Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract - This paper describes a system modeling and controller design method that uses measured signals. The
method normalizes the measured signals to estimate the transfer functions, and applies the DFT to the normalized
signals. The damping coefficient and dominant frequency are estimated from the Fourier spectrum. Subsequently, the
transfer function is computed using the estimated parameters. This application of the proposed method to a DC motor
position control system enables an improvement of the system response properties.
Keywords: control design, discrete Fourier transform, estimation, modeling, system identification
I. INTRODUCTION
System control technology continues to become more complicated as technical innovation progresses. To date, most
controller designs have been based on a linear model. There is, therefore, a need for an accurate means of mathematical
system modeling when designing a controller using a linear model. It is difficult, however, to mathematically model a
system with the required degree of precision. Furthermore, to optimize the performance of a controller, it is necessary to
minimize any uncertainties in the model, such as non-linearity, time lags, and time changes.
Over the last 20 years, the field of signal processing has rapidly grown in step with advances in network
communications technology. In line with this, system identification (SI) algorithms have been developed to analyze a
system based on acquired data [1-6]. These SI algorithms estimate and evaluate a dynamic system model based on
measured data, and can generally be divided into parametric and non-parametric types. The parametric method estimates
parameters in accordance with the structure of a given model. The non-parametric method, on the other hand, estimates
the transfer functions in the Fourier transform of the inputs and outputs [1].
Most SI algorithms allow us to estimate transfer functions or a state space model from the input and output signals.
Algorithms like empirical transfer function estimation (ETFE) allow us to estimate the transfer functions by calculating
the correlation functions or ratios in the Fourier transform of the inputs and outputs [1, 5]. With subspace algorithms like
subspace state space system identification (N4SID), the vectors of measured inputs and outputs can be used to calculate
extended observability matrices, which are used to estimate the state matrices [1,4].
Reference [6] describes an eigensystem realization algorithm for the estimating model parameters for dynamic systems,
while references [7-9] use the Prony method to estimate complex exponential functions. Reference [10] describes a
multi-order SI, while reference [11] reviews the application of several SI algorithms to the estimation of model
parameters. When the input is an ambient signal or cannot be measured, the model parameters should be estimated based
only on the output. To this end, several output-only SI algorithms are currently being developed [12-14].
The transfer function in a linear time invariant system is defined as the ratio between the inputs and outputs. It is
defined as a Laplace transform of the impulse responses when all of the initial conditions are 0 [15, 16]. The approximate
values of the transfer functions and state matrices can be estimated by applying SI algorithms with discrete signals.
Of all the available algorithms, the discrete Fourier transform (DFT) is used most frequently in signal processing.
Reference [17] describes a method for estimating the dominant parameters included in signals obtained from the Fourier
spectrum magnitude and phase. The method of estimating parameters with the DFT offers several advantages, including
the fact that it allows quick estimation of parameters by using the fast Fourier transform (FFT) algorithms.
This paper describes methods for estimating important parameters from measured signals, as well as how they are
used to estimate the transfer functions of a system. First, this paper describes an improvement to the algorithm proposed
in reference [17], and then proposes a means of more accurately estimating the parameters. Using this method, it is
possible to estimate frequencies and damping coefficients, and use the Fourier transform properties of the exponent
functions to estimate the mode magnitudes. These estimated parameters can also be used to determine the transfer
functions of the system.
The unit step responses can be obtained by normalizing the outputs into steady state values. Subsequently, the unit
impulse responses are computed by differentiating the unit step responses. The transfer functions are calculated using the
parameters estimated from the DFT of the unit impulse responses.
The method proposed for estimating the transfer functions involves only very basic calculations, making the overall
process from model estimation to controller design very simple. Thus, it is advantageous if a system has complex
exponential function properties in its outputs as this allows us to estimate the transfer functions and thus design a
controller. Also, the proposed method allows us to easily design a controller for a system, if an appropriate model is not
currently being used in an existing system. The proposed method was applied to a DC motor system to test the efficiency
and reliability of the method.
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -68
0
u (t )
1
t0
(1)
0
(t )
1
t0
(2)
There is a close relationship between the continuous time unit step and unit impulse. The unit impulse function can be
thought of as being the first derivative of the unit step function, as follows:
du (t )
dt
(t )
(3)
Once the impulse response of the LTI system is known, the system output y(t) with any input, u(t), can be found by
using the transfer function.
B. Fourier Transform and Transfer Function
To derive the relationship between the Fourier transform and transfer function of an LTI system, let us consider an
LTI system with an impulse response g(t). For an input u(t), the output can be represented as:
y (t )
g ( )e s d u (t ) G (s )u (t )
(4)
Where G(s) is a complex function and related to the system impulse response.
The response of any LTI system can be characterized by its impulse response g(t), which is defined as the output when
the input is a unit impulse function (t ) .
