Professional Documents
Culture Documents
ROC(A)
Vol. 25, No. 2, 2001. pp. 107-114
ABSTRACT
This paper considers the problem of estimating the speed of an induction motor using a non-linear identification
technique. A discrete-time non-linear identification approach, NARMAX (Non-linear Auto Regressive Moving Average
model with eXogenous inputs), is presented to describe a polynomial modeling between the speed and voltages of
an induction motor for estimating the motor speed. This approach is useful for identifying the non-linear relationship
between the speed and voltages of an induction motor. The feasibility and accuracy of the proposed method are verified
through laboratory tests. This approach will replace the speed sensor used in a speed control closed loop motor system.
In addition, a future robust controller design based on the NARMAX model will apply an innovative and simplified
speed control algorithm for an induction motor. The last research is now underway.
− 107 −
J.J. Liu et al.
techniques, including those which employ linear filtering to ε(k−1), ..., ε(k−ηε)] + ε(k), (1)
estimate linear transfer function (Schoukens, 1990), to esti-
mate the physical parameters (Moons and Moor, 1995), and where y(k), u(k) and ε(k) represent the system output, input,
to estimate the coefficients of linear transfer function based and prediction error, respectively. Also, l is the degree of non-
on measurements of the magnetic force and speed (Gahler and linearity, α is a constant dc level, Fl[.] is some vector valued
Herzog, 1994; Lee et al., 1994), etc. Here, a non-linear iden- non-linear function, and ηu, ηy and ηε represent the number
tification technique, NARMAX (Non-linear Auto Regressive of lags in the input, output and prediction error, respectively.
Moving Average model with eXogenous inputs), is employed The prediction error term ε(k), defined as ε(k) = y(k) − y(k) ,
(Leontaritis and Billings, 1985) to model the relation between is included in the model to accommodate noise, where y(k)
the speed and voltages of an induction motor. is the prediction output. Expanding Eq. (1) by defining the
Our research on developing an identification technique function Fl[.] as a polynomial of degree l gives a representation
for estimating the speed of an induction motor was divided of all the possible combinations of y(k), u(k) and ε(k) up to
into two steps. The objectives of the first step included: (1) degree l. For example, the current output can be presented
choosing the proper parameters of the motor as inputs, cor- as
responding to the output, i.e. speed, by analyzing the governing
equation of the motor, (2) designing and constructing an y(k) = α + θ1y(k−1) + θ2u(k−1) + θ3u(k−1)y(k−1)
induction motor system in order to obtain input/output data,
(3) modeling a NARMAX equation using designated input/ + θ4u(k−1)ε(k−1) + θ5ε(k−1) + ε(k),
output data, and (4) validating the NARMAX model. The
second step involved designing a robust controller for con- by defining p1(k) = y(k−1), p2(k) = u(k−1), p3(k) = u(k−1)y
trolling the speed by adjusting the inputs. This was done by . (k−1), p4(k) = u(k−1)ε(k−1), p5(k) = ε(k−1), p0(k) = 1, and
transforming a non-linear, difference and polynomial NAR- θ0 = α. If N input and output measurements are available,
MAX model in the time domain into the frequency response and if there are M terms in the model, then the above equation
in the frequency domain using an FRF technique. can be written in a matrix form as
This paper will focus on the attained objectives in the
first step and will discuss the research, traditional control of Y = pθ + ε , (2)
speed and sensorless drive motors, and the application of the
NARMAX model to speed estimation. The procedure for where
constructing and validating a NARMAX model of the speed
and voltages of a motor is demonstrated through an experi- YT = [y(1) y(2) ... y(N)]
mental case study. The results of the NARMAX model shown
in this paper show that it can replace the speed sensor, such θ T = [θ0 θ1 ... θM]
as tachometer or encoder in a closed loop speed control motor.
Based on the NARMAX model, an innovative speed control ε T = [ε(1) ε(2) ... ε(N)]
algorithm for induction motors will be presented in the near
future when the robust controller design is completed. Before
this, the motor speed is still controlled by the traditional way.
p 0(1) p 1(1) p M(1)
p 0(2) p 1(2) p M(2)
II. NARMAX Method
. . . .
p= ,
Successful system identification requires correct model- . . . .
ing. In this paper, a NARMAX modeling of identification . . . .
is proposed. For a non-linear system, representing the current p 0(N) p 1(N) p M(N)
output by mapping the previous input, outputs and prediction
error can be done precisely and efficiently using a NARMAX
model. where p represents a term in the NARMAX model, and θ
represents unknown parameters to be estimated. The param-
1. NARMAX Model eter vector θ in Eq. (2) can be estimated using some well-
known methods, such as a least-squares-based or prediction
A wide range of discrete time multiple variable non- error method, Choleski or U-D factorization, the Q-R algorithm,
linear stochastic systems can be represented by the following singular value decomposition or principle component regression.
