Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

LECH M. GRZESIAK AND MARIAN P.

KAZMIERKOWSKI

Exploring the Problems and Remedies

Digital Object Identifier 10.1109/MIE.2007.901483

DIGITAL VISION

8 IEEE INDUSTRIAL ELECTRONICS MAGAZINE FALL 2007

1932-4529/07/$25.002007IEEE

IGH-PERFORMANCE control methods for converter-fed ac motors such as field-oriented control (FOC), direct torque control (DTC), adaptive, and nonlinear or sliding mode control [1][5], [42] require complete information about state and output variables. However, sensors used in feedback loops increase costs and decrease reliability and immunity of the drive system; therefore, they should be avoided. Since the late 1980s, many efforts have been made to reconstruct such state variables of ac motors as rotor or stator flux vectors and rotor angular speed, e.g. [6][12]. Artificial neural networks (ANNs) are well suited for ac drive control and estimation, because of their known advantages, such as the ability to approximate any nonlinear functions to a desired degree of accuracy, learning and generalization, fast parallel computation, robustness to input harmonic ripples, and fault tolerance [3], [4]. These aspects are important in the case of nonlinear systems, like converter-fed ac drives, where linear control theory cannot be directly applied. Additionally,

high-efficiency power electronic converters used for ac motors operate in switch mode, which results in very noisy signals. For these reasons, ANNs are attractive for signal processing and control of ac drives. The commonly used feed-forward ANN (FF-ANN) [Figure 1(a)] as universal nonlinear function approximator is suitable only for steady-state systems. For a system dynamics approximation task, a few modifications of the FF-ANN architecture are commonly used (Figure 1): placement of tapped delay lines (TDLs) at inputs of the FF-ANN [see Figure 1(b)] recurrent ANNs (RANNs) [see Figure 1(c)(d)]. All presented dynamic models [Figure 1(b)(d)] are widely applied as nonlinear models of plants, estimators, and controllers. The motivation for using such an architecture is simple mathematical description and availability of a very efficient training method (e.g., available in the MATLAB Neural Networks Toolbox). However, presented schemes are marked by serious limitations inherited from TDLs and recurrent architecture, which requires that the initially selected sampling time is

applied to the data in learning and working phases. The tapped delay neural architecture [Figure 1(b)] has several limitations caused by sampling and accuracy of measurements [10]. These factors influence rank of a matrix called the input teaching matrix which is essential for neural network parameter tuning. It is quite obvious that the ANN with TDLs shown in Figure 1(b) (also used in an ac drive control system) has to act between two domains marked by very different frequency spectra, at inputs and at output. For instance, a neural speed estimator works in steady state at constant speed, which must be reconstructed from periodical sinusoidal input signals (stator voltage and stator current). Limitations of certain neural schemes based on time instances of periodic signals are discussed in detail in [13] and [14]. The sampling time selected for a continuous signal must fulfil Shannons condition, so it is bounded from the top as T < (1/2fmax ), but it is also bounded from below. If a very small sampling time is selected, then two neighboring columns (each column representing data at given time) of the input teaching matrix become very

Input

ANN

Output

Input DU DU

ANN

Output

(a) Input ANN Input Output DU DU DU DU DU DU DU (c) DU

(b) ANN

Output

(d)

FIGURE 1 Example of ANNs used as models of static or dynamic plant: (a) a simple static FF-ANN, (b) a dynamic model comprised of TDLs and static FF-ANN, (c) a Jordan network, and (d) a recurrent network with TDLs on input and output signals.

FALL 2007 IEEE INDUSTRIAL ELECTRONICS MAGAZINE 9

close. So if their differences are within the accuracy of measurement, then the input teaching matrix may loose its feature of being full column rank, reducing the dimension of the approximation space. The lower bound of sampling time, below which two sinusoidal signals shifted by the sampling time are fully located within accuracy of the measurement, has been determined. Using the concept of indistinguishable signals, it was demonstrated in [10], [14], and [26] that the range of allowed sampling is related to maximal and minimal useful frequencies of the signal and accuracy of the measurement in the following manner: fmax 100 , fmin p (1)

where p is the accuracy of measurement signal in %. This equation expresses that the useful frequency band of every input signal is limited. One has to acknowledge that one value of the sampling time must be selected for all inputs and the output signal. So, if there are signals with different spectrumi.e., different fmin and fmax parametersand the spectrum of one signal is not included in another, then according to (1) it is impossible to select just one T parame-

ter (sampling time) as being appropriate for all the inputs and the output. To avoid such limitation, it is necessary to select either different physical signals or use a nonlinear (or dynamic) transformation (called here preprocessing) of existing signals to achieve the inputs and output spectra confinement. In such cases, new structures for modeling dynamic systems using a neural architecture can be created. This article presents selected examples of ANN-based flux vector and mechanical speed estimators. By using appropriate preprocessing of input signals, the performances of flux vector and speed estimators are considerably improved (compared to estimators based on TDL FF-ANN and RANN) in terms of accuracy and sensitivity to parameter changes. The properties of the discussed estimators are illustrated in an example of the induction motor drive system with the direct torque control and space vector modulation (DTCSVM) scheme, as shown in Figure 2. This system has been selected as a very practical high-performance scheme that can be used for induction and/or permanent magnet synchronous motor drives [3]. All presented experimental oscillograms have been meas-

ured using the laboratory set-up (Figure B) described in the Appendix.

