Automatica 50 (2014) 641–645

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Technical communique

Fast reference governors for second-order linear systems with

constraints and an input time-delay✩
Qinghua Li a , Uroš V. Kalabić b,1 , Ilya V. Kolmanovsky b
a
Space Control and Inertial Technology Research Center, Harbin Institute of Technology, Harbin 150080, China
b
Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI 48109, United States

article info abstract

Article history: This note provides a solution to the constrained command tracking problem using reference governors
Received 1 October 2012 for a class of continuous-time second order linear systems with an input delay and with pointwise-in-
Received in revised form time state and control constraints. The reference governor modifies the command to a closed-loop system
5 August 2013
based on the prediction of whether the system response to constant commands violates the specified con-
Accepted 31 October 2013
Available online 11 December 2013
straints. The solution relies on classical control results for second order linear systems and requires only
checking whether predicted outputs violate the constraints at a small number of time instants (e.g., four
Keywords:
time instants in the single output case). This greatly simplifies the online computation, especially when
Constraints a reference governor is applied to system models that are (slowly) changing in time. The effectiveness of
Predictive control the proposed method is demonstrated by a numerical example.
Time-delay © 2013 Elsevier Ltd. All rights reserved.
Linear systems
Aerospace control

1. Introduction The adjustment of v(t ) occurs at discrete time instants t = kTs ,

where k is an integer and Ts is the sampling time. Let y(t |x(kTs ), v)
The reference governor (see Fig. 1 and Bemporad, 1998; Be- denote the predicted response for t ≥ kTs given the initial state
mporad, Casavola, & Mosca, 1997; Gilbert & Kolmanovsky, 2002; of the system x(kTs ), and the constant command v . The reference
Gilbert, Kolmanovsky, & Tan, 1995; Kapasouris, Athans, & Stein, governor adjusts v toward a desired reference r (kTs ), subject to the
1988; Ohta, Mori, Yukimoto, & Mishio, 2005) is a prediction-based constraint that y(t |x(kTs ), v) ∈ Y for all t ≥ kTs . It is sufficient to
control scheme for enforcing pointwise-in-time state and control perform the prediction over the horizon which is sufficiently long
constraints by modifying the reference command signal to nom- but finite as, under appropriate assumptions, if the constraints are
inal, well-designed closed-loop systems. For other approaches to satisfied over this finite horizon, then they will be satisfied for all
constrained control, see Bernstein and Michel (1995), Blanchini future time (Bemporad, 1998; Gilbert et al., 1995). This prediction
and Miani (2008), Gilbert and Tan (1991), Mayne (2001), and mechanism guarantees that the value of v((k − 1)Ts ) computed
Mayne, Rawlings, Rao, and Scokaert (2000) and references therein. at the previous time instant is feasible at the current time instant
The reference governor modifies the command r (t ) to v(t ) if kTs , and that the constraints are satisfied for all time. We note that
the danger of constraint violation exists based on a state measure- the online prediction can also be performed in the case when the
ment or estimate, x̂(t ). The reference governor is an add-on control model is nonlinear or varying with time (Bemporad, 1998; Gilbert
scheme to the nominal control system and preserves the response
& Kolmanovsky, 2002).
properties of the nominal control system when no danger of con-
In this paper, we focus on simplifications in predicting y(t |x
straint violation exists. It minimally interferes with the closed-loop
(kTs ), v) ∈ Y for a special class of systems which can be adequately
system response to enforce the constraints, when the danger of
represented by second order linear models, i.e., whose state vector
constraint violation exists in response to larger signals.
x(t ) is two dimensional, and which are subject to an input time-
delay. In this special case, we show that to determine if the re-
sponse of the system to a constant input violates the constraints,
✩ The second and third authors acknowledge the support by the National Science
it is sufficient to examine the response only at a small number
Foundation, Award Number 1130160. The material in this paper was not presented
of time instants (i.e., only four time instants in the single output
at any conference. This paper was recommended for publication in revised form by
Associate Editor Chunjiang Qian under the direction of Editor André L. Tits. case). Notwithstanding the effect of system uncertainty because
E-mail addresses: huahit@gmail.com (Q. Li), kalabic@umich.edu (U.V. Kalabić), this is outside of the scope of this technical communique, the prop-
ilya@umich.edu (I.V. Kolmanovsky). erty mentioned above lends itself to a fast procedure for updating
1 Tel.: +1 734 615 9655; fax: +1 734 763 0578. v(t ). This fast procedure is advantageous in situations when the
0005-1098/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.automatica.2013.11.007
642 Q. Li et al. / Automatica 50 (2014) 641–645

