Professional Documents
Culture Documents
Abstract: Automatic train operation technique has gone through developments and advances in urban railways for decades
of years in China, which is now a mature technique that frees the train driver from continuous operation during the running
process. With respect to the main line railways and high-speed railways, demonstration project and tentative endeavors have
been conducted, which are designed to fully exert the capacities of Chinese Train Control System Level 2 and 3 (CTCS-2/3)
systems. This paper addresses a nonlinear feedback adaptive control method for improving the tracking control performance
of automatic train operation, the feature of the proposed method is that the implemented control gains and learning rates of
adaptation laws adjust automatically along with the varying tracking performance during the operation process, which is realized
by a well-constructed continuous differentiable nonlinear feedback function. The proposed method requires no prior precise
information of the operational resistances, and both theoretical analysis and comparative results demonstrate the effectiveness of
the proposed method.
acteristics. Proportion-integral-derivative (PID) control was ATO: automatic train operation BTS: base transceiver station CTC: centralized traffic control
OTE: optical transmission device BRA: balise receiving antenna CCS: comm. control server
the first automatic control algorithm for ATO implemented DMI: driver machine interface CRA: circuits receiving antenna TSRA: temp. speed restriction server
7782
linear feedback technique, originally proposed in our previ- Property 2: Define Nf (s) := Nd (s) · s + N (s), then
ous work [16] for a class of uncertain strict-feedback non- Nf (s) is a monotone increasing function versus s, and
linear systems, to improve the tracking control performance Nf (s) · s ≥ N (s) · s is true.
of automatic train operation process by virtue of a continu-
3 Main Results
ous nonlinear function. Second, a non-quadratic form Lya-
punov function is constructed to prove the closed-loop sta- The adaptive control design will be designed based on the
bility, which guarantees that the tracking error respect to the following coordinates transformation:
target velocity-distance profile can be adjusted to as small
as possible by choosing proper design parameters, and the e(t) =p(t) − pd
(4)
estimation errors of unknown coefficient of Davis equation, ė(t) =v(t) − ṗd = v(t) − vd
which reflects the uncertain operational resistance, are also
where pd and vd denote the desired distance and velocity
guarantee to converge to small regions tunable by design pa-
values derived from the pre-determined profile. To facilitate
rameters.
the control design, we introduce the following filtered error
The rest of this paper is organized as follows. Section 2
variable n(t):
gives some preliminaries and formulated problem. The non-
n(t) = ė(t) + αe(t) (5)
linear feedback based adaptive control is designed in Section
3, where the closed-loop system is also proved to be stable. with α being a positive constant. Differentiating both sides
Section 4 presents comparative results to demonstrate the ef- of (5), one has
fectiveness of the proposed method, some next step works
ṅ =ë + αė
are also briefly introduced in Section 5. Section 6 ends this
paper with conclusions. =u(t) − f (v, t) − v̇d + αė (6)
operation can be given as: for a given reference velocity- ĉ˙ = − δc Nf (n)v 2 + σc ĉ (8c)
distance profile versus time, the actual train can track the
target profile automatically without requiring the precise in- where δa > 0, δb > 0, and δc > 0 are learning rate coeffi-
formation of the operational resistances. For both theoretical cients, σa , σb , and σc are small constants.
and practical purposes, the reference profile needs to be s- One can summarize the above-design procedures as the
mooth enough such that the distance, speed and acceleration following theorem.
values versus time are all bounded. Theorem 1. Consider the train dynamics with unknown pa-
Lemma 1. A nonlinear feedback function has been proposed rameters described by (1). If the designed controller in (7)
in our previous work for a class of strict-feedback nonlinear and the corresponding adaptation laws (8) are implement-
systems, which is briefly introduced here, one refers [16] for ed, then, i), the actual real-time distance and velocity can
more details. be guaranteed to track the desired ones with the adjustable
In order to improve the dynamic performance of con- tracking errors, which can adjusted to arbitrarily small by
trollers and facilitate the stability analysis using Lyapunov choosing proper design parameters, and ii), the closed-loop
theorem, a continuous differentiable nonlinear gain function signals are all kept bounded.
is designed as follows:
⎧ Proof. Substituting (7) into (6) yields the closed-loop dy-
⎨ s, if |s| ≤ Δ namics as
N (s) = [loga (1 − ln a · Δ + ln a · |s|) + Δ] sign(s), n
⎩ ṅ = − k Nf (n) + + â + b̂v + ĉv 2 − f (v, t)
if |s| > Δ m
(2) n
where a > 1, Δ > 0. This designed nonlinear gain function = − k Nf (n) + + ã + b̃v + c̃v 2 (9)
m
holds the following properties.