Stator Flux Estimation Based on ANN


Overview Operation of high-performance ac motor drives (DTC or FOC) depends strongly on stator (or rotor) flux estimation accuracy. Several approaches are used for flux vector estimation: model-based [15], [16], using Kalman filters [17], Luenberger observers [8], [18], [19], and many other closed-loop estimators (e.g., [6], [20]). Nevertheless, in many applications the stator flux is calculated from the so-called voltage model s (t ) = s (t0 ) +
t t0

(v s (t ) (2)

R si s (t )) dt ,

where s denotes stator flux space vector, v s, i s denote stator voltage and cur rent space vectors, respectively, and R s denotes the estimated stator resistance. Accurate knowledge of stator resistance is important for speed sensorless drives operating at a wide speed control range including zero speed in steady states and transients. In order to implement (2), numerous improved

Vdc

Load

(k) us (k) us (k) is ANN


s
(k+1)

LPF

(k) is

(k+1)

ANN-a

ANN-b

Te

DU

Control System (DTC SVM) s ref

DU
ref
m

FIGURE 2 Block scheme of a DTC-SVM induction motor drive applying flux and speed estimators based on ANN.

FIGURE 3 RANN architecture for stator flux estimation.

10 IEEE INDUSTRIAL ELECTRONICS MAGAZINE FALL 2007

integration algorithms have been proposed [6], [7], [21], [22]. These deal with drift and initial condition problems related to pure integration. However, any stator flux estimator which includes back electromotive force integration requires an additional algorithm for Rs adaptation. This is because even relatively small modeling errors due to motor temperature rise can lead to large errors in flux estimation accuracy [6]. Simple RANN Architecture for Estimating Stator Flux Vector To reconstruct flux directly from stator currents and voltages (instead of back EMF), a nonlinear transformation between stator flux and stator voltage and current will be reflected via neural model. From a basic ac motor mathematical model, stator voltage and flux current equations can be written as [24], [25] d (t ) = Rsi s (t ) + v s (t ) . dt s (3) Now, if i s and v s are assumed to be measurable and the model parameters Rs , Lm , and Lr are constant, then the system becomes linear and no typical neural network attributes are needed. Having collected enough data points in time and using a simple linear autoregressive moving average model (ARMA), one can find a linear approximation [i s (k ), v s (k )] s (k + 1) as

Artificial neural networks are well suited for ac drive control and estimation.
a direct pseudo-inverse calculation. However, if some of the parameters change in time, they will influence i s ; therefore, a nonlinear approximation is necessary. The stator resistance variations should be included in a training set; thus, some level of robustness could be achieved. Note that the relationship between (v s (k ), i s (k )), and s (k ) is not a static one. This in turn implies that there is no such function f1 that satisfies s (k) = f1 (v s (k ), i s (k )). One can expand the approximation space to turn the problem into a static one s (k + 1) = f2 (v s (k ), i s (k ), s (k )) . (4) ponents of the stator voltage, stator current, and stator and rotor flux space vectors, respectively. A neural architecture for stator flux vector estimation is presented in Figure 3. The ANN estimates stator flux from the stator voltage and current data. Mathematically, the problem can be classified as a nonlinear dynamic system approximation. The ANN can use only a single-hiddenlayer architecture. Training data can be generated from a simulated mathematical model operated in various working conditions or from the data collected from motor measurements. To train the network, many different methods can be applied. Mostly, backpropagation methods are used. For such a training method, the architecture of the network (number of hidden layers and neurons) must be set up in advance. One can also use incremental learning methods (originated in [27] and [28]), presented for example in [10] and [26], to design an FFANN for function approximation. If an incremental learning method is chosen for ANN design and training (Figure 3), the number of hidden sigmoid neurons is not fixed in advance. At

To build a neural estimator using (4), it is necessary to write down (3) in stationary orthogonal coordinates. Now the mathematical model of flux components is given by d s (t ) = vs (t ) Rsis (t ) dt d s (t ) = vs (t ) Rsis (t ), (5) dt where vs , vs , is , is , s , and (k ) s denote real and imaginary com(k ) (k ) (k ) (k ) (k )

x1

g1 f1

x1

g1 f1

x1

g1 f1

g2 x2 x2 x2

g2

f2

f2

f2

fm xd
(a)

fm xd
(b)

xd

gn
(c)

fm

FIGURE 4 Illustration of incremental approximation concept shown in sequence situation when first, second, and nth neuron are added.

