You are on page 1of 6

Comparison of Estimator-Based Compensation Schemes

for Hydrostatic Transmissions with Uncertainties


Dang Ngoc Danh and Harald Aschemann

Abstract— In this paper, the achievable tracking performance


of a decentralized control scheme of a hydrostatic transmission
is investigated for three alternative estimators in combination
with output feedback linearization: a state and disturbance
observer, an adaptive parameter estimator and a neural net-
work. All these estimators address the given disturbances and
uncertainty in the hydraulic system and contribute to the
robustness of the feedback linearization as well the tracking
accuracy. Firstly, the three control schemes are evaluated by
means of simulations and, secondly, validated by experiments
on a dedicated test rig at the Chair of Mechatronics, University
of Rostock. Fig. 1. Principle configuration of a hydrostatic transmission system.

I. I NTRODUCTION
A hydrostatic transmission system typically consists of
two main components – a hydraulic pump and a hydraulic moreover, are affected by load disturbances and unavoid-
motor, each with variable volumetric displacement – that are able parameter uncertainty. Model uncertainty is physically
connected by hydraulic hoses in a closed hydraulic circuit. related, for example, to variations in the fluid temperature, to
The pump is driven by an electric motor or a combustion kinematic viscosity, to the elasticity of the hydraulic hoses,
engine, which supplies mechanical power to the system. This and the leakage oil flow, cf. [3].
power is converted to hydraulic power – volume flow times
In current industrial practice, gain-scheduled PID con-
pressure – and transmitted to the hydraulic motor, where it
trollers are still the typical choice to control hydrostatic
is converted back to mechanical power at the output shaft.
transmissions, cf. [4]. Nevertheless, if higher performance
Hydrostatic transmissions are used in classical industrial
demands are to be fulfilled, many nonlinear control ap-
applications like heavy working machines, construction and
proaches have proven advantageous and have been success-
agriculture vehicles but also in recent applications such
fully validated in the last decade, see [5] for a general
as power-split gearboxes and off-road vehicles, cf. [1], as
overview and a comprehensive list of references. Regarding
well as wind turbines, cf. [2]. Hydrostatic transmissions
the control structure, both centralized and decentralized
offer many advantages in comparison to purely mechanical
topologies are applicable and enable an accurate tracking
solutions: they provide a continuously variable transmission
control. The decentralized control of the motor bent angle
ratio, a high power density, a directional reversion without
and the motor angular velocity, see [5], outperformed the
changing gears, and it is able to serve as a wearless braking
centralized topology. The achievable tracking performance is
system, cf. [1]. The most popular form of a hydrostatic
higher, and the implementation is simpler in direct compar-
pump and motor involves a design with axial pistons. With
ison to the centralized approach, see [5] for further details.
a changing swash-plate angle of the hydraulic pump and a
changing bent angle of the hydraulic motor, the transmission In this study, the control-oriented model of a hydrostatic
ratio can be varied by means of displacement units. As a transmission is presented in Sec. II. A decentralized scheme
result, both the motor torque and angular velocity of motor discussed in Sec. III serves as the basis for the investigation
output shaft can be adjusted. Beside many advantages like the of alternative estimators – a state and disturbance observer,
flexible geometrical arrangement and the operation principle an adaptive parameter estimator and a neural network –
that let hydrostatic transmissions become a new trend in employed for the feedback linearization of the nonlinear sys-
industry, energy efficiency and control issues still need to be tem. The estimator-based approaches considered in Sec. IV
addressed. From a control point of view, sophisticated control are meant to cancel uncertain nonlinear terms as well as
approaches are required because hydrostatic transmissions to compensate for disturbances. A comparison is performed
are characterized by nonlinear differential equations and, in Sec. V by means of both simulation and corresponding
experiments. Successful experiments at the test rig validate
Dang Ngoc Danh and Harald Aschemann are with the the concept and show that all variants are applicable and,
Chair of Mechatronics, University of Rostock, Justus-von-
Liebig-Weg 6, Rostock D-18059, Germany, {Ngoc.Dang, moreover, allow for an accurate trajectory tracking. Conclu-
Harald.Aschemann}@uni-rostock.de. sions and a short outlook finish the paper in Sec. VI.

