Escolar Documentos
Profissional Documentos
Cultura Documentos
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
art ic l e i nf o a b s t r a c t
Article history: A reduced order state estimator based on recurrent high-order neural networks (RHONN) trained using
Received 23 April 2014 an extended Kalman filter (EKF) is designed for the magnetic fluxes of a linear induction motor (LIM).
Received in revised form The proposed state estimator does not need the mathematical model of the plant. This state estimator is
20 August 2014
employed to obtain the unmeasurable state variables of the LIM in order to use a state feedback
Accepted 11 October 2014
nonlinear controller. A neural inverse optimal control is implemented to achieve trajectory tracking for a
Communicated by Xu Zhao
position reference. Real-time implementation results on a LIM prototype illustrate the applicability of
the proposed scheme.
Keywords: & 2014 Elsevier B.V. All rights reserved.
Neural optimal control
Recurrent neural networks
Neural observers
Linear induction motor
Real-time implementation
Control Lyapunov function
http://dx.doi.org/10.1016/j.neucom.2014.10.031
0925-2312/& 2014 Elsevier B.V. All rights reserved.
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
2 V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Neural networks can be trained by many different algorithms systems that are or can be written in the nonlinear block
[24]; however, most of these algorithms normally encounter controllable form; second, the stability proof on Lyapunov basis
technical problems such as local minima, slow learning and high for the proposed controller, then the use of a reduced order neural
sensitivity to initial conditions, among others [11]. The Extended observer to relax the full state measurement assumption. Finally
Kalman Filter (EKF) form the basis of a second-order neural the real-time implementation of the proposed inverse optimal
network training method [25], where the synaptic weights control scheme and the neural state estimator on a LIM prototype
become the states to be estimated. The EKF training algorithm in order to show the effectiveness of the proposed scheme, for a
provides a recursive optimal estimator for the neural weights. nonlinear system under internal and external disturbances which
Optimal control of nonlinear systems deals with obtaining a model, parameters and uncertainties are considered unknown and
control law for a given system such that a cost functional is with partial state measurements.
minimized [26], Dynamic programming, developed by Bellman In the following, Section 2 presents mathematical preliminaries
[27], is a solution for optimal control which leads to a nonlinear for the neural networks state estimation. Section 3 includes the
partial differential equation named as the Hamilton–Jacobi–Bell- inverse optimal control basis. In Section 4 the RHONN identifier is
man (HJB) equation. Solving this equation is not straightforward: explained. Sections 5 describes the neural inverse optimal control
for systems of dimension higher than two there are no practical application for LIM. Section 6 presents the real-time implementa-
ways to solve such equation [28]. Inverse optimal control is an tion results and Section 7 exposes the respective conclusions.
alternative for optimal control synthesis which avoids the need to
solve the associated HJB equation [28,29]. For the inverse
approach, a stabilizing feedback control law, based on a priory 2. Neural networks state estimation
knowledge of a Control Lyapunov Function (CLF), is designed first
and then it is established that this control law optimizes a cost Through this paper, subindex k is used as the sampling time,
functional. In this paper, we propose to integrate a reduced order with k A f0g [ Z þ .
RHONN state estimator and an inverse optimal control law based In this paper, we consider the discrete-time multiple-input
on a CLF for nonlinear systems. multiple-output nonlinear system
This mentioned control scheme is applied to a linear induction
χ k þ 1 ¼ Fðχ k ; uk Þ ð1Þ
motor (LIM). The LIM is a linear electric actuator on which the
electrical energy is turned into mechanical translational move- where χ k A R is the state of the system, uk A R
n m
is the control
ment; this is, a mobile element on the motor moves linearly with input and f ðχ k Þ : Rn -Rn is a smooth map.
respect to a stationary element [30]. Induction motors have been
widely studied and several control approaches have been applied
2.1. Recurrent high-order neural networks
to them [31–33]. LIMs present advantages with respect to other
types of motors. They develop magnetic forces directly between
For practical situations, the mathematical model of the system
the mobile element and the stationary element, without the need
to be controlled is usually unknown; then, a RHONN identifier is
of physical contact between both elements, which would restrict
employed to obtain a neural model of the system, required for the
the system dynamics. Then, LIM can reach higher speed and
implementation of the control law. The discrete-time RHONN
reduces undesirables vibrations [30]. For these reasons the LIM
employed for identification of a nonlinear system (1) is defined as
has been employed widely in industrial applications such as steel,
textile, nuclear and space industries [34]. However, the most xi;k þ 1 ¼ wi ϕi ðxk ; uk Þ; i ¼ 1; …; n ð2Þ
extensive application for LIMs is for public transportation by
where xi is the state of the ith neuron, wi is the respective online
means of high speed trains. The idea of using linear motors for
adapted weight vector, n is the system state dimension and
mass public transportation is not new. However, the attention is
ϕi ðxk ; uk Þ is given by
focused on the LIM again due to recent developments of smart
2 3
grids. These are complex systems with unknown uncertainties and dij ð1Þ
2 3 6 ∏ yi 7
disturbances [35], which cause that the control synthesis can be 6 j A I1
j
7
ϕ i1 6 7
very difficult to handle with traditional approaches, requiring the 6 7 6 d ð2Þ 7
6 ϕi 7 6 ∏ y i j 7
application of intelligent control ones [17]. 6 27 6 ij 7
ϕi ðxk ; uk Þ ¼ 6 ⋮ 7 ¼ 6 j A I2 7 ð3Þ
Control of Rotational Induction Motor (RIM) has been extensively 6 7 6 7
4 5 6 ⋮ 7
studied, which is not the case for LIM, even if driving principles of both ϕiL 6
6
7
di ðLi Þ 7