FALL 2007 IEEE INDUSTRIAL ELECTRONICS MAGAZINE 11

Considerable improvement of flux vector estimation can be achieved using FF-ANN with dynamic signal preprocessing.
each iteration, one hidden neuron is optimized and added to the network, but all other hidden neuron parameters remain unchanged. Then all output weights are recalculated. The red lines in Figure 4 indicate connection parameters being recalculated when a neuron is added to the network. The iterative process is terminated when final conditions such as error level or number of hidden neurons are met. This type of ANN can approximate any continuous function with any accuracy, provided that one may use as many hidden neurons as needed [10], [27], [28]. However, this ANN architecture (Figure 3) is marked by serious limitations inherited from recurrent ANN. First, it requires that initially selected

sampling time is applied to the data in learning and recall phases, as discussed earlier. Second, there is no guarantee of an ANN estimators stable operation when it is trained in a seriesparallel identification scheme (during the learning process, inputs and outputs signals of plant have been used) and then works in a parallel configuration (only real input signals are available) using an estimated output signal. FF-ANN with Dynamical Preprocessing as a Stator Flux Estimator To improve stator flux estimation accuracy, it is possible to use, rather than the FF-ANN with TDLs or RANN, a simple dynamical preprocessing method

K1u T1u vs K2u T2u v1 FF-ANN vab

K0u T0u vab0

K1u T1u v1 FF-ANN

K2u T2u

K3u T3u

K3u T3u

vs s K1i T1i is i1 K2i T2i s

vbc s K1i T1i ia K2i T2i i1 s

K3i T3i

K3i T3i

is

ib

(a)

(b)

FIGURE 5 An FF-ANN-based stator flux estimator with dynamic signal preprocessing: (a) with ortogonal (, ) components of voltage and current space vector on input and (b) with natural (a, b, c) terminal voltage and phase current on input.

12 IEEE INDUSTRIAL ELECTRONICS MAGAZINE FALL 2007

[29]. It was shown in [29] and [30] that dynamical filtering of ANNs input signals can turn a believed nonfunction modeling problem into a function modeling one. The natural evaluations of FF-ANN with TDLs on the input signals

(vs , vs , is , is ) conduct to the solution presented in [29], which is shown in Figure 5(a). Input signals (vs , vs , is , is ) prompt different low-pass filters (first-order dynamic units). Outputs of filters are connected

to FF-ANN. Vector flux components have been achieved on two outputs of the FF-ANN. Time constants can be chosen by trial and error. The best results are when two filters act roughly as integrators, whereas the third one

1 s [Vs] 0.5 0 0.5 1 1.6 0.5 Error [Vs] 1.7 1.8 1.9 2 2.1

s s [Vs]

1 0.5 0 0.5 1 1.6 0.5 Error [Vs] 1.7 1.8 1.9 2 2.1

2.2

2.2

est s-s

est s-s

0 0.5 1.6

0 0.5 1.6

1.7

1.8

1.9 t [s]

2.1

2.2 (a)

1.7

1.8

1.9 t [s]

2.1

2.2

1 0.8 0.6 0.4 s [Vs] 0.2 0 0.2 0.4 0.6 0.8 1 1 0.5 0 s[Vs] 0.5

s est s -s

est

1 0.8 0.6 0.4 s [Vs] 0.2 0 0.2 0.4 0.6 0.8 1 (b) 1 1 0.5 0 s[Vs] 0.5

s est s -s

est

0 50 20 0 20 0 2 4 6 8 20 0 20 0 2 4 Time (s) 6 8 10 (c) 10 Te 0 2 4 6 8 10 Tload

m [rad/s]

m [Nm]

50

50 0 50 20 0 20 0 2 4 6 8 20 0 20 0 2 4 Time (s) 6 8 0 2 4 6 8

Tload [Nm]

Tload [Nm]

10 Tload

Te [Nm]

Te [Nm]

10 Te

10

FIGURE 6 Performance of DTC-SVM drive: (a) results for programmable LPF as flux estimator, (b) results for neural flux estimator, and (c) ponents, speed, and torque signals.

s com-

FALL 2007 IEEE INDUSTRIAL ELECTRONICS MAGAZINE 13

has a very small time constant and performs only the duty of noise attenuation. The similar flux estimator based on FF-ANN and dynamic signal preprocessing, but uses direct stator electrical signals (a, b, c phase current and terminal voltage), is shown in Figure 5(b) [30]. Note that measured stator voltages and currents are not transformed into components, whereas estimated flux components are expressed in . The Clark transform is automatically formed within the FFANN during the learning process. Each stator electrical signal, (vab , vbc , ia, ib ), excites three different first-order dynamical systems. As shown in [29], those almost dynamically unprocessed signals, combined with others, span an approximation space that turns a modeling problem into a static one. It was found that there exists such a function f3 that satisfies s (t ) = f3 (v1 (t ), . . . , v6 (t ), i1 (t ), . . . , i6 (t )) . (6)

There are several advantages related to that scheme: It is a very simple learning algorithm because a typical feed forward architecture of neural network is used. The resultant estimator is stable (stable dynamics and static nonlinearity connected in series).