978-1-5386-4325-9/18/$31.00 ©2018 IEEE 692

Authorized licensed use limited to: Universitaetsbibl Rostock. Downloaded on August 11,2021 at 17:16:02 UTC from IEEE Xplore. Restrictions apply.
C. Dynamics of the displacement units
The dynamics of the displacement units, which represent
the actuators of the hydrostatic transmission and which are
responsible for changing the swash-plate angle and the bent
axis angle, cover the relationships between normalized tilt
angles and the input signals. They can be described by first-
order lag systems as follows

Fig. 2. Hydrostatic transmission test rig at Chair of Mechatronics,


TuP α̃˙ P + α̃P = kP uP ,
(4)
University of Rostock. TuM α̃˙ M + α̃M = kM uM .
Here, TuP/uM are the time constants. kP,M are proportional
gains and uP/M are normalized input signals of the pump
II. M ATHEMATICAL M ODEL OF H YDROSTATIC
and motor actuators. Due to given limits of the mechanical
T RANSMISSION S YSTEM
design, the actuator dynamics is subject to saturation. In this
The overall system comprises two physical domains: a study, for the sake of simplicity, the saturation is not con-
hydraulic subsystem and a mechanical subsystem, cf. [5]. sidered explicitly in system model but is addressed properly
The hydraulic subsystem characterizes the pressure dynamics by the planning of feasible desired trajectories.
of the hydraulic oil, whereas the mechanical subsystem Combining all subsystems, the overall system can be de-
describes the dynamics of motor output shaft. scribed by a fourth-order nonlinear state-space representation
⎡ ⎤ ⎡ − TuM
1
α̃M + TkuM
M
uM

A. Pressure dynamics α̃˙ M
⎢ α̃˙ P ⎥ ⎢ − T1uP α̃P + TkuP
P ⎥
⎥=⎢ ⎥
uP
Assuming small values of the pump swash-plate angle ⎢ . (5)
⎣ Δ ṗ ⎦ ⎣ 2 ṼP α̃P ωP − 2 ṼM α̃M ωM − qu ⎥

(|αP | ≤ 18o ) and of the motor bent axis angle (αM ≤ 20o ), CH CH CH ⎦
the corresponding volume flow rates of pump and motor can ω̇M − dJVV ωM + ṼJVM Δpα̃M − JτVu
be approximately expressed by
III. D ECENTRALIZED C ONTROL S CHEME
qP = ṼP α̃P ωP ,
(1) In the decentralized control scheme, the overall system dy-
qM = ṼM α̃M ωM .
namics is partitioned into two subsystems: the first subsystem
Here, ωP/M are the angular velocities of pump and motor. corresponds to the differential equation for the motor bent
The variable α̃P , which is limited to the interval α̃P ∈ [−1, 1], axis angle, whereas the second one comprises the remaining
represents the normalized swash-plate angle. The variable last differential equations and serves for the control of the
α̃M ∈ [εM , 1], with εM > 0, denotes the normalized bent axis motor angular velocity.
angle. ṼP/M stand for the volumetric displacements of the
pump and the motor, respectively. For practical reasons,
D Md uM

Hydrostatic Transmission
the pressure dynamics is reduced to the dynamics of the D M - Controller
difference pressure between the high pressure and the low
Abc
F9

pressure side. Assuming symmetric conditions and neglect- 'p


ing the pressure losses in the hydraulic hoses, the differential Dˆ M Observer
ZM
equation for the difference pressure becomes
2  qu 
ZMd
Δ ṗ = ṼP α̃P ωP − ṼM α̃M ωM − . (2) ZM - Controller uP
CH 2
Here, CH denotes the hydraulic capacitance, where qu is a Fig. 3. Decentralized control scheme with an observer for the reconstruc-
disturbance resulting from unknown leakage flows in the tion of unmeasurable parameter.
physical system.

