Journal of Systems Engineering and Electronics

Vol. 29, No. 3, June 2018, pp.580 – 586

Backstepping sliding mode control for uncertain strict-

feedback nonlinear systems using neural-network-based
adaptive gain scheduling

YANG Yueneng* and YAN Ye

College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China

Abstract: A neural-network-based adaptive gain scheduling back- difficult to deal with, the above-mentioned backstepping
stepping sliding mode control (NNAGS-BSMC) approach for a control approach may not achieve the control goal. With
class of uncertain strict-feedback nonlinear system is proposed. respect to this problem, the sliding mode control (SMC)
First, the control problem of uncertain strict-feedback nonlinear
represents an attractive alternative, which has a particular
systems is formulated. Second, the detailed design of NNAGS-
BSMC is described. The sliding mode control (SMC) law is de-
property of insensitivity to parametric uncertainty and ex-
signed to track a referenced output via backstepping technique. ternal disturbance due to the use of a switching control law
To decrease chattering result from SMC, a radial basis function [8 – 13]. Therefore, the back-stepping sliding mode con-
neural network (RBFNN) is employed to construct the NNAGS- trol (BSMC) approach which combines both the merits of
BSMC to facilitate adaptive gain scheduling, in which the gains are backstepping control and SMC, is proposed to address the
scheduled adaptively via neural network (NN), with sliding surface
control problem of uncertain strict-feedback nonlinear sys-
and its differential as NN inputs and the gains as NN outputs. Fi-
nally, the verification example is given to show the effectiveness
tems. However, discontinuous control in SMC is not rea-
and robustness of the proposed approach. Contrasting simulation lizable via the backstepping approach based on Lyapunov
results indicate that the NNAGS-BSMC decreases the chattering stability theory employing the so-called “hard” gain con-
effectively and has better control performance against the BSMC. trol laws, which yields undesirable chattering [14,15]. To
Keywords: backstepping control, sliding mode control (SMC), decrease the chattering, a radial basis function neural net-
neural network (NN), strict-feedback system, chattering decrease. work (RBFNN) is employed to construct the A neural-
network-based adaptive gain scheduling backstepping slid-
DOI: 10.21629/JSEE.2018.03.15
ing mode control (NNAGS-BSMC) to facilitate adaptive
gain scheduling, in which the gains are scheduled adap-
1. Introduction tively via neural network (NN), with sliding surface and
It is well known that the backstepping control technique its differential as NN inputs and the gains as NN outputs
has become one of the most effective control approaches [9,16]. Thus, the proposed control approach, not only takes
for nonlinear systems in strict-feedback form [1 – 4]. The advantage of design flexibility of backstepping control, but
key feature of backstepping design is that it stabilizes also brings into full play SMC’s robustness against para-
the system states through a step-by-step recursive process, metric uncertainties and external disturbances.
which reduces the complexity of the control design and The rest of this paper is organized as follows. In the
guarantees the Lyapunov stability of the closed-loop sys- next section, the control problem is formulated. Section 3
tem [5 – 7]. proposes BSMC and NNAGS-BSMC for uncertain strict-
For more general and challenging problems, involving feedback nonlinear systems. The verification example is
system uncertainties and external disturbances which are presented to demonstrate the effectiveness of proposed
control approach in Section 4. Finally, conclusions are
Manuscript received May 14, 2016.
*Corresponding author. given in Section 5.
This work was supported by the National Natural Science Foundation
of China (11502288), the Natural Science Foundation of Hunan 2. System description and problem
Province (2016JJ3019), the Aeronautical Science Foundation of China formulation
(2017ZA88001) and the Scientific Research Project of National Univer-
sity of Defense Technology (ZK17-03-32). Consider the following uncertain strict-feedback nonlinear
YANG Yueneng et al.: Backstepping sliding mode control for uncertain strict-feedback nonlinear systems using ... 581

system [17,18]: iv) Compute the derivative of e1 :