3. Predicting output response to constant inputs

For the linear system (1)–(2), the state transition formula for
linear systems implies that,
 t0 +τ
x(t ) = eA(t −t0 ) x(t0 ) + eA(t −σ ) Bv(σ − τ )dσ
t0
 t
Fig. 1. Reference governor schematic. + eA(t −σ ) Bv(σ − τ )dσ . (4)
t0 +τ
reference governor is applied to practical systems for which mod- By straightforward algebraic transformations of (4) and assuming
els or constraints vary in time or are learned online through system that v(t ) = v̄ for all t ≥ t0 , we obtain the following prediction
identification and/or adaptation. We present a numerical example equation,
that illustrates our approach to reference governor design and im- 
plementation based on these ideas. A(t −t0 )
x(t ) = −A −1
Bv̄ + e x(t0 ) + e−Aτ A−1 Bv̄
2. Problem formulation and preliminaries

 t0 +τ
Consider the following linear system, + eA(t0 −σ ) Bv(σ − τ )dσ , t ≥ t0 . (5)
t0
ẋ(t ) = Ax(t ) + Bv(t − τ ), x(t0 ) = x0 , (1)
y(t ) = Cx(t ) + Dv(t ) ∈ Y , (2) Let Ci and Di denote the ith rows of the matrices, C and D, i =
1, . . . , ny . Then the response of the ith output has the following
where x(t ), y(t ), and v(t ) are the state, output, and input of the form,
system, respectively. The constraints are prescribed on the output
variable, y(t ). The input time delay is τ ≥ 0. In applications of yi (t ) = γi + Ci eA(t −t0 ) Γ , t ≥ t0 , (6)
reference governors, the system (1)–(2) represents a closed-loop where γ = (−CA −1
B + D)v̄, Γ = Γ (t0 , τ , v̄) = z (t0 , τ ) + Z (v̄, τ ),
system consisting of a plant and a nominal controller, v(t ) is a ref- with,
erence command input to the system, and y(t ) is the output on  t0 +τ
which constraints are imposed and can represent state, output or z (t0 , τ ) = x(t0 ) + eA(t0 −σ ) Bv(σ − τ )dσ ,
control constraints on the variables of the closed-loop system. t0
We make the following assumptions:
Z (v̄, τ ) = e−Aτ A−1 Bv̄.
(A1) the state vector x(t ) in (1) is two-dimensional;
(A2) A in (1) is Hurwitz, i.e., (1) is asymptotically stable;
(A3) output y(t ) has dimension ny > 0 and the set Y is defined by Remark 1. Note that z (t0 , τ ) depends on past history of the com-
box-like inequalities, mand signal, v(t0 + θ ), −τ ≤ θ ≤ 0. In the online implemen-
tation of the prediction (6), the integral term is approximated by
yi,min ≤ yi ≤ yi,max , (3) discretization.
where yi,min and yi,max are given for i = 1, . . . , ny ;
(A4) the time delay τ ≥ 0, is constant and known. 4. Characterizing constraint admissible constant inputs