B. Equation of motion A. Control of normalized bent axis angle


The dynamics of the motor angular velocity is governed A flatness-based controller, see [6] for an overview of
by the following equation of motion this design method, is developed for the normalized motor
bent axis angle, cf. [7]. Solving the first differential equation
JV ω̇M + dV ωM = ṼM Δpα̃M − τu . (3) (5) for the input signal uM and introducing the first time
Here, JV is mass moment of inertia, and dV denotes the derivative as stabilizing control input υM = α̃˙ M results in the
velocity proportional damping coefficient at the rotational inverse dynamics
output shaft. The variable τu stands for an unknown load α̃M + υM TuM
disturbance torque. uM = . (6)
kM

693

Authorized licensed use limited to: Universitaetsbibl Rostock. Downloaded on August 11,2021 at 17:16:02 UTC from IEEE Xplore. Restrictions apply.
The stabilizing control law υM is chosen as three ⎡ ⎤ ⎡ ⎤
ẋ1 x2
ẋ = ⎣ ẋ2 ⎦ = ⎣ x3 ⎦ .
t
(11)
υM = α̃Md + kα0 (α̃Md − α̃M ) + kα1 (α̃Md − α̃M ) dτ, (7) ẋ3 υ
0
Now, the error dynamics can be stabilized by a Lyapunov-
where kα0 > 0 and kα1 > 0 are positive coefficients. The based control as proposed in [8]. In the given case of a ref-
desired trajectory α̃Md is designed in compliance with all erence tracking problem, a Lyapunov function is introduced
constraints and remains in the range [εM , 1]. This control as follows
requires the feedback of α̃M , which it is not directly measur- 2
able; therefore, an observer is used to provide an estimate α̃ˆ M s2 d
V (s) = , s= +λ e = ë + 2λ ė + λ 2 e. (12)
of α̃M , which is used for feedback control and, moreover, for 2 dt
a gain-scheduling in the decentralized control of the motor Here, λ > 0 represents a positive design parameter, and e =
angular velocity. ωM −ωMd is the tracking error based on the desired trajectory
ωMd . The time derivative of s becomes
B. Control of motor angular velocity
... ... ...
Many approaches have already been investigated and val- ṡ = e + 2λ ë + λ 2 ė = ω M − ω Md + 2λ ë + λ 2 ė. (13)
idated for the nonlinear control of angular velocity of the The stabilizing control law υ is now chosen as
motor, cf. [5] for an overview. In this paper, the system ... ...
is first feedback-linearized using estimator-based techniques υ = ẋ3 = ω M = ω Md − 2λ ë − λ 2 ė − ks, (14)
and, subsequently, a Lyapunov-based control is deployed for where k > 0 represents a positive constant. Rearranging (14)
the tracking control. The design is based on the dynamics of yields
the subsystem, where α̃M is considered as a gain-scheduling ... ...
ω M − ω Md + 2λ ë + λ 2 ė + ks = 0, (15)
parameter
⎡ ⎤ ⎡ ⎤ which is equivalent to the first-order dynamics ṡ + ks = 0
α̃˙ P − T1uP α̃P + TkuP
P
uP and guarantees an exponential convergence towards s = 0.
⎣ Δ ṗ ⎦ = ⎢ 2 2 qu ⎥
⎣ CH ṼP α̃P ωP − CH ṼM α̃M ωM − CH ⎦ . (8) The parameters λ > 0 and k > 0 can be employed to adjust
ω̇M − dJVV ωM + ṼJVM Δpα̃M − JτVu the tracking performance.

The output y = ωM represents a flat output of the subsystem,


which can be easily shown by repeated time differentiation. Z
Md 

The control input uP appears first in the third time derivative, ª¬ 0 O 2 2O º¼
Abc
F9

