neuro seminar bldc

Intelligent Automation and Soft Computing, Vol. 13, No. 4, pp. 423-435, 2007 Copyright © 2007, TSI® Press

Printed in the USA. All rights reserved

423

DEVELOPMENT AND IMPLEMENTATION OF A FUZZY-NEURAL

NETWORK CONTROLLER FOR BRUSHLESS DC DRIVES

MUAMMER GÖKBULUT*, BEŞIR DANDIL** AND CAFER BAL* *Firat University

Faculty of Technical Education Department of Electronic and Computer Science

**Firat University

Faculty of Technical Education Department of Electrical Science

ELAZIG/TURKEY Email: [email protected]; [email protected], [email protected]

ABSTRACT—In this paper, a Proportional-Derivative and Integral (PD-I) type Fuzzy-Neural Network Controller (FNNC) based on Sugeno fuzzy model is proposed for brushless DC drives to achieve satisfied performance under steady state and transient conditions. The proposed FNNC uses the speed error, change of error and the error integral as inputs. While the PD-FNNC is activated in transient states, the PI-FNNC is activated in steady state region. A transition mechanism between the PI and PD type fuzzy-neural controllers modifies the control law adaptively. The gradient descent algorithm is used to train the FNN in direct adaptive control scheme. Presented experimental results show the effectiveness of the proposed control system, by comparing the performance of various control approaches including PD type FNNC, PI type FNNC and conventional PI controller, under nonlinear loads and parameter variations of the motor. Key Words: Fuzzy-neural network, brushless DC drives, PD type FNNC, PI type FNNC, PD-I type FNNC.

1. INTRODUCTION Brushless DC (BLDC) motors are widely used in many servo applications in robotics, dynamic

actuation, machine tools and positioning devices, due to their favorable electrical and mechanical characteristics such as high torque to volume ratio, high efficiency and low moment of inertia [1,2]. High accuracy is not usually imperative for most electrical drives, however, in high performance drive applications, a desirable control performance must be provided even while the parameters of the motor and loads are varying during the motion. Conventional constant gain controllers used in the high performance variable speed drives become poor when the load is nonlinear and parameter variations and uncertainties exist. Therefore, control strategy of high performance electrical drives must be adaptive and robust [3]. As a result, interest in developing adaptive control systems for electrical drives has considerably increased during the last decade and several adaptive control schemes for brushless DC motors have been proposed based on linear model [4,5]

As is widely known, both fuzzy logic and neural network systems are aimed at exploiting human-like knowledge processing capability. Neural networks have the powerful capability for learning, adaptation and robustness [6]. Fuzzy systems have the ability to make use of knowledge expressed in the form of linguistic rules. Thus they offer the possibility of implementing expert human knowledge and experience. Hence, fuzzy logic control introduces a good tool to deal with the complicated, nonlinear and ill-defined systems. However, their main drawback is the lack of a systematic methodology for their design and tuning of membership functions’ parameters is a time consuming task. Fuzzy-neural systems combine the advantageous of both neural networks and fuzzy logic systems. Neural networks provide connectionist

424 Intelligent Automation and Soft Computing

structure and learning abilities to the fuzzy logic systems, the fuzzy logic systems provide the neural networks with a structural framework with high-level fuzzy IF-THEN rule thinking and reasoning [7].

During the past decade, neural network-based fuzzy systems have become attractive and applied to the control of nonlinear systems using direct adaptive control [8], model reference control [9] and hybrid control techniques [10]. In the field of electrical drives, fuzzy-neural network control is applied to induction motors in [11] and brushless DC motors in [12], and used in [13] to update the control gain of the sliding mode position controller for an induction motor drive. A fuzzy-neural network controller is augmented with an IP controller in [14], PD controller in [15] and an adaptive controller in [16]. Furthermore, a recurrent fuzzy-neural network controller is applied to a dynamical system in [17] and a linear induction motor drive in [18]. PD and PI type fuzzy controllers for the direct torque control of induction motor drives are presented in [19].