⎧
⎨ ẋi = xi+1 , i = 1, 2, . . . , n − 1 ė1 = −k1 e1 + e2 .
ẋn = f (x) + g(x)u + d(x, t) (1) Step 2 Design of virtual control α2
⎩
y = x1 i) Compute the tracking error e2 = x2 − α1 ;
ii) Select the optional parameter k2 > 0;
where x = [x1 , x2 , . . . , xn ]T ∈ Rn , u ∈ R and y ∈ R
iii) Design the virtual control α2 = −k2 e2 − e1 + α̇1 ;
are the state variables, system input and output, respec-
iv) Compute the derivative of ė2 = e3 − k2 e2 − e1 .
tively, which are all assumed to be available for measure-
Step i Design of virtual control αi
ment; f (x) and g(x) are smooth nonlinear functions and
i) Compute the tracking error ei = xi − αi−1 ;
d(x, t) includes parametric uncertainties and external dis-
ii) Select the optional parameter ki > 0;
turbances.
iii) Design the virtual control αi = −ki ei −ei−1 + α̇i−1 ;
Assumption 1 For the referenced signal yr , it is as-
iv) Compute the derivative of ei : ėi = ei+1 −ki ei −ei−1 .
sumed that yr is bounded and with n − 1 bounded deriva-
Step n Design of the control law
tives, i.e., there exists a positive constant Y > 0 such that
i) Compute the tracking error en = xn − αn−1 ;
[yr , yr(1) , . . . , yr(i) , . . . , yr(n) ]  Y (2) ii) Select the optional parameter c > 0;
iii) Define the following sliding surface s = cen−1 +en ;
where  ·  denotes the Euclidean norm. iv) Select the optional parameters ρ > 0, λ > 0;
Assumption 2 The parametric uncertainties and exter- v) Design the reaching law ṡ = −ρs − λsign(s);
nal disturbances are assumed to be bounded such that vi) Design the control input u:

u = g −1 (x)[−c(−kn−1 en−1 − en−2 + en ) − en−1 −

d(x, t)  D (3)
f (x) + α̇n−1 − ρs − λsign(s)].
where D is a positive constant.
Remark 1 In order to design the BSMC, the Assump- Step n + 1 Termination
tion 2 given in (3) is employed. If the tolerance of control error is satisfied, terminate the
algorithm and output u. Otherwise, go to Step 1.
The control problem of the current paper is stated as:
The stability of the closed-loop system is proven via the
given any initial states x1 (0) and x2 (0), and the referenced
following theorem.
signal yr , consider the nonlinear system defined by (1), de-
Theorem 1 For the nonlinear system (1), if the sliding
sign an appropriate law such that the system output can
surface is chosen as
track the referenced signal asymptotically, i.e.,
s = cen−1 + en (5)
lim |y − yr | = lim |e| = 0 (4)
t0 →∞ t0 →∞
and the control input is designed as
where yr denotes the referenced signal, e is the tracking
u = g −1 (x)[−c(−kn−1 en−1 − en−2 + en ) − en−1 −
error.
f (x) + α̇n−1 − ρs − λsign(s)] (6)
3. Control design
then it guarantees that all the signals of the closed-loop
3.1 BSMC
system are bounded.
The current subsection presents a BSMC approach for the Proof Select the following Lyapunov function candi-
control problem [17 – 19]. The detailed design is described date:
n−1
step-by-step as follows. 1 2 1 2
V = e + s . (7)
Algorithm 1 BSMC 2 i=1 i 2
Inputs: Differentiating (7) with respect to time and using the
i) referenced output yr ; derivatives of ei and s, it is obtained that
ii) system output y.
n−1
 n−1