indicating a relative degree of three. With the introduction of


new state variables x1 = ω, x2 = ω̇, x3 = ω̈ and substitutions ª¬ O 2 2O 1º¼ k 

X

according to (8), the system can be stated in controller 

normal form
ªZMd º
«Z Md » >Z , Z M , ZM @
T

ẋ1 = x2 «¬ZMd »¼ M

ẋ2 = x3
2ṼM2 α̃M
2 2ṼM2 α̃M
2 dV Fig. 4. Stabilizing control law based on a Lyapunov function.
ẋ3 = − x1 − + x2 (9)
JV CH TuP JV CH JV TuP
In the sequel, several estimators are investigated that allow
dV 1 ṼM α̃M 2ṼM α̃MṼP ωP kP for addressing uncertainties in the feedback linearization
− + x3 − qu + uP .
JV TuP JV CH TuP JV CH TuP approach: a state and disturbance observer, an adaptive
Obviously, the relative degree equals the subsystem order parameter estimator and a neural network that is trained
and, hence, no internal dynamics exists. Provided that all online. All of these estimators are capable of rendering the
system parameters and the disturbance input qu are perfectly feedback linearization robustly stable. Thereby, a superior
known, the system can be easily feedback-linearized by com- behaviour is expected in comparison to classical schemes
pensating all nonlinearities. This is achieved by introducing that assume a perfect model knowledge.
a new input υ according to IV. E STIMATOR -BASED F EEDBACK L INEARIZATION
2ṼM2 α̃M
2 2ṼM2 α̃M
2dV The system is feedback-linearized by using an inverse
υ =− x1 − + x2
JV CH TuP JV CH JV TuP dynamics of the form
dV 1 ṼM α̃M 2ṼM α̃MṼP ωP kP uP = g · [υ − f (x, qu )] . (16)
− + x3 − qu + uP .
JV TuP JV CH TuP JV CH TuP
(10) The Lyapunov-based design of the stabilizing control law
Solving this definition for the physical input uP yields the (14) achieves a good tracking behaviour if the system non-
corresponding inverse dynamics. As a result, the feedback- linearities are compensated for appropriately. This requires
linearized system (9) becomes an integrator chain of length a reconstruction of the uncertain function f (x, qu ), which

694

Authorized licensed use limited to: Universitaetsbibl Rostock. Downloaded on August 11,2021 at 17:16:02 UTC from IEEE Xplore. Restrictions apply.
depends on the state vector x, the leakage volume flow qu Solving (22) for the unknown function Φ leads to
as disturbance and unavoidable model uncertainty. Similar
Φ (ym , τ, u) = τ̇ − Hẏm . (22)
ideas have been investigated in work of [9] on model-free
control. Consequently, the current value of f (x, qu ) must Substituting τ̇ and ẏm from (5) into (22) and considering the
be determined online according to the changes in the state integrator disturbance model τ̇d = 0 results in
variables, the external disturbances and the given model ⎡ ⎤
− TuM
1
α̃M + TkuM
M
uM
uncertainty. A nonlinear reduced-order observer is used to
⎢ − 1 α̃ + kP u ⎥
Φ (ym , τ, u) = ⎢ ⎣
TuP P TuP P ⎥

0
uM (23)

Hydrostatic Transmission
Abc
F9

 2 0 
q
ṼP α̃P ωP − C2H ṼM α̃M ωM − CHu
X 
g
uP −HT CH dV .
− JV ωM + ṼJVM Δpα̃M − JτVu

Dˆ M Asymptotic stability of the linearized error dynamics is
f (x, qu ) achieved by placing all the eigenvalues of the Jacobian in
>ZM , Z M , ZM @
T the left complex half-plane
x
∂ Φ (ym , τ, u) . 4
det sI − = ∏ (s + si ) . (24)
Fig. 5. Estimator-based feedback linearization for the trajectory tracking ∂τ i=1
of the motor angular velocity.
With the four design specifications s1/2/3/4 > 0 – corre-
estimate the unknown state variables and external disturbance sponding to eigenvalues with a negative real part –, the four
affecting the system. The estimated state variables – α̃M/P – unknown observer gains in the matrix H can be directly de-
are required in all three control structures under consideration termined from a comparison of the characteristic polynomial
because direct measurements are not available at the test rig. with the desired one according to (24).