In this paper, PD type and PI type fuzzy-neural network controllers are described and then, a PD-I type fuzzy-neural network controller for brushless DC drives is presented. The presented control approach consists of PD type FNNC at transient state and PI type FNNC at steady state to achieve satisfied performance under steady state and transient conditions. A simple and smooth transition mechanism depending on the tracking error is described. The features of the presented FNN controller are highlighted by comparing the various FNN control approaches. The gradient descent algorithm is used to train the FNN on-line in direct adaptive control scheme using the simulation model of the drive, and then trained FNNC is used for experiments. The speed control performance of the proposed FNNC is evaluated under the parameter and load variations of the motor using the experimental setup including the DS-1104 control card. The BLDC drive system is summarized in section 2. The description of the FNN controller is explained in section 3. Experimental results showing the performance of the control system under the various operating conditions are given in section 4.

2. BLDC DRIVE SYSTEM A brushless DC machine is basically synchronous machine with a permanent magnet in the rotor

circuit. The armature windings, which are mounted on the stator, are electronically switched according to position of the rotor. The state space model of a BLDC motor referred to the rotor rotating reference frame is given in Eqn (1). Details of this model can be found in [20,21].

⎥⎥⎥

⎦

⎤

⎢⎢⎢

⎣

⎡

⎥⎥⎥⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢⎢⎢⎢

⎣

⎡

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

⎥⎥⎥⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢⎢⎢⎢

⎣

⎡

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−

+

−

−−

−−−

=

LTdvqv

000J2

P00

0L

10

00L

1

diqi

0100

0J

B0

J8

2P3

00L

R

0LL

R

diqi

dt

d

θλ

ω

λω

θω ω

(1)

Where, id and iq are the direct and quadrature components of the stator current, vd and vq are the direct and quadrature components of the stator voltage, R is the stator resistance, L is the stator inductance in dq reference frame, λ is the magnitude of the flux linkage established by the rotor magnet and ω is the electrical rotor speed, θ is the rotor position, P is the number of poles, TL is the load torque (Nm) and, J and B are inertia and friction of the motor respectively. The model is based on the assumptions that the air-gap is uniform, rotor induced currents are neglected due to the high resistivity of magnet and stainless steel, and the motor is supplied by a three phase sinusoidal input source. Furthermore, BLDC motor has a surface mounted permanent magnet, and thus the d-axis inductance is assumed to be equal to the q-axis inductance. With the vector control (or field oriented control), stator currents are decomposed into the flux and torque components which can be controlled independently. The output of the speed controller is the quadratic current component iq

* (torque current) for the vector controlled BLDC drives, and direct current component id

* (exciting or flux current) is set to zero to avoid demagnetization of the permanent magnet on the rotor.

Block diagram of BLDC drive system is shown in Figure 1, which consists of a BLDC motor loaded with a DC generator, current controlled PWM voltage source inverter, vector control mechanism, and a

Development and Implementation of a Fuzzy-Neural Network Controller for Brushless DC Drives 425

speed control loop. The current control and PWM modulation in Figure 1 can be different depending on the types of current controllers to ensure good current regulation. The control algorithm is realized by a PC including DSPACE-1104 signal processor kit. Coordinate translation, current control and PWM generation are also implemented digitally.

I*q

I*d=0

Fuzzy-neural

networkcontroller

ω∗

PWMinverter BLDC

ω

Ib

θ

Enc. DCrectifier

CurrentController&

PWMModulator

L

C220 V50 Hz

LimiterTa TcTb Ia

Digital filter&

dθ/dt

PC and DS-1104 signal processor kit

+

-

Figure 1. Block diagram of the BLDC drive system.

2. A FUZZY-NEURAL NETWORK CONTROLLER FOR BLDC DRIVES Using the vector control technique and assuming ideal current control, i.e., iq

*=iq and id*=id, the

simplified block diagram of BLDC drive system including the proposed FNN controller can be represented as shown in Figure 2. Where, Te is the electric torque, KT is the torque constant, iq

* is the torque current

command and BJs

1)s(G p += is plant transfer function.

ωI*q

TL

BJs+1

KT

Teω∗ y6

e

2y 1Π

Π

Π

Π

Π

Π

Ν

Ν

Ν

Ν

Ν

Ν

Π

2

3

16

17

18

Σ

3y 4y5y

TransitionMechanism

1y

∫e

e∆+

-+

-

Figure 2. Simplified block diagram of the proposed control system.


