained solution, according to the application proposed in this paper, was a soil stratification in three layers, as illustrated in figure 12. Table 3. Estimated values (second spacing). -1.0 0.8754 2.1611 3.5778 4.4342 5.1306 5.6908 6.1551 6.5873 a4 = 16m; (a4)/ ̂ 12 = 0.2726 h (m) h(m) K a4 -0.1 -0.2 -0.3 -0.4 -0.5 -0.6 0.1871 2.9935 -0.7 0.3432 5.4907 -0.8 0.4232 6.7714 -0.9 0.4862 7.7799 7.0035 -1.0 0.5383 Soil Plane 6.0041 m -0.1 -0.2 -0.3 -0.4 -0.5 -0.6 -0.7 -0.8 -0.9 h(m) 1 Thus, it is obtained the value of 3: a3 = 8m, with resistivity (a3) = 11058 .m a4 =16m, with resistivity (a4) = 5026 .m h (m) a3 0.2701 0.4472 0.5543 0.6413 0.7114 0.7694 0.8234 0.5 Fig 11. – Curves h x K (second spacing). Repeating the previous procedures for the second space of the curve, it is generated from the neural network developed (ANN-2) the values presented in table 3, being adopted the values following: K 0 k d1 dˆ2 0.6266 4.7183 18759 m 0.6266 4.7183 d1 dˆ2 8575 22265 1 2 a3 = 8m; (a3)/ ̂ 12 = 0.5997 -0.5 0.6266 m ˆ 12 0 h First Layer 1=8575 .m Second Layer 2=22265 .m 8.6123 3=2708 .m The interception of the curves is shown in Figure 11 whose values are given by: Fig. 12. – Soil stratification in three layers. h2 = 6.0041 m ; K = - 0.7439 From the Figure 12, it is verified that the resistivity obtained to first layer was 1 = 8575 .m with a depth of 0.6266 meters. The resistivity obtained to second layer was 2 = 22265 .m, having a depth of 6.0041 meters. Finally, for the third layer was obtained a resistivity 3 = 2708 .m. Table 4 compares the results obtained by the neural network with the values obtained from Pirson’s method. 7 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 681 Pirson’s Method for Stratication in Serveral Layers Earth Using Neural Network José Francisco Rodrigues, Renato Crivellari Creppe, José Alfredo Covolan Ulson, Paulo José Amaral Serni, Rogério Andrade Flauzino defects of phase A to ground, a fault impedance of five ohms, and a 90-degree angle of fault incidence. This approach also allows for the study to be extended to other types of configuration, for variations in the type of fault and in fault resistance values, as well as to locate the distance of the fault’s point in relation to the substation, factors that other studies have aimed at. Table 4. Results comparison. 1 2 3 H1 H2 Neural Application 8575 .m 22265 .m 2708 .m 0.6266 m 6.0041 m Two Layers method 8600 .m 21575 .m 3103 .m 0.6400 m 5.6400 m ACKNOWLEDGMENTS Thanks the São Paulo State University - Research ProRectory /UNESP for approval this research project. As observed in this table, it is verified that the results obtained by the neural network are very close to the results provided by Pirson’s method. The simplicity and efficiency of the proposed neural application indicates that the methodology can be used as an efficient alternative to estimate the variables related to the electric grounding project. REFERENCES [1] G. Kindermann, J. M. Campagnolo “Aterramento Elétrico”, 3 ed., Editora Sagra-DC Luzzatto, 1995. 8. CONCLUSIONS [2] Z.L. Kovács, Redes Neurais Artificiais, 4 ed., Edição Academica:São Paulo, 2006. After In this paper, an approach using artificial neural networks was developed with the objective of automating the processes of soil stratification for electric grounding, because the conventional processes usually depend on the experience and the planner's sensibility. The application of artificial neural networks was shown efficient to estimate values, which were before obtained through graphs. In function of this fact, the process becomes faster due to the simple and efficient type that the network offers for the processing of data picked from experiments. Thus, the use of artificial neural networks besides providing a new method for the process of soils stratification, it presents results that are similar to the real values. The method utilized here is independent of the fault conditions, and of those of the system. This indicates the easy adaptation of ANNs to this type of study, without requiring the use of aerial electric power line equations to locate the fault point, and requiring only the use of a high performance software program to simulate faults and a good analysis to extract the characteristics of the signals. Another advantage is the easy implementation of this locating algorithm in Operation and Control Centers. The methodology employed here was designed for a configuration containing three sections, simulation of [3] C. M Leite, M. L. Pereira Filho “Técnicas de Aterramentos Elétricos”, 2 ed., Editora Officina de Mydia, 2001. [4] S. Haykin “Neural Networks – A Comprehensive Foundation”, Macmillan. Englewood Cliffs – NJ, USA, 1994. [5] B. Kosko “Neural Networks and Fuzzy Systems – A Dynamical Systems Approach to Machine Intelligence”, Prentice-Hall, Englewood Cliffs – NJ, USA, 1992. [6] M. S. Bazaraa, C. M. Shetty “Nonlinear Programming – Theory and Algorithms”, 3 ed., John Wiley & Sons, New York – NY, USA, 2006. [7] Matlab, The Language of Technical Computing, The MathWorks Inc., version 5.2, Natick, Massachussetts, Usa, 1998. 8 Proceedings of the 9th Brazilian Conference on Dynamics Control and their Applications Serra Negra, SP - ISSN 2178-3667 682