Contrary to an uncompensated pure integrator, there is no drift problem due to LPFs presence. There are different sampling times available in learning and recall mode (contrary to the tapped delay or recurrent architecture). A significant level of robustness to stator resistance variations is achieved. The Matlab/Simulink model of the DTC-SVM drive is shown in Figure A in the Appendix. The performance of the DTC-SVM drive with ANN stator flux estimator (Figure 5) is compared with the programmable low-pass filter (PLPF) algorithm presented in [22], [43], and [44]. After off-line supervised training, ANN estimators have been implemented in the DTC drive. Figure 6 shows selected signals in the DTC-SVM drive. A stator resistance rise in the amount of 30% has been used. Figure 6(a) (with PLPF estimator) shows clearly that such identification error leads to significant dynamics deterioration and even nonexclusion of unstable behavior. A contrary ANN-based estimator with dynamic signal preprocessing with the same conditions assures stable operation and decent dynamics [Figure 6(b)]. Experimental Verification The block diagram of the experimental system is shown in the Appendix (Figure B). The ANN was learned with the help of patterns taken from tested

drives equipped with a conventional PLPF-based estimator. Estimated resistancethe one present in the algorithmic estimatorwas updated simultaneously with imposed stator resistance variations. If there is an error in the stator resistance (Rs ) of about 30%, a response of the system to step change of reference speed visibly deteriorates [Figure 7(a)]. Under similar conditions, the drive that takes advantage of feedback signals provided by the ANN estimator with dynamic signal preprocessing reverses smoothly [Figure 7(b)]. So the idea of dynamical preprocessing at the inputs of the neural approximator was successfully implemented and shows that common problems related to recurrent and tapped delayed ANNs can be avoided.

Selected Problems of ANN-Based Speed Estimation


Overview In the past few years, many speed-sensorless techniques have been proposed to cope with the speed sensing problem [5]. Developed speed estimation algorithms are more or less parameter dependent and/or computationally time-consuming, so further investigation is justified. Speed estimators are designed on the basis of the very common belief that information about actual speed is contained in

[rad/s]

[rad/s]

50 0 50 0 0.5 1 1.5 2 2.5 3 3.5 4 20 0 20 20 0 20 0 0.5 1 1.5 2 2.5 (a) 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4

50 0 50 0 0.5 1 1.5 2 2.5 3 3.5 4 20 0 20 20 0 20 0 0.5 1 1.5 2 2.5 (b) 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4

4.5
est T e [Nm] est Te

Tref [Nm]

Tref [Nm]

4.5 Tref

4.5 Tref

4.5 Te
est

T e [Nm]

est

4.5

4.5

Time (s)

Time (s)

FIGURE 7 Performance of the experimental drive: (a) with incorrect stator resistance (Rs ) identification (PLPF estimator) and (b) with ANN-based estimator (visible improvement compare to the PLPF estimator).

14 IEEE INDUSTRIAL ELECTRONICS MAGAZINE FALL 2007

easily accessible electrical signals. Available solutions can be divided into two main groups. The first group includes methods that should be regarded as algorithmic ones. These methods exploit a mathematical model of an induction motorextended Luenberger observer [18], sliding-mode observers [33], extended Kalman filter (EKF) [17], model reference adaptive systems (MRAS) [16], indirect flux detection by online reactance measurement (INFORM) [2], high-frequency signal injection [34], low-frequency signal injection [12], slip calculation [16], pseudoinversion [10], and their mutations. Unfortunately, any mathematical modelings introduce some simplification, which in turn entails deterioration of estimation. On the other hand, there is a set of solutions that does not take advantage of any mathematical model of discussed plant. This group consists of estimators based on ANN. In such a case, only inputs and outputs are known, whereas nonlinear relationships between them are not. ANN solutions can incorporate online trained nets [11], [31], [35] or off-line trained ones [36], [37]. These estimators naturally possess robustness to noise and parameter disturbances. However, most popular ANNs dedicated to system dynamics approximation, consisting of a multilayer FF-ANN with additional TDLs or recurrences [Figure 1(b)(d)], cause problems and limitations that have been briefly discussed earlier. Therefore, the speed reconstruction problem using as inputs current and voltage signals has exactly the same limitation like flux vector estimation. Therefore, it is also necessary to use a nonlinear transformation (called here preprocessing) of existing signals to achieve the inputs and output spectra confinement. Speed Estimators with Nonlinear Preprocessing and Neural Function Approximator Nonlinear preprocessing of the electrical signals should be applied to solve problems occurring with TDLs. Its task is to provide input signals with a spectrum similar to the speed signal. At the same time, a dimension

FF-ANN vs vs is is Nonlinear Preprocessing m

FIGURE 8 Speed estimator for ac motor with nonlinear preprocessing of input signals and neural function approximator.

of the approximation space is enlarged and the number of delay units can be reduced even to zero [10], [13]. There is a large number of nonlinear functions useful at this stage of preprocessing [see (7)]. The nonlinear preprocessing is performed in order to create a set of signals marked by suitable spectrum. These transformations are designed with the aid of cross and dot products of stator voltage v s and current i s space vectors, giving for instance the following equations: u1 = abs(v s ) = u2 = abs(i s ) = u3 = Re(v s i ) s = is vs + is vs u4 = Im(v s i ) s = is vs is vs vs u5 = Re is is vs + is vs = 2 2 is + is u6 = Im = . . . un = f(is , vs , is , vs ) . (7) vs is is vs is vs
2 is2 + is 2 2 vs + vs 2 2 is + is

the neural speed estimator with nonlinear input signal preprocessing is shown in Figure 8. It is a difficult task to take the best of the signal results (including most of the important

isa

isb

vsa

vsb

Transformation abc/

is

is

vs

vs

Nonlinear Preprocessing SelfOrganizing PCA

Online FF-ANN

Offline m FIGURE 9 Improved speed estimator for ac motor with nonlinear preprocessing, neural (PCA) linear preprocessing and neural function approximator.