First-order Sugeno type fuzzy-neural network controller is adopted for the speed control of brushless DC motor in this study. For a first order Sugeno FNN, a common rule set with two fuzzy if-then rules is given as follows:

2221222221

1211111211

rx qx pf then ,B is x and A is x Ifrx qx pf then ,B is x and A is x If

++=++=

These rules show that the Sugeno type FNNC can be designed as an adaptive PD, PI or PID type controller. PD-FNNC uses the speed error and the change of the error as inputs and produces the torque current command. PD-FNNC has favorable transient response characteristics; however, significant steady state error occurs when the load torque is applied since it has no integral mechanism. Steady state errors can be reduced by long training period for the FNNC; however, this may result in high gains which cause noise problems in control systems. Drawbacks of the PD-FNNC can be overcome with the PI-FNNC which includes an integral compensator at the output. Steady state response of the PI-FNNC is acceptable; however, tuning of the integral gain has a considerable effect on the control performance such as overshoot and settling time. PI type FNNC presented in this study uses the speed error and integral of the error as inputs and produces the torque current command, which eliminates an additional integral compensator. In addition, PD-I type FNNC is proposed in this paper to cope with the transient and steady state problems, which consists of PD-FNNC in transient states, and PI-FNNC in steady state region. A simple transition mechanism depending on the tracking error is provided. The activation mechanism modifies the control law adaptively and thereby achieving high performance control for both transient and steady state. The activation mechanism is as follows:

⎪⎩

⎪⎨⎧

≤−

>−

thr

thr

eifactivatedisFNNCPI

eifactivatedisFNNCPD

ω

ω

The threshold value of the speed ωthr is chosen as 200 rpm, according to maximum speed change at the full load condition. To highlight the merits of the proposed controller, all of the controllers explained in this section are implemented experimentally under various load conditions.

2.1 Description of FNN Controller A six-layer fuzzy-neural system based on Sugeno fuzzy models as shown in Figure 2 is adopted to

implement the fuzzy-neural network controllers in this study. The FNN controller comprises of input (y1), membership (y2), rule (y3), normalization (y4), inference (y5) and the output (y6) layers. FNN inputs are defined in the first layer. Speed error, the change of the error and the integral of the error are the inputs of the FNN controller. The input layer transmits input signals to the next layer. For every node i in this layer, inputs and output are expressed as,

1inet1

iy,1ix1

inet

(t)e3x,e(t)2x,e(t)1x

==

=== ∫∆ (2)

where, 1inet is the net input and 1

ix is the ith input of the first layer. The second layer includes the

membership functions to express the input/output fuzzy linguistic rules. Membership layer calculates the degree of membership functions for the input values. Sigmoid functions are adopted for the small and large values of the inputs. Generalized Bell functions are used for the other input values. For the jth node in this layer,

( ))2

jexp(net1

1)2j(net2

jf2jy,ijm2

ixij2jnet

+==−−= σ (3)


2jnet1

1)2j(net2

jf2jy,

i2b

ij

ijm2ix

2jnet

+==

−=

σ (4)

where, ijσ , ijm and bi are the parameters of the membership functions in the jth node of the ith input

variable 2ix . As the values of these parameters change, membership functions varies accordingly, thus

exhibiting the most suitable forms of membership functions for fuzzy set. The weights between the input and membership layer are assumed to be unity. Three membership functions are used for each input. Therefore, 9 nodes in the membership layer are included. The third layer of the FNNC represented by Π includes the fuzzy rule base and the nodes in this layer determine the fuzzy rules. In total, 27 nodes in this layer are required for three inputs FNNC; however, 18 rules are sufficient for this study, depending on the selected controllers such as PD-FNNC or PI-FNNC. Consequently, each node in the rule layer multiples the two fixed weight incoming signals from the membership layer. The output of this layer are given as,

3knet3

ky,3jx

j3jkw3

knet == ∏ (5)

where, 3jkw is assumed to be unity and 3

jx represents the jth input to the layer 3. The fourth layer labeled

N is the normalization layer and it determines the certainty of the fuzzy rules. It calculates the ratio of the kth rule’s firing strength to the sum of all rule’s firing strength:

∑

=

k4kx

4kx4

ky (6)

The fifth layer which has adaptive nodes gives the certainty of rules. For the first-order Sugeno fuzzy-neural systems, the output of this layer denoted by the numbered rectangles is the product of the normalized firing strength of layer 4 and an adaptive node function given by

)kreakqek(p4

ky5ky ++= ∆ (7.a)

)krebkqek(p4

ky5ky ++= ∫ (7.b)

which depends on active controller, i.e., PD or PI type FNNC. Where, pk, qk and rk are the output function parameters of the FNNC to be determined by training. The sixth layer is the output layer and produces an output as the sum of all incoming signals. Hence, the output of this layer

∑==k k

*qi

5y6oy (8)

is the torque current command for the motor. FNN controller has two groups of modifiable parameters which are the membership and the output functions parameters. Parameters of the FNNC are modified using the backpropagation algorithm to minimize the performance index E

2e21E = (9)


where the speed tracking error is ωω −= *e . ω∗ is the reference speed and ω is the actual shaft speed. For the output parameters, the gradient of the performance index can be derived as follows:

ex

xey

py

yy

pE

k

4k

4k14

k1

k

5k

5

61

k ∑==

∂∂

∂∂

=∂∂ δδδ (10)

ex

xey

qy

yy

qE

k

4k

4k14

k1

ak

5k

5

61

ak

∆δ∆δδ∑

==∂∂

∂∂

=∂∂

(11.a)

∫∑∫ ==∂∂

∂∂

=∂∂ e

xx

eyqy

yy

qE

k

4k

4k14

k1

bk

5k

5

61

bk

δδδ (11.b)

∑

==∂∂

∂∂

=∂∂

k

4k

4k14

k1

k

5k

5

61

k xx

yry

yy

rE δδδ (12)

where, local error 1δ is defined as 61

ye

eE

∂∂

∂∂

∂∂

=ω

ωδ . 6y∂

∂ωin this equation should be calculated using

the motor dynamics. The gradient of complex nonlinear dynamic systems cannot be found explicitly. In this study, instead of the system gradient, delta adaptation rule proposed in [9] is used

eAe1 &+=δ (13)

where A is a positive number. In similar way, the gradient of the cost function for the parameters of the membership functions can be written as,

)y(1)ym(xy

x

x-xfE 2

j2jij

2i

j

3k2

k

4k

4k

k

4k

k1

ij

−−

⎟⎠

⎞⎜⎝

⎛=

∂∂ ∏

∑

∑δ

σ (14)

)y(1)y(y

x

x-xf

mE 2

j2jij

j

3k2

k

4k

4k

k

4k

k1

ij

−−

⎟⎠

⎞⎜⎝

⎛=

∂∂ ∏

∑

∑σδ (15)

where fk is first order linear output function of the FNN in Eqn. (7).

3. EXPERIMENTAL RESULTS The fuzzy-neural control system proposed in this study is implemented using the dSPACE-DS1104

signal processor control card and the results are compared with various control approaches, under the various load conditions and mechanical parameter variations of the motor. The nameplate and parameters of BLDC motor are: Permanent Magnetic Brushless Servomotor, Type: BLQ 42S30, R=11.05 Ω, L=21.5 H, J=0.0001kg.m2, P=6, nominal torque Tn=0.9 Nm, counter emf Ez=39 Volt/1000 rpm, nominal supply voltage 220 V, nominal current In=1.7 A maximum speed n=3000 rpm, maximum torque and current Tmax=3.5 Nm Imax=6.8 A. The parameters of the current controllers are tuned to obtain a favorable current


regulation. The control algorithm, current control and PWM modulation are realized by a PC with dSPACE-1104 control card. dSPACE-DS1104 control card allows user to construct the system in MATLAB/Simulink and then to convert the model files to real-time codes using the Real-Time Workshop of the MATLAB/Simulink and Real-Time Interface (RTI) of dSPACE-DS1104 control card [22]. FNNC is trained using the simulation model which is verified by the experimental data obtained from the motor and then the trained FNNC is used in experiments as shown in Figure 3. The RTI software which is called as dSPACE RTI1104 comprises of four sub-libraries, including some sub-blocks which provide the connection between Simulink and physical equipment such as; digital-analog converter, analog-digital converter, incremental encoder interface and various pulse with modulation units. These blocks are added to Simulink libraries by RTI. Real time values of the physical systems’ variables can be assigned to the user defined variables using the dSAPCE-Control Desk Developer (CDD) software. Thus the graphical user interface can be designed by the user to observe the real time values of the variables or to change the input variables such as reference speed.