A. Nonlinearity compensation using a disturbance observer B. Nonlinearity compensation by online parameter estima-
tion
In the first approach – a compensation scheme based
The inverse dynamics for the online parameter estimation
on a disturbance observer (DO) as proposed in [10] – the
is modified in the following form
measured angular velocity, their first two time derivatives
and the estimate q̂u for a unknown volume flow qu are used 3

in the feedback linearization uP = â0 · g · υ + ∑ âi · fi (x) + â4 . (25)


i=1
uP = g · [υ − f (x, q̂u )]. (17) The parameters in the inverse dynamics (25) – âi , i ∈
{0, 1, 2, 3, 4} – are considered to be unknown in the sequel
The output equation of the observer reads and are estimated online by methods proposed in [8]. The
  
x̂e H1 z1 nominal model with a vanishing disturbance qu = 0 corre-
= ym + , (18) sponds to the values âi = 1, i ∈ {0, 1, 2, 3} and â4 = 0. The
τ̂d H2 z2
appropriate modification of (10) and the solution for uP leads
where ym = [Δp, ωM ]T denotes the vector of measurable to
output variables. The vector x̂e contains the estimates for JV CH TuP ṼM α̃M
the normalized tilt angles xe = [α̃P , α̃M ]T , and τ̂d is the uP = â0 υ + â1 ωM
2ṼM α̃MṼP ωP kP ṼP ωP kP
estimate
 T of the
 disturbance vector τ̂d = [qu , τu ]T . The matrix
T TuPṼM α̃M CH dV
H = H1 , HT2 contains the observer gains and is chosen as +â2 + ω̇M (26)
ṼP ωP kP 2ṼM α̃MṼP ωP kP
T
h11 0 h31 0 CH TuP dV JV CH
H= . (19) +â3 + ω̈M + â4 ,
0 h22 0 h42 2ṼM α̃MṼP ωP kP 2ṼM α̃MṼP ωP kP
 T
The internal state vector z = zT1 , zT2 of the reduced-order with the state vector x = [x1 , x2 , x3 ]T = [ωM , ω̇M , ω̈M ]T . Ob-
observer is obtained from the solution of the following state viously, (26) can be written in the form of (25)
equations uP = â0 · g · υ + â1 · f1 (x) + â2 · f2 (x) + â3 · f3 (x) + â4 . (27)
Φ (ym , τ̂, u) .
ż =Φ (20)
 T T T  T T T Note that this representation is linear in the parameters to
Here, τ̂ = x̂e , τ̂d provides the estimates for τ = xe , τd be estimated. Following the suggestion in [8], the adaptation
and u = [uP , uM ]T represents the control input vector. The laws are chosen as follows
vector function Φ follows from the design condition that the â˙0 = −γ0 · sign(â0 ) · s · υ ,
steady-state observer error vanishes
â˙i = −γi · sign(â0 ) · s · fi (x), i = 1, 2, 3 , (28)
τ̃˙ = τ̇ − τ̂˙ = 0 = τ̇ − Hẏm −Φ
Φ (ym , τ, u) . (21) â˙4 = −γ4 · sign(â0 ) · s.

695

Authorized licensed use limited to: Universitaetsbibl Rostock. Downloaded on August 11,2021 at 17:16:02 UTC from IEEE Xplore. Restrictions apply.
Here, the adaptation rates γ0/1/2/3/4 should be chosen rea- compensation. All these structures use state estimates for
sonably small, see [8] for more details. The variable s the unmeasurable tilt angles, the same stabilizing control
corresponds to the argument of the Lyapunov function V (s) law – however with individual parametrizations. The time
in (12). derivatives of the motor angular velocity are obtained by
numerical differentiation from output measurements.
C. Nonlinearity compensation by a neural network In this paper, for the sake of simplicity, the feasible desired
Starting point the nonlinearity compensation is again trajectories for α̃Md and ωMd are generated in such a way
the system representation in controller normal form (9). as to avoid actuator saturation and to fulfil the requirements
Moreover, the desired trajectory yd = [ωMd , ω̇Md , ω̈Md ]T is regarding differentiability.
generated satisfying the condition yd   Q, where Q is
a known bound. Under the realistic assumption that the
disturbance volume flow qu is bounded as well, the nonlinear
function f (x, qu ) can be approximated by a neural network
in the form of a multilayer perceptron (MLP), cf. [11] and
[12]. This leads to the following inverse dynamics
uP = g · [υ − F (x)] . (29)
Fig. 6. Desired trajectory for the motor tilt angle α̃Md
A single output with one hidden layer neural network can
be stated in a general form as follows
     T 
F(x) = σ wT σ vT1 xa , ..., σ vTL xa . (30)