Given the first set of six signals (7), one can enlarge this set by including their product ratios and powers [13]. The basic structure of

FALL 2007 IEEE INDUSTRIAL ELECTRONICS MAGAZINE 15

information) from preprocessing. In [13] and [32], few methods of final selection from candidate signals have been described and discussed. Speed Estimator with Nonlinear Preprocessing and Neural Implementation of Principal Component Analysis and Function Approximator Further improvement of the speed estimator can be achieved by adding an

extra system that can automatically select the best signals from the large number of candidates [29]. The structure of such a speed estimator (Figure 9) consists of three stages connected in series: a nonlinear preprocessing, a linear preprocessing (self-organizing principal component analysis, or PCA) and nonlinear function approximator (FF-ANN) stage. The second stage is designed to supply optimal signals for the neural function

approximator. The goal of the second stage is to decorrelate variables and to maximize eigenvalues of the autocorrelation matrix at the same time. Such a problem can be solved by PCA [32]. The PCA model generates a new set of a few decorrelated variables called principal components (PCs). The purpose of the PCA is to derive a small number of decorrelated linear combinations of a set of zero-mean variables while retaining as much of the

APPENDIX
The Simulink simulation model of a DTC-SVM drive with neural flux and speed estimators is shown in Figure A. The block scheme of the laboratory set-up used for experimental verification is shown in Figure B. It consists of a 1.5-kW, six-pole ac motor which is fed by an IGBT voltage source inverter. For control tasks, a dSpace DS1103 card programmed in the Matlab/Simulink Real-Time Workshop and monitored via the ControlDesk interface has been used. The control and estimation algorithms run at a computation time of 0.5 ms. The angular speed of the drive is measured with an encoder.

Reference_Flux

PI Flux Controller

I_s_alfa [gamma] gamma U_s_alfa_r U_s_d U_s_beta_r U_s_q U_s_beta_r U_s_beta U_s_alfa_r U_s_alfa I_s_ beta Us_alfa Torque Speed Us_beta Psi_r_alfa

[Us_Is]

[abs_Psi_s]

[Torque] [Speed]

Reference Speed Signal Builder

+
k1

PI Speed Controller

+
k2

PI Torque Controller

[Psi_r] Transformation Inverter Load Torque Signal Builder Psi_r_beta Psi _s_alfa T_load Psi _s_beta ac Motor

[Speed] MS2 [Torque]

[Psi_s]

[Speed_est]

[Psi_s]

gamma

[gamma]

In
[Psi_s_est] MS1 abs(Psi_s) [abs_Psi_s] U1 U2 Calculation U1 U2 [Us_Is] In U3 U4 U5 U6 Linear Dynamic Prprocessing Neural Network 1 Stator Flux Estimator MS3 p{1} y{1} [Psi_s_est] [Us_Is] In U3 U4 U5 U6 Nonlinear Preprocessing Neural Network 2 Speed Estimator p{1} y{1} [Speed_est]

[Us_Is] [Psi_s] [Speed]

[Us_Is] [Psi_s] [Psi_r] [Speed]

[Torque] Training_Data.mat To File1 [Psi_s_est] [Speed_est]

Visualization_Data.mat To File2

FIGURE A Simulink model of a DTC-SVM drive with neural flux and neural speed estimators.

16 IEEE INDUSTRIAL ELECTRONICS MAGAZINE FALL 2007

information from the original variables as possible [38]. This, in turn, implies reduction of a number of signals in comparison to the number of nonlinear transformations. The dimensionality reduction is one of the basic features of PCA. The PCA model involves secondorder statistics and guarantees data decorrelation with simultaneous denoising [38]. PCA is normally done by analytically solving an eigenvalue problem of the input correlation func-

The tapped delay neural architecture has several limitations caused by sampling and accuracy of measurements.
tion. The eigenvectors must be put in decreasing order with respect to eigenvalues. However, in [38] it was demonstrated that PCA can be accomplished by a single-layer, linear self-organizing neural network trained with a modified Hebbian learning rule [37], [39], [40]. Several different types of stopping rules have been developed to determine the number of extracted components from a given analysis. One of the most popular is a percentage of variance criterion in which successive eigenvectors are extracted until some absolute percentage of the total variance has been explained (commonly used with typical threshold set to 95%). Such a type of PCA implementation is shown in Figure 9 as the second stage of signal preprocessing for a neural speed estimator. Computer test results of a DTC-SVM drive operated with a feedback signal from a neural speed estimator is shown in Figure 10. One can see the very good performance of drive systems. Experimental Verification The experimental has been performed on drive system shown in Figure B with data given in the Appendix. The control and estimation algorithms have been running with a sampling time of 0.5 ms. The drive was tested at random speed and load profiles. The speed of the drive was measured additionally with an encoder. After offline training, the nonlinear ANN was implemented in the DSP board. Figure 11 shows some experimental results of a drive operated with speed estimator of Figure 9(b).