Vq

Vd

Iqref

w

0

wref

abc --> d-q

In1Out1

Speed Calc

RT I Data

1PWMCnt

PI

PI_q

PI

PI_d

NFC

Tz

z-1Integrator

0Idref

ADC

Ib

ADCIa

Enc position

Enc delta position

ENC_POS

Vd

Vq

teta_e

da

db

dc

Duty Cycle

Duty cy cle a

Duty cy cle b

Duty cy cle c

PWM Stop

DSP_PWM

eiq

eid

iq

id

Figure 3. MATLAB/Simulink implementation of experimental blocks using the RTI of the DS-1104 control card.

Some experimental results are provided to demonstrate the effectiveness of the proposed fuzzy-neural controller under different operating conditions of the motor. For comparison, PD and PI type FNNC mentioned in this paper are tested for the same operating conditions. Furthermore, results of the conventional PI controller are presented; however, parameters of the PI controller are tuned for each load condition to obtain the best result. It is clear that the constant gain controllers do not provide favorable results for different load conditions. To test the loads characteristic of the controllers, a DC generator which produces torque proportional to the speed is coupled to the motor. Results are obtained by the designed graphical user interface.

Figure 4.a shows the input membership functions before the training process for all controllers. Figure 4.b, 4.c and 4.d show the input membership functions after the training for the P, D and I controllers respectively. In this study, three membership functions are used for each input and this yields nine rules for PD and nine rules for PI controllers. As shown in experimental results, the desired control performance is obtained using this membership functions. If the numbers of input membership functions (number of rules) are increased, the structural and computational complexity increases dramatically and this is not desired for real control applications. Figure 5 shows the change of the objective function during the training process.

In electrical drives, the most important parameters affecting the sensitivity are inertia (J) and viscous friction (B). The main aim of this study is to reduce the sensitivity of the system to these parameters’ variations and external disturbances. Therefore, the robustness of the control system is tested under inertial-frictional variations and unwanted disturbances.


-30 -20 -10 0 10 20 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-30 -20 -10 0 10 20 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-30 -20 -10 0 10 20 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

-30 -20 -10 0 10 20 300

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 4. Initial and final membership functions of the proposed fuzzy-neural network controller.

First, inertia of the motor is increased approximately thirteen times of the nominal inertia and friction of the motor is increased approximately six times of the nominal friction The experimental results given in Figure 6 displays the motor speed and torque current for the tuned PI, PD, PI and PD-I FNNC. The tuned PI produces 7% overshoot and no steady state error (Figure 6.a). As expected, PD-FNNC controller has a steady state error which is negligible and can only be seen by zooming with no overshoot (Figure 6.c). PI-FNNC produces 11% overshoot, no steady state error and longer settling time (Figure 6.e). As shown in Figure 6.g, a better performance is obtained from the PD-I type FNNC which provides no overshoot and no steady state error. The PI type controllers have approximately the same rise time less than PD-FNNC; however, they have an overshoot resulting in longer settling time. On the other hand, the proposed PD-I FNNC provides a better performance as seen in Figure 6.g. It should be noted that the noise in torque currents of the integral based controllers is less than the derivative based controllers as shown in Figure 6(b), (d), (f) and (h).

In the second experiment, the controllers are tested under increased inertia and friction with the speed dependent load produced by the DC generator. The maximum value of the load is 90% of the nominal value. The speed tracking responses are shown in Figure 7. It is observed that the PD-FNNC produces

(c) (d)

(b)(a)

Initial membership functions for all controllers Final membership functions for P controller

Final membership functions for D controller Final membership functions for I controller


0 50 100 150 200 250 300 350 40010

-2

10-1

100

101

102

Figure 5. Variation of the root mean squares error.

significant steady state errors since it has no integral mechanism (Figure 7.b), and the tuned PI controller has 3% overshoot (Figure 7.a) and PI-FNNC has 6% overshoot (Figure 7.c). The proposed PD-I FNNC provides a better performance with no steady state error and no overshoot as seen in Figure (7.d). As the system is loaded, rising times for the each algorithm increase compared to Figure 6.

