This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY 1

Equivalence Among Flat Filters, Dirty

Derivative-Based PID Controllers, ADRC,
and Integral Reconstructor-Based
Sliding Mode Control
Hebertt Sira-Ramírez, Eric William Zurita-Bustamante , and Congzhi Huang, Senior Member, IEEE

Abstract— We explore the equivalence among flat filters, Sliding mode control (SMC) is also a paradigmatic method
dirty derivative-based proportional integral derivative (PID) that is frequently exhibited as the typical robust control
controllers, active disturbance rejection control, and integral scheme’s alternative. One of the disadvantages, aside from the
reconstructor-based sliding mode control, in the context of SISO
second-order, perturbed, pure integration systems. This is the associated chattering responses, lies in the need to measure,
prevailing paradigmatic class of differentially flat systems or for sliding surface synthesis, the entire state vector of the
feedback linearizable systems. The equivalence is valid beyond system [5], [6]. To circumvent this need, state observers have
the second-order pure integration systems. However, PID con- been proposed and extensively used, in spite of the lack of
trollers of such plants do not make much sense without imposing a clear-cut separation principle, for nonlinear systems [6].
assumptions that will move the considerations out of the pure
integration systems case. The equivalence among the rest of the In [7], an SMC scheme is presented, which avoids the need
controllers is valid for any finite order pure integration system. to explicitly use observers for the state vector. An integral
reconstructor-based sliding surface synthesis procedure is pro-
Index Terms— Active disturbance rejection control (ADRC),
extended state observers (ESOs), flat filters, flatness, proportional posed for the simplified input-to-flat output model of the feed-
integral derivative (PID) control, sliding mode control (SMC). back linearizable nonlinear systems. The scheme is naturally
related to the delta–sigma modulation implementation of an
I. I NTRODUCTION average designed, classical, feedback controller.
A research trend in robust control techniques based on
T HE proportional integral (PI) and PI derivative (PID)
controllers are, no doubt, the simplest and most ubiqui-
tous control algorithms extensively used nowadays in industry
disturbance observers is represented by the active disturbance
rejection control (ADRC) methodology (see [4], [8]–[10]).
ADRC is based on the possibility of online estimating, via
and academic labs. These controllers are used in a wide
a suitable observer, the combined effects of endogenous and
variety of instances, including motor drives, magnetic and
exogenous disturbances affecting the plant behavior. The pre-
optic memories, automotive components, flight control sys-
scribed control law proceeds to approximately cancel the plant
tems, robotics, and instrumentation (see [1], [2]). With the
disturbances via a direct use of the observed disturbance [11].
advancement in technology and control theory, a vast variety
Typically, a linear extended state observer (ESO) of the
of schemes have been developed to deal with the PID tuning
Luenberger type is specified on the basis of the simplified
tasks and to efficiently improve the controllers’ performance
version of the plant. This consists of a perturbed chain of
(see [3]). Robust control, in general, and disturbance rejection
integrations with known or unknown control input gain [12].
techniques (see [4]) have become powerful alternatives in
An alternative algorithm that circumvents the need for
industrial and academic settings without the benefit of a well-
asymptotic ESOs is represented by the flat filtering con-
established relation with the PID control schemes. Often,
trol (FFC) schemes. Such technique constitutes a slightly
comparisons are demanded and reported between every new
modified, robust, version of generalized PI (GPI) controllers,
robust controller design technique and PID control.
introduced in [13]. The linear FFC effectively attenuates the
Manuscript received March 22, 2019; accepted May 15, 2019. Manuscript effects of total additive disturbance inputs, arising from the
received in final form May 24, 2019. This work was supported by Conacyt, neglected state-dependent nonlinearities, un-modeled dynam-
CVU under Contract 270591. Recommended by Associate Editor X. Chen.
(Corresponding author: Eric William Zurita-Bustamante.) ics, and external unknown perturbations (see [14], [15]).
H. Sira-Ramírez and E. W. Zurita-Bustamante are with the Department All the seemingly unrelated controllers described earlier
of Electrical Engineering, Mechatronics Section, CINVESTAV-IPN, Mexico enjoy an underlying equivalence in the context of the output
City 07360, Mexico (e-mail: [email protected]; [email protected]).
C. Huang is with the School of Control and Computer Engineering, feedback control of perturbed chains of integrators. This equiv-
North China Electric Power University, Beijing 102206, China (e-mail: alence is quite clear from the frequency domain viewpoint.
[email protected]). In this paper, we establish the equivalence among dynami-
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. cal feedback controllers synthesized via: 1) flat filters [11];
Digital Object Identifier 10.1109/TCST.2019.2919822 2) reduced-order ESO (ROESO)-based ADRC; and 3) integral
1063-6536 © 2019 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.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

reconstructor-based sliding mode controllers. For the particular

case of second-order plants, the equivalence includes PID
controllers with the derivative control part computed via
the popular “dirty derivative” approximation (see [16]). The
equivalence is, thus, relevant for controlling the interrelated
sets of second-order systems, such as mechanical systems,
robotic systems, electromechanical systems, and some power
electronics systems.
Section II presents a brief revisiting of classical con-
trol systems, in terms of transfer functions block diagrams,
which are pertinent to the main developments in this paper. Fig. 1. Control system block diagram.
Section III explains, in the context of second-order perturbed
pure plants, the flat filters as a natural outcome of the integral
reconstructor-based feedback control for linear pure integra- eu (s) = u(s) − u ∗ (s) (4)
tion systems. The stability issues as well as the disturbance y(s) = G(s)(u(s) + ξ(s)). (5)
attenuation features are examined in a frequency domain
setting, including sensitivity, complementary sensitivity, and One obtains
   
open-loop transfer functions. In Section IV, we continue to G(s) 1
e y (s) = ξ(s) + η(s). (6)
prove that the “dirty derivative”-based PID controller yields 1 + G(s)C(s) 1 + G(s)C(s)
the same closed-loop transfer function than the flat filter con-
troller. Section V demonstrates that a ROESO-based ADRC The transfer function 1/(1 + G(s)C(s)) is called the sensi-
yields precisely the same closed-loop transfer function than tivity function. It is denoted by S(s), and it plays a major role
the studied FF- and PID-based controllers. Section VI con- in the frequency domain robust control considerations.
siders the integral reconstructor-based sliding surface coor- From e y (s) = y(s) − y ∗ (s) + η(s), one also has
   
dinate function for the SMC of pure integration, switched, ∗ G(s) G(s)C(s)
systems. The scheme is found to be equivalent to that of an y(s) = y (s)+ ξ(s)− η(s).
1 + G(s)C(s) 1+G(s)C(s)
FFC implemented through a delta–sigma modulator. On the (7)
average, the studied SMC scheme produces the same closed-
loop transfer function as the FF, PID, and ADRC controllers. The transfer function G(s)C(s)/(1+G(s)C(s)) is addressed
Section VII is devoted to the experimental implementation of as the complementary sensitivity function, denoted by T (s).
the four control schemes for an output reference trajectory Its name stems from the fact that S(s) + T (s) = 1 for all s.
tracking control task on a nonlinear system constituted by The open-loop transfer function is, simply, the product:
a “boost” switched regulated power converter. Section VIII G(s)C(s). The transfer function between the disturbance
presents the conclusions of this paper and indicates the possi- input, ξ(s), and the output tracking error, e y (s), is defined as
ble avenues for the development of the equivalence to the class G(s)/(1+G(s)C(s)) and addressed as the (input) disturbance-
of MIMO static or dynamically feedback linearizable systems. driven closed-loop transfer function. Clearly, e y (s) =
Appendix A contains generalities about average pulsewidth [G(s)/(1 + G(s)C(s))]ξ(s) whenever η(s) is either zero or
modulation (PWM)-controlled systems, while Appendix B not considered.
contains all the generalizations of the results found for second-
order plants to the case of n-D, perturbed, pure integration III. F LAT F ILTERS
systems.
Consider the perturbed second-order system

