Sensorless Speed and Direct Torque Control of
Surface Permanent Magnet Synchronous
Machines using an Extended Kalman Filter
Vasile COMNAC, Member, IEEE, Mihai CERNAT, Florin MOLDOVEANU, Ioan DRAGHICI
Abstract--The paper describes the development of a sensorless
speed and direct torque control system for surface permanent
magnet synchronous machines (SPMSM). An Extended
Kalman Filter (EKF) is used for the stator flux and
mechanical state estimation of the SPMSM. The observer was
developed based on non-linear model of the synchronous
machine that employs a α − β stationary reference frame. The
torque controller was designed taking into consideration a
model of the machine in the rotating reference frame attached
to the stator flux space vector. The electromagnetic torque and
stator flux, are controlled by non-linear hysteresis controllers;
their output signals are used for generating the control signals
of inverter, based on switching table. The angular speed
controller is linear of I + PI type
Index Terms-- Adjustable speed drives, Converter control,
DTC, Flux control, Kalman filter, Non-linear control,
Permanent magnet motors, PM motor drives, Sensorless
control, Synchronous machines, Torque control, Variable
speed drives, Vector control.
I. INTRODUCTION
There are basically two techniques for the instantaneous
torque control of high-performance variable speed drives:
vector control and direct torque control (DTC). The main
advantages of DTC are: absence of coordinate
transformations, absence of separate voltage modulation
block and of voltage decoupling circuit, reduced number of
controllers and s. o. This solution requires knowledge of the
stator flux, electromagnetic torque and also, angular speed
and position of the rotor. In last few years a strong interest
arises in the alternative PMSM sensorless control, using
Fig. 1.
only stator voltage and current measurement, based on state
observers.
Luenberger and Gopinath type state observers yield very
good results in the estimation of state quantities of linear
systems. Yet, in applications of these estimation procedures
in driving systems with a.c. machines, weaker performances
are achieved, the main cause being the strongly non-linear
model of a.c. machines.
The EKF is a state observer specially designed for non-
linear systems. Literature features numerous applications of
Kalman filters in the estimation of mechanical states, in
order to achieve sensorless control of asynchronous
machines [5], or d.c.-brushless machines [6], or permanent
magnet synchronous machines [2, 8, 9]. In [7] the EKF is
used for the estimation of the mechanical quantities of a
synchronous machine with electromagnetic excitation and
starting cage.
This paper presents the development of an EKF specially
designed for sensorless control of the direct torque
controlled SPMSM. The simplified block diagram of the
extended Kalman filter-based control system is shown in
Fig. 1. Here the superscript * (star) denote reference
quantities. The control of the SPMSM drive is carried out
with a conventional VSI-fed DTC system that uses two
hysteresis controllers for directly selecting (in each
sampling period) of the optimal voltage switching vector
(there are six non-zero and two zero vectors). The I+PI
speed controller gives the reference of the electromagnetic
torque (me∗ ) .
The Kalman filter generates the estimated values of the
stator linkage flux, actual angular velocity of the rotor and
position of the stator flux phasor, using the measured values
of the currents and voltages from the motor terminals.
II. THE KALMAN FILTER
In principle, the Kalman filter is a state observer that
establishes the best approximation by minimization of the
square error for the state variables of a system, subjected at
both its input and output to random disturbances. If the
dynamic system of which the state is being observed is non-
linear, then the Kalman filter is called an extended one
(EKF).The development of the Kalman filter is closely
linked to the stochastic systems. The linear stochastic
systems are described by relations:
x! (t ) = A x (t ) + B u(t ) + w (t ), x (t 0 ) = x 0 , (1)
y (t ) = C x (t ) , (2)
z (t ) = y (t ) + v (t ) , (3)
where x , y , u , A , B , C have the significance known
from deterministic systems [1, 2]. In these expressions,
w (t ) represents the vector of disturbances applied at the
system input, while z (t ) is the vector of the measurable
outputs, affected by the random noise v (t ) . It can be
considered that vector w (t ) includes, besides the input
disturbances, some uncertainties referring to the process
model. It will be assumed that the vector functions w (t )
and v (t ) are not correlated and zero-mean stochastic
processes.