Property 1: The function N (s) is a continuous differen- where ã = â−a, b̃ = b̂−b, c̃ = ĉ−c are the estimation errors
tiable strictly-monotone increasing function versus its argu- of Davis coefficient, respectively. Consider the Lyapunov
2 2 2
ment s, and its derivative is function as V = N (n)n + (1/2) ãδa + b̃δb + c̃δc , and its
derivative along (8) and (9) can be calculated as
1, if |s| ≤ Δ
Nd (s) = −1 ˙
(1 − ln a · Δ + ln a · |s|) , if |s| > Δ ãã˙ b̃b̃ c̃c̃˙
(3) V̇ =Nf (n)ṅ + + +
δa δb δc
7783
104 0.005
Nonlinear feedback AC
Linear feedback AC
20 0
10-8
-0.005 5
0
-0.01 -0.005 0
-5
2
-0.01
-0.015
-10
-0.015
-0.02 130 130.5 131 131.5
-0.02 10-4
0
-0.025 -0.025
-5
-0.03 -0.03
-10
-0.035
-0.035
-15
0 20 1960 1970 1980 1990 2000
0 0 -0.04
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
t (s) t (s)
Fig. 2: Reference velocity and distance profiles. (a) Comparative filtered tracking error n(t) of Case 1.
10-3
0.5
Nonlinear feedback AC
Nf (n)n
= − k Nf2 (n) +
Linear feedback AC
+ ãNf (n) + b̃vNf (n) 0
m
˙ -0.5 10-3 10-7
ãã˙ b̃b̃ c̃c̃˙ 0.5
+ c̃v 2 Nf (n) + + + -1 0
15
δa δb δ 10
c -0.5
N (n)n -1.5 5
= − k Nf2 (n) +
f
− σa ãâ − σb b̃b̂ − σc c̃ĉ -1
0
m -2 -1.5 300 350 400
(10) -2 10-4
0
-2.5 -2.5
-4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Nf (n)n σa 2 σb 2 σc 2
V̇ ≤ − k Nf2 (n) +
t (s)
− ã − b̃ − c̃
m 2 2 2 (b) Comparative filtered tracking error n(t) of Case 2.
σa 2 σb 2 σc 2
+ a + b + c Fig. 3: Trajectories of filtered tracking errors n(t).
2 2 2
kN (n)n σa 2 σb 2 σc 2
≤− − ã − b̃ − c̃
m 2 2 2
σa 2 σb 2 σc 2 There exists some time moment t = T , the term e−T t ≈ 0
+ a + b + c when t > T , in this sense, the
2 2 2 ultimate bound
ofthe
k P
filtered error n is |n| ≤ max T Δ , min P
≤ − min , σa δa , σb δb , σc δc V + P (11) T ,Δ .
m
˜
Similarly,
bounds of A are
the transient and ultimate
σa 2 σb 2 σc 2
with P := 2 a + 2 b + 2 c . From (11), one has 2δA V (0)e−T t + 2δTAP
and 2δA P
, respectively,
T
where A ∈ {a, b, c}. It is also noticeable that the above-
V (t) ≤V (0)e− min{ m ,σa δa ,σb δb ,σc δc }t
k
mentioned bounds can be adjusted to as small as possible
P by choosing proper design parameters. Other closed-loop
+ k (12) signals are guaranteed to be bounded by similar arguments
min m , σa δa , σb δb , σc δc
in previous works [17–20], which is not discussed in details
where V (0) is the initial value of V (t). When the filtered here. The proof completes.