II. C LASSICAL F REQUENCY D OMAIN C ONSIDERATIONS ÿ = γ u + ξ(t) (8)

γ
We refer our considerations to the following control system G(s) = 2 , η(t) = 0 (9)
s
block diagram (see Fig. 1).
Let η denote the output measurement noise that, typically, where ξ(t) is a smooth, unknown, exponentially integrable,
is of high-frequency nature. As above, ξ represents the input time-varying, signal and γ is a nonzero scalar gain. Suppose
disturbance, regarded as a low-frequency signal, and y ∗ (t) it is desired to track a, given, smooth output reference tra-
is the output reference trajectory, considered to be a low- jectory y ∗ (t). Define u ∗ (t), the nominal control input for the
frequency signal. The nominal input, u ∗ , satisfies the nominal perturbation-free system, as u ∗ (t) = ÿ ∗ (t)/γ . Define, also,
input–output relation y ∗ = G(s)u ∗ . In the classical tradition the output trajectory tracking error as e y = y − y ∗ (t). Note
of transfer function algebra, one gets the following basic that s 2 e y (s) = γ eu (s) + ξ(s).
definitions and relations in the frequency domain: Let {k3 , k1 , k1 , k0 } be a set of constant design parameters.
The following feedback law, written in the frequency domain,
ym (s) = y(s) + η(s) (1) for the input error eu = u − u ∗ (t):
e y (s) = ym (s) − y ∗ (s) = y(s) − y ∗ (s) + η(s) (2)  
1 k2 s 2 + k1 s + k0
eu (s) = − e y (s) = −C(s)e y (s) (10)
eu (s) = −C(s)e y (3) γ s(s + k3 )
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 3

is a classical compensation network, addressed here as a flat transfer function, G(s)/(1 + G(s)C(s)), acting on the distur-
filter (see, also, [11, Appendix B]). By direct substitution of bance signal ξ(s)
(10) in the perturbed tracking error plant, the disturbance-  
s(s + k3 )
driven closed-loop transfer function is found to be e y (s) = 4 ξ(s). (17)
s + k3 s 3 + k2 s 2 + k1 s + k0
 
s(s + k3 ) The design coefficients in the characteristic polynomial of the
e y (s) = 4 ξ(s). (11)
s + k3 s 3 + k2 s 2 + k1 s + k0 closed-loop system
Clearly, the set of ki s is primarily determined on the basis pcl (s) = s 4 + k3 s 3 + k2 s 2 + k1 s + k0 (18)
of a stable characteristic polynomial. The design parameters
are to be chosen so that pcl (s) is a Hurwitz polynomial.
to attenuate the effects of the disturbance input, ξ(·), on the
The coefficients may be obtained by the direct comparison
output tracking error, e y (·), are discussed at the end of this
with the coefficients of a classical desired stable fourth-order
section.
polynomial pd (s) with 0 < ζ ≤ 1, ωn > 0
Flat filters are intuitively valid. They are, also, naturally
2
derived from considerations, including integral reconstructors p(s) = s 2 + 2ζ ωn s + ωn2
(see [13]), consisting of structural estimates of the pure
= s 4 +4ζ ωn s 3 + 4ωn2 ζ 2 +2ωn2 s 2 +4ωn3 ζ s +ωn4 . (19)
integration system phase variables under disturbance free
conditions. We now derive the flat filter (10) from that context. Alternatively, the Routh–Hurwitz stability criterion, applied
Thus, the first-order time derivative of the tracking error to the closed-loop characteristic polynomial pcl (s), yields the
satisfies following conditions on the flat filter parameters:
 t
k3 >0
ė y = γ eu (ρ)dρ + ė y (0) =: 
ė y + ė y (0). (12) k3 k2 − k1
0 >0
t k3
We address the quantity  ė y := γ 0 eu (ρ)dρ as the integral k3 k2 k1 − k12 + k32 k0
reconstructor of ė y . As an estimate of ė y , it exhibits a constant >0
error dependent upon the unknown initial condition ė y (0). k3 k2 − k1
k0 > 0.
One may use integral reconstructors on any linear state-
based feedback control scheme, provided that the constant These conditions guarantee that all the polynomial roots are
error is compensated via integral action of the output tracking located in the left half of the complex plane. Typically,
error. the disturbance-driven closed-loop transfer function exhibits
Then, consider the integral reconstructor-based feedback large attenuation both at very low and high frequencies with
controller, where an additional iterated integral control action a maximum amplitude at the bandwidth frequency. In order
is provided in order to gain some possible robustness when to diminish this maximum amplitude, a high gain factor is
the unknown additive input disturbance effect is present introduced in the coefficients of transfer function in the form
 t of a small parameter, denoted by . As this parameter is

γ eu = −k3 ė y − k2 e y − k1 e y (ρ)dρ made smaller, the mid-frequency attenuation becomes larger
0 in absolute value, while the bandwidth is enlarged to include
 t ρ
−k0 e y (ρ1 )dρ1 dρ. (13) higher frequencies. One considers
0 0 ⎡  ⎤
k3
G(s) s s + 
The controller is rewritten in an implicit manner for the control =⎣ ⎦ . (20)
error eu as 1 + G(s)C(s) s 4 + k3 s 3 + k22 s 2 + k13 s + k04
   
  t 
The enhanced attenuating effects of the parameter  on the
γ eu + k 3 eu (ρ)dρ disturbance-driven closed-loop transfer function are depicted
 0  t  t ρ  in the Bode diagram shown in Fig. 2.
= − k2 e y + k1 e y (ρ)dρ + k0 e y (ρ1 )dρ1 dρ . In the Bode diagram shown in 2, the disturbance-driven
0 0 0
closed-loop transfer function
(14)  
G(s) s(s + k3 )
In Laplace transform terms, we have = 4 (21)
1 + G(s)C(s) s + k3 s 3 + k2 s 2 + k1 s + k0
   
k3 k1 k0 exhibits, at low frequencies, an asymptote of slope of
γ 1+ eu (s) = − k2 + + 2 e y (s) (15) +20 dB/dec, while at high frequencies, it exhibits an asymp-
s s s
tote decreasing with a slope of −40 dB/dec. This implies
that is significant attenuation of the low-frequency input disturbance,
 
