358 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO.

1, JANUARY 2019

Discrete-Time Direct Model Reference Adaptive

Control Application in a High-Precision Inertially
Stabilized Platform
Ke Deng , Shuang Cong , Senior Member, IEEE, Dejie Kong, and Honghai Shen

Abstract—This paper studies the practical use of I. INTRODUCTION

discrete-time direct adaptive control in a high-precision
NERTIALLY stabilized platform (ISP) is the core device
inertially stabilized platform’s turbulence isolation system
for the purpose of enhancing isolation performance. Under
low-frequency and low-velocity environments, the velocity-
I used in navigation, guidance, and measurement applications.
It is used to isolate carrier turbulence and stabilize the optical
stabilized loop of the platform shows severe nonlinear char- equipment, so the equipment can keep its orientation or main-
acteristic; therefore, its isolation performance is limited. In
previous research, we constructed a nonlinear model of the
tain line-of-sight [1]. It has a wide range of use in both civil and
velocity-stabilized loop by using an improved Stribeck fric- military applications [2], [3] such as aircraft, ships, and electron
tion model, and also designed a feed-forward compensation telescopes. Furthermore, high-precision ISP is one of the most
strategy. Both have obtained outstanding performance dur- important components in the present day sophisticated military
ing practical experiments. However, errors still exist as dis- equipment. The stabilization accuracy of the optical axis is one
turbance of the unmodeled part and environmental change
causes the system’s parameter to vary. To solve this, a
of the key performance indexes of ISP, and it directly impacts
novel discrete-time direct model reference adaptive con- its tolerance toward carrier turbulence. This is currently reduced
trol based on nonlinear friction compensation is introduced to microradian magnitude (μrad). ISP is a complicated control
to the original proportional-integral control system. An im- system, and it is affected by various disturbances, for example,
proved projection algorithm and a recursive least-square al- nonlinear frictions, gyrodrift, and mechanical resonance. Under
gorithm with fading memory are respectively used to design
the adaptive law. By using a turbulence observer to provide
the assumption that the hardware of ISP is fixed, the key to fur-
a reference signal, both types of controls are applied to the ther improve accuracy lies within analyzing errors and how they
carrier turbulence isolation system. Results of the practical affect the system, and then designing a strategy to compensate
experiments prove that model reference adaptive control them. Accurate modeling of the platform is the most effective
can further enhance the system’s isolation ability. method to analyze the errors.
Index Terms—Adaptive control, inertially stabilized ISP requires high stabilization precision and quick real-time
platform (ISP), model reference adaptive control, nonlinear feedback [4], especially because the optical equipment it carries
friction, turbulence isolation. already faces the problem of time lapse during identification
and analysis of the target. Due to its high practical usage, a
linear controller with a simple design and structure, such as
proportional-integral-differential (PID) compensator, is widely
Manuscript received September 24, 2017; revised January 25, 2018
used on ISP. However, its ability to further increase stabilization
and March 7, 2018; accepted April 6, 2018. Date of publication April 27, precision faces two limitations. First, under low-frequency and
2018; date of current version August 31, 2018. This work was supported low-velocity environment, nonlinear friction causes the qual-
in part by the Projects of Ultra-Precision Control System Joint Laboratory
of USTC under Grant KD1012210167 and in part by the National Natural
ity of the control system to significantly decrease [5]. Sec-
Science Foundation of China under Grant 61573330. (Corresponding ond, when environment change occurs, it is unable to overcome
author: Shuang Cong.) the control error that is brought by parameter variations of the
K. Deng is with the State Key Laboratory of Air Traffic Manage-
ment System and Technology, Nanjing 210001, China (e-mail: dengke@
system.
mail.ustc.edu.cn). (e-mail:, [email protected]). For the first limitation, nonlinear modeling and friction com-
S. Cong is with the , Department of AutomationUniversity of Sci- pensation are effective methods to rectify it. Currently, friction
ence and Technology of China, Hefei 230027, China (e-mail: scong@
ustc.edu.cn). (e-mail:, [email protected]).
compensation is a necessary process in high-precision control
D. Kong is with the , Department of Optoelectronic Engineer- system [6], and it has been split into model-based and model-
ingChangchun University of Science and Technology, Changchun free. The key to model-based friction compensation is to select
130022, China (e-mail: [email protected]). (e-mail:, kongde-
[email protected]).
the proper friction model and to identify important parameters
H. Shen is with the , Key Laboratory of Aviation Optical Imaging [7]. Various types of friction models [8] are currently being stud-
and MeasurementChangchun Institute of Optics, Fine Mechanics and ied, and are split into static friction model and dynamic friction
Physics,, Chinese Academy of Sciences, Changchun 130022, China
(e-mail: [email protected]). (e-mail:, [email protected]).
model. The static friction models, i.e., Stribeck model [8], have
Digital Object Identifier 10.1109/TIE.2018.2831181 a simple and practical structure, but they face problems such

0278-0046 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
DENG et al.: DISCRETE-TIME DIRECT MODEL REFERENCE ADAPTIVE CONTROL APPLICATION IN A HIGH-PRECISION ISP 359