error n(t) satisfies |n(t)| ≤ Δ, N (n)n = n2 , or else, 4 Simulation Results
N (n)n = [loga (1 − ln a · Δ + ln a · |n|) + Δ] |n| ≥
Δ|n|. As a component of the Lyapunov function, In this section, simulation and comparative results are p-
N (n)n ≤ V is known to be always true. As a
resented to demonstrate the effectiveness of the proposed
method. The following traditional adaptive controller will
result, n satisfies |n| ≤ V (0)e−T t + P
T with
k be utilized for unprejudiced comparison:
T := min m , σa δa , σb δb , σc δc if |n| ≤ Δ, if
not, |n| ≤ V (0)e
−T t
+ TPΔ . In short, the bound- u(t) = − kn + â + b̂v + ĉv 2 + v̇d − αė (13a)
Δ
ed value
of the filtered error n is obtained as |n| ≤ â˙ = − δa [n + σa â] (13b)
V (0)e−T t
max P
+ T Δ , min P
V (0)e−T t + T , Δ . ˙
Δ b̂ = − δb nv + σb b̂ (13c)
7784
10-3
0.01 2
Nonlinear feedback AC
0.005 Linear feedback AC Nonlinear feedback AC
1 Linear feedback AC
0
-0.005 10-4 0
0.01
10-3 10-4
-0.01 0 2
0
-1 0
-1
-0.015 0
-0.01 -2
-2
-0.02 -3 -2
-0.02
-2
-4 -4
-0.025 -0.03
-5 -3
-4 -6
-0.03 -0.04 -6 0 50 100 150 1950 1960 1970 1980 1990 2000
0 10 20 30 1960 1980 2000
-0.035 -4
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
t (s) t (s)
(a) Comparative velocity tracking errors. (a) Comparative velocity tracking errors.
0.04 0.025
Nonlinear feedback AC Nonlinear feedback AC Linear feedback AC
Linear feedback AC 0.02
0.02
0.015
0 0.01
0.01 0.01
0 0.01
0.005
-0.02 0.005
-0.02 0.005
0 0 0
-0.04 -0.04
0 -0.005 -0.005
-0.06 -0.01
-0.01 -0.01
-0.06 -0.005
-0.08 -0.015
-0.015 -0.02
-0.08 -0.1 -0.01 -0.02
-0.02 0 20 40 60 80 1960 1980 2000
0 20 40 1960 1980 2000
-0.1 -0.025
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
t (s) t (s)
(b) Comparative distance tracking errors. (b) Comparative distance tracking errors.
Fig. 4: Comparative results of Case 1 (Red solid line: er- Fig. 5: Comparative results of Case 2 (Red solid line: er-
ror using nonlinear feedback adaptive control; Blue dot line: ror using nonlinear feedback adaptive control; Blue dot line:
error using linear feedback adaptive control). error using linear feedback adaptive control).
7785
ties, etc. The prior knowledge of these factors can not [4] S. Gao, H. Dong, Y. Chen, B. Ning, and G. Chen, “Adaptive
be obtained off-line even if enough historical data are and robust automatic train control systems with input satu-
available. The core idea of the RDCM mixed-driven ration,” Control and Intelligent Systems, vol. 41, no. 2, pp.
lies in that the mathematical model of high-speed trains 103–111, 2013.
is established by the organic integrity of CM based on [5] H. Dong, S. Gao, B. Ning, and L. Li, “Extended fuzzy logic
controller for high speed train,” Neural Computing and Appli-
historical data and RD reflecting the real-time opera-
cations, vol. 22, no. 2, pp. 321–328, 2013.
tional environments. Theorematically speaking, such a [6] X. Xu, H. Yi, and Y. Wang, “Optimization of train control sys-
modeling method merges the merits of data-driven and tem using ctcs2+ato in intercity railways (in chinese),” Rail-
model-driven based control. Only in this way can the way signalling and communication, vol. 53, no. 3, pp. 12–16,
accurate model of high-speed trains with complex envi- 2017.
ronments and high speed be obtained. [7] K. Ichikawa, “Application of optimization theory for bounded
2) Artificial intelligence (AI)-based ATO control. Intel- state variable problems to the operation of train,” Bulletin of
ligent control on the basis of control theory, comput- JSME, vol. 11, no. 47, pp. 857–865, 1968.
er science, artificial intelligence and opsearch produces [8] P. Gruber and M. Bayoumi, “Suboptimal control strategies
several intelligent methods, including fuzzy sets, neu- for multilocomotive powered trains,” IEEE Transactions on
ral networks, expert systems, and so on. Our next step Automatic Control, vol. 27, no. 3, pp. 536–546, 1982.
[9] S. Yasunobu, S. Miyamoto, and H. Ihara, “Fuzzy control for
attempt will be AI-based online optimization and con-
automatic train operation system,” in IFAC/IFIP/IFORS Con-
trol to satisfy the dynamic constraints of operational s- ference on Control in Transportation Systems (4th). Control
tationarity, online scheduling, energy-saving criterions. in transportation systems, 1984.