1 k2 s 2 + k1 s + k0 ξ(s), and good high-frequency measurement noise rejection
eu = − e y (s). (16) with a good low-frequency output reference tracking features.
γ s(s + k3 )
At intermediate frequencies, the smaller the factor ,
The disturbance-driven closed-loop system results in a the larger the bandwith, and the higher the attenuation of the
tracking error determined by the attenuating features of the disturbance input around the filter bandwidth frequency.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

IV. D IRTY D ERIVATIVE -BASED PID C ONTROL OF

S ECOND -O RDER P URE I NTEGRATION S YSTEM
A dirty derivative-based PID controller for the tracking error
dynamics
ë y = γ eu + ξ(t) (22)
is given, under perfect measurements [η(t) = 0], by
 
1 s kI
eu (s) = − kD + kP + e y (s). (23)
γ s+α s
A straightforward manipulation reveals that the controller may
also be expressed as a flat filter
 
1 (k P +k D )s 2 +(αk P +k I )s +αk I
eu (s) = − e y (s) (24)
γ s(s + α)
with
Fig. 2. Typical closed-loop frequency response for FF-controlled double
integrator. α = k3 , k P + k D = k2 , αk P + k I = k1 , αk I = k0 .
Given a dirty derivative-based PID controller design, an FF
controller is readily computed from the previous identification.
A term-by-term identification of the closed-loop characteris-
tic polynomial with those of the preferred fourth-order Hurwitz
polynomial pd (s) = (s 2 + 2ζ ωn s + ωn2 )2 , 0 < ζ ≤ 1, ωn > 0,
yields the following stable closed-loop PID controller design:
ωn3 (16ζ 2 − 1)ωn2
kI = , kP =
4ζ 16ζ 2
 2
ωn
kD = 2ζ ωn + , α = 4ζ ωn .
4ζ
On the other hand, given an FF controller of the form
 
1 k2 s 2 + k1 s + k0
eu = − ey . (25)
Fig. 3. Frequency response of the open-loop transfer function for FF and γ s(s + k3 )
double integrator.
An equivalent dirty derivative-based PID controller is obtained
as
k0 k1 k3 −k0 k3 (k2 k3 −k1 )+k0
kI = , k P = 2
, kD = , α = k3 .
k3 k3 k32
In both cases, the disturbance-driven closed-loop system is
e y (s)
 
s(s + α)
= 4 ξ(s)
s + αs 3 + (k P + k D )s 2 + (αk P + k I )s + αk I
 
s(s + k3 )
= 4 ξ(s). (26)
s + k3 s 3 + k2 s 2 + k1 s + k0

V. ADRC S CHEME
Consider the state representation of the second-order output
Fig. 4. Sensitivity transfer function for the FF and double integrator. reference tracking error plant, ë y = γ eu + ξ(t), with tracking
error, e y = y − y ∗ (t) (i.e., η(t) = 0)
χ̇1 = χ2 (27)
The open-loop transfer function corroborates large magni-
χ̇2 = γ eu + ξ(t) (28)
tude at low frequencies, thus indicating good input disturbance
rejection and good low-frequency reference trajectory tracking e y = χ1 . (29)
features; this performance is shown in Fig. 3. A ROESO is obtained on the basis of the “fake measurement”
The sensitivity transfer function Bode plot (see Fig. 4) of the tracking error time derivative ė y = χ2 = χ̇1
reveals good output measurement noise rejection at low
frequencies. χ̂˙ 2 = γ eu + z + λ1 (χ2 − χ̂2 ) (30)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 5

ż = λ0 (χ2 − χ̂2 ). (31) we have

With the following output-dependent, state and distur- 1
eu = − [γ1 (ė y − e2 ) + γ0 e y + ξ(t) − eξ ]. (51)
bance coordinate transformation, the reduced-order extended γ
observer is easily implementable without the fake measure- One obtains, after the substitution of the derived controller in
ment χ2 : the tracking error dynamics, ë y = γ eu + ξ(t)
ζ̂2 = χ̂2 − λ1 e y (32) ë y + γ1 ė y + γ0 e y = γ1 e2 + eξ . (52)
η̂ = z − λ0 e y . (33)
Reverting to the frequency domain, and using the previously
One obtains found expressions for e2 (s) and eξ (s), one finds that
 
ζ̂˙2 = γ eu + η̂ − λ1 ζ̂2 + λ0 − λ21 e y (34) s(s + γ1 + λ1 )
e y (s) = ξ(s). (53)
η̂˙ = −λ0 ζ̂2 − λ1 λ0 e y (35) (s 2 + λ1 s + λ0 )(s 2 + γ1 s + γ0 )
with the tracking error phase variable estimate and input Upon the identifications
disturbance estimate, given by
γ1 + λ1 = κ3
χ̂2 = ζ̂2 + λ1 e y (36) γ0 + γ1 λ1 + λ0 = κ2
z = η̂ + λ0 e y . (37) γ0 λ1 + γ1 λ0 = κ1
A similar state and disturbance coordinate transformation on γ0 λ0 = κ0 . (54)
the state and disturbance variables of the (reduced) tracking
The ROESO-based ADRC disturbance-driven closed-loop
error system
transfer function is found to be
ζ2 = χ2 − λ1 e y (38)  
s(s + κ3 )
η = ξ(t) − λ0 e y (39) e y (s) = 4 ξ(s). (55)
s + κ3 s 3 + κ2 s 2 + κ1 s + κ0
produces a system description of the form Equation (54) is exactly the same disturbance-driven closed-
ζ̇2 = γ eu + η − λ1 ζ2 + λ0 − λ21 e y (40) loop transfer function found for the flat filter-controlled
second-order, perturbed, integrator system. It is also coincident
η̇ = ξ̇ (t) − λ0 ζ2 − λ1 λ0 e y . (41) with that obtained by the dirty derivative-based PID controller
The state and the disturbance estimation errors satisfy acting on the same plant system. Notice that given a set
of stable coefficients, K = {κ0 , κ1 , κ2 , κ3 }, there may be
ζ2 − ζ̂2 = (χ2 − λ1 e y ) − (χ̂2 − λ1 e y ) multiple solutions for the corresponding sets = {γ0 , γ1 } and
= χ2 − χ̂2 =: e2 (42)  = {λ0 , λ1 }. The map in (54) is only locally invertible, pro-
η − η̂ = (ξ(t) − λ0 e y ) − (z − λ0 e y ) vided that the sets and  are not equal. This prompts high-
gain ROESOs and a moderate gain phase variable feedback
= ξ(t) − z =: eξ . (43)
gains controller. Equation (53) indicates how to proceed, via
The dynamics of the estimation error and the disturbance stable factorization, to obtain an equivalent stable ROESO-
estimation error is given by based ADRC or a stable “dirty derivative” PID controller from
a given FFC.
ė2 = eξ − λ1 e2 (44)
ėξ = ξ̇ − λ0 e2 (45)
VI. I NTEGRAL R ECONSTRUCTOR -BASED
that is S LIDING M ODE C ONTROLLER
ë2 + λ1 ė2 + λ0 e2 = ξ̇ . (46) Consider the switched controlled second-order, perturbed,
integrator with no output noise
In the frequency domain, we write
  ÿ = u + ξ(t), u ∈ {0, 1}. (56)
s
e2 (s) = 2 ξ(s) (47)
s + λ1 s + λ0 The tracking error e y = y − y ∗ (t) evolves in accordance with
and from the fact that eξ = ė2 + λ1 e2 , we have
  ë y = eu + ξ(t). eu = u − u ∗ (t). (57)
s(s + λ1 )
eξ (s) = se2 (s) + λ1 e2 (s) = 2 ξ(s). (48) The integral reconstructor of the output tracking error time
s + λ1 s + λ0
derivative, ė y , for the unperturbed, pure integration system,
The ROESO-based ADRC controller is then given by ë y = eu , u ∈ {0, 1}, is given by
1  t
eu = − [γ1 χ̂2 + γ0 χ1 + z]. (49)
γ ė y = eu (ρ)dρ. (58)
0
Using the facts that
Clearly, the estimate is off by an unknown initial condition
z = ξ − eξ , χ̂2 = 
ė y = ė y − e2 (50)
ė y = 
ė y + e˙y (0). (59)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