kind of motors are similar. However, recently different control
i
4 ∏ yi j 5
j
j A I Li
techniques have been developed for LIM. For instance in [18] an
adaptive backstepping-sliding mode controller is proposed, in [19] a with Li being the respective number of high-order connections,
fuzzy sliding-mode controller is implemented in a field-programmable I 1 ; I 2 ; …; I Li is a collection of non-ordered subsets of dimension
gate array, and then in [20] a robust controller is proposed to relax 1; 2; …; n þ m, m is the number of external inputs, dij is a non-
disturbances requirements with the fusion of an integral-proportional negative integer and yi is defined as follows:
position control and neural network to estimate disturbances. In [21] 2 3 2 3
is established a field-oriented control scheme, considering the end yi1 Sðx1 Þ
6 ⋮ 7 6 7
effect. For those controllers the design is developed for continuous- 6 7 6 ⋮ 7
6 7 6 7
time and implemented experimentally for position trajectory tracking. 6 yi 7 6 Sðxn Þ 7
6 n 7 6 7
Although, those controllers are robust to uncertainties, they require yi ¼ 6 7 ¼6 7 ð4Þ
6 yin þ 1 7 6 u1 7
previous knowledge of plant model and/or plant parameters at least 6 7 6 7
6 ⋮ 7 6 ⋮ 7
their nominal values. Besides, the progresses in digital equipment have 4 5 4 5
yin þ m um
attracted considerable efforts to the design of high performance
discrete-time controllers for continuous-time plants, which has not In (4), uk ¼ ½u1 ; …; um T is the input vector to the RHONN and
been studied deeply as the continuous-time ones [18–21]. SðÞ is defined by
The main contributions of the paper are: first a novel inverse
optimal neurocontroller with control gain matrix reduction for SðxÞ ¼ tanhðγ xÞ ð5Þ
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3
with γ a positive constant. The neural weights updated by means and ϕj are defined as in (3). The output estimation error is given by
of the EKF-based training algorithm are described by the following
ek ¼ yk y^ ð13Þ
equations:
wi;k þ 1 ¼ wi;k þ ηi K i;k ei;k ð6Þ and the state estimation error is defined as
x~ k ¼ xk x^ ð14Þ
K i;k þ 1 ¼ P i;k H i;k M i;k ð7Þ
The neural state estimator defined by Eq. (12), trained with the
P i;k þ 1 ¼ P i;k K i;k H Ti;k P i;k þ Q i ð8Þ EKF (6), ensures that the output estimation error (13) and the state
with estimation error (14) are semiglobally uniformly ultimately
h i1 bounded (SGUUB) [17]. For a detailed analysis, see [17].
M i;k ¼ Ri þ H Ti;k P i;k H i;k
ei;k ¼ χ i;k xi;k ð9Þ
3. Inverse optimal control
where wi is the ith weight vector, ei is the neural identification
error, P i A RLi Li is the weight estimation error covariance matrix, Li
For the reminder of this paper, let us consider the following
is the respective number of neural network weights, χ A Rn is the
affine discrete-time nonlinear system
plant vector state, x A Rn is the neural network output, K A RLi n is
the Kalman gain matrix, Q i A RLi Li is the neural network weight χ k þ 1 ¼ f ðχ k Þ þgðχ k Þuk ð15Þ
estimation noise covariance matrix, R A Rnn is the measurement
noise covariance, and H i A RLi n is a matrix of the derivatives of the where χ k A Rn is the state of the system, uk A Rm is the control
network outputs with respect to all trainable weights as follows: input and f ðχ k Þ : Rn -Rn and gðχ k Þ : Rn -Rnm are smooth maps.
Without loss of generality, system (15) is supposed to have an
∂xi;k T equilibrium point at χ k ¼ 0.
H i;k ¼ ð10Þ
∂wi;k An inverse optimal control is employed in order to achieve
trajectory tracking of a position reference. For the state feedback
As additional parameter, the rate learning ηi is introduced
control synthesis, the following assumption is established:
in (6), such that 0 o ηi r 1. Usually P i;k , Qi and Ri are initialized as
diagonal matrices.
Assumption 1. The full state xk of system (15) is available. □
2.2. Reduced-order neural state estimator If the full state vector is not completely measurable, Assump-
tion 1, may be fulfilled by means of a state estimator considering
The characteristic robustness of the neural networks is desir- the discrete-time separation principle for nonlinear systems,
able for the design of a state estimator. With the aim of reducing which is explained in Section 3.3.
the computational complexity of this algorithm, a reduced order Let us propose a control Lyapunov function (CLF) as
neural state estimator may be employed, which uses the available
states measures in order to estimate the unmeasurable system Vðχ k Þ ¼ 12 χ Tk P χ k ; P ¼ PT 4 0 ð16Þ
states. In the following, a neural state estimator with a
in order to ensure stability of system (15), which will be achieved
Luenberger-type structure is described. This state estimator is
by defining an appropriate matrix P. Instead of solving the
based on the RHONN, and then it is named the recurrent high
associated HJB equation, the inverse optimal control synthesis is
order neural observer (RHONO) [17]. Which only requires to know
based on the knowledge of V ðχ k Þ.
the system state order, without the need of a nominal system
The inverse optimal control law for system (15) with (16) as CLF
model [36].
is defined as
The reduced-order neural state estimator scheme consists of a
neural identifier, as in (2), for the measured system states and a 1 ∂Vðχ k þ 1 Þ
uk ¼ R 1 ðχ k Þg T ðχ k Þ
neural state estimator for the unmeasurable ones, such that a 2 ∂χ k þ 1
complete neural model of the system is obtained [17]. Let assume 1
the nonlinear system (1) can be written as 1 1
¼ Rðχ k Þ þ g T ðχ k ÞPgðχ k Þ g T ðχ k ÞPf ðχ k Þ ð17Þ
2 2
xa;k þ 1 ¼ f a ðxa;k ; xb;k ; uk Þ
xb;k þ 1 ¼ f b ðxa;k ; xb;k ; uk Þ where Rðχ k Þ ¼ RT ðχ k Þ 4 0 is a weighting matrix whose entries can
yk ¼ xa;k ð11Þ be functions of the system state or can be fixed, this is, it can be
T selected as Rðχ k Þ ¼ R. Since P and Rðχ k Þ are positive definite and
where
xa;k ¼ x1;k …xp;k is the measurable-states vector, xb;k ¼ symmetric matrices, the existence of the inverse in (17) is ensured
T
xp þ 1;k …xn;k is the unmeasurable-states vector and f a ðÞ and f b ðÞ [29]. Now, the control synthesis is to obtain an appropriate matrix
are unknown nonlinear functions. The reduced-order neural state P for (17). The following theorem establishes a sufficient condition
estimator is designed with the following structure [17]: for matrix P on inverse optimal control.