MOTOR DATA:

Induction cage motor Sf100L6K: P N = 1.5 kW, U N = 220 V, I N = 6.8 A, f N = 50 Hz, cos N = 0.75, N = 0.76, nN = 930 rpm, Rs = 1.54 , Rr = 1.29 , Xs = 31.56 , Xr = 30.43 , Xm = 28.74
INVERTER DATA:

IPM-IGBT module: V dc = 330 V, switching frequency = 10 kHz, I out = 50 A

Rectifier
ac Grid

Inverter

Current and Voltage Measurement

Power Circuit Encoder M ac Motor M Load ENC

est DTC Torque and Flux Controller

est Neural Flux Estimator

is Neural Speed Estimator est S2 m

Speed Controller ref PC Computer (ControlDesk) ref

dSPACE DS1103 Output: Screen Visualization

Speed Calculation

PWM-SVM
us, i s

Flux Estimator PLPF


S1

Conclusions
This article presented problems and selected improvements related to the application and implementation of ANN-based flux vector and mechanical speed estimators for control of high-performance pulse width modulated (PWM) inverter-fed induction motor drives. Key conclusions include the following:

Input: Keyboard Parameters Set Up Configuration Set Up

FIGURE B Block diagram of experimental set-up DTC-SVM drive for testing of flux and speed estimators.

FALL 2007 IEEE INDUSTRIAL ELECTRONICS MAGAZINE 17

By using appropriate preprocessing of input signals, the performances of flux vector and speed estimators are considerably improved in terms of accuracy and sensitivity to parameter changes.

m [rad/s]

100 0 100 5 0 5 20 0 20 10 0 10 0 2 4 6 8 10 Time (s) 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20

FIGURE 10 Simulation results of DTC-SVM drive with speed feedback signal taken from a neural speed estimator (as shown in Figure 8).

50 es [rad/s] 0 50 50 m [rad/s] 0 50 5 [rad/s] 0 5

10

20

30

40

50

60

70

80

10

20

30

40

50

60

70

80

Recurrent ANNs can be used for flux vector estimation only under the assumption of very fast sampling equal in learning and recall mode. Considerable improvement of flux vector estimation can be achieved using FF-ANN with dynamic signal preprocessing (Figure 5). This estimator poses the following advantages: very simple learning algorithm, stable operation, no drift problems due to LPFs presence, different sampling times available in learning and recall mode (contrary to the tapped delay or recurrent architecture), and some level of robustness to stator resistance variations. For reliable mechanical speed estimation, an FF-ANN with nonlinear signal preprocessing has been described. Two-stage preprocessing (nonlinear preprocesing and linear PCA) of four terminal signals is combined with FF-ANN. The first one guarantees enlarged approximation space using a linearly independent set of input signals based on cross and dot products of motor voltage and current vectors. The second one (linear stage, PCA) takes advantage of self-organizing principal component analysis to optimize the approximation space without loss of information. Proposed estimators are very convenient to implement. The presented experimental oscillograms measured in a 1.5-kW laboratory induction motor drive verify estimators based on ANN with nonlinear preprocessing of input signals. Achieved results can be easily extended to DTC-SVM permanent magnet synchronous motor drives. It is expected that, thanks to continuous developments in digital signal processing technology, ANN-based techniques will have a strong impact on drive control, estimation, and monitoring in the coming decades.

T load [Nm]

Te [Nm]

Error m [%]

Biographies
0 10 20 30 40 Time (s) 50 60 70 80

FIGURE 11 Experimental waveforms of speed estimation based on seven principal components (estimator structure shown in Figure 9).

Lech M. Grzesiak graduated from the Electrical Engineering Faculty of Warsaw University of Technology in 1976. He received the Ph.D. in 1985 and the Dr.Sc. degree in 2002, respectively, from the same university. Since 1977, he