Here, σ represents an activation function, which is specified


below. Moreover, w is a weighting vector belonging to the
output layer, whereas V = [v1 , ..., vL ] is a weighting matrix of
the hidden
 layer. The augmented input vector is denoted by
T
xa = 1, xT with 1 corresponding to the bias of the neurons.
In this paper, a linear activation function is employed for the Fig. 7. Desired trajectory for the motor angular velocity ωMd [rad/s]
output layer, and a sigmoid function is used for hidden layer.
The neural network for nonlinearity approximation can now Fig. 8 depicts simulation results for the unmeasurable state
be described in more detail as follows variable α̃M , which needs to be estimated by the observer.
It can be seen that the estimate matches quite well the
L  T 
F (x) = ∑ w jσ v j xa , (31) simulated values. Consequently, it can be employed for
feedback control.
j=1

with σ (xi ) = (1 + e−xi )−1 . L is the total number of neurons,


w j the j −th element of w, and v j denotes the j −th column
of V that contains the weights of j − th neuron. The weights
of the neural network can be collected in an augmented
weighting matrix

V 0
Θ= . (32)
0 w Fig. 8. Comparison of simulated and estimated values for α̃Md .
Under the assumption that Θ  θm holds, where θm is a
known upper bound, the neural network can be tuned in such Fig. 9 shows the comparison of the tracking errors for the
a way that both the tracking performance and the estimator motor angular velocity for the three alternatives used for an
stability are guaranteed, see [11] for details. The tuning rules estimator-based nonlinearity compensation: neural network
as proposed by [11] are given by (NN), online parameter estimation (PE) and disturbance
      observer (DO). To render the simulation results realistic,
ẇ j = M σ vTj xa −σ  vTj xa v j T xa s−κ |s| Mw j , white noise is added to the measured angular velocity and
  (33)
v̇ j = Nsxa w j σ  vTj xa −κ |s| Nv j , the measured pressure in the simulation model, respectively.
The simulation shows that all three approaches for hydro-
where M, N and κ are positive constants.
static transmission yield similar results, the tracking errors
V. C OMPARISON OF THE THREE ESTIMATOR SCHEMES are small. A numerical comparison is stated in Table I.
A. Simulation results B. Experimental results
The implemented control structures represent a combi- The controllers are implemented on the hydrostatic trans-
nation of the stabilizing tracking control law (14) and mission test rig and evaluated in experiments. The corre-
one of three approaches to an estimator-based nonlinearity sponding results for the tracking errors are depicted in Fig. 10

696

Authorized licensed use limited to: Universitaetsbibl Rostock. Downloaded on August 11,2021 at 17:16:02 UTC from IEEE Xplore. Restrictions apply.
TABLE I TABLE II
C OMPARISON OF ERROR MEASURES ( SIMULATION RESULTS ). C OMPARISON OF ERROR MEASURES ( EXPERIMENTAL RESULTS ).

NN PE DO NN PE DO
Max. error in rad/s 0.2712 0.2874 0.2803 Max. error in rad/s 0.5469 0.7215 0.5148
RMS error in rad/s 0.06186 0.08527 0.07067 RMS error in rad/s 0.1492 0.2044 0.1621