T
x^ k ¼ x^ a;k x^ b;k
T Theorem 1 (Sanchez and Ornelas [26]). Consider the system (15). If
x^ a;k ¼ x^ 1;k …x^ p;k there exist a matrix P ¼ P T 4 0 such that the following inequality
T
x^ b;k ¼ x^ p þ 1;k …x^ n;k holds:
x^ i;k þ 1 ¼ wTi ϕi ðxa;k ; x^ b;k ; uk Þ þ g i ek 1
V f χ k 14 P T1 χ k R χ k þ P 2 χ k P 1 χ k r χ Tk Q χ k
x^ j;k þ 1 ¼ wTj ϕj ðxa;k ; x^ b;k ; uk Þ þ gj ek 1 T
where V f χ k ¼ 2f χ k Pf χ k V χ k , P 1 χ k ¼ g T χ k Pf χ k , P 2 ¼
y^ k ¼ C x^ k ð12Þ
2g χ k Pg χ k , V χ k as defined in (16) and Q ¼ Q 4 0; then the
1 T T
where xi represents the neural identifier dynamics, xj represents equilibrium point χ k ¼ 0 of system (15) is globally exponentially
the reduced-order neural state estimator dynamics and wi, wj, ϕi stabilized by the control law (17).
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
4 V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Moreover, this control law is inverse optimal in the sense that it z2k þ 1 ¼ K 2 z2k þ B2 ðχ 1k ; χ 2k Þz3k
minimizes the meaningful functional given by ⋮
1 zrk þ 1 ¼ f ðzk Þ þBr ðχ k Þuðzk Þ
r
ð28Þ
J χk ∑ l χ k þ uTk R χ k uk ð18Þ
k¼0
If the new error variables (zk) go to zero, according to Theorem
with l χ k ¼ V χ k þ 1 þ V χ k uknT R χ k unk: 1, χ 1k will converge to χ 1δ;k as desired. Then, uðzk Þ can be selected as
For the detailed proof of Theorem 1, we refer the reader to [26]. an inverse optimal control input (17) in order to stabilize (28).
For trajectory tracking, it is convenient to apply a special proce-
dure to system (15), as explained in the following section. 3.2. Neural inverse optimal control
3.1. Nonlinear block controllable form The RHONN (2) can be employed to identify the system in the
NBC form (19). Then, this RHONN has the form
A special state representation of system (15), which is referred
x1k þ 1 ¼ W 1k ϕ ðχ 1k Þ þ W 01 χ 2k
1
as the Nonlinear Block Controllable (NBC) form, can be used to
x2k þ 1 ¼ W 2k ϕ ðχ 1k ; χ 2k Þ þ W 02 χ 3k
2
decompose the control law synthesis into a number of sub-
problems of lower order. The NBC form consists of r blocks [37]: ⋮
r 1
χ 1k þ 1 ¼ f 1 ðχ 1k Þ þB1 ðχ 1k Þχ 2k xrk 1 r 1
þ1 ¼ Wk ϕ ðχ 1k ; χ 2k ; …; χ rk 1 Þ þ W 0r 1 χ rk
χ 2k þ 1 ¼ f 2 ðχ 1k ; χ 2k Þ þ B2 ðχ 1k ; χ 2k Þχ 3k xrk þ 1 ¼ W rk ϕ ðχ k Þ þ W 0r uðχ k Þ
r
ð29Þ
⋮
where and W are weight matrices and ϕ are function vectors
Wik 0i i
χ rk þ 11 ¼ f r 1 ðχ 1k ; χ 2k ; …; χ rk 1 Þ þ Br 1 ðχ 1k ; χ 2k ; …; χ rk 1 Þχ rk defined as in (2).
χ rk þ 1 ¼ f r ðχ k Þ þ Br ðχ k Þuðχ k Þ ð19Þ The RHONN-model (29) can be employed to synthesize the
inverse optimal control law. Hence, the exact parameters of the
where χ k A R , χ k ¼ ½χ χ
n
χ χ A R ; j ¼ 1; …; r; nj denotes
1T
k
rT T
k ⋯ k ,
2T j nj
system (19) are not needed. This scheme is called the neural
the order of each rth block and χ ¼ ½χ j1 χ j2 ⋯χ jn j .
j
inverse optimal control in the sense that it uses the neural weights
Many systems can be represented in the NBC form after a to synthesize the corresponding control inputs [26].
nonlinear transformation [37]. In general, electromechanical sys-
tems are or can be easily written in the NBC form. 3.3. Separation principle for discrete-time nonlinear systems
For trajectory tracking of first block in (19), let define the
tracking error as Consider a discrete-time nonlinear system (15) for which a
z1k ¼ χ 1k χ 1δ;k ð20Þ reduced-order state estimator (12) is applied. The following
proposition is established.