Output: the control input u for system (1).
V̇ = ei ėi + sṡ = − ki e2i + en−1 en +
Step 1 Design of virtual control α1
i=1 i=1
i) Compute the tracking error e1 = y − yr ;
ii) Select the optional parameter k1 > 0; s[c(−kn−1 en−1 −en−2 +en )+f (x)+g(x)u− α̇n−1 +d].
iii) Design the virtual control α1 = −k1 e1 + ẏr ; (8)
582 Journal of Systems Engineering and Electronics Vol. 29, No. 3, June 2018

Substituting (6) into (8) yields Define μ = min{2k1 , . . . , 2(kn − c), 2ρ}, and then (9)
n−1 can be expressed as follows:

V̇ = − ki e2i + en−1 (s − cen−1 )+ n−1 
i=1 1 1
V̇  −μ e2i + s2 = −μV. (13)
s(−en−1 − ρs − λsign(s) + d) = i=1
2 2
n−1

− ki e2i − ce2n−1 − ρs2 − λ|s| + ds. (9) 
i=1 Remark 2 It is proven that all states and signals are
Based on Assumption 2, we have bounded, and it is concluded that the BSMC designed as
n−1
 (6) ensures Lyapunov stability of the nonlinear system (1).
V̇ = − ki e2i − ce2n−1 − ρs2 − λ|s| + ds 
i=1
3.2 NNAGS-BSMC
n−1
 The designed BSMC provides an effective control ap-
− ki e2i − ce2n−1 − ρs2 − λ|s| + D|s|. (10) proach for the system (1). However, after the system state
i=1 reaches the sliding phase, chattering phenomena is brought
By selecting the optional gain, λ satisfies the following about near the sliding surface, that is the major drawback
condition: of the BSMC. Because the control law (6) includes the
λ>D (11) sign function, the repeated switching of the sliding sur-
and we have face induces chattering, and the gain of sign function de-
n−1
termines the chattering intensity. To address this problem,

V̇  − ki e2i − ce2n−1 − ρs2 − (λ − D)|s|  an RBFNN is adopted to construct the NNAGS-BSMC to
i=1 facilitate adaptive gain scheduling in this subsection (its
block diagram is depicted in Fig. 1), in which the gains are
n−1

− ki e2i − ce2n−1 − ρs2 . (12) scheduled adaptively, with sliding surface s and its differ-
i=1
ential ṡ as NN inputs and the gain λ as NN output.

Fig. 1 Block diagram of NNAGS-BSMC

The design of RBFNN is described in detail as follows The following membership function is selected:
[20 – 26]. 
s − μi 2
Algorithm 2 RBFNN hi (s) = exp
σi2
Inputs:
where μi and σi are the mean and the standard deviations
i) sliding surface s; of the Gaussian function in the ith term, respectively.
ii) its differential ṡ. Step 3 Calculating the output
Calculate the gain of the sign function:
Output: gain λN of sign function.
n

Step 1 Defining the input layer λN = wi hi
The sliding surface and its differential are defined as NN i=1
net inputs, i.e., s = [s, ṡ]T ; where wi is the weight connecting the ith RBF μi and the
Step 2 Selecting the membership function output, and n is the number of RBFs.
YANG Yueneng et al.: Backstepping sliding mode control for uncertain strict-feedback nonlinear systems using ... 583

Step 4 Design of learning algorithm is effective for tracking the referenced output. Fig. 2(c)
i) Select the energy function displays the control input of BSMC. The control input
switches continuously, and the chattering phenomenon is
1 2
E= e obvious due to the use of a switching control law.
2
where e = y − yr is the tracking error;
ii) Design the learning algorithm based on the back
propagation method:

Δw ≈ −γe(−g −1(x)sign(s)h(s))sign(wT h(s)).