transmission that is subject to external disturbances and


model uncertainty.
VI. C ONCLUSIONS AND O UTLOOK
This paper presents a decentralized control approach for a
hydrostatic transmission, where the motor tilt angle and the
Fig. 9. Comparison of the tracking errors eωMd [rad/s] for three alternative
motor angular velocity serve as controlled outputs. Here, the
approaches to an estimator-based nonlinearity compensation (simulation unmeasurable tilt angles of both pump and motor are recon-
results). structed with a disturbance observer. Whereas the flatness-
based control of the motor tilt angle is straightforward, the
focus is on the compensation of uncertain terms in the
and the numerical values can be found in Table II. As
feedback-linearizing control of the motor angular velocity. A
comparison of the three alternative compensation approaches
– an observer-based approach, an online parameter estima-
tion and a neural network – shows that all of them enable
a comparable and highly accurate tracking behaviour. The
smallest RMSE values in both simulations and experiments
were obtained with the neural network, followed by the state
and disturbance observer and, finally, the online parameter
Fig. 10. Tracking errors eωMd [rad/s] for the alternative approaches to an estimation.
estimator-based nonlinearity compensation (experimental results). In the near future, the parallel use with identified distur-
bance models as feedforward control action will be investi-
indicated by the experimental results, all three compensation gated. This measure relieves the estimators, which only have
approaches work pretty well, and they can keep the tracking to cope with the deviations from the feedforward term.
error small.
R EFERENCES
The approach with a disturbance observer requires a
complete model of the hydrostatic transmission, and the [1] H. Schulte, “Control-oriented modeling of hydrostatic transmission
using takagi-sugeno fuzzy systems,” in International Fuzzy Systems
control performance depends fully on the accuracy of the Conference, IEEE, 2007.
system model, see earlier research in [5] for an overview. [2] B. Dolan and H. Aschemann, “Control of a wind turbine with a
The approach based on an online parameter estimation hydrostatic transmission - an extended linearisation approach,” in 17th
International Conference on Methods and Models in Automation and
also relies on a system model. This model, however, is Robotics (MMAR) (2012), Miedzyzdroje, Poland, 27-30 Aug 2012, pp.
parametrized by a set of adaptable parameters that are 445–450.
optimized online. This adaptation mechanism supports the [3] A. Nawrocka and J. Kwasniewski, “Advanced algorithm for stabiliza-
tion rotational speed in hydrostatic transmission,” Intern. Journal of
controller at tracking the desired trajectory despite external Mechanics, no. 4, pp. 164–168, 2006.
disturbances and model uncertainty. This concept has been [4] S. Stoll, M. Kliffken, M. Behm, and X. Wang, “Regelungskonzepte für
validated successfully at the test rig and the obtained results hydrostatische Antriebe in mobilen Arbeitsmaschinen (in German),” at
– Automatisierungstechnik, p. 5 (2007) 2, 19-21 Aug. 2007.
confirm the functionality of the nonlinearity compensation. [5] H. Sun, Decentralized Nonlinear Control for a Hydrostatic Drive Train
This approach also works well with a less accurate model with Unknown Disturbances. PhD thesis, University of Rostock,
and can cope with large disturbances and model uncertainty Shaker, 2015.
[6] M. Fliess, J. Levine, P. Martin, and P. Rouchon, “Flatness and defect of
– at a higher level in comparison to the approach using the nonlinear systems: Introductory theory and examples,” Intern. Journal
disturbance observer. of Control, no. 6, pp. 1327–1361, 1995.
For the nonlinearity compensation based on a neural [7] H. Aschemann and H. Sun, “Decentralised flatness-based control
of a hydrostatic drive train subject to actuator uncertainty and dis-
network, the state-independent part of the model is replaced turbances,” in IEEE Intl. Conference on Methods and Models in
by a neural network, which is trained online to compensate Automation and Robotics MMAR, Miedzyzdroje, Poland, 2013.
both disturbances and uncertainty affecting in the system. [8] E. Slotine and W. Li, Applied nonlinear control. Prentice Hall, 1991.
[9] M. Fliess and C. Join, “Model-free control,” Int. Journal of Control,
This approach turned out to allow for an equivalent tracking pp. 2228–2252, 12 2013.
performance in comparison with the others variants. Con- [10] B. Friedland, Advanced Control System Design. Prentice Hall, 1996.
sidering the whole structure of this approach, it can be [11] F. L. Lewis, S. Jagannathan, and A. Yesildirek, Neural Network
Control of Robots and Nonlinear Systems. London: Taylor and
clearly seen that a complete model of the system is not Francis, 1999.
required. Moreover, this nonlinearity compensation showed a [12] O. Nelles, Nonlinear System Identification: From Classical Ap-
high robustness and proved to be applicable to a hydrostatic proaches to Neural Networks and Fuzzy Models. Springer, 2001.

697

Authorized licensed use limited to: Universitaetsbibl Rostock. Downloaded on August 11,2021 at 17:16:02 UTC from IEEE Xplore. Restrictions apply.

You might also like