where χ 1δ;k is the desired trajectory signal. Taking one step ahead in
(20) we have Proposition 1. Given a dynamic system state χi, a desired trajectory
for the state χ δ;i and a state estimator output χb i , the following
z1k þ 1 ¼ f ðχ 1k Þ þ B1 ðχ 1k Þχ 2k χ 1δ;k þ 1
1
ð21Þ inequality holds [38]:
Eq. (21) is viewed as a block with state and the state χ 2k is z1k J χ δ;i χ i J r J χ i χ
b i J þ J χ δ;i χ
bi J
considered as a pseudo-control input; desired dynamics can be
where χ δ;i χ i is the state tracking error, χ i χb i is the estimation
imposed as follows:
error and χ δ;i χ
b i is the tracking error for the state estimator. □
z1k þ 1 ¼ f ðχ 1k Þ þ B1 ðχ 1k Þχ 2k χ 1δ;k þ 1 ¼ K 1 z1k
1
ð22Þ
Proposition 1, can be established because of the separation
where K 1 ¼ diagfk11 ; …; kn1 1 g is a Schur matrix such that K 1 z1k is a principle for discrete-time nonlinear systems [39]; then, the
stable dynamic. Then, the desired behavior of χ2k is calculated as trajectory tracking problem can be separated into two parts:
χ 2δ;k ¼ ðB1 ðχ 1k ÞÞ 1 ðχ 1δ;k þ 1 f 1 ðχ 1k Þ þ K 1 z1k Þ ð23Þ
Minimization of χ i χ^ i , which can be achieved by the proposed
neural state estimator trained with the EKF algorithm.
Proceeding along the same way as for the first block, a second
Minimization of χ δ;i χ^ i , which is obtained by the inverse
variable in the new coordinates is defined as
optimal control law.
z2k ¼ χ 2k χ 2δ;k ð24Þ
In order to relax the need of Assumption 1, the proposed
and the desired behavior for χ3k can be calculated. Taking these
inverse optimal controller is combined, under the separation
steps iteratively, the last new variable is defined as
principle, for discrete-time nonlinear systems, with a reduced
zrk ¼ χ rk χ rδ;k ð25Þ order neural state estimator, to allow the applicability of the
proposed scheme to systems without the complete access to the
Taking one step ahead yields state variables measurements.
zrk þ 1 ¼ f ðχ k Þ þBr ðχ k Þuðχ k Þ χ rδ;k þ 1
r
ð26Þ
3.4. Reduction of matrix P dimension
Then, the control law can be selected as
uðχ k Þ ¼ u1;k þ u2;k As mentioned before, for inverse optimal control an appro-
1
r
priate selection of the matrix P is required, which is an n n
¼ Br ðχ k Þ f ðzk Þ f ðχ k Þ þ χ rδ;k þ 1 þuðzk Þ
r
ð27Þ
matrix where n is the dimension of the system state vector; then,
By means of this change of variables, system (19) can be depending on the system, the matrix P can be very large, increas-
represented as ing the calculation complexity of the control law. In order to
facilitate the design of the inverse optimal control, a reduction of
z1k þ 1 ¼ K 1 z1k þ B1 ðχ 1k Þz2k the required dimension of P is proposed.
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 5
Let us consider a system in the form (28). This system has been into account the end effects [40,41] which can be despised
designed such that the first r 1 state blocks have stable depending on the dimensions of the LIM. This model is described
dynamics; then, the inverse optimal control law can be employed by the following equations:
to stabilize only the last state block. Then, P is a nr nr matrix, qk þ 1 ¼ qk þ vk T
where nr o n is the dimension of the rth block. This proposition is
formalized in the following theorem. vk þ 1 ¼ ð1 k2 TÞvk k3 TF L k1 T λα;k ρ1 iα;k k1 T λβ;k ρ2 iα;k
Theorem 2. Let us consider a dynamic system as in (28), to which þ k1 T λα;k ρ2 iβ;k k1 T λβ;k ρ1 iβ;k
the following inverse optimal control is applied:
1
1 1 T
uðzk Þ ¼ R þ Br P Br
T r
Br P f ðzk Þ ð30Þ λα;k þ 1 ¼ ð1 k6 TÞλα;k þ k4 Tvk ρ1 iα;k k4 T ρ1 iα;k þ k5 T ρ2 iα;k
2 2 þ k4 T ρ2 iβ;k k4 Tvk ρ2 iβ;k þk5 T ρ1 iβ;k
T
where P ¼ P 4 0 has dimension nr nr , being nr the last state block
dimension of (28). If there exists a matrix P such that the following
inequality holds λβ;k þ 1 ¼ ð1 k6 TÞλβ;k þ k4 Tvk ρ2 iα;k k4 T ρ2 iα;k k5 T ρ1 iα;k
k4 T ρ1 iβ;k þ k4 Tvk ρ1 iβ;k þk5 T ρ2 iβ;k
1
r T r T
1 T 1 1
f ðzk ÞP f ðzk Þ zrk P zrk P 1 R þ P 2 P 1 r zTk Q zk ð31Þ
2 4 2
where iα;k þ 1 ¼ ð1 þ k9 TÞiα;k k7 T λα;k ρ2 k8 T λα;k vk ρ1
T r
P 1 ¼ Br P f ðzk Þ;
T
P 2 ¼ Br P Br þ k7 T λβ;k ρ1 k8 T λβ;k vk ρ2 k10 Tuα;k
T
and Q ¼ Q 4 0; then, the equilibrium point χ k ¼ 0 of system (28) is
stabilized by control law (30). Moreover, this control law minimizes iβ;k þ 1 ¼ ð1 þ k9 TÞiβ;k þ k8 T λα;k vk ρ2 k7 T λα;k ρ1
the cost function given by (18).
k7 T λβ;k ρ2 k8 T λβ;k vk ρ1 k10 Tuβ;k ð32Þ
Proof. For proof, see Appendix A.
where
np Lsr
3.5. Control scheme ρ1 ¼ sin ðnp qk Þ ρ2 ¼ cos ðnp qk Þ k1 ¼
Dm Lr
np Lsr Rm 1
Fig. 1 represents the proposed scheme of the state estimator, k1 ¼ k2 ¼ k3 ¼
Dm Lr Dm Dm
identifier and controller for the LIM position trajectory tracking.