When s j is a complex number, G(s) is the transfer function of the corresponding system. Consider the
specific case in which 0 , so that s j , and consequently, e st is of the form e j . The transfer function is referred
to as the frequency response of the system and is given by:
G ( j )
g ( )e j d
(5)
f (t )e jt dt
(6)
If f(t)=g(t), for all , F ( ) G( ) . If function f(t) is the unit impulse response, the frequency response of the
transfer function can be obtained as the Fourier transform of f(t) [18].
C. Discrete Fourier Transform of Complex Exponential Function
If the complex exponential function is assumed to be a sinusoidal signal multiplied by a decaying exponential, then
the discrete form of the complex exponential function can be represented by:
y ( n) Ae n cos( 0 n 0 )
(7)
where A is the amplitude, is the damping coefficient, 0 is the frequency, and 0 is the phase.
The complex exponential function consists of the exponential function, ye (n) e n multiplied by the cosine function
yc (n) A cos(0 n 0 ) .
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -69
(8)
Here, Ye ( ) and Yc ( ) are the Fourier transforms of the exponential function and the cosine function, respectively.
The * indicates the convolution of the two functions.
If frequency / 0 , the DFT of equation (7) using frequency shifting properties can be represented as follows
(Appendix A.1).
Y ( )
Ae j 0 Ye ( 0 ) Ye ( 0 ) Y
2
(9)
If 0 , then 1 , Ye ( 0 ) 0 , and hence the DFT of the complex exponential function can be represented by:
Y (0 )
1
AYe0e j 0 Y 0 0
2
(10)
Here, Ye0 is the Fourier transform of the exponential function at frequency 0 . When the number of data is N, Ye0
can be represented by:
N 1
Ye0
(11)
n0
Thus, the magnitude of the Fourier transform of the complex exponential function Y 0 can be represented by:
Y 0
1
AYe0
2
(12)
(13)
0 arg max (Y )
(14)
(15)
Using this equation, the damping coefficient can be directly obtained from the Fourier spectrum magnitude. However,
a spectrum magnitude corresponding to 70.7% of the peak spectrum should be calculated to estimate the damping
coefficient. If the damping coefficient is small, the peak spectrum will be much greater than the left or right spectrum.
Thus, it can be considered as being linear between the right spectrum and peak spectrum. Let the first left spectrum of the
peak spectrum be Yl , and the frequency corresponding to this be l . The damping coefficient can be induced from the
DFT of the complex exponential function as follows (Appendix A.2):
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -70
(02 l2 )Yl
2l Y20 Yl 2
(16)
With r , the damping coefficient estimated from the right spectrum in the same way, the damping coefficient can be
estimated from:
l
r
(17)
2
This equation gives the average value of the damping coefficient as estimated from the left and right spectra. The
Fourier spectrum is not symmetrical due to the spectrum leaking or low frequencies. Thus, any estimation errors can be
reduced by the average value of the damping coefficient.
C. Mode Magnitude Estimation
Once the damping coefficient and the magnitude of a Fourier spectrum are known, it is possible to calculate the mode
magnitude of a complex exponential function. The mode magnitude from equation (12) can be represented by:
A
2Y 0
Ye0
(18)
where Y 0 is the magnitude of the peak spectrum at 0 . Ye0 is the magnitude of the Fourier spectrum of the
exponential function at 0 .
D. Mode Phase Estimation
The mode phase is the same as the phase of the Fourier spectrum in equation (13), and hence can be represented by:
0 0
(19)
Up to this point, a method of estimating parameters from the DFT was described. Fig. 1 shows the relationship
between the parameters and Fourier spectrum, as described above.
Fig. 1. The relationship between the parameters and the Fourier spectrum
y (t )
y ss
(20)
As the steady state value of the normalized output signal ys (t ) is 1, ys (t ) is a unit step response. The transient and
steady state characteristics do not change because the signals are divided by the steady state value yss .
On the other hand, a unit impulse response can be obtained by differentiating a unit step response as in equation (3).
Thus, the unit impulse response can be represented by:
yi (t )
dys (t )
dt
(21)
where yi (t ) is the output for the unit impulse input. Thus, yi (t ) involves the inherent characteristics of a system.
Consequently, the inherent characteristics of a system can be determined by estimating the important parameters for that
signal.
C. Transfer Function Estimation
The transfer function is defined as being the Laplace transform of a unit impulse response. Thus, if the dominant
parameters of the output signal are known, the transfer function can be estimated. A complex mode consisting of a
damping coefficient and a frequency is defined as follows:
s j0
(22)
A complex exponential function consists of a complex mode, a mode magnitude, and a mode phase. Thus, a system
transfer function can be obtained by estimating these parameters from a measured output signal.