as a lack of precision and switching of model equations. While In this paper, we will develop a novel discrete-time direct
dynamic models, i.e., Lugre model [9], are more precise, their MRAC method to overcome the turbulence isolation problem
structure and parameter tuning are more complicated. Model- of ISP’s velocity-stabilized loop. By adopting an upgraded strat-
free friction compensation, which is based on disturbance ob- egy, this MRAC method is designed to overcome errors due to
server [10], combines friction and other disturbance into overall disturbance of the unmodeled part and environmental change
disturbance [11]. Through the aid of an observer to inhibit the based on the original PI control system, in addition to the pre-
overall disturbance, the effect of friction is then removed. vious work on nonlinear friction compensation. Unlike other
For the second limitation, adaptive control can be introduced MRAC methods, it constructs a schema through model match-
to reduce the impact of time-varying parameters. Model refer- ing, in which a class of parameter-optimization algorithms can
ence adaptive control (MRAC) is an important type of adaptive be used to obtain adaptive laws. Improved projection algorithm
control; it has been widely used and studied both theoretically and recursive least-square algorithm with fading memory are
and practically. MRAC has developed from being applied to respectively used to design the law. After stability analysis and
continuous-time systems to now discrete-time systems. From simulations, practical experiments have been conducted with the
the aspect of design, it has already developed from an optimal aid of the refine-designed turbulence observer, which is used to
local parameter method, i.e., MIT method, to two main types provide a reference signal. All this would assist the system to
of stability theorem: Lyapunov stability theorem and Popov hy- achieve a better isolation performance, which is the main contri-
perstability theorem. MRAC is equivalent to the nonlinear time- bution of this paper. In addition, this is the first time the discrete-
varying feedback system, and Lyapunov function is a powerful time direct MRAC method has been applied to high-precision
tool in solving its stability. The second theorem of Lyapunov is ISP with microradian magnitude.
more frequently used to design MRAC, and this method is fur- This paper is organized as follows. In Section II, we review
ther divided into direct and indirect types. In direct MRAC, the previous work on nonlinear modeling and friction compen-
it directly amends the parameter in the controller, while in sation. In Section III MRAC is designed, and then we conduct
indirect MRAC, it estimates the parameters of the model of stabilization analysis and simulation. In Section IV, we perform
controlled process, and then amends the controller’s parameter practical experiments on ISP using MRAC. Finally, a conclusion
through obtaining the correlation between the controller’s and is drawn in Section V.
the model’s parameters. Studies have shown that the order of
adaptive law in indirect MRAC’s is lower than direct MRAC;
II. NONLINEAR MODELING AND FEED-FORWARD
however, analyzing its stability is more difficult [12]. In recent
COMPENSATION
years, MRAC has been widely studied and developed, espe-
cially on robustness. L1 adaptive control (L1-AC) [13] adopts a Here, we describe our previous work on nonlinear friction
strictly stable filter to filter the input signals, which is used to ob- modeling and feed-forward compensation based on original PI
tain faster convergence. However, recent researches have shown control system; these provide the foundation of our MRAC.
that when compared to standard MRAC, the stability margin In our previous research, we have concluded that the nonlinear
of robustness has actually been reduced [14]. By incorporating friction is the main factor that influences isolation performance
a proportional-integral (PI) compensator in the feedback loop in low-frequency and low-velocity environment. To resolve it, a
of MRAC architecture [15], the controller achieves guaranteed nonlinear model of velocity-stabilized loop was constructed and
transient and steady-state performance with improved robust- the corresponding friction compensation strategy was designed
ness. An improved discrete-time MRAC is designed to cope and verified. The specific about this research is published in [23].
with automotive device’s nonlinear and discontinuous dynam- Two axes four-gimbal ISP consists of inner/outer azimuth
ics [16], and it extends the class of the minimal control synthesis gimbal and inner/outer elevation gimbal. Within the inner
algorithms for discrete-time systems. Adding a saturation func- gimbal lies the velocity-stabilized loop, which is the key com-
tion and a PI controller, a modified version of variable structure ponent toward isolating turbulence. It uses rate gyros as the
MRAC (VS-MRAC) schema is developed, and it is applied feedback sensors, and its performance has a direct effect on
to the fault-tolerant control [17]. A predictive model reference the stabilization precision of ISP. The performance indicator
adaptive system (MRAS) speed estimator based on the finite is based on its capability to isolate carrier turbulence. Assume
control set-model predictive control (FCS-MPC) principle is isolation φ is
proposed for sensorless induction motor [18]. With an external
filter, the MRAC method is proposed to not only stabilize the φ = |ω + ω 2 |m ax /|ω|m ax = |ω 1 |m ax /|ω|m ax (1)
error system but also guarantee its transient performance for
a class of nonlinear systems with uncertain parameters, distur- where ω 1 is the angular velocity vector of the platform related
bance, and unmodeled dynamics [19]. In addition, neural net- to the inertia space, ω 2 is the angular velocity vector of the
works have been introduced into adaptive control architecture to platform relative to the carrier, and ω is the turbulence angular
improve both transient and steady-state performance [20]–[22]. velocity vector of the carrier relative to inertia space, where
As shown, there are a lot of different design methods for MRAC; ω 1 = ω + ω 2 . Assume that the reference signal is zero, which
however, how to fuse them into practical application is a hurdle. means maintaining an orientation. Hence, the lower the value of
This hurdle is especially difficult in a high-precision control φ, the better the ISP is able to isolate turbulence. It is noteworthy
system. that these max values in (1) should be measured when the ISP
360 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 1, JANUARY 2019

is in stable operation and the sampling time span is at least one

period of turbulence signal.