Finally, Figure 8 shows the performances of the controllers when 90% load disturbances is applied. It can be seen from the figures that the PD-I type FNNC is more robust than the others, when the step load is applied. The PD-FNNC produced larger steady state error when compared to the no load condition. In order to have a better view, similar speed tracking performances are shown in Figure 9 with enlarged y axis.

5. CONCLUSION In this paper, a PD-I type fuzzy-neural controller is proposed for a BLDC drive to improve the control

performance of the drive system at transient and steady state conditions. An alternative approach is presented for fuzzy-neural control and the effectiveness of the proposed control system is shown experimentally under the parameter and load variations. From the experiments, it can be concluded that the PD type FNNC produces significant steady state errors since it has no integral mechanism, PI type FNNC has an overshoot and more settling time and, even when the conventional PI controller is tuned for each load condition, it still produces an overshoot.

REFERENCES 1. M.A. Rahman and P. Zhow, “Analysis of brushless permanent magnet synchronous motors”, IEEE

Transactions on Industrial Electronics, 43, pp.256-267, 1996. 2. A. Consoli and A. Raciti, “Analysis of permanent magnet synchronous motors”, IEEE Trans. on

Industry Ap., 27, pp.350-354, 1991. 3. Z.H. Akpolat and M. Gokbulut, “Discrete time reaching law speed control of electrical drives”,

Electrical Engineering, 85, pp.53-58, 2003. 4. M.A. El-Sharkawi, “Development and implementation of high performance variable structure

tracking control for brushless motors”, IEEE Trans. on Energy Conversion, 6, pp.114-119, 1991. 5. A.A. El-Samahy, M.A. El-Sharkawi, and S.M. Sharaf, “Adaptive multi-layer self-tuning tracking

control for DC brushless motors”, IEEE Trans. on Energy Conversion, 9, pp.311-316, 1994. 6. S. Haykin, Neural networks, A comprehensive foundation, Macmillan publishing, New York, USA,

1994.

Epoch number

Root mean squares error


Figure 6. Experimental results for the increased inertia and friction. (a) speed and (b) torque current tracking performance of the PI controller tuned for this load. (c) speed and (d) torque current tracking performance of the PD-FNNC (e) speed and (f) torque current tracking performance of the PI-FNNC. (g) speed and (h) torque current tracking performance of the PD-I FNNC.

Speed (rpm)

Speed (rpm)

Speed (rpm)

Speed (rpm)

Torque current

Torque current

Torque current

Torque current time (s) time (s)

time (s) time (s)

time (s) time (s)

time (s)time (s) (g) (h)

(e) (f)

(c)

(b)

(d)

(a)


Figure 7. Speed tracking performance (a) PI controller tuned for this load, (b) PI-FNNC (c) PD-FNNC, (d) PD-I FNNC.

7. J-SR. Jang, C-T. Sun, and E. Mizutani, Neuro-Fuzzy and Soft Computing, Prentice Hall, Upper Saddle River NJ, USA, 1997.

8. F. Da and W. Song, “Fuzzy neural networks for direct adaptive control”, IEEE Transactions on Industrial Electronics, 50, pp.507-513, 2003.

9. Y.C. Chen and C.C. Teng, “A model reference control structure using a fuzzy neural network”, Fuzzy Sets and Systems, 73, pp. 291-312, 1995.

10. B. Lazerini, L.M. Reyneri, and M. Chiaberge, “A neuro-fuzzy Approach to hybrid intelligent control”, IEEE Transactions on Industry Applications, 35, pp. 413-425, 1999.

11. B. Dandil, “Robust speed control of induction motors using neuro-fuzzy controllers”, PhD Thesis, Firat University, 2004.

12. A. Rubai, D. Ricketts, and D. Kanham, “Development and implementation of an adaptive fuzzy-neural network controller for brushless drivers”, IEEE Transaction on Industry Applications, 38, pp.441-447, 2002.