We propose the following integral reconstructor-based slid- feedback control error. This control error is addressed as the
ing surface coordinate function: equivalent control error, denoted by eu,eq , and defined as
           
(2) (3) (2)

σ̂ = ė y +λ3 e y +λ2 e y +λ1 e y +λ0 ey eu,eq +λ3 eu,eq +λ2 e y +λ1 e y +λ0 e y = 0.

(60) (66)
where the following notation is being used: The equivalent control error is, thus, implicitly obtained from
   t  ρ1  ρ j −1 the relation
( j)
φ := ··· φ(ρ j )dρ j · · · dρ1 . (61)       
(2)
0 0 0
eu,eq +λ3 eu,eq = − λ2 e y +λ1 e y +λ0 ey
The linear combination of the iterated output integrations
will compensate, in sliding mode closed loop, the constant (67)
ˆ and it will exhibit further integral action
estimation error in ẏ, which is readily solvable in the frequency domain
feedback in the light of the effects of the unknown disturbance  
input signal ξ(t). λ2 s 2 + λ1 s + λ0
eu,eq (s) = u eq (s) − u ∗ (s) = − e y (s)
The estimated and sufficiently compensated sliding surface s(s + λ3 )
coordinate function, σ̂ , is then given by (68)
   
σ̂ = eu + λ3 e y + λ2 ey i.e., the equivalent control input is also described as a classical
    controller by
(2) (3)  
+ λ1 e y + λ0 ey . (62) λ2 s 2 + λ1 s + λ0
u eq (s) = u ∗ (s) − e y (s). (69)
s(s + λ3 )
The time derivative of σ̂ is readily found to be The equivalent control, u eq , is the virtual average smooth
    feedback control law responsible for ideally maintaining the
d σ̂ (2)
= u −u ∗ (t)+λ3 ė y +λ2 e y +λ1 e y +λ0 ey . sliding motion on σ̂ = 0 when the motions start, precisely,
dt on the zero level set of the sliding surface coordinate function.
(63) In the perturbed case, ë y = eu + ξ(t), the perturbed
system dynamics satisfies, under closed-loop, ideal sliding
The necessary and sufficient condition for the existence of
mode conditions
a sliding regime on the zero level set, σ̂ = 0, is given by
σ̂ (d/dt)σ̂ < 0. This yields the switching control law ë y = eu,eq + ξ(t). (70)
1
u = (1 − sign σ̂ ), σ̂ = 0. (64) Substituting the average controller expression on the ideal
2 sliding mode dynamics, one obtains, in the frequency domain
A sliding regime locally exists on σ̂ = 0 if and only if σ̂ σ̂˙ < 0  
λ2 s 2 + λ1 s + λ0
for σ̂ = 0 s 2 e y (s) = − e y (s) + ξ(s) (71)
     s(s + λ3 )
(2)
∗
0 < u (t)− λ3 ėy +λ2 e y +λ1 e y +λ0 ey < 1. that is
 
s(s + λ3 )
e y (s) = 4 ξ(s). (72)
Under ideal sliding conditions, the invariance conditions, s + λ3 s 3 + λ2 s 2 + λ1 s + λ0
σ̂ = 0, (d/dt)σ̂ = 0, are satisfied. From (d σ̂ /dt) = 0, one
The disturbance-driven closed-loop transfer function induced
obtains
     on the average by the SMC scheme is the same as those
(2)
obtained with the FF [see (17)], the dirty derivative PID
u eq = u ∗ (t) − λ3 ė y + λ2 e y + λ1 e y + λ0 ey .
[see (26)], and the ADRC control schemes [see (55)] for the
(65) second-order perturbed plant.
Notice that the expression for d(σ̂ )/dt also leads to the
This control is called the equivalent control input. It plays following relation:
an important role in the (average) characterization of a sliding
d σ̂ 1
regime. In fact, the previously found existence condition is = u − u av = (1 − sign σ̂ ) − u av (73)
also interpreted on σ̂ = 0, as dt 2
where u av , given in terms of the time derivative of the tracking
0 < u eq < 1. error, is simply
    
Using, once again, the estimate of the output time derivative, (2)
∗
ẏ, as its faulty but dully compensated integral reconstructor, u av = u (t) − λ3 ė y + λ2 e y + λ1 e y + λ0 ey .

ė y = ( eu ), one redefines, on σ̂ = 0, the virtual smooth
(74)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 7

Fig. 5. Delta–sigma modulator.