A. Nonlinear Modeling
To have an insight into friction, we have conducted the non-
linear modeling of velocity-stabilized loop.
The velocity-stabilized loop of the inner azimuth gimbal is
used as an example. By using the balance equations of the Fig. 1. Carrier turbulence isolation system with nonlinear friction feed-
forward compensation.
driving torque of the dc torque motor and the armature voltage,
and by considering the effect of the current loop and neglecting
the armature inductance, the model of the velocity-stabilized Combining the nonlinear friction feed-forward compensation
loop is obtained and simplified. It can be divided into a first- strategy (NFFCS) into the actual carrier turbulence isolation
order linear system part and a nonlinear friction part Tf (k). system, we have produced control system (see Fig. 1). The
According to the classic Stribeck model [8], friction torque turntable stimulates the angular velocity ω during carrier turbu-
only correlates to angular velocity. However, actual friction lence. ISP uses reference signal uref = 0 to achieve line-of-sight
torque is also affected by the gimbal’s position. Distinguish- stabilization.
ing different movement directions and different half-planes, the During practical experiments, the azimuth control system re-
complete movement of the inner azimuth gimbal consists of four quires the inner and outer azimuth gimbals to work in coordi-
submovements. Based on above, an improved Stribeck nonlin- nation. The movement of the outer azimuth gimbal follows the
ear friction model with high precision is proposed. inner gimbal and provides expansion of rotation range and wind
Combining the nonlinear friction model with the linear resistance isolation. The controller of the inner gimbal consists
model, we have obtained the nonlinear model of the velocity- of PI controller, notch filter, and bound of control quantity. Ten
stabilized loop, which has 24 parameters that need to be groups of experiments with the same turbulence (position sig-
identified. Based on the designed identification signal, we nal of 3°, 1/6 Hz) were separately conducted on the original PI
thereby utilize the genetic algorithm (GA) within the optimiza- control system and control system with the nonlinear friction
tion toolbox in MATLAB. The fitness function of GA minimizes compensation (PI+NFFCS). Results prove that when compared
the root-mean-square error (RMSE) between the model output to the PI, the PI+NFFCS has shown enhanced performance.
and measured output. We obtain the results at 95% confidence Average isolation has decreased from 23.01% to 13.40%, so the
level. With average RMSE of 0.134 °/s, it is only 3.8% of the isolation performance has increased by 41.76%. It shows that
absolute maximum amplitude of the input signal. Last, model the compensation strategy significantly inhibits nonlinear fric-
verifications are conducted by another four sets of data and some tion, and increases the control system’s isolation ability against
simulation tests. The details of the parameters of model, the carrier turbulence. The details of the parameters of the con-
configurations of GA, and the process of verification are in [23]. troller, the configurations of NFFCS, and the results analysis of
experiments are in [23].
B. Friction Feed-Forward Compensation
III. DESIGN OF DISCRETE-TIME DIRECT MRAC
To overcome the negative effect of nonlinear friction, we have
By using NFFCS the isolation performance has been signif-
conducted the friction feed-forward compensation.
icantly improved, but there is still room for further improve-
Based on the nonlinear model above, the friction feed-forward
ments. As a low-pass filter has been used, the main factor that
compensation can be conducted by extracting the nonlinear fric-
now affects the performance is parameter change caused by the
tion part. However, adjustment needs to be made when designing
unmodeled part and environmental change. Additionally, ac-
the model-based compensation strategy. First, to facilitate the
cording to the isolation system in Fig. 1, it is equivalent to the
addition to the control quantity, the compensation quantity of
velocity ω2 to track the reverse value of the carrier turbulence ω,
friction is converted to control voltage Tf (k)/Ke , where Ke is
ω2 → −ω when uref = 0. Hence, we can design discrete-time
the equivalent model of the amplifier. Second, because the non-
direct MRAC to further increase its isolation performance and
linear friction model has several equations, the compensation
ensure that it has stability.
quantity would change between them during the actual usage,
causing the system to tremble. To prevent the tremble causing
A. Model Matching
mechanical resonance, the compensation quantity is smoothly
filtered. Last, due to the unmodeled part and other nonlinear Transforming the turbulence isolating problem to the turbu-
disturbances, differences exist between friction model and real lence tracking problem, during the design of the MRAC, we
friction. Under compensated friction would lead to steady-state can then set the reverse value of the carrier turbulence −ω as
error, while over compensation would lead to limited cycle os- the reference signal. After friction compensation, the nonlinear-
cillation [24]. To prevent oscillation causing the system to be un- ity of the original closed-loop control system has nearly been
stable, an attenuation factor γ is introduced to make sure that the suppressed. We assume that the closed-loop system approxi-
entire system is maintained at slightly under-compensation state. mates to a linear model B(z)/A(z). The reference model is set
DENG et al.: DISCRETE-TIME DIRECT MODEL REFERENCE ADAPTIVE CONTROL APPLICATION IN A HIGH-PRECISION ISP 361

Fig. 2. Discrete-time direct MRAC in turbulence tracking system.

as Br (z)/Ar (z). Select feed-forward regulator plus feed-back

controller as model matching, and set them as T (z)/R(z) and
S(z)/R(z)¸respectively. We then obtained the MRAC structure
(see Fig. 2).
Set the degrees of A(z), B(z), Ar (z), Br (z), S(z), R(z),
and T (z) as n, m, nr , mr , nS , nR , and nT , respectively. As
for the controlled process, assume n > m, A(z), and B(z) are
coprime, B(z) is stable, and its highest order coefficient b0 is
known. As for the reference model, assume Ar (z) is monic and
stable and the time-delay of the reference model and controlled
process is the same, which means d = n − m = nr − mr .
In the tracking system in Fig. 2, we hope the system out-
put ω2 (k) can be equal to the reference input −ω(k) at all the
time; therefore, set reference model as 1/z n k , where nk is the Fig. 3. Discrete-time direct MRAC based on the PA/RLS in turbulence
computer time-delay and is set as one. According to the fitted tracking system.
curves under some linear models’ parameter identifications, the
PI+NFFCS+Plant closed-loop control system shows high lin-
earity. Based on several experimental tests, we assume that a Substituting (2) and (5) into (4), we can obtain S(z) = s0
first-order linear model can approximate it. Then, we have = −a1 and R(z) = r0 = b0 . Therefore, the feed-forward regu-
 lator and feed-back controller are
B(z)/A(z) = b0 /(z + a1 )
. (2) 
Br (z)/Ar (z) = 1/z T (z)/R(z) = 1/r0 = 1/b0
. (6)
If the output ω2 (k) is equal to reference model output ωr (k), S(z)/R(z) = s0 /r0 = −a1 /b0
then the transfer function should satisfy
From Fig. 2 and (6), we can obtain the control law as
B(z)T (z)/(A(z)R(z) + B(z)S(z)) = Br (z)/Ar (z) (3)
where if A(z), B(z), Ar (z), and Br (z) are known, then the u(k) = −(1/r0 ) [s0 (k)ω2 (k) − (−ω(k))] (7)
problem turns into solving S(z), R(z), and T (z).
To make the derivation easier, set T (z) = Br (z) = 1, then where r0 = b0 , and have previously assumed that b0 is known.
S(z) and R(z) still need to satisfy Therefore, when system parameters are changed, we can adjust
the parameter s0 to update the control law. Then, by using (6) to
A(z)R(z) + B(z)S(z) = Ar (z)B(z). (4) adjust the controller, we achieve direct adaptive control. Projec-
tion algorithm (PA) and recursive least-square (RLS) algorithm
Set S(z) and R(z) as
have been used to optimally estimate the parameter s0 . The de-