Rr Lsr Rr
The RHONO employs the input signals of the LIM model and the k4 ¼ np Lsr k5 ¼ k6 ¼
Lr Lr
available state data (χ a;k ) to estimate the unmeasurable magnetic
fluxes of the LIM to complete the state vector χk. The RHONN Lsr Rr Lsr np L2r Rs þ L2sr Rr
k7 ¼ k8 ¼ k9 ¼
identifier is attached to the RHONO and acts in series-parallel with Lr ðL2sr Ls Lr Þ L2sr Ls Lr Lr ðL2sr Ls Lr Þ
the real LIM model, which is considered unknown, and calculates L2 Rs þ L2sr Rr Lr
the corresponding neural weights for identification. The neural k9 ¼ r 2 k10 ¼
Lr ðLsr Ls Lr Þ L2sr Ls Lr
inverse optimal control (NIOC) block uses these neural weights to
synthesize the input signals for the LIM model, stabilizing the
error between the current position of the LIM and the position
reference and minimizing the respective cost functional. and λβ;k are the magnetic fluxes, iα;k and iβ;k are the primary sector
currents, uα;k and uβ;k are the input voltages, Rs is the primary
sector resistance, Rr is the secondary sector resistance, Lsr is the
4. Linear induction motor identification magnetizing inductance, Ls is the primary sector inductance, Lr is
the secondary sector inductance, FL is the load perturbation, Rm is
4.1. LIM discrete-time model the viscous friction coefficient, Dm is the secondary sector mass, np
is the number of poles pairs and T is the sampling time.
The LIM discrete-time mathematical model employed in this
paper is described on the α β frame. This model does not take 4.2. RHONN model
x2;k þ 1 ¼ w21 Sðλα;k Þ2 þ w22 Sðλβ;k Þ2 þ w23 wf Sðvk Þ2 þ2wf ðw21 Sðλα;k Þρ2
w22 Sðλβ;k Þρ1 Þiα;k þ 2wf ðw21 Sðλα;k Þρ1 þ w22 Sðλβ;k Þρ2 Þiβ;k
x3;k þ 1 ¼ w31 Sðvk Þ þ w32 Sðλα;k Þ þ w33 Sðλβ;k Þ þw34 Sðiα;k Þ þ w35 uα;k
x4;k þ 1 ¼ w41 Sðvk Þ þ w42 Sðλα;k Þ þ w43 Sðλβ;k Þ þw44 Sðiβ;k Þ þw45 uβ;k
x5;k þ 1 ¼ w51 Sðqk Þ þ w52 vk ð33Þ
where wij are the online adjustable network weights, exc-
ept wf ¼ 0:001, w14 ¼ 0:001, w15 ¼ 0:001, w35 ¼ 0:02178, w45 ¼
Fig. 1. Reduced order neural state estimator and control scheme. 0:02178 and w52 ¼ 0:001 which are fixed weights for
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
6 V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎
controllability purposes. SðÞ is defined as in (5) with γ ¼ 1. Each 5.1. Change of variables
state in the RHONN model identifies a corresponding state of the
LIM model: x1;k for the velocity, x2;k for the flux magnitude, x3;k and The RHONN identifier (33) fulfills the NBC form with the three
x4;k for the currents alpha and beta respectively, and x5;k for the blocks:
system position. At this stage, it is supposed that all the states " # " #
x1;k x3;k
measures are available. x1k ¼ x5;k ; x2k ¼ ; x3k ¼ ð37Þ
x2;k x4;k
The decoupled EKF training algorithm is applied to update the
trainable neural weights. These weights are separated in five For (33), taking into account Assumption 2, the change of
different vectors according to each neural state; this is, variables can be performed so that the system with the desired
w1 ¼ ½w11 w12 w13 , w2 ¼ ½w21 w22 w23 , w3 ¼ ½w31 w32 w33 w34 , dynamics is represented as
w4 ¼ ½w41 w42 w43 w44 and w5 ¼ w51 and the EKF algorithm is
implemented for each vector. Initial values of the neural weights z1k þ 1 ¼ K 1 z1k þ W 01 2
k zk
are randomly selected. The matrices Qi and R for the Kalman filter k ðχ k Þzk
z2k þ 1 ¼ K 2 z2k þ W 02 3
þ wα5 Sðvk ÞSðγ iβ;k Þρ2 þ L1 ek þ 0:002ρ2 iα;k þ 0:002ρ1 iβ;k þ w44 Sðiβ;k Þ
" #
w35 0
xβ;k þ 1 ¼ wβ1 Sðγ xβ;k Þ þ wβ2 Sðγ iα;k Þð ρ2 þ ρ1 Þ W 03
k ¼ ð39Þ
0 w45
þ wβ3 Sðγ iβ;k Þðρ1 ρ2 Þ þ wβ4 Sðvk ÞSðγ iα;k Þρ2 Trajectory tracking can now be achieved for the position
reference χ 1δ;k .
þ wβ5 Sðvk ÞSðγ iβ;k Þρ1 þL2 ek þ 0:002ρ1 iα;k þ 0:002ρ2 iβ;k
ð35Þ 5.2. Controller synthesis
where xα;k ¼ λ^
α;k is the estimated of the state λα;k , xβ;k ¼ λ^
β;k is the In (38) z1k and z2k will have stable dynamics if the third block
estimated of λβ;k , SðÞ is defined as in (5) with γ ¼ 0:01, state z3k tends to zero when k-1 (Theorem 2). Then, a control law
L1 ¼ L2 ¼ ½0:1 0:01 0:01 0:01 and has to be designed such that it stabilizes z3k þ 1 in (38).