Note that on σ̂ = 0, the average control input, u av , coincides

with the implicitly described equivalent control, u eq , obtained
Fig. 6. Block diagram of the flat filter controller with – modulator.
from the equivalent control input error, eu , as explained earlier.
Define the scalar variable e as e := −σ̂ . From (73), we have
the following relation:
The second-order state average dynamics of the flat output
de 1 is simplified to a pure, perturbed, second-order integration
= u av − u, u = (1 + sign e). (75)
dt 2 system with a nonlinear, state- and disturbance-dependent,
Equation (75) precisely describes the popular delta–sigma gain which is measurable and only approximately cancellable
modulator (see Fig. 5) fed by the average control input u av and in the feedback scheme. The disturbances are represented by
producing, as an output, the switched signal u whose average the neglected nonlinearities and the effects of an unknown
coincides with the provided control input. The delta–sigma current demand term, taken here to be of discontinuous nature.
modulator is a translator, taking a continuous input signal u av We experimentally test the performance of the three first
into a binary-valued signal u = (1/2)(1 + signe) ∈ {0, 1}. alluded controllers and find their responses to be identical.
The delta–sigma modulator is a 1-D device that implements These controllers are implemented via a static PWM scheme,
average designed (continuous) control inputs on a switched directly taking the average FF generated control input signal
input system. The integral reconstructed sliding surface coordi- as the required duty ratio signal (see Appendix A). The sliding
nate function σ̂ is, thus, reduced to the state e of an exogenous mode controller, provided with an integral reconstructed slid-
analog-to-binary converter. This error is known as the coding ing surface, is directly implemented from a flat filter cascaded
error. with a delta–sigma modulator. The average dynamics, under
Indeed, a sliding regime exists on e = 0 on its own sliding mode conditions, coincides with those obtained from
right, and moreover, the input magnitude limitations demanded the three previously tested dynamic feedback controllers.
by the delta–sigma modulator precisely coincide with the Consider the following load perturbed model of a boost
existence conditions demanded on the integral reconstructed converter with measurable states i, v [17]:
sliding surface coordinate function.
Consider the sliding mode existence condition eė < 0 on di
L = −uv + E, u ∈ {0, 1} (77)
e=0 dt
  dv v
1 C = ui − − I (t) (78)
eė = e u av − (1 + sign(e)) . (76) dt R
2
For e > 0, one has that eė = e(u av − 1) must be negative. where i is the inductor current, v is the output capacitor
Hence, u av must be less than one, i.e., u av < 1. For e < 0, voltage, E is the constant input voltage, R is the load resistor,
eė = eu av ; hence, u av > 0. A sliding regime exists on the L and C are, respectively, the inductance and capacitance.
coding error state value, e = 0, if and only if I (t) is the unknown external input disturbance acting as a
nonenergy-injecting load disturbance; I (t) > 0 ∀t, i.e., I (t)
0 < u av < 1. always extracts the energy from the converter system. u is the
The scheme shown in Fig. 6 depicts that a sliding regime switched control input.
based on the integrally reconstructed sliding surface coordinate A flat output for the unperturbed system is given by the
function for the pure integration system is actually constituted normalized total stored energy
by an FF- or ADRC-based average controller implemented
1
through a delta–sigma modulator. F= (Li 2 + Cv 2 ). (79)
2
VII. B OOST C ONVERTER E XAMPLE Indeed
To emphasize that the obtained results are relevant in ⎡ ⎤
1  
the control of nonlinear SISO feedback uncertain feedback (Li 2 + Cv 2 )
⎢ ⎥ F
linearizable (i.e., flat) SISO systems, we take the case of a (i, v) = ⎣ 2 ⎦ = (80)
v 2
Ḟ
popular switched “boost” power converter. The average state Ei −
R
model of this converter is a nonlinear system. Consider its
flat output (the average total stored energy) as the output to is a local diffeomorphism used in the feedback linearization
be regulated in an output reference trajectory tracking setting. of the unperturbed system.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

The flat output F, defined by (79), satisfies the following TABLE I

perturbed relation: G AIN PARAMETERS FOR THE H URWITZ P OLYNOMIALS
 
d2 F E 2i v i I (t) E2 2v 2
2
= −u v + + + + 2
dt L RC C L R C
   
perturbed input gain endogenous

2v I 2 (t) ν I (t) d I (t)

+ I (t) + + −v (81) were implemented into a DSP F28335 controller card. The
RC
 C  RC dt 
sample time period of the card was set to be 1 μs. It is
exogenous
desired to have F(t) track a given, smooth, output reference
= −γ (i, v, I (τ ))u + ξ(i, v, I (t)). (82) trajectory F ∗ (τ ). The controllers for each one of the schemes
The scalar nonlinear function, γ (i, v, I (t)), is a disturbance- are expressed as
dependent input gain and, hence, not available for use in  
1 (γ0 +γ1 λ1 +λ0 )s 2 +(γ0 λ1 +γ1 λ0 )s +γ0 λ0
the controllers. It turns out that this input gain is strictly eu (s) = − e y (s)
γ s(s +(γ1 +λ1 ))
positive around any feasible prescribed flat output trajectory
AD RC (83)
[i (t), v(t), I (t) > 0]. A state- and disturbance-dependent time  
1 k2 s 2 +k1s +k0
coordinate transformation based on this nonlinear scalar input eu (s) = − e y (s), F F (84)
gain function, γ (i, v, I (t)), yields a unit input gain time-scaled γ s(s +k3 )
 
relation for the flat output, i.e., defining 1 (k P +k D )s 2 +(αk P +k I )s +αk I
eu (s) = − e y (s)
γ s(s +α)
dτ 2 = γ (i, v, I (t))dt 2 .
P I D (85)
The (Lie–Bäcklund) time-scaled system is of the form
where the relations among the controller gains are given by
d2 F
= −u + ξ(τ ) k3 = γ1 + λ1 , k2 = γ0 + γ1 λ1 + λ0 , k1 = γ0 λ1 + γ1 λ0
dτ 2
k0 = γ0 λ0
where ξ(τ ) is the quotient of the unknown total disturbance k0 k1 k3 − k0 k3 (k2 k3 − k1 ) + k0
and the strictly positive disturbance-dependent input gain γ (·). kI = ; k P = , kD =
k3 k32 k32
For a given feedback control input u, the stability features
of the closed-loop, scaled, system are exactly the same as α = k3 . (86)
those of the unscaled system. The corresponding flat output In the ADRC scheme, the characteristic polynomial pc1 (s) for
trajectories are, just, homotopic deformations of each other. the ROESO and for the corresponding feedback controller pc2
For this reason, the uncertain control input gain, γ (i, v, I (t)), are specified by
may be, simply, ignored.
It is desired to have the flat output, F(t), track a given, pc1 (s) = s 2 + λ1 s + λ0 (87)
smooth, output reference trajectory, F ∗ (t), which is character- pc2 (s) = s 2 + γ1 s + γ0 . (88)
ized by rest-to-rest (equilibrium) maneuvers, linked by smooth
The polynomials pc1 and pc2 guarantee the stability of
interpolating time-polynomial functions, say, of the Bèzier
the unperturbed closed-loop second-order integration system.
type.
For the perturbed system, sufficiently large magnitude roots,
The ADRC, FF, and PID controllers will take the synthe-
located in the left half of the complex plane, establish the
sized control input u as a duty ratio function to be implemented
convergence of the flat output, F, toward a small vicinity of
via a high sampling frequency, PWM transducer (static analog-
the desired output reference trajectory F ∗ (t). The polynomials
to-binary-valued converter) (see the Appendices A and B).
pc1 and pc2 are, therefore, chosen as the Hurwitz polynomials
However, the integral reconstructor-based sliding surface cor-
of the form, pc1 = (s 2 + 2ζ1 ω1 s + ω12 ) and pc2 = (s 2 +
responding to the previously proposed sliding mode controller
2ζ2 ω2 s + ω22 ), where
will be implemented as an FF controller in combination with
a delta–sigma modulator (dynamic analog-to-binary-valued λ1 = 2ζ1 ω1 ; λ0 = ω12 ; γ1 = 2ζ2 ω2 ; γ0 = ω22 .
converter), as explained in Section VI.
For the experimental setup, the gains, γ1 , γ0 , λ1 , and
λ0 , along with the equivalence relations (86) were used to
A. Experimental Setup demonstrate the equivalence among the different controllers.
An experimental platform was built to illustrate the equiva- The values of the Hurwitz polynomials are shown in Table I.
lence between the ROESO-based active disturbance rejection A rest-to-rest reference trajectory tracking task was spec-
controller, the dirty derivative-based PID controller, the flat ified, which took the total stored energy from the initial
filter controller, and the integral reconstructor-based sliding value F1 = 0.1859 [w] toward a second equilibrium value
mode controller. A boost converter was designed with the F2 = 0.6338 [w] during the time interval [t1 , t2 ] = [7, 10]
following parameters, L = 3 mH, C = 2200 μF, R = 150 , [s]. The flat output was kept at this equilibrium value for a
and E = 12 V. The PWM (20 kHz), the Bèzier polynomial, few seconds. Then, the stored energy F was taken toward
and the feedback control laws (ADRC, FF, PID, and SMC) a new final equilibrium value F3 = 0.4401 [w] during the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 9