S(z) = s0 z n S + s1 z n S −1 + · · · + sn S tailed control structure is shown in Fig. 3, where ŝ0 (k) is the
. (5) estimated value of s0 .
R(z) = r0 z n R + r1 z n R −1 + · · · + rn R
362 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 1, JANUARY 2019

B. Adaptive Law Based on the Improved PA

PA is a type of stochastic gradient algorithm, and it is used to
solve the minimum value problem when there are constraints.
According to Fig. 3, we set the cost function as
1
JPA (ŝ0 (k), ŝ0 (k − 1), k) = |ŝ0 (k) − ŝ0 (k − 1)|2
2
+ λ [ω2 (k) − b0 u(k − d) − ω2 (k − d)ŝ0 (k)] (8)
where λ is the Lagrange multiplier and d is the time delay, which
is d = n − m = 1. The cost function is split into two halves: the
first half is used to obtain the minimum error of the parameter,
while the latter half is constrained.
According to the cost function, we can deduce the ŝ0 (k)
recurrence equation as
ŝ0 (k) = ŝ0 (k − 1) + (1/(ω2 (k − 1))
Fig. 4. Tracking error ω 2 (k) − (−ω(k)) curve during the beginning
× [ω2 (k) − b0 u(k − 1) − ω2 (k − 1)ŝ0 (k − 1)] . (9) 60 ms in simulation. (a) MRAC based on improved PA. (b) MRAC based
on fading memory RLS.
There are some shortcomings when (9) is used in ISP. Its
incremental factor 1/ω2 (k − 1) will lead to the control law
to cause strong tremble when velocity changes directions. To where selecting a large positive value of P (0) can eliminate
improve the algorithm, we add the iterative formula ρ(k) the influence caused by initial value. The smaller the λRLS , the
ρ(k) = λPA ρ(k − 1) + ω2 (k − 1)2 (10) stronger the ability to track changes in the parameter; however,
if it is too small, the parameter will present large fluctuation.
where λPA is the forgetting factor, and λPA ∈ [0.9, 0.999]. Set
the initial ρ(0) = 1. Amend the incremental factor in (11) into
ω2 (k − 1)/ρ(k), we obtained the following formula: D. Stability Analysis and Simulations

ŝ0 (k) = ŝ0 (k − 1) + (ω2 (k − 1)/ρ(k)) The system shown in Fig. 3 is a compound control system.
It consists of the original PI closed-loop control system, the
× [ω2 (k) − b0 u(k − 1) − ω2 (k − 1)ŝ0 (k − 1)] (11) NFFCS, and the MRAC. Its stability is the primary concern, but
where ρ(k) under the condition that ρ(0) = 1 ensures that the we also expect the illustration to be simple.
increment will not abruptly change. At the same time, the for- The MRAC’s controlled process is the PI+NFFCS+Plant.
getting factor λPA gives the algorithm the ability to track the After debugging parameters, the original PI closed-loop control
time-varying parameter. system can be stable during runtime. Adding NFFCS with a
slight adjustment will not affect its stability, as the feed-forward
C. Adaptive Law Based on the Fading Memory RLS compensation does not change the solutions of the characteristic
equation. Hence, the controlled process is stable.
The RLS algorithm minimizes sum of squared errors and is The method behind how to prove the stability of MRAC
widely used in engineering. Applying it to Fig. 3, we set the is to build error dynamic equations for model matching and
cost function with exponential window parameter estimation. Then, designing a Lyapunov function that

k combines the two errors and proving that the difference equation
−i
JRLS (ŝ0 (k), k) = λkRLS (ω2 (i) − b0 u(i − d) is below zero. The proof process is complex; however, it is
i=d supported by the related theory [25]. We will only provide the
− ω2 (i − d)ŝ0 (k))2 (12) conclusion here: In both types of discrete-time direct MRAC,
when the reference input −ω(k) is bounded, the control system
where k ≥ d and d = n − m = 1. λRLS is the forgetting factor, is Lyapunov stable.
and λRLS ∈ [0.9, 0.999]. To test the performance of the both types of MRAC under
The introduction of the forgetting factor eliminates the data nonlinear disturbance, we have designed an exaggerated simu-
saturation problem and allows the algorithm to have the ability lation experiment in which the quantity of friction compensa-
to track the time-varying parameter. According to (12), we can tion is reduced by half. The simulation system is set up based on
derive the recursive equation of ŝ0 (k) Fig. 3, and the parameters used are given in Sections II.B, III.B,
⎧ (k −1)ω 2 (k −1) and III.C. During simulation, input −ω(k) is given by velocity
⎪
⎪ ŝ0 (k) = ŝ0 (k − 1) + λ P+P
⎨ RLS (k −1)ω 2 (k −1) 2 with π ◦ /s , 1/6 Hz. To obtain the optimal performance under
× [ω2 (k) − b0 u(k − 1) − ω2 (k − 1)ŝ0 (k − 1)] each MRAC, we have used trial and error and found the proper
⎪
⎪