18 IEEE INDUSTRIAL ELECTRONICS MAGAZINE FALL 2007

has been employed at Warsaw University of Technology, currently as a professor. He was also codirector of the Centre of Excellence on Power Electronics Intelligent Control for Energy Conservation (PELINCEC) from 20032005. He is an associate editor of IEEE Transactions on Industrial Electronics since 2004. His main research interests and publications are in the theory and application of control, generally dedicated for electric drives, power electronics, and energy generating systems. From 1994, his research interest has focused on development and applications of neural systems. He is an author and coauthor of a variety of papers related to this subject. He is a Senior Member of the IEEE. Marian P. Kazmierkowski received the M.S., Ph.D., and Dr.Sci. degrees in electrical engineering from the Institute of Control and Industrial Electronics (ICIE), Warsaw University of Technology, Warsaw, Poland, in 1968, 1972, and 1981, respectively. Since 1987, he has been a professor and director of ICIE. He was also head of PELINCEC from 20032005 (European Framework Program V) at ICIE, Warsaw University of Technology, Poland. He coedited (with R. Krishnan and F. Blaabjerg) and coauthored the compendium Control in Power Electronics (Academic Press, 2002). He received an Honorary Doctorate degree from Aalborg University in 2004 and the Dr. Eugene Mittelmann Achievement Award from the IEEE Industrial Electronics Society in 2005. He was the editor-in-chief of IEEE Transactions on Industrial Electronics (20042006). He is past-chair of the IEEE Poland Section.

[6] J. Holtz and J. Quan, Drift- and parameter-compensated flux estimator for persistent zero-statorfrequency operation of sensorless-controlled induction motors, IEEE Trans. Ind. Applicat., vol. 39, no. 4, pp. 10521060, July-Aug. 2003. [7] J. Hu and B. Wu, New integration algorithms for estimating motor flux over a wide speed range, IEEE Trans. Power Electron., vol. 13, no. 5, pp. 964977, Sept. 1998. [8] J. Maes and J.A. Melkebeek, Speed-sensorless direct torque control of induction motors using an adaptive flux observer, IEEE Trans. Ind. Applicat., vol. 36, no. 3, pp. 778785, MayJune 2000. [9] J. Zhao and B.K. Bose, Neural-network-based waveform processing and delayless filtering in power electronics and AC drives, IEEE Tran. Ind. Electron., vol. 51, no. 5, pp. 981991, Oct. 2004. [10] B. Beliczy n ski and L. Grzesiak, Induction motor speed estimation: Neural versus phenomenological model approach, Neurocomput., vol. 43, no. 1-4, pp. 1736, 2002. [11] S.-H. Kim, P.T.-S. Ark, J.-Y. Yoo, and G.-T. Park, Speed-sensorless vector control of an induction motor using neural network speed estimator, IEEE Trans. Ind. Electron., vol. 48, no. 3, pp. 609614, 2001. [12] V.-M. Leppnen, Low-frequency injection-based speed sensorless control of induction motors Applicability and implementation aspects, in Proc. EPE Conf., 2003 [CD-ROM]. [13] L.M. Grzesiak, B. Beliczy ski, and B. Ufnalski, n Input preprocessing in tapped delay neural architecture for induction motor speed estimation, in Proc. EPE Conf., 2003 [CD-ROM]. [14] B. Beliczyn ski, On input discretisation processes for tapped delay neural archirecture, in Proc. Int. Conf. Artificial Neural Networks Genetic Algorithms (ICANNGA), 2003, pp. 2832. [15] P.L. Jansen and R.D. Lorenz, A physically insightful approach to the design and accuracy assessment of flux observers for field oriented induction machine drives, IEEE Trans. Ind. Applicat., vol. 30, no. 1, pp. 101110, Jan.-Feb. 1994. [16] P. Vas, Sensorless Vector and Direct Torque Control (Monographs in Electrical and Electronic Engineering, vol. 42). London: Oxford Univ. Press, 1998. [17] Sensorless control with Kalman filter on TMS320 fixed-point DSP, Texas Instruments Europe, Literature no. BPRA057, 1997, Dallas, TX. [18] H. Kubota and K. Matsuse, Speed sensorless field-oriented control of induction motor with rotor resistance adaptation, IEEE Trans. Ind. Applicat., vol. 30, no. 5, pp. 12191224, Sept.Oct. 1994. [19] H. Kubota, K. Matsuse, and T. Nakano, DSPbased speed adaptive flux observer of induction motor, IEEE Trans. Ind. Applicat., vol. 29, no. 2, pp. 344348, Mar.Apr. 1993. [20] D. Casadei, G. Serra, and A. Tani, Steady-state and transient performance evaluation of a DTC scheme in the low speed range, IEEE Trans. Power Electron., vol. 16, no. 6, pp. 846851, Nov. 2001. [21] G. Andreescu and A. Popa, Flux estimator based on integrator with DC-offset correction loop for sensorless direct torque and flux control, in Proc. Int. Conf. on Electrical Machines (ICEM), 2002, [CD-ROM]. [22] M. Hinkkanen and J. Luomi, Modified integrator for voltage model flux estimation of induction motors, in Proc. IEEE Ind. Electronics Society Conf. (IECON), vol. 2, Nov. 2001, pp. 13391343. [23] L.M. Grzesiak and B. Beliczynski, Simple neural cascade architecture for estimating of stator and rotor flux vectors, in Proc. EPE99 Conf., 1999, [CD-ROM]. [24] M.P. Kazmierkowski and H. Tunia, Automatic Control of Converter-Fed Drives. Warsaw, Poland: Elsevier-Amsterdam and PWN, 1994. [25] W. Leonhard, Control of Electric Drives. New York: Springer-Verlag, 1985.