13. R.-J. Wai and F.-J. Lin, “Fuzzy neural network sliding mode position controller for induction motor drive”, IEE Proc. Electrical Power Appl. 146, pp.297-308, 1999.

14. F.J. Lin, R.J. Wai, and H.P. Chen, “A PM synchronous servo motor drive with an on-line trained fuzzy neural network controller”, IEEE Transactions on Energy Conversion, 13, pp. 319-325, 1998.

15. M.J. Er and Y. Gao, “Robust adaptive control of robot manipulators using generalized fuzzy neural networks”, IEEE Transactions on Industrial Electronics, 50, pp. 620-628, 2003.

16. F.J. Lin and R.J. Wai, “Adaptive fuzzy-neural network control for induction spindle motor drive”, IEEE Trans. on Energy Conversion 17, pp.507-513, 2002.

17. C.H. Lee and C.C. Teng, “Identification and control of dynamic systems using recurrent fuzzy neural Networks”, IEEE Transactions on Fuzzy Systems, 8, pp.349-366, 2000.

Speed (rpm) Speed (rpm)


time (s)time (s)

time (s) time (s)(c) (d)

(a) (b)


Figure 8. Speed variations for the 0.9 p.u. step load disturbance with increased inertia and friction: (a) PI controller tuned for this load, (b) PI-FNNC , (c) PD-FNNC, (d) PD-I FNNC.

18. F.J. Lin and R.J. Wai, “Hybrid control using recurrent fuzzy neural networks for linear induction motor servo drive”, IEEE Transactions on Fuzzy systems, 9, 2001, pp.102-115, 2001.

19. Y.S. Lai and L.C. Lin, “New hybrid fuzzy controller for direct torque control induction motor drives”, IEEE Transactions on Power Electronics, 18, pp.1211-1219, 2003.

20. P. Pillay and R. Krishnan, “Modelling, simulation and analysis of permanent magnet motor drives, Part-1: The permanent magnet synchronous motor drive”, IEEE Trans. on Industry Ap., 25, pp. 265-273, 1989a.

21. F. Pillay and R. Krishnan, “Modelling, simulation and analysis of permanent magnet motor drives, Part-2: The brushless DC motor drive”, IEEE Trans. on Industry Ap., 25, pp.274-279, 1989b.

22. DS1104 Controller Board Features, dSPACE GmbH,Germany, 2003.

ABOUT THE AUTHORS M. Gökbulut received the B.Sc and M.Sc degree in electrical and electronics education from Gazi University, Ankara, Turkey, in 1980 and 1988 respectively, and the PhD degree in electronics engineering from Erciyes University, Kayseri, Turkey, in 1998. He is currently associate professor and the head of the department of electronics and computer education at Firat University, Faculty of Technical Education, Elazig, Turkey. His research interests include research and development of intelligent control systems (neural networks and fuzzy logic), electrical drives control and adaptive control systems.



time (s)time (s)

time (s) time (s)(c) (d)

(a) (b)


Figure 9. Speed variations for the 0.9 p.u. step load disturbance with increased inertia and friction: (a) PI controller tuned for this load, (b) PI-FNNC , (c) PD-FNNC, (d) PD-I FNNC.

B. Dandil received the B.Sc, M.Sc and PhD degrees in electrical and electronic engineering from Firat University, Elazig, Turkey, in 1992, 1998 and 2004, respectively. Currently, he is a lecturer at the same university. He research interests include high performance control of induction motors, identification and control of nonlinear systems including artificial neural network and fuzzy control.

C. Bal received the B.Sc and M.Sc degrees in electronic and computer education from Firat University, Elazig, Turkey, in 1997 and 2002 respectively. He is currently working towards to the Ph.D degree in electrical and electronic engineering and he is a lecturer at the same university. He research interests include speed sensorless control of electrical drives, artificial neural networks and fuzzy control, educational tools and virtual laboratories.

time (s)

time (s)

time (s)

Speed (rpm)

Speed (rpm)

Speed (rpm)

(d)(c)

(a) (b)

Speed (rpm)

time (s)

neuro seminar bldc

Documents