⎩ P (k) = 1 2
P (k − 1) − P (k −1) ω 2 (k −1)
2
parameters as λPA = 0.99 and λRLS = 0.95. Fig. 4 shows the
λR L S λR L S +P (k −1)ω 2 (k −1) 2 tracking error curve ω2 (k) − (−ω(k)) during the first 60 ms.
(13) We can observe the adaptive processes of the two MRACs. The
DENG et al.: DISCRETE-TIME DIRECT MODEL REFERENCE ADAPTIVE CONTROL APPLICATION IN A HIGH-PRECISION ISP 363

Fig. 5. Carrier’s turbulence isolation system with discrete time direct

MRAC.

Fig. 6. 3-D structure graphs of the ISP used in experiments.

convergence rate of the MRAC based on fading memory RLS
is faster than that based on enhanced PA, but the initial error is
larger. isolation system, the carriers’ turbulence ω(k) is provided by
Furthermore, we have observed the curve under stable state the turntable from the outside. Therefore, we should construct
and found that fluctuation occurs when the velocity of the gimbal a turbulence observer.
is close to zero. Through calculation of one period with 6000 ms,
the isolation of the MRAC based on PA is 1.36%, and that of the
MRAC based on RLS is 1.53%. The results prove the feasibility B. Design of the Carrier’s Turbulence Observer
of the both MRAC.
According to the ISP’s active isolation principle and (1), when
only considering the azimuth turbulence, it is known that ω =
IV. EXPERIMENTS ω1 − ω2 . Note that ω2 is the inner azimuth gimbal’s rotation
Even after applying the MRAC designed to the ISP, we would rate relative to the carrier. In reality, we can only obtain ω̂2
still need to solve three key problems. First, we need to covert by the resolver, which is the inner azimuth gimbal’s rotation
the turbulence tracking system in Fig. 3 into the original iso- rate relative to the outer gimbal. Obviously, most of the time,
lation system. Second, the carrier turbulence, which is pro- ω̂2 = ω2 . To obtain the real-time turbulence signal for MRAC,
vided by turntable with predefined amplitude and frequency, we approach this by looking at the coupled kinematic equations.
has no sensor to provide the real-time value of the turbulence The ISP’s three-dimensional (3-D) structure graphs are pre-
for MRAC’s reference signal. Hence, we need to design a real- sented in Fig. 6. From inside out, it consists of the inner azimuth
time turbulence observer to overcome it. Last, to facilitate the gimbal IAG (loading optical devices), the inner elevation gim-
setup of the initial parameters of the MRAC, we need to per- bal IEG, the outer elevation gimbal OEG, and the outer azimuth
form identification on the general controlled process, which is gimbal OAG.
PI+NFFCS+Plant. Based on all above-mentioned problems, According to this, we place the ISP’s motion decomposition
we then conduct experiments to examine the MRAC. on the five coordinates within the 3-D space. From outward
inward, they are carrier coordinate OXB YB ZB , OAG’s coor-
A. Convert to Turbulence Isolation System With MRAC dinate OXA YA ZA , OEG’s coordinate OXE YE ZE , IEG’s co-
ordinate OXe Ye Ze , and IAG’s coordinate OXa Ya Za . Within a
When implementing the MRAC from turbulence tracking sys- small period Δt, assume that the OAG rotates θA around axis
tem to isolation system in practical work, we need to adjust the OZB , the OEG rotates θE around axis OYA , the IEG rotates θe
inner and outer azimuth gimbals’ joint control system. For the around axis OYE and the IAG rotates θa around axis OZe . The
convenience of this section, we will temporary ignore the outer relationships between them are OZB = OZA , OZe = OZa ,
gimbal control system’s part, and single out the inner gimbal’s and OYA = OYE = OYe . Hence, we obtained Fig. 7.
MRAC system, as shown in Fig. 5. Assume the velocity vectors relative to initial space of the
As shown in the figure, the input signal uref = 0. The refer- carrier’s turbulence, OAG, OEG, IEG, and IAG, are ω B , ωA ,
ence signal of MRAC is −ωobs (k), which is the reverse of the ωE , ωe , and ωa , respectively. TBA , TAE , TEe , and Tea are defined
turbulence’s observation value. PI+NFFCS+Plant is the gen- as transformation matrixes of the five coordinates from outside
eral controlled process. According to Fig. 3, the feed-forward to inside. Hence, we obtained the following kinetic equations:
regulator is G(k) = 1/b0 , and it is fixed. The feed-back con-
troller is F (k) = ŝ0 (k)/b0 , and it is adjustable. The reference 
model is 1/z. The adaptive mechanism varies according to the ωA = TBA ωB + [ 0 0 θ̇A ]T , ωE = TAE ωA + [ 0 θ̇E 0 ]T
estimation method of PA or RLS. The output ω1 (k) is the ve- ωe = TEe ωE + [ 0 θ̇e 0 ]T , ωa = Tea ωe + [ 0 0 θ̇a ]T
locity of the gimbal relative to the inertial space. Within the (14)
364 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 1, JANUARY 2019

observer is designed as the following equation:

ωobs = ω = (ω1 − ω̂2 )/(cos θe cos θE − sin θe sin θE ) − ω̂2A
(18)
where ω1 and ω̂2 are provided by the flexible rate gyro and the
resolver fixed on the IAG, respectively. ω̂2A = θ̇A is provided
by resolver fixed on the OAG. θe and θE are provided by re-
solvers fixed on the IEG and OEG, respectively. Additionally,
some filters are designed for these measured values. During ex-
periments, we compared ωobs with the set turbulence signal.
The observer is verified, but we will not get into details here.