2 3 2 3 Control law (27) can now be applied taking f ðχ k Þ ¼ f ðχ k Þ,
r 3
e1;k vk x1;k
6 e 7 6 i x 7 Br ðχ k Þ ¼ W 03;k and uðzk Þ as in (17). Then, the inverse optimal control
6 3;k 7 6 α;k 3;k 7
ek ¼ 6 7 6 7
6 e4;k 7 ¼ 6 iβ;k x4;k 7 ð36Þ law applied to (38) is as follows:
4 5 4 5
1
qk x5;k uðχ k Þ ¼ W 03;k f ðzk Þ f ðχ k Þ þ χ 3δ;k þ 1 þ uðzk Þ
3 3
e5;k ð40Þ
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 7
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
8 V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎
-0.1
displays the mean square error and it standard deviation of the
-0.2
proposed controller against a conventional sliding mode controller
-0.3 [42]. It is important to note that the controller in [42], requires
-0.4 previous knowledge of fluxes dynamics, the parameters to design
a conventional observer and the estimation of the load torque.
-0.5
0 10 20 30 40 50 60 70 These information are not required for the scheme proposed in
Time (s) this paper due to the use of the neural observer. Besides, it is
important to note that even if the sliding mode controller is not
Fig. 7. Real-time motor position with a 50 cm sinusoidal reference.
new, for comparison purposes, both experiments are performed
using the same prototype.
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 9
1
Appendix A For system (A.1), f ðzk Þ ¼ K 1 z1k þ B1 ðz1k Þz2k ; taking this fact into
account, the following term is analyzed
1T 1
Proof. The proof of Theorem 2 is performed for a two-blocks f ðzk ÞP 11 f ðzk Þ z1T 1
k P 11 zk
T
system;
¼ K 1 z1k þ B1 ðz1k Þz2k P 11 K 1 z1k þ B1 ðz1k Þz2k
1
z1k þ 1 ¼ f ðzk Þ ¼ K 1 z1k þ B1 ðz1k Þz2k z1T 1
k P 11 zk
2
z2k þ 1 ¼ f ðzk Þ þ B2 uðzk Þ ðA:1Þ ¼ z1T T 1 1T T 1 1 2
k K 1 P 11 K 1 zk þ zk K 1 P 11 B ðzk Þzk
1T 2T 1T 1T
The general case proof follows straightforward from this case. þ z2T 1 2
k B ðzk ÞP 11 K 1 zk þ zk B ðzk ÞP 11 B ðzk Þzk
First, let suppose that the control goal is to stabilize system (A.1) z1T 1
k P 11 zk
employing a full-order matrix P for inverse optimal control. Then, T 1T T 1 1 2
¼ z1T 1
k ðK 1 P 11 K 1 P 11 Þzk þ zk K 1 P 11 B ðzk Þzk
P is selected as 1T 2T 1T 1T
" # þ z2T 1
k B ðzk ÞP 11 K 1 zk þ zk B ðzk ÞP 11 B ðzk Þzk
2
P 11 P T21 " T #
P¼ ðA:2Þ K 1 P 11 K 1 P11 K T1 P 11 B1 ðzk Þ
P 21 P 22 ¼ zTk 1T
zk
B ðzk ÞP 11 K 1 B ðzk ÞP 11 B1 ðzk Þ
1T
where P 11 ¼ P T11 A Rn1 n1 , P 21 A Rn2 n1 and P 22 ¼ P T22 A Rn2 n2 , being ¼ zTk Q P zk ðA:8Þ
ni the dimension of the ith state block. Then, the inverse optimal
control law is defined as
1 Note that QP is a symmetric matrix. Substituting (A.8) in (A.7),
1 1 then
uk ¼ R þ g T ðzk ÞPgðzk Þ g T ðzk ÞPf ðzk Þ
2 2
1 2T 2
" # ! 1 2 ðf ðzk ÞP 22 f ðzk Þ z2T 2
k P 22 zk Þ
1 1h i P
11 P T21 0
1
¼ Rþ 0 B 2T
14 f
2T
ðzk ÞP 22 B2 R þ 12 B2T P 22 B2
2
B2T P 22 f ðzk Þ
2 2 P 21 P 22 B2
" #" 1 # r 12 zTk Qzk 12 zTk Q P zk ¼ 12 zTk Q zk ðA:9Þ
h i P P T21 f ðzk Þ
2T 11
0 B
P 21 P 22 2
f ðzk Þ where Q ¼ Q þQ P is a symmetric matrix. Let us select Q such that
1
Q is a positive definite matrix.
1 1 1 2 Then, if there exists P ¼ P 22 such that inequality (31) holds,
¼ R þ B2T P 22 B2 B2T P 21 f ðzk Þ þ B2T P 22 f ðzk Þ ðA:3Þ
2 2 then P 21 ¼ 0 can be selected and the control law (30) achieves
system stabilization. □
Let us consider the case of stabilizing only the last state block of
(A.1). Employing the matrix P A Rn2 n2 , the following inverse References
optimal control law is obtained
1 [1] H. Wang, B. Chen, C. Lin, Adaptive neural trackingcontrol for a class of
2
u k ¼ 12 R þ 12 B2T P B2 B2T P f ðzk Þ ðA:4Þ stochastic nonlinear systems, Int. J. Robust Nonlinear Control 24 (7) (2014)
1262–1280.
[2] F.J. Lin, R.J. Wai, C.H. Lin, D.C. Liu, Decoupled stator-flux-oriented induction
motor drive with fuzzy neural network uncertainity observer, IEEE Trans. Ind.
Comparing (A.3) and (A.4) it is possible to conclude that if Electron. 47 (2) (2000).