NARMAX model: The present study employed an orthogonal estimator algorithm
to conduct parameter estimation (Korenberg et al., 1998; Bil-
y(k) = α + Fl[y(k−1), ..., y(k−ηy), u(k), ..., u(k−ηu), lings and Leontaritis, 1981, 1982).
− 108 −
Nonlinear ID of Motor Speed
where
y 2(k) = Σ g 2i w 2i (k) + ε 2(k) ,
i =1
(8)
N
where ε(k) is assumed to be a zero mean white noise sequence
Σ w i(k)p j(k) which is not correlated with the input and output data records.
α ij = k =1
N
Σ w 2i (k)
k =1
The mean-squared prediction error will be the maximum error
when no terms are included in the model, that is nθ = np +
nε = 0. As a result, we have [ε(k)2]nθ = 0 = y2(k). Equation
for j = 1, 2, ..., nθ; i = 1, 2, ..., j − 1, j, (8) shows that the reduction in the mean squared error, achieved
by including the ith term, giwi(k), in the auxiliary model of
and in this case the orthogonality property holds, i.e., Eq. (3), is g 2i w 2i (k). Expressing this quantity as a fraction of
the total mean squared error yields the ERR for the ith term
wi(k)wj(k) = 0, i ≠ j. (5) as
− 109 −
J.J. Liu et al.
R s R r(1 – σ ) L mR r
– – 0
σL s σL s σ L sL 2r
i γs i γs
Rs
p i δs = 0 – 0 i δs
Fig. 1. Vector control diagram of the induction motor. σL s
λ γr λ γr
L mR r Rr
0 –
Lr Lr
Therefore, a strategy for evaluating the correctness and validity
of the model is necessary. If validation shows that the model
is not good, then some of the design variables should be
changed, and the identification procedure should be redone. v γ′ s
Two model validation strategies, one using the model predicted 1
+ v′ , (12)
output and the other using a model validity test, will be σ L s δs
discussed in the following. 0
For a non-linear system, the residuals usually can not λγr, λδr : γ-δ-axis rotor flux;
be predicted from the linear and nonlinear correlation of past
inputs and outputs. This will be true if the following correlation Ls, Lr : stator, rotor inductance;
tests are passed (Billings and Chen, 1989):
σ : total leakage factor;
φεε(τ) = δ(τ); φu ε (τ) = 0; φuε(τ) = 0; φεεu(τ) = 0;
2' 2'
− 110 −
Nonlinear ID of Motor Speed
− 111 −
J.J. Liu et al.
Fig. 5. (a) System output and model predicted output. (b) Difference between
the outputs predicted using the system and model.
y(k) = 0.21u(1,k−1)u(2,k−1)+0.09u(1,k−2)u(2,k−1)
+0.038u(1,k−3)u(2,k−1)+0.016u(1,k−4)u(2,k−1)
+0.007u(1,k−5)u(2,k−1)+0.003u(1,k−6)u(2,k−1)
+0.001u(1,k−7)u(2,k−1)+0.014u(2,k−1)0.14ε
Fig. 6. Model validity test results obtained in the experimental study. (the
×(k−9)+0.12ε(k−2), upper left figure is first then right, and the lower figure is next,
according to the order φεε, φu1ε, φu2ε, φε2u1, φε2u2, φu12'ε, φ(u1u2)'ε,
where u(1, k−a) is the input v γ′ s of lag a, and u(2, k−b) is the φu22'ε, φ(u1u2)'ε2, φ(u1u1)'ε2, φ(u2u2)'ε2)
− 112 −
Nonlinear ID of Motor Speed
and the auto-correlation between errors. The dash lines for to design controllers in frequency domain, and the design
the upper and lower bounds imply a deviation band of 5%. satisfies a pre-specified output tolerance in time domain.