From the statistic point of view, the stochastic processes
w (t ) and v (t ) are characterized by the covariance matrices
Q and R respectively.
It is further assumed that the initial state x 0 is a vector of
random variables, of mean x 0 and covariance P0 , not
correlated with the stochastic processes w (t ) and v (t ) over
the entire interval of estimation.
The covariance matrices Q , R and P0 that characterize
the noise sources of system (1)–(3) are, by definition,
symmetrical and positively semi-definite, of dimensions
(n × n) , (m × m) and (n × n) respectively, where n and m
represent the number of state and output variables,
respectively.
For linear time invariant systems, the following relations of
recurrent computation describe the general form of the
implementation algorithm of Kalman filters:
K k = Pk k −1C
T
(C Pk k −1C
T
+ R) −1 ,(4)
xˆ k = xˆ k + K k ( y k − C xˆ k
k k −1 k −1 ) , (5)
Pk k = ( I n − K k C ) Pk k −1 , (6)
where the quantity Ts represents the sampling period, and
Ad is the matrix of the discrete linearised system,
expressed by:
Ad = I n + T s A . (9)
In these relationships, the (n × m) matrix K represents the
Kalman gain, P represents the covariance state matrix and
I n is the (n × n) unit matrix. In the recurrent computation
relationships, the subscript index notations of type k k − 1
show that the respective quantities (state vectors or their
covariance matrices) are computed for sample k , using the
values of similar quantities from the previous sample.
For non-linear stochastic systems, the dynamic state model
is described by the following expressions:
x! (t ) = f ( x (t ), u(t ), t ) + w (t ) , (10)
y (t ) = h ( x (t ), t ) , (11)
z (t ) = h ( x (t ), t ) + v (t ) , (12)
where f and h are (n × 1) and (m × 1) function vectors,
respectively. Matrices A and C of the EKF structure,
dependent now upon the state of the system, are determined
by using relationships:
∂ f ( x (t ), u(t ), t )
A( xˆ (t ), t ) = , (13)
∂ x T (t ) x (t )= xˆ (t )
∂ h( x (t ), t )
C ( xˆ (t ), t ) = . (14)
∂ x T (t ) x (t )= xˆ (t )
Additionally to the fact that matrices A and C have now
became dependent on the state of the system, the algorithm
suffers some further changes, which affect relationships (5)
and (7), now expressed as:
xˆ k k = xˆ k k −1 + K k [ y k − h ( xˆ k k −1 )] , (15)
xˆ k +1 k = xk k + Ts f ( xˆ k k , uk ) . (16)
The Kalman filter algorithm is initiated by adopting
adequate values for the covariance matrix of the initial
state P0 = P0 −1 , as well as for the weighting matrices Q
and R , the latter two being constant during the estimation.
III. MACHINE MODEL FOR EKF DEVELOPMENT
The studied SPMSM is assumed to have a symmetrical
three-phase, Y-connected, isolated neutral point stator
winding. The on-line implementation computations of the
EKF algorithm are carried out based on an dynamic model
of the SPMSM expressed in the stationary reference frame,
α − β , attached to the stator.