P 21 ¼ 0 and P 22 ¼ P , then uk ¼ u k ; this is, stabilizing only the last [3] Z. Wang, D.W.C. Ho, X. Lin, State estimation for delayed neural networks, IEEE
Trans. Neural Netw. 16 (1) (2005).
system block is equivalent to stabilizing the full system. [4] S. Yin, G. Wang, Robust PLS approach for KPI related prediction and diagnosis
In the following, a sufficient condition is established such that against outliers and missing data, Int. J. Syst. Sci. 45 (7) (2014) 1375–1382.
the submatrix P21 can be selected as zero and the inverse optimal [5] J.A. Ruz-Hernandez, E.N. Sanchez, Dionisio A. Suarez, Designing an associative
memory via optimal training for fault diagnosis, in: Proceedings of the
control achieves system stabilization. From Theorem 1, the matrix
International Joint Conference on Neural Networks 2006, Vancouver, BC, 2006.
P must be selected such that [6] A.J. Krener, A. Isidori, Linearization by output injection and nonlinear
observers, Syst. Control Lett. 3 (1) (1983).
T T
2 f ðzk ÞPf ðzk Þ zk Pzk
1
[7] S. Nicosia, A. Tornambe, High-gain observers in the state and parameter
estimation of robots having elastic joins, Syst. Control Lett. 13 (4) (1989) 20.
T 1 T
14 f ðzk ÞPgðzk Þ R þ 12 g T ðzk ÞPgðzk Þ g ðzk ÞPf ðzk Þ [8] I. Salgado, L. Fridman, O. Camacho, I. Chairez, Discrete time super-twisting
observer for 2n dimensional systems, in: 8th International Conference on
r 12 zTk Qzk ðA:5Þ Electrical Engineering Computing Science and Automatic Control, Merida,
Yucatan, Mexico, October 2011.
T
with Q ¼ Q 4 0. For the NBC-form system (A.1), employing a P [9] J.A. Moreno, M. Osorio, A Lyapunov approach to second-order sliding mode
controllers and observers, in: Proceedings of the 47th IEEE Conference on
matrix defined as in (A.2) with P 21 ¼ 0 Decision and Control, Cancun, Q. Roo, Mexico, December 2008.
[10] G.A. Rovithakis, M.A. Chistodoulou, Adaptive Control with Recurrent High-
1T 1 2T 2
1
2 f ðzk ÞP 11 f ðzk Þ þ f ðzk ÞP 22 f ðzk Þ z1T 1 2T 2
k P 11 zk zk P 22 zk Order Neural Networks, Springer-Verlag, Berlin, Germany, 2000.
1 [11] A.Y. Alanis, E.N. Sanchez, A.G. Loukianov, M.A. Perez, Real-time recurrent
2T 2
14 f ðzk ÞP 22 B2 R þ 12 B2T P 22 B2 B2T P 22 f ðzk Þ neural state estimation, IEEE Trans. Neural Netw. 22 (3) (2011).
[12] R.P. Landim, B.R. de Menezes, S.R. Silva, W.M. Caminhas, On-line neo-fuzzy-
r 12 zTk Qzk ðA:6Þ neuron state observer, in: Proceedings of the 6th Brazilian Symposium on
Neural Networks, Rio de Janeiro, 2000.
[13] S.H. Kim, T.S. Park, J.Y. Yoo, G.T. Park, Speed sensorless vector control of an
and then, induction motor using neural network speed estimation, IEEE Trans. Ind.
Electron. 48 (3) (2001).
2T 2
1
2 f ðzk ÞP 22 f ðzk Þ z2T 2
k P 22 zk [14] B. Karanayil, M.F. Rahman, C. Grantham, Stator and rotor resistance observer
1 for induction motor drive using fuzzy logic and artificial neural networks, IEEE
2T 2
14 f ðzk ÞP 22 B2 R þ 12 B2T P 22 B2 B2T P 22 f ðzk Þ Trans. Energy Convers. 20 (4) (2005).
[15] F. Abdollahi, H.A. Talebi, R.V. Patel, A stable neural network observer with
1T 1
r 12 zTk Qzk 12 f ðzk ÞP 11 f ðzk Þ z1T 1
ðA:7Þ application to flexible joint manipulators, in: Proceedings of the 9th Interna-
k P 11 zk
tional Conference on Neural Information Processing, Singapore, 2002.
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i
10 V.G. Lopez et al. / Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎
[16] S.M. Gadoue, D. Giaouris, J.W. Finch, Sensorless control of induction motor Alma Y. Alanis, received the B.Sc degree from Instituto
drives at very low and zero speeds using neural network flux observers, IEEE Tecnologico de Durango (ITD), Durango Campus, Dur-
Trans. Ind. Electron. 56 (8) (2009). ango, Durango, in 2002, the M.Sc. and the Ph.D. degrees
[17] A.Y. Alanis, E.N. Sanchez, M. Hernandez, L.J. Ricalde, Discrete-time reduced in electrical engineering from the Advanced Studies and
order neural observer for linear induction motors, in: 2011 IEEE Symposium Research Center of the National Polytechnic Institute
on Computation Intelligence Applications in Smart Grid, Paris, France, April (CINVESTAV-IPN), Guadalajara Campus, Mexico, in 2004
2011. and 2007, respectively. Since 2008 she has been with
[18] F.J. Lin, P.-H. Shen, S.-P. Hsu, Adaptive backstepping sliding mode control for University of Guadalajara, where she is currently a chair
linear induction motor drive, IEE Proc. Electr. Power Appl. 149 (3) (2002). professor in the Department of Computer Science. She is
[19] F.J. Lin, D.-H. Wang, P.-K. Huang, FPGA-based fuzzy sliding-mode control for a also a member of the Mexican National Research System
linear induction motor drive, IIEE Proc. Electr. Power Appl. 152 (5) (2005). (SNI-1). Her research interest centers on neural control,
[20] F.J. Lin, R.J. Wai, Robust control using neural network uncertainty observer for backstepping control, block control, and their applica-
linear induction motor servo drive, IEEE Trans. Power Electron. 17 (2) (2002). tions to electrical machines, power systems and robotics.