3) Integration of online rescheduling (OR) and ATO. [10] C. Chang, D. Xu, and H. Quek, “Pareto-optimal set based
In most existing literatures, the studies of reschedul- multiobjective tuning of fuzzy automatic train operation for
ing and ATO are conducted separately. The online mass transit system,” IEE Proceedings-Electric Power Appli-
rescheduling problem is to set the arrival and departure cations, vol. 146, no. 5, pp. 577–583, 1999.
moments and routes discretely, while the ATO control [11] E. Khmelnitsky, “On an optimal control problem of train op-
is to design real-time control signals continuously. The eration,” IEEE transactions on automatic control, vol. 45,
core idea of integration of OR and ATO lies in that the no. 7, pp. 1257–1266, 2000.
rescheduling decisions are made with real-time status [12] Q. Song and Y. Song, “Adaptive control and optimal pow-
er/brake distribution of high speed trains with uncertain non-
of trains and the implemented control signals in ATO
linear couplers,” in Control Conference (CCC), 2010 29th
are designed with explicit considerations of real-time Chinese. IEEE, 2010, pp. 1966–1971.
OR decisions. Such an integration can be tentatively [13] S. Wen, J. Yang, A. B. Rad, and P. Hao, “Multi-model direc-
attempted by hybrid modeling method to capture the t generalised predictive control for automatic train operation
essence of discrete OR and continuous ATO, and virtu- system,” IET Intelligent Transport Systems, vol. 9, no. 1, pp.
al marshalling method is a promising methodology to 86–94, 2014.
achieve the autonomous ATO under OR framework. [14] S. Li, L. Yang, and K. Li, “Robust output feedback cruise
control for high-speed train movement with uncertain param-
6 Conclusions
eters,” Chinese Physics B, vol. 24, no. 1, p. 010503, 2014.
In this paper, a nonlinear feedback adaptive control [15] Y. Zhao, T. Wang, and H. R. Karimi, “Distributed cruise con-
scheme is developed for improving the tracking control per- trol of high-speed trains,” Journal of the Franklin Institute,
formance of automatic train operation by utilizing a contin- vol. 354, no. 14, pp. 6044–6061, 2017.
uous differentiable nonlinear feedback function, where the [16] S. Gao, H. Dong, and B. Ning, “Neural adaptive dynamic sur-
face control for uncertain strict-feedback nonlinear systems
implemented control gains and learning rates of adaptation
with nonlinear output and virtual feedback errors,” Nonlinear
laws adjust automatically with the varying states. The pro-
Dynamics, vol. 90, no. 4, pp. 2851–2867, 2017.
posed method requires no prior information of operational [17] S. Gao, B. Ning, and H. Dong, “Fuzzy dynamic surface con-
resistances. Both theoretical analysis and comparative re- trol for uncertain nonlinear systems under input saturation via
sults are given to demonstrate the effectiveness of the pro- truncated adaptation approach,” Fuzzy Sets and Systems, vol.
posed method. Future theoretical research is also briefly in- 290, pp. 100–117, 2016.
troduced, and future practical work is to apply and validate [18] S. Gao, H. Dong, B. Ning, and X. Sun, “Neural adaptive con-
the proposed method in real applications. trol for uncertain mimo systems with constrained input via in-
tercepted adaptation and single learning parameter approach,”
References Nonlinear Dynamics, vol. 82, no. 3, pp. 1109–1126, 2015.
[1] S. Yasunobu, “Automatic train operation by predictive fuzzy [19] S. Gao, H. Dong, S. Lyu, and B. Ning, “Truncated adapta-
control,” Industrial applications of fuzzy control, pp. 1–18, tion design for decentralized neural dynamic surface control
1985. of interconnected nonlinear systems under input saturation,”
[2] S. Gao, H. Dong, B. Ning, Y. Chen, and X. Sun, “Adap- International Journal of Control, vol. 89, no. 7, pp. 1447–
tive fault-tolerant automatic train operation using rbf neural 1466, 2016.
networks,” Neural Computing and Applications, vol. 26, pp. [20] S. Gao, H. Dong, B. Ning, and X. Yao, “Single-parameter-
141–149, 2015. learning-based fuzzy fault-tolerant output feedback dynamic
[3] S. Gao, H. Dong, Y. Chen, B. Ning, G. Chen, and X. Yang, surface control of constrained-input nonlinear systems,” In-
“Approximation-based robust adaptive automatic train con- formation Sciences, vol. 385, pp. 378–394, 2017.
trol: An approach for actuator saturation,” IEEE Transaction-
s on Intelligent Transportation Systems, vol. 14, no. 4, pp.
1733–1742, 2013.
7786