This shows that the model validity test results are all inside There are three major disadvantages in applying the
the 95% confidence band and indicates that the fitted model vector control theory to induction motors to control speed:
is almost unbiased and has correctly captured the system (1) For the purpose of speed control, three controllers are
dynamics. The excellent output predicted by the model and employed to keep the rotor flux and to control the stator
validity test results reveal that the multi-input single output current according to the speed. As a result, a higher
(MISO) second degree NARMAX model can sufficiently drive cost and more complex structure are added.
represent the speed dynamics of an induction motor. (2) Rotor flux is hard to be kept at a steady value by a
controller due to noise. Therefore, the linearity of the
V. Discussion and Conclusions vector control destroys.
(3) At low speed, the current and flux have low values and
In Section IV, we presented the parameters of the are easily disturbed. The behavior of an induction motor
NARMAX model for the induction motor assigned to the at low speed deviates greatly from expectations.
system. How can reasonable parameters for the NARMAX Therefore, the designs of the controllers, the predictions of
model, l, ηy, ηu and ηε, be assigned in physical applications? the electric parameters and the responses of the whore system
The non-linear order l depends on the non-linearity of the are inaccurate. Consequently, the performance of the motor
strong or weak intensity, and the values of ηy, ηu and ηε depend deteriorates.
on the available input, output and error. The other requirement For the sake of improving the disadvantages of the
for model parameter assignment is numerical accuracy. Simply sensorless control motor, an identification strategy has been
increasing the number of parameters in the polynomial ex- presented here in which voltages and speed are used as inputs
pansion to achieve the desired prediction accuracy will, in and outputs to create a model. Modeling by means of data
general, result in an excessively complex model and possibly collection over a wide range of speed improves the accuracy,
in numerical ill-conditioning. Simulation has shown that, reduces the complexity, increases the immunity to noise, and
usually, less than ten key terms in the NARMAX model are fewer controllers are needed. The main contribution of this
dominant, and that the remainder can be deleted with little paper has been to extend the NARMAX methodology to the
distortion subsequent in the prediction accuracy of the model application of speed estimation for induction motors. The
(Billings and Fadzil, 1985). Here, in this motor system, if results obtained here have demonstrated that the estimated
Eq. (12) is linear, and the measurement Eq. (13) is non-linear model can sufficiently capture all the system dynamics over
with two states multiplied by each other, then the non-linear the desired operating range. In practical applications, the
order assigned to l = 2 is appropriate. If the available input number of times that the speed sensor is measured is the lag
and output measurements are sufficient for a NARMAX assigned by the parameter in the NARMAX model, i.e., ηy
model, then the key terms can be determined through numeri- = 9 in this experimental study, so the sensor is discarded. In
cal constraint with ERR ranking. other words, the sensor is employed only at the beginning or
The popular sensorless drive motor has been realized before the beginning, and the NARMAX model with a
by applying the estimation theory to the governing equation measurement of motor voltages can be used to estimate the
in the state space. The estimation theory includes least-squares, remaining speeds. After all, voltage measurement is easier
maximum likelihood, Kalman filtering etc. The governing than speed measurement. Furthermore, for a black-box motor
equation is corrupted by noise, as mentioned in Section I, and system or a more complicated one, not much a priori infor-
the speed estimation obtained using the estimation theory also mation is available, but a NARMAX model can represent the
deteriorates. The proposed method using the NARMAX system by simply starting with the needed inputs and outputs.
model to estimate the motor speed constructs the governing A design of a closed loop speed controller based on the
equation based on experimental data, including the effect of NARMAX model is being studied now, and a PC-based
noise. The prediction error in the NARMAX model is defined controller can possibly be used to integrate the whole operating
as the difference between the real and model output, and this process. This paper has discussed the value of the model based
term takes into consideration noise in the sense of its physical on validation test results. The NARMAX model approach has
meaning. The mathematical NARMAX model can represent been shown capable of accurately estimating the speed through
a physical motor accurately to estimate the speed. non-linear identification. Eventually, a robust controller will
A robust controller design is the final goal of this research. be designed and will be presented in a forthcoming paper.
Glass and Franchek (1999) gives a promise of success of the
robust control using a describing function representation and Acknowledgment
loop shaping approach of a non-linear model. In other words,
The authors would like to thank the National Science Council,
once the NARMAX model has been developed, the model R.O.C., for financially supporting this research under contract NSC 87-2218-
is mapped into a describing function so that it can be used E-014-005.
− 113 −
J.J. Liu et al.
非線性識別技術於感應馬達轉速估測之應用
摘 要
本論文以一個非線性識別技術進行感應馬達之轉速估測。此離散非線性之識別方法稱為NARMAX。藉由馬達轉
速及電壓間之多項式模式以估測馬達轉速。此方法在識別馬達轉速及電壓間之非線性關係是可行的,且可以取代現行
感應馬達轉速控制迴路中之轉速量測器。此外,基於此理論獲得之結果,正研究設計一控制器以作為轉速控制之用。
相較於現行感應馬達轉速控制理論,此轉速控制提供一創新且簡便方式。
− 114 −