C. Identification of the General Controlled Process

To obtain b0 and initial ŝ0 (k) for the MRAC, the model of
general controlled process needs to be identified.
From Section II and [23], we know that PI+NFFCS+Plant
shows a high linearity. To approach the general controlled pro-
cess of MRAC, we use the following first-order linear model:
ω2 (k) = âω2 (k − 1) + b̂u(k − 1) (19)
Fig. 7. Decomposition of coupled motion of the multi-gimbal ISP. where â and b̂ are parameters to be identified.
Since the frequency of carrier turbulence in experiments is
where the matrixes TBA , TAE , TEe , and Tea are 1/6 Hz, the identification signal is designed as
⎡ ⎤
cos θA sin θA 0 u(k) = (1/2)π cos(2π · 1/6k) + (1/3)π cos(2π · 1/3k)
TBA = ⎣ − sin θA cos θA 0 ⎦ , + (1/6)π cos(2π · 1/2k) (20)
0 0 1
⎡ ⎤ where the unit of the input signal is °/s.
cos θE 0 − sin θE For higher accuracy identification, we then differentiate the
TAE = ⎣ 0 1 0 ⎦
parameters â and b̂ in (19) during ISP gimbal’s positive and neg-
sin θE 0 cos θE ative movements. Take the inner azimuth gimbal for example,
⎡ ⎤ ⎡ ⎤
cos θe 0 − sin θe cos θa sin θa 0 set the X-axis as the optical axis of optic-electronic equipment
TEe = ⎣ 0 1 0 ⎦ , Tea = ⎣− sin θa cos θa 0⎦. when the gimbal is at the center 0° position; set the Y-axis as the
sin θe 0 cos θe 0 0 1 range of the rotation angle, and it is between ±4 ◦ . The X-axis
(15) is perpendicular to the Y-axis. Thus, Y ≤ 0 indicates that the
position is on the left half plane, and Y > 0 indicates that the
Combining (14) and (15), the velocity vector ωa is position is on the right half plane. Suppose the speed ω2 is the
 T positive direction of movement when the gimbal moves from
ωa = Tea TEe TAE TBA ωB + Tea TEe TAE 0 0 θ̇A the left half plane to the right half plane, that is, the negative
 T  T direction is under the opposite condition. Hence, the complete
+ Tea TEe 0 θ̇E 0 + Tea 0 θ̇e 0 movement of the inner azimuth gimbal is consisted of two sub-
 T movements. We use “+” and “−” to label the respective positive
+ 0 0 θ̇a . (16) and negative directions of the gimbal’s motions, respectively.
When only considering the azimuth motion relative to iner- In practice, we place the ISP on a horizontal desktop. Then,
tial space, the corresponding component of vector ωa is needed. the identification signal (20) is input into the PI+NFFCS control
According to Fig. 5, ωa becomes ω1 , θ̇a is the rotatory velocity system of the inner azimuth gimbal. The I/O data of the system
ω̂2 (k), the turbulence’s azimuth component is ωB z = ω. There- is obtained for parameter identification use.
fore, ω1 is Using the method in Section II, the GA toolbox of MAT-
LAB is also employed for parameter identification, and the re-
ω1 = (ωB x cos θA + ωB y sin θA ) sults are shown in Table I. The RMSE of the model is 0.1532
°/s, which only takes up 4.88% of the input signal’s absolute
× (sin θe cos θE + cos θe sin θE ) amplitude π °/s. The results provide useful information for
+ (ω + θ̇A )(cos θe cos θE − sin θe sin θE ) + ω̂2 . (17) the initial tuning of the MRAC. According to (5) and (6), we
can set b0 = (b̂+ b̂− )1/2 = 0.9827, and initial value of ŝ0 (k)
During experiments, the 3-D turntable is set to only provide as ŝ0 (0) = (â+ â− )1/2 = 7.07E − 7. This will help the MRAC
azimuth turbulence; hence, ωB x = ωB y = 0. Simplifying (17), method to fast implement line-of-sight stabilization, avoiding
we can obtain the azimuth turbulence ω. Hence, the turbulence large initial oscillation.
DENG et al.: DISCRETE-TIME DIRECT MODEL REFERENCE ADAPTIVE CONTROL APPLICATION IN A HIGH-PRECISION ISP 365

TABLE I
RESULTS OF THE PARAMETERS IDENTIFICATION FOR GENERAL
CONTROLLED PROCESS

Fig. 9. Results of the turbulence isolation system with PA-MRAC.

(a) ω 1 (k) curve. (b) ŝ0 (k) curve.

TABLE III
ISOLATIONS OF THE TURBULENCE ISOLATION SYSTEM WITH RLS-MRAC

isolation performance and the corresponding ŝ0 (k) are shown

Fig. 8. Two axes four-gimbal ISP and 3-D flight simulation turntable in in Fig. 9. Through calculation, the isolation is 8.10%.
actual experiments.