Step 5 Termination
If the chattering is decreased and the tolerance of con-
trol error is satisfied, terminate the algorithm and output
λN . Otherwise, go to Step 1.
Remark 3 The mean and standard deviations of the
kernel layer are updated in the similar way. The gain λN
can be scheduled adaptively via RBFNN with switching
sliding surface and its differential as inputs.
Based on the designed RBFNN, the NNAGS-BSMC is
designed as follows:

uN = g −1 (x)[−c(−kn−1 en−1 − en−2 + en ) − en−1 −

f (x) + α̇n−1 − ρs − λN sign(s)]. (14)

4. Verification example
In the current subsection, simulation results are presented
to illustrate the effectiveness and robustness of the de-
signed control approaches. Take the following nonlinear
system for instance:
⎧
⎨ ẋ1 = x2
x2 = 4.5 cos x1 + (12 cos x1 )u + 0.4 sin(π/100t) .
⎩
y = x1

(15)
The referenced output is selected as yr =
0.5 sin(π/15t), the external disturbance is selected as
0.02 sin(π/100t), and the initial state is x1 (0) = 0.035
and x2 (0) = 0. Simulations are performed by using
BSMC, BSMC using saturation function (SF-BSMC) and
NNAGS-BSMC, respectively.
4.1 Simulation of BSMC
The optional parameters of BSMC are chosen as ρ = 50,
Fig. 2 Simulation results of BSMC
λ = 15. Simulation results of BSMC are shown in Fig. 2.
Fig. 2(a) shows the referenced output (dashed) and system
4.2 Simulation of SF-BSMC
output (real line). The system output y with an initial value
of 0.035 increases to the referenced output yr within 2 s. Replacing the sign function with a saturation function is
Fig. 2(b) shows the control errors. The error decreases to one of the most common solutions to the chattering prob-
zero after 2 s. Fig. 2(a) and Fig. 2(b) indicate that BSMC lem. The sign function of BSMC given by (6) is replaced
584 Journal of Systems Engineering and Electronics Vol. 29, No. 3, June 2018

with a saturation function, and the SF-BSMC is expressed

as follows:

u = g −1 (x)[−c(−kn−1 en−1 − en−2 + en ) − en−1 −

f (x) + α̇n−1 − ρs − λsat(s)]

s
where sat(s) = , δ = 0.02.
|s| + δ
The other optional parameters of SF-BSMC are the
same as those of BSMC. Simulation results of SF-BSMC
are shown in Fig. 3. Fig. 3(a) shows the referenced output
(dashed) and system output (real line), and Fig. 3(b) shows Fig. 3 Simulation results of SF-BSMC
the control errors. Similar to Fig. 2(a) and Fig. 2(b), the
system output tracks the referenced output effectively, and
the error decreases to zero after 2 s. Fig. 3(c) displays the 4.3 Simulation of NNAGS-BSMC
control input of SF-BSMC. It is shown that the chattering The following Gaussian function is adopted to construct
of control input has been decreased, which illustrates the the RBFNN:
performance of chattering decrease of the SF-BSMC. 
s − μ2
h(s) = exp
σ2

where μ = 1.5, σ = 0.25.

The optional parameters are chosen as ρ = 15, γ = 0.8.
Different from BSMC and SF-BSMC, the gain of the sign
function is not a constant, and it is scheduled online. Simu-
lation results of NNAGS-BSMC are shown in Fig. 4.
Fig. 4(a) shows the referenced output (dashed) and sys-
tem output (real line), and Fig. 4(b) shows the control er-
rors. Similar to Fig. 2(a) and Fig. 2(b), the system output
tracks the referenced output effectively, and the error de-
creases to zero after 2 s. Fig. 4(c) displays the control input
of NNAGS-BSMC. Compared with BSMC, the chattering
of control inputs has been decreased effectively. Fig. 4(d)
displays the gain of the sign function, which is scheduled
via the RBFNN adaptively.
YANG Yueneng et al.: Backstepping sliding mode control for uncertain strict-feedback nonlinear systems using ... 585

5. Conclusions
This paper has successfully demonstrated the development
and application of the NNAGS-BSMC for uncertain strict-
feedback nonlinear systems with parametric uncertainties
and external disturbances. The BSMC is designed firstly,
then, an RBFNN is used to design the NNAGS-BSMC
for adaptive self-gain-scheduling to decrease the chatter-
ing results from the sign function. Through the compar-
ing results, it demonstrates that the NNAGS-BSMC de-
creases the chattering effectively and has better control per-
formances against BSMC.