ADRC based on a ROESO, and, more strikingly, average

closed-loop behavior of SMC with sliding manifolds based
on a suitable, stable, linear combinations of tracking velocity
integral reconstructor and a sufficient number of iterated output
tracking error integral compensation. The underlying setting is
that of the frequency domain. The controllers are applicable
to nonlinear second-order systems. All these controllers were
shown to have exactly the same disturbance-driven closed-
loop transfer function, excited by the unknown disturbance
input. The rest of the classical transfer functions, such as
open-loop, sensitivity, and complementary sensitivity transfer
function, are all equal. The equivalence goes beyond the
second-order system, at least, among flat filters, ADRC, and
SMC. PID controllers for high-order, pure integration, systems
do not make sense. An illustration of this equivalence has been
presented in terms of the control of a perturbed nonlinear boost
converter circuit viewed as a perturbed second-order integrator
with an unknown input gain.

Fig. 7. Trajectory tracking for flat output in the boost converter. A PPENDIX A
AVERAGE R ESPONSE OF PWM-C ONTROLLED S YSTEMS
Consider the nonlinear SISO switched system
ẋ = f (x) + g(x)u, x ∈ Rm , u ∈ {0, 1} (89)
y = h(x) ∈ R (90)
with f and g as smooth vector fields and controlled, for output
reference trajectory tracking purposes, by means of a PWM
implementation of a smooth dynamical nonlinear feedback
controller generating the duty ratio function, μ(·), as a smooth
function of the output tracking error e y = y − y ∗ (t) and of the
state z of the dynamical controller, i.e., μ(t) = μ(e y (t), z(t)).
We assume that provisions have been taken to have
μ(t) ∈ (0, 1) without saturations. The PWM strategy is based
upon the following sampled process:
μ = μ(e y , z), ż = ϕ(z, e y ), z ∈ Rq (91)
and

1, for tk ≤ t < tk + μ(e y (tk ), z(tk ))T
u=
Fig. 8. Average input control response for the boost converter. 0, for tk + μ(e y (tk ), z(tk ))T ≤ t < tk + T (= tk+1 ).
The closed-loop controlled system is described by
interval [t2 , t3 ] = [12, 14]. The performance of the several 
f (x) + g(x), for tk ≤ t < tk + μ(e y (tk ), z(tk ))T
controllers is shown in Fig. 7 for these maneuvers. Approx- ẋ =
imately, at the time instants t = 15.4 [s] and t = 17.4 [s], f (x), for tk + μ(e y (tk ), z(tk ))T ≤ t < tk + T.
an external and unknown switched static load was connected (92)
at the converter’s output terminals. The load was emulated
The sampled system may be exactly described, in an implicit
with a resistor in parallel with a second switched resistor. For
manner, by
all the controllers, the total energy has a remarkably coincident
 tk +μ(tk )T
tracking performance (see Fig. 8).
x(tk+1 ) = x(tk ) + [ f (x(τ )) + g(x(τ ))] dτ
tk
VIII. C ONCLUSION  tk +T
In the context of the perturbed, pure integration systems, + f (x(τ ))dτ (93)
tk +μ(tk )T
a complete equivalence has been established among four seem-  tk +T  tk +μ(tk )T
ingly unrelated control strategies, namely, flat filters, as derived = x(tk ) + f (x(τ ))dτ + g(x(τ ))dτ
from GPI controllers, traditional PID control based on the tk tk
so-called “dirty derivative” high-pass filter approximation, (94)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

where μ(tk ) stands for μ(e y (tk ), z(tk )). We define the average by time-polynomials whose coefficients are dependent on the
PWM system as that obtained from the idealized closed- initial conditions. Thus
 
loop responses when the sampling frequency, 1/T , approaches (i)
 (n−i)
infinity or the sampling period, T , is ideally shrunk to zero. ey =γ eu + p(O(i − 1))
We have
x(tk + T ) − x(tk ) where p(O(i − 1)) stands for a generic time-polynomial of
ẋ(t) = lim order, at most, i − 1, with a constant parameter being trivially
T →0 T
tk →t identified with a time-polynomial of order zero, i.e.,
 tk +T
1 p(O(i − 1)) = α0 + α1 t + · · · + αi−1 t i−1 .
= lim f (x(τ ))dτ
T →0 T tk
tk →t The phase variables’ estimates may then be used in the
 tk +μ(tk )T design of any state feedback controller based on the phase
1
+ lim μ(tk )g(x(τ ))dτ (95) variables, provided that a suitable linear combination of output
T →0 μ(tk )T tk
tk →t error iterated integrals is added to the controller expression.
= f (x(t)) + μ(t)g(x(t)) Since the highest order time-polynomial estimation error
= f (x(t)) + [μ(e y (t), z(t))]g(x(t)). (96) is ascribed to ėy , it will be required, at least, n − 1 iterated
integrals of the output error in the controller. However, to have
The average PWM-controlled system trajectories are those that an extra degree of compensation, we propose a combination,
are ideally generated by the duty ratio as if it were a smooth including a maximum of n iterated output tracking error
control input. The duty ratio has a similar interpretation as the integrals. We
 then propose 
equivalent control in sliding mode controlled systems.  (n−1)
γ eu = − γ2n−1 e y  
+ · · · + γn+2 ÿ + γn+1 ẏ
In switched controlled systems, a static PWM strategy,
based on the smoothly generated duty ratio function, represents     
(n−1) (n)
the natural analog-to-digital conversion for the control input − γn e y + · · · + γ1 e y + γ0 ey
whose ideal features may be recovered via infinitely fast
    
sampling. This is how the ideal sliding dynamics is obtained (n−2)
on the basis of the equivalent control. A delta–sigma modu- = −γ γ2n−1 eu + · · · + γn+2 eu
lator represents a dynamical implementation of the smoothly  
generated average control input. (n−1)
+γn+1 eu
A PPENDIX B
    
E XTENSIONS TO n TH -O RDER S YSTEMS (n−1) (n)

A. Flat Filter Control for nth-Order Plant − γn e y + · · · + γ1 e y + γ0 ey .

Consider the tracking problem y → y ∗ (t) on the nth-order (97)
perturbed pure integration system y (n) = γ u+ξ(t). The output
tracking error system satisfies
From here, it is clear that eu satisfies the following implicit
e(n)
y = γ eu + ξ(t), γ > 0 description:
    
(n−2)
(e y := y − y ∗ (t), eu = u − u ∗ (t) = u − [y ∗ (t)](n) ). For the γ eu + γ2n−1 eu + · · · + γn+2 eu
derivation of the FF controller, we use the unperturbed version
 
of the system, i.e., ξ(t) = 0 ∀t, and, then, proceed to examine (n−1)
the effects of the neglected signal on the closed-loop system. +γn+1 eu
Define     
  (n−1) (n)
(n−1)
ey =γ u = − γn e y + · · · + γ1 e y + γ0 ey .
 