E. Experiments of the Discrete-Time Direct MRAC

TABLE II Based on Fading Memory RLS
ISOLATIONS OF THE TURBULENCE ISOLATION SYSTEM WITH PA-MRAC
The system’s control law is also given as (21) and the recur-
sive estimation equation of ŝ0 (k) is given by (13). Set λRLS =
0.95, P (0) = 1000, b0 = 0.9827, and ŝ0 (0) = 7.07 × 10−7 .
Ten groups of experiments on such control system (labeled
as RLS-MRAC) are conducted and the results are shown in
Table III. The average isolation is 9.68%. The isolation perfor-
mance and corresponding ŝ0 (k) in one experiment are shown in
Fig. 10. According to calculation, the isolation is 8.56%.
D. Experiments of the Discrete-Time Direct MRAC
Based on Improved PA F. Experiment Conclusion
According to Fig. 5, (6), and (7), we obtain the control law as Referring to Section II.B and Tables II and III, the average
isolations of PI, PI+NFFCS, PA-MRAC, and RLS-MRAC in
u(k) = −(1/b0 ) [ŝ0 (k)ω2 (k) − (uref − ωobs (k))] (21)
experiments are 23.01%, 13.40%, 9.11%, and 9.68%, respec-
where the iterative equations for ŝ0 (k) are (10) and (11). Refer- tively. Compared to the PI+NFFCS, the isolation performances
ring to the simulation results, we set λPA = 0.99 and ρ(0) = 1. of the PA-MRAC and RLS-MRAC have increased by 32.01%
Set b0 = 0.9827 and ŝ0 (0) = 7.07 × 10−7 . and 27.76 %, respectively. It proves that the MRAC designed can
In experiments, the ISP is fixed on the 3-D flight simulation further enhance the high-precision ISP’s isolation performance.
turntable, as shown in Fig. 8. The input azimuth turbulence is Moreover, for the PA-MRAC, the maximum of ŝ0 (k) in
provided by the turntable with sinusoidal position signal, which Fig. 9(b) is 0.048, the minimum is −0.099, and the range is
can then be converted to the velocity ω(k) with π °/s and 1/6 Hz. 0.147. For the RLS-MRAC, those in Fig. 10(b) are 0.0030,
The output is ω1 (k) with unit °/s. −0.0061, and 0.0091, respectively, which means that the adjust-
Ten groups of experiments are done to the combined con- ment range of the PA-MRAC is larger. By normalizing ŝ0 (k) of
trol system (labeled as PA-MRAC) and the results are given in the two MRACs with the same mean of 1, the standard devia-
Table II. The average isolation is 9.11%. In one experiment, the tions of the PA-MRAC and the RLS-MRAC are 19.33 and 0.92,
366 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 66, NO. 1, JANUARY 2019