References
[1] CHUTIPHON P. Adaptive backstepping finite-time sliding
mode control of spacecraft attitude tracking. Journal of Sys-
tems Engineering and Electronics, 2015, 26(4): 826 – 839.
[2] ZHENG X M, LI P, LI H Y. Adaptive backstepping-based
NTSM control for unmatched uncertain nonlinear systems.
Journal of Systems Engineering and Electronics, 2015, 26(3):
557 – 564.
[3] WEI Y H, CHEN Y Q, LIANG S, et al. A novel algorithm
on adaptive backstepping control of fractional order systems.
Neurocomputing, 2015, 165: 2312 – 2322.
[4] CONG B L, LIU X D, CHEN Z. Disturbance observer based
time-varying sliding mode control for uncertain mechanical
system. Journal of Systems Engineering and Electronics, 2012,
23(1): 108 – 118.
[5] CHUNG C W, CHANG Y T. Backstepping control of multi-
input non-linear systems. IET Control Theory and Applica-
tions, 2013, 7(14): 740 – 741.
[6] CAYERO J, MORCEGO B, NEJJARI F. Modelling and adap-
tive backstepping control for TX-1570 UAV path tracking.
Aerospace Science and Technology, 2014, 39: 342 – 351.
[7] YANG Z X, ZHAO G S, RONG H J, et al. Adaptive back-
stepping control for magnetic bearing system via feedforward
networks with random hidden nodes. Neurocomputing, 2016,
174(22): 109 – 120.
[8] YANG Y N, YAN Y. Decoupled nonsingular terminal sliding
mode control for affine nonlinear systems. Journal of Systems
Engineering and Electronics, 2016, 27(1): 192 – 200.
[9] DEBBACHE G, GOLEA N. Neural network based adaptive
sliding mode control of uncertain nonlinear systems. Journal
of Systems Engineering and Electronics, 2012, 23(1): 119 –
128.
Fig. 4 Simulation results of NNAGS-BSMC
[10] YANG Y N, WU J, ZHENG W. Trajectory tracking for an au-
tonomous airship using fuzzy adaptive sliding mode control.
Journal of Zhejiang University – Science C, 2012, 13(7): 534 –
543.
Remark 4 i) The BSMC, SF-BSMC and NNAGS-
[11] YANG Y N, YAN Y. Neural network approximation-based
BSMC all address the control problem of strict-feedback nonsingular terminal sliding mode control for trajectory track-
nonlinear system with parametric uncertainties and exter- ing of robotic airships. Aerospace Science and Technology,
nal disturbances effectively. ii) The chattering is the major 2016, 42: 50 – 57.
[12] YANG Y N, YAN Y. Attitude regulation for unmanned quadro-
drawback of BSMC. iii) The SF-BSMC reduces the chat- tors using adaptive fuzzy gain-scheduling sliding mode con-
tering by replacing the sign function with the saturation trol. Aerospace Science and Technology, 2016, 54: 208 – 217.
function. iv) The gain of NNAGS-SMC is scheduled adap- [13] YANG Y N. A time-specified nonsingular terminal sliding
mode control approach for trajectory tracking of robotic air-
tively with the sliding surface and its differential according
ships. Nonlinear Dynamics, 2018, 92: 1359 – 1367.
to the designed RBFNN, which reduces the chattering ef- [14] YANG Y N, WU J, ZHENG W. Attitude control for a station-
fectively. keeping airship using feedback linearization and fuzzy sliding
586 Journal of Systems Engineering and Electronics Vol. 29, No. 3, June 2018