This paper assumes an output signal to be a complex exponential function. Thus, the unit impulse response is also a
complex exponential function. There are usually a number of modes in output signals involving disturbances and noise.
However, if the output of the system is a complex exponential function, the dynamic characteristics of the system affect
one dominant mode, such that the system can be modeled as a second-order system.
Fig. 2 is a block diagram of a second-order system with unit feedback. In this block diagram, R (s ) and Y (s) are an
input and an output, respectively, and E (s) and U (s ) are an error signal and a controller output signal, respectively.
Finally, Gc (s) is a controller transfer function.
n2
s 2 n s n2
2
(23)
The damping coefficient and frequency 0 can be obtained by applying DFT to the output signal. The dominant mode,
damping ratios, and natural frequencies can be computed from these parameters. The natural frequency n is given by:
n 2 02
(24)
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -72
(25)
Thus, the damping ratios and natural frequencies can be computed from a complex mode estimated from an output of
the system.
D. Controller Design
In this paper, the important parameters related to measured signals are estimated, and then they are applied to model
estimation for a simple second-order system. Even if an actual system can only be expressed as a very complicated model,
the system can be modeled as a second-order system when the system responses are in the form of a complex exponential
function.
The controller for a second-order system can be designed by using time and frequency responses. The related design
methods are described in numerous publications [15, 16]. Thus, this paper does not provide a detailed description of the
method of designing a controller for a second-order system, nor other types of controllers.
V. EXPERIMENTS AND DISCUSSION
The methods of parameter estimation and simple modeling proposed in this paper were tested with actual devices.
Before applying the proposed method to an actual system, the method was applied to a test model to determine the
correct solution.
A. Application to a Test System
Assuming that there is a system with a natural frequency of 1.0 [Hz] and a damping coefficient of 0.2, the unit
feedback transfer function of the system will be as follows.
39.513
(26)
G ( s) 2
s 0.402 s 39.513
With the unit step input, it is possible to obtain a unit step response, as shown in Fig. 3(a). By differentiating this
response, the unit impulse response can be obtained, as shown in Fig. 3(b).
39.526
2
s 0.427s 39.526
(27)
TABLE I
COMPARISON BETWEEN THE ORIGINAL AND ESTIMATED PARAMETERS
f[Hz]
[rad/s] n[rad/s]
Original
1.000
6.283
6.286
0.200
0.032
Estimated 1.000
6.283
6.287
0.211
0.034
B. DC Motor Modeling and Control
The experiment was carried out by attaching an inertial load (model 9126-00) to a DC Motor system (model 8211-00).
Fig. 5 shows the equipment. The step inputs are applied to the obtained step responses. These are normalized to give the
unit step response shown in Fig. 6(a). The unit step response is differentiated to obtain the unit impulse response shown
in Fig. 6(b).
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -74
parameter 1.661
10.437
10.447
0.455
0.044
The transfer function can be computed from the estimated parameters, as follows:
G ( s)
109.1398
s 2 0.9193s 109.1398
(28)
Table 3 shows the parameters of the PD controller. In the design of the controller, the gain KP is fixed to 1, and the
value KD is selected to give a damping ratio of 100%.
TABLE III
PARAMETERS OF THE CONTROLLER
KP
KD
1.0000
0.2009
Fig. 8 shows the results of the experiment with the step inputs before and after the application of the PD controller.
With the addition of the PD controller, the system responses become better as the damping ratio improves, overshoot is
greatly decreased, and the rise time and settling time are also improved.
F e j 0 t y (t ) Y ( 0 )
(A.1)
where F indicates the Fourier transform. From this equation, the Fourier transform of the function multiplied by a
cosine function is:
F y (t ) cos( 0t )
1
Y ( 0 ) Y ( 0 )
2
(A.2)
Let the function yc (t ) be a cosine function consisting of frequency 0 and phase 0 . The following equation can thus
be established.
yc (t ) A cos(0t 0 )
(A.3)
A cos(0t 0 )e jt dt
(A.4)
Yc ( ) Ae j0 0 F[cos(0 t )]
(A.5)
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -76
1
Ae j 0 Ye ( 0 ) Ye ( 0 )
2
(A.6)
(A.7)
A 2 2
( 2 02 2 ) 2 (2 ) 2
(A.8)
Then, if the frequency 0 at 0 is determined, the magnitude of the Fourier spectrum can be represented
simply as:
Y 0
A
2
(A.9)
A 2 l2
( 2 02 l2 )2 (2l )2
(A.10)
From equations (A.9) and (A.10), the damping coefficient can be induced as:
(02 l2 )Yl
2l Y20 Yl 2
(A.11)
If the peak, left, and right spectra can be determined from the Fourier spectrum magnitude, the damping coefficient
can be computed by substituting them into equation (A.11).
_________________________________________________________________________________________________
2015, IJIRAE- All Rights Reserved
Page -77