[26] B. Beliczynski and L. Grzesiak, Dynamic models and learning: application of neural approach to estimation of stator and rotor fluxes in an induction motor, in Proc. Conf. Engineering Application of Neural Networks, Warsaw, Sept. 1999, pp. 110116. [27] A.R. Barron, Universal approximation bounds for superpositions of a sigmoidal function, IEEE Trans. Inform. Theory, vol. 39, no. 3, pp. 930945, 1993. [28] L.K. Jones, A simple lemma on greedy approximation in Hilbert space and convergence rates for projection pursuit regression and neural network training, in Annals of Statistics. Beachwood, OH: Institute of Mathematical Statistics, 1992. [29] L.M. Grzesiak and B. Ufnalski, Neural stator flux estimator with dynamical signal preprocessing, in Proc. IEEE Conf. AFRICON, 2004, pp. 11371142. [30] L.M. Grzesiak and B. Ufnalski, DTC drive with ANN-based stator flux estimator, in Proc. 11th EPE05, Sept. 1114, 2005, Dresden, Germany, [CD-ROM]. [31] D. L. Sobczuk, Application of ANN for control of PWM inverter fed induction motor drives, Ph.D. thesis, Warsaw University of Technology, Faculty of Electrical Engineering, 1999. [32] B. Ufnalski, Application of artificial neural networks for estimation of rotor angular speed and stator flux in cage induction motor drive, Ph.D. thesis, Warsaw University of Technology, Faculty of Electrical Engineering, 2005 (in Polish). [33] A. Derdiyok, A novel speed estimation algorithm for induction machines, Elect. Power Systems Res., vol. 64, no. 1, pp. 7380, 2003. [34] R.D. Lorenz, Advances in electric drive control, in Proc. IEEE Int. Electric Machines and Drives Conf. (IEMDC), 1999, pp. 916. [35] L. Ben-Brahim, S. Tadakuma, and A. Akdag, Speed control of induction motor without rotational transducers, IEEE Trans. Ind. Applicat., vol. 35, no. 4, pp. 844850, 1999. [36] T. Orlowska-Kowalska and P. Migas, Analysis of the induction motor speed estimation quality using neural networks of different structure, Archives Elect. Eng., vol. L, no. 4, pp. 411425, 2001. [37] A. Cichocki and S. Amari, Adaptive Blind Signal and Image Processing: Learning Algorithms and Applications. New York: Wiley, 2002. [38] A. Hyvrinen and E. Oja, Independent component analysis: A tutorial, Helsinki University of Technology, Laboratory of Computer and Information Science, 1999 [Online]. Available: http://www.cis.hut.fi/projects/ica/ [39] P. Paplin ski, Generalized Hebbian learning and its application in dimensionality reduction, Faculty of Computing and Information Technology, Department of Digital Systems, MONASH Australias International University, Technical Rep. 97-2, 1997. [40] NeuroDimension, NeuroSolutions 4.2 Manual, 2003 (free evaluation copy) [Online] Available: http://www.nd.com/ [41] K. Simoes and B.K. Bose, Neural network based estimation of feedback signals for a vector controlled induction motor drive, IEEE Trans. Ind. Applicat., vol. 34, no. 3, pp. 620629, 1995. [42] G.S. Buja and M.P. Kazmierkowski, Direct torque control of PWM inverter-fed ac motorsA survey, IEEE Trans. Ind. Elect., vol. 51, no. 4, pp. 744757, Aug. 2004. [43] B.K. Bose and N.R. Patel, A sensorless stator flux oriented vector controlled induction motor drive with neuro-fuzzy based performance improvement, in IEEE IAS Conf. Rec., 1997, pp. 393400. [44] L. Borges, B.K. Bose, and J. Pinto, Recurrent neural network based implementation of programmable cascaded low pass filter used in stator flux synthesis of vector controlled induction motor drive, IEEE Trans. Ind. Elect., vol. 46, pp. 662665, June 1999.

References
[1] M.P. Kazmierkowski and T. Kowalska-Orlowska, ANN based estimation and control in converterfed induction motor drives, in Soft Computing in Industrial Electronics, S.J. Osaka and L. Sztandera, Eds. Heidelberg, Germany: Physica Verlag, 2001. [2] P. Vas, Artificial-Intelligence-Based Electrical Machines and Drives: Application of Fuzzy, Neural, Fuzzy-Neural, and Genetic-Algorithm-Based Techniques. London: Oxford Univ. Press, 1999. [3] M.P. Kazmierkowski, R. Krishnan, and F. Blaabjerg, Control in Power ElectronicsSelected Problems. New York: Academic, 2002. [4] B.K. Bose, Modern Power Electronics and AC Drives. Pearson Education, Englewood Cliffs, NJ: Prentice-Hall, 2002. [5] K. Rajashekara, A. Kawamura, and K. Matsue, Sensorless Control of AC Motor Drives, Speed and Position Sensorless Operation. New York: IEEE Press, 1996.

FALL 2007 IEEE INDUSTRIAL ELECTRONICS MAGAZINE 19

You might also like