(2) (98)
(n−2)
ey =γ u
Resorting to the Laplace transform domain, one can solve for
.. the control error term and immediately obtain the flat filter
.   controller as  
 (n−2) 1 γn s n + · · · + γ1 s + γ0

ë y = γ u eu (s) = − e y (s).
γ s(s n−1 + γ2n−1 s n−2 + · · · + γn+1 )
  (99)
(n−1)
ėy = γ u . The closed-loop system for the perturbed system is readily
obtained upon substitution of the control input error expression
These estimates are called structural estimates or integrally in the error system description in the frequency domain
reconstructed estimates of the output error phase variables, 1 1
(n−1)
(ė y , ë y , . . . , e y ). In fact, the structural estimates are off e y = n eu + n ξ(s) (100)
s s
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 11

that is present in the structural phase variable estimates need suffi-

  cient output tracking error iterated integral compensation. One
s(s n−1 + γ2n−1 + · · · + γn+1 )
s n−2
e y (s) = ξ(s). (101) then needs to incorporate, in the sliding surface description,
s 2n + γ2n−1 s 2n−1 + · · · + γ1 s + γ0
the effect of, at least, n − 1 integrators. One has to add two
Clearly, all the frequency response features of the disturbance- more compensation terms with extra iterated integrators. One
driven closed-loop transfer function, the sensitivity function, of them is justified to obtain the effect of n − 1 iterated
and the complementary sensitivity function remain essentially integrators of e y when the sliding surface coordinate function
valid regarding good low-frequency output reference trajectory is time differentiated for obtaining the equivalent control
tracking quality, low-frequency input disturbance attenuation, error. The second one bestows the same degree of robustness
and high-frequency output measurement noise rejection. achieved before, in the FF controller, by adding a term,
including n + 1 output error iterated integrations.
The sliding surface coordinate function is then given by
B. Flat Filters and Integrally Reconstructed Sliding Mode      
Control for nth-Order Plants (2) (n−1)
σ̂ = γ eu +γ2n−1 eu +· · ·+γn+2 eu
Consider the output reference trajectory tracking problem
of a, given, smooth signal y ∗ (t) by means of an nth-order  
perturbed pure integration system +γn+1 e y + γn ey
   
y (n) = γ u + ξ(t), e(n)
y = γ eu + ξ(t), γ > 0.
(n) (n+1)
+ · · · + γ1 ey + γ0 ey .
Here, we assume, without loss of generality, that the control
input is a binary-valued signal representing a switch position. (106)
We assume, hence, that u ∈ {0, 1}. Under ideal sliding dynamics, the invariance conditions,
As, before, for the derivation of the controller, we use the σ̂ = 0 and σ̂˙ = 0, are enforced. The last condition implicitly
unperturbed tracking error system e(n) y = γ eu , and once the defines the so-called equivalent control input error eu as
controller structure is found, we examine the effect of the     
(n−2)
disturbance input, ξ(t), in the feedback controlled system.
γ eu + γ2n−1 eu + · · · + γn+2 eu
Let p(O j ) denote
 j an arbitrary time polynomial of degree j ,
i.e., p(O j ) = i=0 αi t i with p(O(0)) being a constant. The     
(n−1) (n)
structural estimates of the phase variables (ė y , . . . , e(n−1)
y ), + γn+1 ė y +γn e y +· · ·+γ1 e y +γ0 ey = 0.
associated with the output error e y , are given by the following
definitions: (107)
 
 (n−1) Substituting the term ė y by its structural estimate,
e(n−1)
y = γ e u + p(O(0)) =: e y + p(O(0)) (102)  (n−1)
γ( eu ), given that enough iterated integral compensation
 
(2) has been added in the expression to compensate for its

ey(n−2)
=γ eu + p(O(1)) =: e(n−2)
y + p(O(1)) accompanying n − 2th-order time-polynomial, one obtains
    
(n−2)
(103)
γ eu + γ2n−1 eu + · · · + γn+2 eu
..
.    
(n−1) (n−1)
ė y = γ eu + p(O(n − 2)) =: ėy + p(O(n −2)). +γn+1 eu
    
(104) (n−1) (n)
+ γn e y + · · · + γ1 e y + γ0 ey = 0.
A sliding mode controller for the switched, unperturbed, pure
integration, system, e(n)
y = eu , based on the phase variables (108)
integral reconstructors, is specified by a sliding surface coor- The equivalent control input error, denoted by eu,eq and
dinate function denoted by σ̂ and defined as defined as eu,eq = u eq − u ∗ (t), is immediately obtained, from
(n−1)  (n−2) this last relation, after using the Laplace transform in the

σ = ey + γ2n−1 e y + · · · + γn+2 ėy + γn+1 e y
      expression
(n) (n+1)  
+ γn e y + · · · + γ1 e y + γ0 ey . 1 γn s n + γn−1 s n−2 + · · · + γ1 s + γ0
eu (s) = − e y (s)
γ s(s n−1 + γ2n−1 s n−2 + · · · + γn+1 )
(105) (109)
which entirely coincides with the flat filter controller expres-
The reason behind putting a linear combination of n + 1 sion found ealier.
output tracking error iterated integral compensators in σ̂ The equivalent control u eq , valid only on σ̂ = 0 while
resides in the fact that the classical time-polynomial errors enforcing the invariance condition σ̂˙ = 0, is interpreted as a
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

virtual or average control input that smoothly drives the system Let the scalar variable e be defined as e = −σ̂ , then the above-
response constrained to the sliding manifold, σ̂ = 0. In our mentioned equations are those of a delta–sigma modulator
case, this is a manifold defined on an extended tracking error accepting the continuous signal u av as an (analog) input and
phase space, including the coordinates represented by a finite producing the switched signal u = 0.5(1+sign(e)) as a binary-
number of iterated integrals of the output error. valued (digital) output
The sliding regime existence is assessed from the necessary 1
and sufficient condition σ̂ (d/dt)σ̂ < 0. From the expression ė = γ [u av − u] , u = (1 + sign(e)) .
2
d
σ̂ σ̂
dt   C. Flat Filters and Active Disturbance Rejection Control in
   
(n−2) nth-Order Plants
= σ̂ γ eu + γ2n−1 eu + · · · + γn+2 eu Consider the tracking error pure integration perturbed
  system
(n−1)
+ γn+1 eu e(n) ∗
y = eu + ξ(t), e y = y − y (t) := e1 .
    