mode control. International Journal of Innovative Computing, nonlinear dynamic system. IEEE Trans. on Industrial Electron-
Information and Control, 2012, 8(12): 8299 – 8310. ics, 2014, 61(2): 1099 – 1112.
[15] WANG L, SHENG Y Z, LU X D. High-order sliding mode at- [25] QASIM M, KANJIYA P, KHADKIKAR V. Artificial-neural-
titude controller design for reentry flight. Journal of Systems network-based phase-locking scheme for active power filters.
Engineering and Electronics, 2014, 25(5): 848 – 858. IEEE Trans. on Industrial Electronics, 2014, 61(8): 3857 –
[16] VALTIERRA R, JESUS R T, OSORNIO R A, et al. Detec- 3866.
tion and classification of single and combined power quality [26] YANG Y N, YAN Y. Neural network gain- scheduling slid-
disturbances using neural networks. IEEE Trans. on Industrial ing mode control for three-dimensional trajectory tracking of
Electronics, 2014, 61(5): 2473 – 2482. robotic airships. Journal of Systems and control Engineering,
[17] YIN X X, LIN Y G, LI W, et al. Adaptive sliding mode back- 2016, 229(6): 529 – 540.
stepping pitch angle control of a variable-displacement pump
controlled pitch system for wind turbines. ISA Transactions, Biographies
2015, 58: 629 – 634.
[18] LUA C H, HWANG Y R, SHEN Y T. Backstepping sliding
mode tracking control of a vane-type air motor X-Y table mo- YANG Yueneng was born in 1984. He received
tion system. ISA Transactions, 2011, 50: 278 – 286. his B.S. and M.S. degrees in civil engineering
[19] YANG Y N, YAN Y, ZHU Z L, et al. Chattering-free sliding from National University of Defense Technology,
mode control for strict-feedback nonlinear systems using back- Changsha, China, in 2006 and 2008, respectively,
stepping technique. Proc. of the International Conference on and his Ph.D. degree in aeronautical and astronau-
Mechanical, System and Control Engineering, 2017: 53 – 57. tical science and technology from National Uni-
[20] LI S, JIANG X Q. RBF neural network based second-order versity of Defense Technology, Changsha, in 2013.
sliding mode guidance for Mars entry under uncertainties. He is now a lecturer in College of Aerospace Sci-
Aerospace Science and Technology, 2015, 43: 226 – 235. ence and Engineering, National University of Defense Technology. His
[21] LIN F, LU K, KE T, et al. Reactive power control of three- research interests include nonlinear control, sliding mode control and
phase grid-connected PV system during grid faults using Tak- intelligent control.
agi – Sugeno – Kang probabilistic fuzzy neural network con- E-mail: [email protected]
trol. IEEE Trans. on Industrial Electronics, 2015, 62(9):
5516 – 5528. YAN Ye was born in 1971. He received his B.S.,
[22] YANG F Q, LI Z, LI H Y. Method of neural network modu- M.S. and Ph.D. degrees in aeronautical and astro-
lation recognition based on clustering and Polak-Ribiere algo- nautical science and technology from National Uni-
rithm. Journal of Systems Engineering and Electronics, 2014, versity of Defense Technology, Changsha, China, in
25(4): 671 – 580. 1994, 1997 and 2006, respectively. He is currently
[23] LIU Z G, WANG J Z, ZHAO J B. Neural network adaptive serving as a professor of College of Aerospace Sci-
sliding mode control for permanent magnet synchronous mo- ence and Engineering, National University of De-
tor. Electric Machines and Control, 2009, 13(4): 290 – 295. fense Technology. His research interests include air-
[24] HAN S I, LEE J M. Fuzzy echo state neural networks and fun- craft design, guidance, navigation and control.
nel dynamic surface control for prescribed performance of a E-mail: yeyan [email protected]