The state variables of the model deterministic part are
included in vector:

xˆ k +1 = Ad xˆ k + Ts B uk , (7) x = [Ψα Ψβ ω θ] T , (17)

k k
where quantities Ψα and Ψβ are the direct and quadrature
Pk +1 k = Ad Pk k AdT + Q , (8)
components of the stator linkage flux space phasor in
relation to the stator reference frame, α − β , and quantities (Ψα − Ψe cos pθ) / L s 
ω and θ represent the rotor (mechanical) angular velocity y = h ( x (t )) =  . (24)
 (Ψβ − Ψe sin pθ) / L s 
and position, respectively. It must to note that rotor position
Based on the previous elements, the EKF can now be built
information is necessary only for normal behavior of the
and applied to the SPMSM. By derivation of the vector
EKF. The input quantities of the process (SPMSM in our
function (21) in relation to the state vector, and then by
case) are the components of the vector
applying relation (9), matrix Ad of the discrete linearised
u = [u α u β ml ] T , (18)
system will be obtained, described by the expression:
where quantities u α and uβ are the direct and quadrature  Rs Ψ 
components of the stator voltages space phasor, and 1 − Ts L 0 0 − pTs Rs e sin pθ 
L
 s s 
quantity ml represents the load torque. In order to ensure  Ψe
cos pθ  (25)
Rs
Ad=  0 1 − Ts 0 pTs Rs
an interface of the EKF with quantities directly measurable Ls Ls 
at the machine terminals, the model used for estimation  0 0 1 0 
 
components u α and uβ are expressed in dependence on the  0 0 Ts 1 
terminal voltages using relationships:
Matrix C of the EKF algorithm will follow, considering
u α = u a , uβ = (u a + 2u b ) / 3 , (19) relation (14), by derivation of vectorial function h of
where u a and u b are the instantaneous values of the phase expression (24) in relation to the state variables:
stator voltages. 1 Ψ 
L 0 0 p e sin p θ 
Under these circumstances, the deterministic part of the Ls
dynamic model of the SPMSM written in the standard form C= s . (26)
 0 1 0 − p Ψe cos p θ
specified by relation (10) will be [2]:  
 Ls Ls 
f ( x (t ), u(t )) = [ f1 f 2 f 3 f 4 ]T , (20)
where: IV. THE CONTROL SYSTEM STRUCTURE
R R
f 1 = − s Ψα + s Ψe cos pθ + u α , (21 a) The simplified block diagram of the extended Kalman
Ls Ls filter-based control system is shown in Fig. 1. The control
Rs R system, consist of two parallel loops for direct torque and
f2 = − Ψβ + s Ψe sin pθ + u β , (21 b) stator flux control and an outer loop for the linear control of
Ls Ls
the rotor angular velocity.
f 3 = (−k m Ψα sin pθ + k m Ψβ cos pθ − D ω − ml ) / J , (21 c)
f4 = ω , (21 d) A. The Direct Torque and Stator Flux Control (DTFC)
k m = 3 p Ψe /(2 L s ) , (22) This part of the control system contains two non-linear,
hysteresis controllers and a functional block labeled
where Rs and Ls are the stator resistance and inductance
“Switching table”, that generate binary signals S a , S b and
respectively, Ψ e is the linkage flux of the permanent
magnet, J is the inertia, D represent the viscous damping S c applied to the inverter branches.
factor and ml is load torque.
Selection of the output quantities is the most important
stage in the development of the EKF proposed in this paper.
These quantities will have a double expression. Thus, in
order to ensure an interface as simple as possible of the
EKF with the process, the output quantities comprised by
vector:
y = [ i α iβ ] T , (23)
will be adopted, where iα and iβ are the components of
the space vector of the currents of the fixed two-phase
reference frame, that can be obtained as a function of the
measurable quantities ia and ib with relationships similar For S i = 1 , the transistors of the branch i = a, b, c ,
to expressions (19). Ti and Ti − (Fig. 2), are switched on and off, respectively
+
On the other hand, the output quantities (23) will be
expressed in dependence on the state variables (17), thus and for S i = 0 the two transistors are inverse switched. The
obtaining the non-linear output function of the deterministic switching state of the inverter ( S a S b S c ) is achieved in two
part of the state variable model: steps. Firstly, based on outputs d Ψ and d m of the non-
linear controllers and angle λ of the stator flux vector,
according to Table 1 [2, 3] a stator voltage space phasor is ε ω ( n ) = ω * ( n ) − ω ( n) , (27)
selected.