The assumption (A1) is restrictive. In certain applications, higher Given (3), the set of admissible constant commands v̄ at time t0
order systems may be reduced to second order using model or- satisfies,
der reduction. This is the case, for instance, if the closed-loop sys-
tem is designed with pole placement techniques such that it has yi,min ≤ γi + Ci eA(t −t0 ) Γ ≤ yi,max , (7)
a dominant closed-loop eigenvalue pair and the rest of eigenval- where t ≥ t0 , i = 1, . . . , ny . For second order systems, the infinite
ues are placed significantly further in the left half plane and are set of inequalities (7) can be easily checked using the following
much faster. As shown in Kalabić, Kolmanovsky, Gilbert, and Buck- results.
land (2012), reference governors can be designed based on such
reduced order models by appropriately tightening the output con- Proposition 1. Suppose that A is diagonalizable and has two real
straints and imposing appropriate constraints on changes in v(t ). and unequal eigenvalues, λ1 and λ2 . Let A = P Σ P −1 , where Σ =
The presence of the time delay in the input channel enlarges the diag(λ1 , λ2 ), and consider,
class of systems to which the subsequent results apply, e.g., to
higher order closed-loop systems that are remodeled as second or- yi (t ) = γi + Ci eA(t −t0 ) Γ (8)
der with time delay. The assumption (A2) is reasonable because
= γi + ci1 b1 eλ1 (t −t0 ) + ci2 b2 eλ2 (t −t0 ) , (9)
the reference governor is applied to closed-loops systems that are
asymptotically stable by design; this assumption also implies that where
A is invertible. The assumption (A3) is reasonable because con-  
straints can be redefined to make it hold. For instance, a constraint, b1
ci1 ci2 = Ci P , P −1 Γ = .
 
(10)
−1 ≤ y1 + y2 ≤ 1 can be rewritten as −1 ≤ ỹ1 ≤ 1 where ỹ1 = b2
y1 + y2 is a redefined output. Finally, (A4) assumes a known con-
Then,
stant time-delay.
The focus of this paper is on developing a scheme for predicting sup yi (t ) = max Π (t0 ) and inf yi (t ) = min Π (t0 ), (11)
whether the condition y(t |x(kTs ), v) ∈ Y holds and on using it for t ≥t0 t ≥ t0

reference governor implementation. We will demonstrate that it where either,

is sufficient to examine a small finite number of time instants t to
make this determination. Π (t0 ) = {γi , γi + ci1 b1 + ci2 b2 }, (12)
 Q. Li et al. / Automatica 50 (2014) 641–645 643

or if sgn ci1 b1 ̸= sgn ci2 b2 , ci1 b1 ̸= 0, and ci2 b2 ̸= 0, then, Noting that yi (t ) = eα(t −t0 ci1 b1 ejθ (t −t0 ) + ci2 b2 e−jθ (t −t0 is a
) )
 

product of decaying exponential, eα(t −t0 ) , and a 2π /θ -periodic

Π (t0 ) = {γi , γi + ci1 b1 + ci2 b2 , γi + ci1 b1 eλ1 t + ci2 b2 eλ2 t },
∗ ∗
(13) function, the infimum and supremum if achieved at the extrema
where t > t0 is the solution to ẏi (t ) = 0 and sgn is the algebraic
∗ ∗ must be achieved during the first period. The possibility of t1∗ < 0,
sign function. however, implies the need to also consider t3∗ .
Proof. Without loss of generality, we can assume t0 = 0. The val-
Remark 3. If h1 = 0, h2 ̸= 0, then extremal time instants, if they
ues, yi (0) = γi + ci1 b1 + ci2 b2 , and limt →∞ yi (t ) = γi , are poten-
exist, must satisfy the equation e−jθ (t −t0 ) h2 = 0. This equation can-
tial extreme points because they correspond to extreme points in
the prediction horizon. All other extreme points need to satisfy ex- not be satisfied since a pure rotation transformation of a non-zero
tremality conditions. The necessary first order condition, that there complex number cannot produce zero. Consequently, in this case,
is a solution to ẏi (t ) = ci1 b1 λ1 eλ1 t + ci2 b2 λ2 eλ2 t = 0, t > 0, is sat- Π (t0 ) = {γi , γi + ci1 b1 + ci2 b2 }.
isfied if and only if sgn ci1 b1 ̸= sgn ci2 b2 , ci1 b1 ̸= 0, and ci2 b2 ̸= 0.
Remark 4. Because a closed form solution to ẏi (t ) = 0 does not,
Remark 2. The t ∗ of Proposition 1, if it exists, can be easily found. in general, exist for higher order systems, the above results are
For instance, if λ1 > λ2 , then, limited to systems of second order. A similar and computationally
feasible approach applied to higher order systems could utilize an
−ci2 b2 λ2 approximate numerical solution to yi (t ) = 0 with constraint tight-
 