[7] Z. Jamaludin, H. V. Brussel, and J. Swevers, “Friction compensation of

an XY feed table using friction-model-based feedforward and an inverse-
model-based disturbance observer,” IEEE Trans. Ind. Electron., vol. 56,
no. 10, pp. 3848–3853, Oct. 2009.
[8] B. Armstrong-Helouvry, P. E. Dupont, and C. C. De Wit, “A survey
of models, analysis tools and compensation methods for the control of
machines with friction,” Autom., vol. 30, no. 7, pp. 1083–1138, Jul. 1997.
[9] J. Y. Yao, W. X. Deng, and Z. X. Jiao, “Adaptive control of hydraulic actua-
tors with LuGre model-based friction compensation,” IEEE Ind. Electron.
Soc., vol. 62, no. 10, pp. 6469–6477, Oct. 2015.
[10] W. S. Huang, C. W. Liu, P. L. Hsu, and S. S. Yeh, “Precision control
and compensation of servomotors and machine tools via the disturbance
observer,” IEEE Trans. Ind. Electron., vol. 57, no. 1, pp. 420–429, Jan.
2010.
[11] J. Q. Han, “From PID to active disturbance rejection control,” IEEE Trans.
Ind. Electron., vol. 56, no. 3, pp. 900–906, Mar. 2009.
[12] X. J. Xie and J. L. Li, “A systematic analysis approach to discrete-time
indirect model reference adaptive control,” Acta Autom. Sin., vol. 33,
no. 11, pp. 1170–1175, Nov. 2007.
Fig. 10. Results of the turbulence isolation system with RLS-MRAC. [13] C. Cao and N. Hovakimyan, “Stability margins of L1 adaptive control
(a) ω 1 (k) curve. (b) ŝ0 (k) curve. architecture,” IEEE Trans. Autom. Control, vol. 55, no. 2, pp. 480–487,
Feb. 2010.
[14] P. A. Ioannou and A. M. Annaswamy, “L1 -adaptive control: Stability,
robustness, and interpretations,” IEEE Trans. Autom. Control, vol. 59,
no. 11, pp. 3075–3080, Nov. 2014.
respectively. It shows that variability of ŝ0 (k) in PA-MRAC [15] R. Dey, S. K. Jain, and P. K. Padhy, “Robust closed loop reference MRAC
is stronger, and it has better ability to track the time-varying with PI compensator,” IET Control Theory Appl., vol. 10, no. 18, pp. 2378–
parameters. Hence, the PA-MRAC has better effects than the 2386, Dec. 2016.
[16] U. Montanaro, A. di Gaeta, and V. Giglio, “Robust discrete-time
RLS-MRAC. MRAC with minimal controller synthesis of an electronic throttle body,”
IEEE/ASME Trans. Mechatron., vol. 19, no. 2, pp. 524–537, Apr. 2014.
[17] L. E. G. Castanon, A. V. Martinez, and R. C. Reynoso, “MRAC-based
V. CONCLUSION fault tolerant control of a SISO real process application,” IEEE Latin
Amer. Trans., vol. 13, no. 8, pp. 2545–2550, Aug. 2015.
Based on the previous work on nonlinear friction modeling [18] Y. B. Zbede, S. M. Gadoue, and D. J. Atkinson, “Model predictive MRAS
and compensation, this paper proposed a type of discrete-time estimator for sensorless induction motor drives,” IEEE Trans. Ind. Elec-
direct MRAC to solve the problem of time-varying parameters. tron., vol. 63, no. 6, pp. 3511–3521, Oct. 2016.
[19] W. C. Sun, Y. F. Zhang, Y. P. Huang, H. J. Gao, and O. Kaynak, “Transient-
It has employed enhanced PA and the RLS with fading memory performance-guaranteed robust adaptive control and its application to pre-
to respectively design the adaptive law. With the aid of the well- cision motion control systems,” IEEE Trans. Ind. Electron., vol. 63, no. 10,
designed turbulence observer, the MRACs have been applied pp. 6510–6518, Oct. 2016.
[20] Z. Wang, R. Q. Lu, F. R. Gao, and D. R. Liu, “An indirect data-driven
into the carrier turbulence isolation system, and the experiment method for trajectory tracking control of a class of nonlinear discrete-time
results show they have further improved isolation performance. systems,” IEEE Trans. Ind. Electron., vol. 64, no. 5, pp. 4121–4129, Oct.
Due to the original controller’s upgraded strategy, the proposed 2017.
[21] M. B. Rădac, R. E. Precup, E. M. Petriu, and S. Preitl, “Iterative data-
MRAC based on NFFCS has strong reliability and a high prac- driven tuning of controllers for nonlinear systems with constraints,” IEEE
tical value. Trans. Ind. Electron., vol. 61, no. 11, pp. 6360–6368, Nov. 2014.
[22] Y. Yang, S. N. Balakrishnan, L. Tang, and R. G. Landers, “Electrohy-
draulic control using neural MRAC based on a modified state observer,”
ACKNOWLEDGMENT IEEE/ASME Trans. Mechatronics, vol. 18, no. 3, pp. 867–877, Jun. 2013.
[23] S. Cong, K. Deng, D. J. Kong, and H. H. Sheng, “Isolation control for
K. Deng thanks all the researchers at the Key Laboratory of inertially stabilized platform based on nonlinear friction compensation,”
Aviation Optical Imaging and Measurement, Changchun, China, Nonlinear Dyn., vol. 84, no. 3, pp. 1123–1133, May 2016.
[24] D. Putra, H. Nijmeijer, and N. van de Wouw, “Analysis of undercompen-
for their help during the experiments. sation and overcompensation of friction in 1DOF mechanical systems,”
Automatica, vol. 43, no. 8, pp. 1387–1394, Aug. 2007.
[25] S. Akhtar and D. S. Bernstein, “Lyapunov-stable discrete-time model
REFERENCES reference adaptive control,” Int. J. Adapt. Control Signal Process., vol. 19,
[1] J. M. Hilkert, “Inertially stabilized platform technology: Concepts and no. 10, pp. 745–767, Dec. 2005.
principles,” IEEE Control Syst., vol. 28, no. 1, pp. 26–46, Feb. 2008.
[2] Z. Hurak and M. Rezac, “Image-based pointing and tracking for inertially
stabilized airborne camera platform,” IEEE Trans. Control Syst. Technol., Ke Deng was born in Meishan, China, in 1985.
vol. 20, no. 5, pp. 1146–1159, Sep. 2012. He received the B.E degree in automation and
[3] M. K. Masten, “Inertially stabilized platforms for optical imaging sys- the Ph.D. degree in systems engineering from
tems,” IEEE Control Syst., vol. 28, no. 1, pp. 47–64, Feb. 2008. the University of Science and Technology of
[4] J. Debruin, “Control systems for mobile Satcom antennas,” IEEE Control China, Hefei, China, in 2009 and 2016, respec-
Syst., vol. 28, no. 1, pp. 86–101, Feb. 2008. tively.
[5] M. W. Sun, Z. H. Wang, Y. K. Wang, and Z. Q. Chen, “On low-velocity He is currently an Engineer with the State Key
compensation of brushless DC servo in the absence of friction model,” Laboratory of Air Traffic Management System
IEEE Trans. Ind. Electron., vol. 60, no. 9, pp. 3897–3905, Sep. 2013. and Technology, Nanjing, China. His research
[6] J. Amin, B. Friedland, and A. Harnoy, “Implementation of a friction interests include nonlinear modeling analy-
estimation and compensation technique,” IEEE Control Syst., vol. 17, sis, advanced control strategies, and systems
no. 4, pp. 71–76, Aug. 1997. engineering.
DENG et al.: DISCRETE-TIME DIRECT MODEL REFERENCE ADAPTIVE CONTROL APPLICATION IN A HIGH-PRECISION ISP 367

Shuang Cong (M’07–SM’12) received the B.E. Honghai Shen received the Master’s degree in
degree in automation from Beijing University of mechanical and electrical engineering from the
Aeronautics and Astronautics, Beijing, China, in Chinese Academy of Sciences, Beijing, China,
1982, and the Ph.D. degree in system engineer- in 2001.
ing from the University of Rome La Sapienza, In 2001, he joined the Changchun Insti-
Rome, Italy, in 1995. tute of Optics, Mechanics and Physics, Chinese
She is currently a Professor with the Depart- Academy of Science, Changchun, Jilin, China,
ment of Automation, University of Science and where he is currently the Director of the Key
Technology of China, Hefei, China. Her research Laboratory of Aviation Optical Imaging and Mea-
interests include advanced control strategies for surement, Changchun, China. His research in-
motion control, fuzzy logic control, neural net- terests include control and optimization algo-
works design and applications, robotic coordination control, and quan- rithms applied to motor drives systems.
tum systems control.

Dejie Kong received the Master’s degree in

power electronics and electric drives from Shan-
dong University, Jinan, China, and the D.E. de-
gree in optical engineering from the University of
Chinese Academy of Sciences, Beijing, China,
in 2002 and 2014, respectively.
From 2005 to 2009, he was an Assistant Engi-
neer with the Changchun Institute of Optics Fine
Mechanics and Physics. From 2009 to 2014, he
was an Assistant Professor with the Changchun
Institute of Optics Fine Mechanics and Physics.
Since 2017, he has been an Associate Professor with the Changchun
University of Science and Technology , Changchun, China, where he is
mainly engaged in inertial platform stabilization platform control, as well
as motor control and drive research.