Table 1: Switching logic of the three-phase inverter, for x ω (n ) = x ω (n − 1) +
Teω
[ε ω (n ) − ε ω (n − 1 )], (28)
different positions of the stator flux vector 2 Trω
dq 1 -1 -1 1
λ m e* (n ) = k rω [ x ω (n ) − ω (n )] , (29)
dm 1 1 -1 -1
where: ε ω , x ω - the auxiliary variables associated to the
-30º÷30º U2 U3 U5 U6
angular velocity loop error and to the integrator
30º÷90º U3 U4 U6 U1
respectively; k rω , Trω - the tuning parameters of the angular
90º÷150º U4 U5 U1 U2 velocity controller, n - the index of the sampling period
150º÷210º U5 U6 U2 U3 Teω . The tuning parameters of the speed controller are
computed with the following relations [2]:
210º÷270º U6 U1 U3 U4
k r ω = J /(4 Tμ k ω ), Tr ω = 8 Tμ , (30)
270º÷330º U1 U2 U4 U5

where Tµ is the time constant of the first order delay

Then, with this space vector, based on Table 2, is achieved element that models the control loop of the electromagnetic
the switching state of the inverter. torque.
Table 2: Relations between switching states and voltage
space vectors (VSV). V. KALMAN FILTER-BASED CONTROL SYSTEM ANALYSIS
VSV Sa Sb Sc In order to analyze the dynamic behavior of the control
U1 1 0 0 system with a Kalman filter, simulations were conducted in
U2 1 1 0
C++. Normal operation circumstances were taken into
U3 0 1 0
U4 0 1 1 consideration, the estimated quantities being connected as
U5 0 0 1 reaction quantities to the digital current and angular
U6 1 0 1 velocity controllers.
U7 0 0 0 The feedback quantities, me and Ψ s , are obtained trough
U8 1 1 1 numerical computation, based on relationships:

B. Angular Velocity Controller
The angular velocity controller has an important influence
on the entire control system performance, both during
transient and steady state regimes. Good performance could
be achieved using an I type control law for the angular
∗
velocity reference quantity ω , and a PI control law for the
angular velocity feedback ω respectively. The I+PI
controller will operate with both these laws. The main
advantage of an I+PI controller is the achievement of good
transient performance. At the same time, due to the unitary
discrete pole, a null stationary error with respect to the
perturbation of the load torque (including the friction torque
of the machine) is achieved.
Fig. 3.
The block diagram of this controller, using z-transfer
functions, is presented in Fig.3. It can be noticed that the
integration uses the trapeze method.
The controller algorithm, obtained by applying the inverse
Z transform, is described by recurrent equations:
m e = 3 p (Ψ
ˆ i −Ψ
α β
ˆ i ), Ψ = Ψ
β α s
ˆ 2 +Ψ
α
ˆ2 .
β
2
The model of the simulated machine considers the load
torque and the viscous friction torque: ml N = 1 Nm,
DN =0.002 Nms/rad. The rated data related to the motor
employed for simulation were: p =2, Rs N = 0.98 Ω ,
LsN =18. mH, Ψ eN =0.174 Wb; J N =0.006 Kg m2. The
synchronous machine is fed from a PWM converter of 12
kHz frequency. The sampling periods are of 2 ms for the
angular velocity controller and of 0.1 ms for the stator flux
and electromagnetic torque control loops and also for the
Kalman filter. It is assumed that the reference quantity for
the angular velocity has a trapezoidal waveform, specific
for servo-drives with incremental motion. As previously
shown, in order to optimize the dynamic behavior of the
Kalman filter, it is required to adopt adequate values for the
elements of the square matrices P0 , Q and R . In the paper
the following values were adopted: P0 = 0.01 I 4 , Q =
0.00002 I 4 and R =0.9 I 2 . In order to highlight the
performances of the Kalman filter control system, at first
the behavior of the direct torque digital control system in
absence of a Kalman filter is presented.