1
∗
t = ln . (14) ening on yi (t ). The development of this method is outside the scope
λ1 − λ2 ci1 b1 λ1
of this short note but possible approaches may exploit a predic-
We have the following result for the case in which A is not diago- tor–corrector root finding (Kolmanovsky, Kalabić, & Gilbert, 2012)
nalizable. and rational approximations of eigenvalues, in order to obtain a
polynomial root finding problem.
Proposition 2. If A is not diagonalizable and has one real eigenvalue,
λ1 , then (9) becomes yi (t ) = γi +(ci1 b1 + ci2 b2 + ci1 b2 (t − t0 ))eλ1 (t −t0 ) , Remark 5. Because ÿi (t ) is easily computed, second order neces-
assuming Σ is in Jordan canonical form. Consequently, sary conditions (ÿi (t ∗ ) ≥ 0 in Propositions 1 and 2 or ÿi (tm
∗
) ≥ 0 in
Π (t0 ) = {γi , γi + ci1 b1 + ci2 b2 , Proposition 3 for a minimizer or the same but with opposite sign
for a maximizer) can be applied to further restrict candidate max-
γi + (ci1 b1 + ci2 b2 + ci1 b2 (t ∗ − t0 ))eλ1 (t
∗ −t )
0
}, (15) ima and minima of the response.
where the last element is included if there exists t ∗ > t0 that solves
ẏi (t ∗ ) = 0. 5. Reference governor implementation
Proof. Without loss of generality, we can assume t0 = 0. The
The scalar reference governor adjusts v(kTs ), k ∈ Z+ , where
first two points are included in Π (t0 ) for the same reasons as in
Ts > 0, is the update period of the reference governor, based on
the proof of Proposition 1. The first order necessary condition, that
the solution of scalar minimization problem,
there exist t > 0 such that ẏi (t ) = (λ1 (ci1 b1 + ci2 b2 + ci1 b2 t )+ ci1 b2 )
eλ1 t = λ1 yi (t ) + ci1 b2 eλ1 t = 0, is satisfied only if ci1 b2 ̸= 0. κ(k) → max, (22)
v(kTs ) = v((k − 1)Ts ) + κ(k)(r (kTs ) − v((k − 1)Ts )), (23)
Proposition 3. Suppose that A has two complex conjugate eigenval-
ues, λ1,2 = α ± jθ , with α < 0, θ > 0. Let A = P Σ P −1 , where y(t |x(kTs ), v(kTs )) ∈ Y , ∀t ≥ kTs , κ(k) ∈ [0, 1] (24)
Σ = diag(λ1 , λ2 ), and consider,
where y(t |x(kTs ), v(kTs )) denotes the predicted response for t >
yi (t ) = γi + Ci eA(t −t0 ) Γ (16) kTs given the initial state of the system, x(kTs ), and the constant
command, v . The constraints (24) are rewritten based on (7) and
= γi + ci1 b1 eλ1 (t −t0 ) + ci2 b2 eλ2 (t −t0 ) . (17) simplified based on the results of Propositions 1–3. Essentially, (24)
needs to be checked at t = kTs as t → ∞, and at a small number of
Define two complex numbers, h1 = α + (jθ ) and h2 = α ci2 b2
ci1 b1 ci1 b1
time instants identified in Propositions 1, 2, or 3. Note that the re-
− ci b2 (jθ ). Suppose that h1 ̸= 0 and consider the time instants,
2
cursive feasibility if κ(k) = 0 is always preserved by the reference
mπ governor.
    
1 h2 h2
∗
tm = arctan im − , re − + , Numerical strategies for solving (22)–(24) exploit several ap-
2θ h1 h1 θ
proaches. For instance, bisections on the value of κ(k) can be used
m ∈ Z+ , (18) (Gilbert & Kolmanovsky, 2002). Alternatively, the incremental step
reference governor strategy can be used to distribute the updates
and the set,
of v(kTs ) over time (Tsourapas, Sun, & Stefanopoulou, 2009).
Π (t0 ) = {γi , γi + ci1 b1 + ci2 b2 , We summarize the bisections-based algorithm as follows:

γi + ci1 b1 eλ1 tm + ci2 b2 eλ2 tm },

∗ ∗
m = 0, 1, 2, (19) (1) set κmin = 0, κmax = 1;
(2) set κM = 12 (κmin +κmax ) and, vcand = v((k − 1)Ts )+κM (r (k)Ts −
where each of the last three elements is added only if the associated v((k − 1)Ts ));
∗
time instant, tm are well-defined and positive. Then, (3) check the conditions: yi,min ≤ yi (t |x(kTs ), vcand ) ≤ yi,max , i =
1, . . . , ny , based on the results of Propositions 1–3. If these
sup yi (t ) = max Π (t0 ) and inf yi (t ) = min Π (t0 ). (20)
t ≥t0 t ≥ t0 conditions are satisfied, then set κmin = κM ; otherwise set
κmax = κM ;
Proof. The proof follows from the observation that extrema of (4) if |κmin − κmax | ≤ ε , where ε > 0 is a stopping threshold, then
yi (t ) = γi + eα(t −t0 ) ci1 b1 ejθ(t −t0 ) + ci2 b2 e−jθ(t −t0 ) are the roots of,

stop; otherwise go to (2).

ẏi (t ) = α eα(t −t0 ) ci1 b1 ejθ(t −t0 ) + ci2 b2 e−jθ(t −t0 ) The computational benefits of the proposed approach become ap-
 
parent in practical situations when the system dynamics or con-
+ (jθ )eα(t −t0 ci1 b1 ejθ(t −t0 ) − ci2 b2 e−jθ(t −t0
) )
= 0.
 
(21) straints are slowly changing or become known at ‘‘run time’’, e.g.,
644 Q. Li et al. / Automatica 50 (2014) 641–645

Fig. 4. Control signal response with the reference governor (solid) and without the
Fig. 2. The position set-point and the output of the reference governor. reference governor (dot-dashed) plotted against constraints (dashed).

4.5·10−5 , β = 1.92, γ = 1.99, m = 1.54, c = 0.0659, k = 38.94,

and d0 = 0.0102. The control constraint due to current limits is
given by,

0 ≤ u(t ) ≤ umax = iβmax = 0.3. (27)

The constraint on x1 (t ) prevents collision of the mass and the elec-
tromagnet is given by,
x1 (t ) ≤ 0.0086. (28)
Consider the nonlinear stabilizing feedback control law,
1
u(t ) = u(x1 , x2 , v) = (d0 − x1 )γ (kv − cd x2 ), (29)
α
where cd = 0.1. The control input constraints (27), (29) are nonlin-
ear; hence we resort to its linearization. The linearized constraint
at the discrete time instant t has the form,
y1,min ≤ y1 = C1 x + D1 v ≤ y1,max ,
∂u ∂u ∂u
 
Fig. 3. Position response with the reference governor (solid) and without the C1 = , D1 = ,
reference governor (dot-dashed) plotted against constraints (dashed). ∂ x1 ∂ x2 ∂v
∂u ∂u
for aircraft transitioning between different altitudes, speeds, and y1,max = umax + x1 (t ) + x2 ( t )
angles of attack. In this case, if the reference governor based on ∂ x1 ∂ x2
∂u
online simulations and bisections is used (Gilbert & Kolmanovsky, + v(t − 1) − u(x1 (t ), x2 (t ), v(t − 1)),
2002), then, based on numerical experience, the proposed ap- ∂v
proach is an order of magnitude faster. An alternative approach of ∂u ∂u
covering the flight envelope by multiple maximal output admissi- y1,min = 0+ x1 ( t ) + x2 ( t )
∂ x1 ∂ x2
ble sets computed offline requires a substantially larger amount of ∂u
memory for implementation. Furthermore, we note that the ap- + v(t − 1) − u(x1 (t ), x2 (t ), v(t − 1)).
proach proposed in this paper implicitly takes advantage of the
∂v
exact (i.e. continuous-time) maximal output admissible set rather The partial derivatives in these expressions are evaluated at the
than its discrete-time approximation. present values of x1 (t ), x2 (t ), and v(t − 1). These partial deriva-
tives are easily computed given the closed form expression of the
6. Numerical example control law (29). Note that the resulting linearized constraint is
time-varying. The constraint (28) on the position can be restated
To illustrate the implementation of the reference governor as follows,
based on the proposed method, we consider an example of a y2,min ≤ y2 = C2 x + D2 v ≤ y2,max ,
second order electromagnetically actuated mass–spring–damper
0 , D2 = 0,
 