[21] G. Kang, K. Nam, Field-oriented control scheme for linear induction motor
with the end effect, IEE Proc. Electr. Power Appl. 152 (6) (2005).
[22] J. Hopfield, Neurons with graded responses have collective computational
properties like those of two state neurons, Proc. Natl. Acad. Sci. USA 81 (10)
(1984). Edgar N. Sanchez obtained the BSEE from Universidad
[23] A.Y. Alanis, E.N. Sanchez, A.G. Loukianov, G. Chen, Discrete-time output Industrial de Santander (UIS), Bucaramanga, Colombia
trajectory tracking by recurrent high-order neural network control, in: in 1971, the MSEE from CINVESTAV-IPN (Advanced
Proceedings of the Conference on Decision and Control 2006, San Diego, CA, Studies and Research Center of the National Polytechnic
USA, December 2006. Institute), Mexico City, Mexico, in 1974 and the Docteur
[24] S. Haykin, Neural Networks A Comprehensive Foundation, 2nd ed., Pearson Ingenieur degree in Automatic Control from Institut
Prentice Hall, Upper Saddle River, NJ, USA, 2005. Nationale Polytechnique de Grenoble, France in 1980.
[25] S. Haykin, Kalman Filtering and Neural Networks, John Wiley and Sons, Inc., He was granted an USA National Research Council
New York, NY, USA, 2001. Award as a research associate at NASA Langley Research
[26] E.N. Sanchez, F. Ornelas, Discrete-time Inverse Optimal Control for Nonlinear Center, Hampton, Virginia, USA (January 1985–March
Systems, CRC Press, Boca Raton, FL, USA, April 2013. 1987). His research interest center in Neural Networks
[27] R.E. Bellman, Dynamic Programming, Princeton University Press, Princeton, NJ, and Fuzzy Logic as applied to Automatic Control sys-
USA, 1957. tems. He has been advisors of 12 Ph.D. thesis and 33 M.
[28] R.A. Freeman, P.V. Kokotovic, Robust Nonlinear Control Design. State space and Sc Thesis. Since January 1997, he is a professor of CINVESTAV-IPN, Guadalajara
Lyapunov Techniques, Birkhauser, Boston, MA, USA, 1996. Campus, Mexico.
[29] F. Ornelas, A.G. Loukianov, E.N. Sanchez, Discrete-time nonlinear systems
inverse optimal control: a control Lyapunov approach, in: IEEE Multiconfer-
ence on Systems and Control, Denver, CO, USA, September 2011.
[30] T. Wildi, Electrical Machines, Drives and Power Systems, 5th ed., Prentice Hall, Jorge Rivera received the B.Sc. degree in Electronic
Upper Saddle River, NJ, USA, 2002. Engineering from the Technological Institute of the Sea,
[31] J. Thomas, A. Hansson, Speed tracking of a linear induction motor-enumerative Mazatlan, Mexico, in 1999, and the M.Sc. and Ph.D.
nonlinear model predictive control, IEEE Trans. Control Syst. Technol. 21 (5) degrees in Electrical Engineering from the Advanced
(2013) 1956–1962. Studies and Research Center, National Polytechnic
[32] M. Feemster, P. Aquino, D.M. Dawson, A. Behal, Sensorless rotor velocity Institute, Guadalajara, Mexico, in 2001 and 2005,
tracking control for induction motors, IEEE Trans. Control Syst. Technol. 9 (4) respectively. Since 2006, he has been with the Electro-
(2001). nics Department in the University of Guadalajara. His
[33] G. Rubio-Astorga, J.D. Sanchez-Torres, J. Canedo, A.G. Loukianov, High-Order research interests focus on sliding mode control, reg-
sliding mode block control of single-phase induction motor, IEEE Trans. ulation theory, discrete-time nonlinear control systems,
Control Syst. Technol. 22 (5) (2014) 1828–1836. and their applications to electrical machines.
[34] H. Toliyat, G.B. Kliman, Handbook of Electric Motors, 2nd ed., CRC Press, Boca
Raton, FL, USA, 2004.
[35] G.K. Venayagamoorthy, Smart grid and electric transportation, in: 12th
International IEEE Conference on Intelligent Transportation Systems, October
2009, St. Louis, MO, USA, pp. 1–2.
[36] E. Sanchez, A. Alanis, A. Loukianov, Discrete Time High Order Neural Control,
Springer, New York, NY, USA, 2008.
[37] A.G. Loukianov, Robust block decomposition sliding mode control design,
Math. Probl. Eng. 8 (1) (2002) 4–5.
[38] B.S. Leon, A.Y. Alanis, E.N. Sanchez, E. Ruiz-Velazquez, F. Ornelas-Tellez, Inverse
optimal neural control for a class of discrete-time nonlinear positive systems,
Int. J. Adapt. Control Signal Process. 26 (1) (2012) 20.
[39] W. Lin, C.I. Byrnes, Design of discrete-time nonlinear control systems via
smooth feedback, IEEE Trans. Autom. Control 39 (11) (1994).
[40] H.A. Hairik, M. Hassan, Dynamic model of linear induction motor considering
the end effects, Iraq J. Electr. Electron. Eng. 5 (1) (2009).
[41] L. Radzevicius, E. Matkevicius, The generalized model of the linear induction
motor, Electron. Electr. Eng. 190 (7) (2006).
[42] A.G. Loukianov, J. Rivera, J.M. Cañedo, Discrete-time sliding mode control of an
induction motor, in: Proceedings IFAC’02, Barcelone, Spain, 2002.
Please cite this article as: V.G. Lopez, et al., Real-time implementation of neural optimal control and state estimation for a linear
induction motor, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.10.031i