 a) b)
Fig. 4

a) b)
Fig.5
Fig. 4 presents the waveforms of the mechanical quantities mismatch of the inertia, damping factor and load torque
(rotor angular velocity, electromagnetic torque), of the related to the rated values: J = 0.7 J N , D = 0.5 D N ,
amplitudes of the reference and real stator linkage flux and ml = 0.5 ml N (Fig. 7) and J = 1.4 J N (Fig. 8). In these
of the a stator phase linkage flux Ψsa = Ψα , obtained by a
DTFC system without EKF, for precise tuning of the figures, Δω = ω − ωˆ , Δθ = θ − θˆ are the estimation errors of
velocity controller. The same quantities, obtained for the the angular velocity and position, respectivelly. The scales
control system with an EKF are presented in fig. 5. used in all figures presented above and below are the
A comparison of figures 4 and 5 shows insignificant follwing: t − 66 ms/div , ω ∗ , ω, ω
ˆ − 32 rads −1 /div ,
modifications in the behavior of the dynamic control
system. The wave forms from Fig. 6 have been obtained for Δω − 2 rads −1 / div , Δθ = 10 " / div , m e − 4 Nm/div ,
a mismatch of the model electrical parameters values ˆ , Ψ ,Ψ
Ψs∗ , Ψs , Ψ ˆ − 0.16 Wb/div ,
s sa sa
( R s = 1.7 R s N , Ls = 0.7 L s N , Ψ e = 0.8 Ψ eN ) related to the
rated values used for tuning of the speed controller and for
EKF. Figures 7 and 8 was obtained for a simultaneous

a) b)
Fig.6
 a)
Fig.7
a)
Fig.8
VI. CONCLUSION
The paper develops a mechanical sensorless digital control
system of the rotor angular velocity for an PMSM . The
estimation of the mechanical states quantities of the drive
system is made using an extended Kalman filter. The paper
studies the interaction between the observer and closed loop
control system. The simulations show that the dynamic
behavior of the Kalman filter-based control system is good
even if the electric and mechanic parameters of the machine
are different from the nominal values used for the design of
the controllers and the Kalman filter and also for an
incorrect initialization of the estimated position.
VII. REFERENCES
[1] Vas, P., Sensorless Vector and Direct Torque Control, Oxford
University Press (UK), 1998.
[2] Comnac, V., Contributions to the Study of Incremental Motion
Drive Systems with Permanent Magnet Synchronous Machines, PhD
thesis, (In Romanian), Transilvania University of Brasov (Romania),
1998.
[3] Comnac, V., Cernat, M., Moldoveanu, F., Draghici, I., Cernat, R.
M., Variable Structure Control of Surface Permanent Magnet
Synchronous Machines (SPMSM), Proceedings of the International
Conference on Power Conversion and Intelligent Motion, PCIM‘99,
Nuremberg (Germany), June 1999, pp. 351-356.
b)
b)
[4] Moerschell I., Tursini M., A New Vector Control Scheme for
Inverter-fed Permanent Magnet Synchronous Motor Using DSP,
Proceedings of 4th European Conference on Power Electronics and
Applications, Firenze (Italy) 1991, Vol. 4, pp. 683-688.
[5] Beierke, S.; Vas, P.; Simor, B.; Stronach, A. F., DSP-Controlled
Sensorless AC Vector Drives Using Extended Kalman Filter,
Proceedings of the International Conference on Power Conversion
and Intelligent Motion, PCIM‘97, Nuremberg (Germany), June 1997,
pp. 31-41.
[6] Kettle, P., Murray, A., Moynihan, F., Sensorless Control of a
Brushless DC Motor Using an Extended Kalman Estimator. .
Proceedings of the International Conference on Power Electronics
and Motion Control PCIM ‘98, Nuremberg (Germany), May 1998,
pp.385-392
[7] Liu, S.; Stiebler, M., State Estimation of PWM Inverter’s Fed
Synchronous Motor by Using Stochastic Filtering Techniques.,
Proceedings of the International Conference on Electrical Machines
ICEM’90, Cambridge-Massachusetts (USA), 1990,1206-1211.
[8] Comnac, V.; Moldoveanu, F.; Cernat, M.; ,Draghici, I.; Cernat, R.
M.: Sensorless Control of Interior Permanent Magnet Synchronous
Machine Using Stochastic Filtering Techniques. Proceedings of the
7th International Conference on Optimization of Electrical and
Electronic Equipment OPTIM 2000, Brasov Romania, 11 – 12 May
2000, pp. 619-624.
[9] Dhaouadi, R., Mohan, N., Application of Stochastic Filtering to a
Permanent Magnet Synchronous Motor Drive System without
Electro-Mechanical Sensors. Proceedings of International
Conference on Electrical Machines ICEM’90, Cambridge-
Massachusetts (USA), 1990, pp.1225-1230