system (EAMSD) from Gilbert and Kolmanovsky (2002). The open- C2 = 1
loop system is described by the equations, y2,max = 0.0086, y2,min = −∞.
ẋ1 = x2 , (25) Figs. 2–4 show the responses. Fig. 2 shows the finite-time con-
k c α u vergence of the set-point, v(t ), to the desired constant reference.
ẋ2 = − x1 − x2 + , u = iβ , (26) Figs. 3 and 4 show adherence to state and control constraints, re-
m m m d 0 − x1 spectively. At 6 s, the outputs are still oscillating and have yet
where x1 is the position, x2 is the velocity, u is the control input and to settle to the reference value; however, the trajectories remain
i denotes the current applied to the coil. The parameters are α = within constraints.
 Q. Li et al. / Automatica 50 (2014) 641–645 645

The fast computations of the proposed reference governor are Bernstein, D. S., & Michel, A. N. (1995). A chronological biography on satu-
an advantage given that one of the constraints is time-varying and rating actuators. International Journal of Robust and Nonlinear Control, 5(5),
375–380.
has to be handled online. Blanchini, F., & Miani, S. (2008). Set-theoretic methods in control. New York:
Birkhäuser.
Gilbert, E., & Kolmanovsky, I. (2002). Nonlinear tracking control in the presence
7. Conclusion
of state and control constraints: a generalized reference governor. Automatica,
38(12), 2063–2073.
In this paper, we have studied the design of reference governors Gilbert, E. G., Kolmanovsky, I., & Tan, K. T. (1995). Discrete-time reference governors
for second order continuous-time linear systems with constraints and the nonlinear control of systems with state and control constraints.
International Journal of Robust and Nonlinear Control, 5, 487–504.
and an input time-delay. A new method of reference governor de-
Gilbert, E. G., & Tan, K. T. (1991). Linear systems with state and control constraints:
sign is proposed which requires calculating the predicted output at the theory and application of maximal output admissible sets. IEEE Transactions
a very small number of time instants to judge constraint violation. on Automatic Control, 36, 1008–1020.
Kalabić, U., Kolmanovsky, I., Gilbert, E., & Buckland, J. (2012). Reduced order
The approach takes advantage of the characteristics of the second
reference governor. In Proc. IEEE conf. decision and control (pp. 3245–3251).
order linear systems and reduces the computing time of the refer- Maui, HI, Dec.
ence governor algorithm, while being able to account for an input Kapasouris, P., Athans, M., & Stein, G. (1988). Design of feedback control systems for
delay. The effectiveness of the proposed method is shown in a nu- stable plants with saturating actuators. In Proc. IEEE conf. decision and control
(pp. 469–479). Austin, TX, Dec.
merical simulation. Future research will broaden the applicability
Kolmanovsky, I., Kalabić, U., & Gilbert, E. (2012). Developments in constrained
of the proposed approach by linking it to reduced order reference control using reference governors. In Proc. IFAC conf. nonlinear model predictive
governor results (Kalabić et al., 2012). We will also generalize these control (pp. 282–290). Noordwijkerhout, Netherlands, Aug.
results to networked systems and consider the development of ap- Mayne, D. Q. (2001). Control of constrained dynamic systems. European Journal of
Control, 7(2–3).
proximate techniques for higher order systems.
Mayne, D. Q., Rawlings, J. B., Rao, C. V., & Scokaert, P. O. M. (2000). Con-
strained model predictive control: stability and optimality. Automatica, 36(6),
References 789–814.
Ohta, Y., Mori, K., Yukimoto, K., & Mishio, R. (2005). On-line reference management
Bemporad, A. (1998). Reference governor for constrained nonlinear systems. IEEE for discrete-time servo systems under state and control constraints. In Proc. conf.
Transactions on Automatic Control, 43(3), 415–419. IEEE industrial electronics society (pp. 183–188). Raleigh, NC, Nov.
Bemporad, A., Casavola, A., & Mosca, E. (1997). Nonlinear control of constrained Tsourapas, V., Sun, J., & Stefanopoulou, A. (2009). Incremental step reference
linear systems via predictive reference management. IEEE Transactions on governor for load conditioning of hybrid fuel cell and gas turbine power plants.
Automatic Control, 42(3), 340–349. IEEE Transactions on Control Systems Technology, 17(4), 756–767.