(n−1) (n) We adopt, for the reduced-order observer, the n − 1-D system
+ γn e y +· · ·+γ1 e y +γ0 ey state representation, including an artificial error derivative
measurement, (e2 = ẏ − ẏ ∗ (t) = ė1 )
<0
ė2 = e3 (112)
one obtains, using eu = u − u ∗ (t), the control input switching ..
policy .
1 ėn−1 = en (113)
u= (1 − sign(σ̂ )). (110) ėn = eu + ξ(t) (114)
2
Define, in general, an average control input, u av , as e2 = ė1 . (115)
     A ROESO, which takes e2 = ė1 as the artificially measured
(n−2)
∗
u av = u (t) − γ2n−1 eu + · · · + γn+2 eu output, including, also, one extra output integration, is given
  by
(n−1)
d
+ γn+1 eu e2 = e3 + λn−1 (ė1 − e2 ) (116)
dt
     d
1 (n−1) (n) e3 = e4 + λn−2 (ė1 − e2 ) (117)
− γn e y + · · · + γ1 e y + γ0 ey dt
γ ..
(111) .
d

en−1 = en + λ2 (ė1 − e2 ) (118)
with eu = u − u ∗ (t). The average control u av coincides with dt
u eq whenever the invariance conditions, σ̂ = 0 and σ̂˙ = 0, are d
en = eu + z + λ1 (ė1 − e2 (119)
satisfied and eu = eu,eq . dt
A sliding regime exists on σ̂ = 0 with the help of the d
z = λ0 (ė1 − e2 ). (120)
derived control input switching policy if and only if dt
Define
σ̂ {γ [u − u av ]} < 0, (γ > 0).
η2 = e2 − λn−1 e1 (121)
This implies that for σ̂ > 0 (i.e., when u = 0),
η3 = e3 − λn−2 e1 (122)
−γ (u av ) < 0, and for σ̂ < 0 (i.e., when u = 1),
γ (1 − u av ) > 0. ..
.
In summary, a sliding regime exists on σ̂ = 0 if and only
ηn = en − λ1 e1 (123)
if u av satisfies

ζ = z − λ0 e1 . (124)
0 < u av < 1.
The ROESO is, thus, proposed to be
On the sliding surface, with the invariance conditions holding, d
this last condition translates into η2 = η3 − λn−1 η2 + λn−2 − λ2n−1 e1 (125)
dt
0 < u eq < 1. d
η3 = η4 − λn−2 η2 + (λn−3 − λn−2 λn−1 )e1 (126)
dt
We write ..
.
1
σ̂˙ = γ [u − u av ] , u = (1 − sign(σ̂ )). d
η
n−1 = η
n − λ2 η2 + (λ1 − λ2 λn−1 )e1 (127)
2 dt
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 13

d
ηn = eu + ζ − λ1 η2 + (λ0 − λ1 λn−1 )e1 (128) From here, it easily follows that:
dt  
d s

ζ = −λ0 η2 − λ0 λn−1 e1 . (129) 2 (s) = n ξ(s) (149)
dt s + λn−1 s n−1 + · · · + λ1 s + λ0
The estimates of the original tracking error phase variables  
may be computed from the following expressions: s(s + λn−1 )
3 (s) = n ξ(s) (150)
s + λn−1 s n−1 + · · · + λ1 s + λ0
e1 = e1 = e y
..
e2 = 
η2 + λn−1 e1 (130) .
 
e3 = 
η3 + λn−2 e1 (131) s(s n−2 + λn−1 s n−3 + · · · + λ2 )
n (s) = ξ(s) (151)
.. s n + λn−1 s n−1 + · · · + λ1 s + λ0
.
 
en = η̂n + λ1 e1 (132) s(s n−1 + λn−1 s n−2 + · · · + λ2 s + λ1 )
ξ (s) = ξ(s)
s n + λn−1 s n−1 + · · · + λ1 s + λ0
z = ξ̂ = 
ζ + λ0 e1 . (133)
(152)
Defining also, for the original tracking error system
which implies large attenuation of the low-frequency input
η2 = e2 − λn−1 e1 (134) disturbance signal, ξ(t), in the estimation errors of the tracking
η3 = e3 − λn−2 e1 (135) error phase variables as well as in the input disturbance
.. estimation error.
. The phase variables tracking errors ei , i = 1, . . . , n, satisfy
ηn = en − λ1 e1 (136)
e1 = ê1 = e y (153)
ζ = ξ(t) − λ0 e1 . (137) e2 = 2 + e2 (154)
One readily obtains ..
.
η̇2 = η3 − λn−1 η2 + (λn−2 − λ2n−1 )e1 (138) en = n + en (155)
η̇3 = η4 − λn−2 η2 + (λn−3 − λn−2 λn−1 )e1 (139) and
..
. ξ = z + ξ . (156)
η̇n−1 = ηn − λ2 η2 + (λ1 − λ2 λn−1 )e1 (140)
An ADRC controller can be synthesized that stabilizes the
η̇n = u + ζ − λ1 η2 + (λ0 − λ1 λn−1 )e1 (141) output reference trajectory tracking errors phase variables, e j ,
ζ̇ = −λ0 η2 − λ0 λn−1 e1 . (142) and tries to remove, in the feedback law, the perturbation input
ξ(t) by means of the disturbance estimate z.
Let i = ηi − ηi , i = 2, 3, . . . , n, and ζ = ζ − 
ζ = ξ − z. e1 = e1 . A ROESO-based ADRC controller, for
Finally, let 
Clearly, from the definitions of the phase variable tracking output stabilization purposes, is proposed as
error estimates, one has

n 
n 
n
i = ηi − ηi = ei − ei , i = 2, 3, . . . , n eu = −ξ̂ − γ j −1 ej = −ξ̂ − γ j −1e j + γ j −1(e j − ej )
j =1 j =1 j =2
ζ = ξ(t) − ξ̂ = eξ = ξ − z.
 n 
n
The estimation errors of the tracking error phase variables, = −ξ̂ − γ j −1e j + γ j −1 j (157)
ηi − η̂i , and that of the disturbance estimation error, eζ = j =1 j =2
ζ − ζ̂ = ξ − z, are seen to satisfy
where e j is the ( j − 1)th-order time derivative of the output
˙2 = 3 − λn−1 2 (143) tracking error e y = e1 .
The nth-order time derivative of the output tracking error
˙3 = 4 − λn−2 2 (144)
satisfies
..
. 
n 
n
e(n)
y = (ξ − ξ̂ ) − γ j −1 e j + γ j −1 j
˙n−1 = n − λ2 2 (145)
j =1 j =2
˙n = (ξ − z) − λ1 2 (146) n n
= eξ − γ 0 e1 − γ j −1 e j + γ j −1  j . (158)
˙ζ = −λ0 2 . (147)
j =2 j =2
In other words, the estimation error 2 satisfies the linear
perturbed differential equation Let q(s) denote the following characteristic polynomial,
q(s) = s n + λn−1 s n−1 + · · · + λ1 s + λ0 , and let p(s)
(n) (n−1)
2 + λn−1 2 + · · · + λ1 ˙2 + λ0 2 = ξ̇ . (148) denote the closed-loop control characteristic polynomial
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

p(s) = s n + γn−1 s n−1 + · · · + γ1 s + γ0 . Using the above- where

mentioned expressions, one obtains, in the Laplace transform
n(s) = s(s n−1 + k2n−1 s n−2 + · · · + kn+1 )
domain,
d(s) = s 2n + k2n−1 s 2n−1 + · · · + kn+1 s n+1 + · · · + k1 s + k0
n n−1
p(s)e y (s) = (s + γn−1 s + · · · + γ0 )e y (s)
 n     which is clearly perfectly matched with the ROESO-based
n j (s) n n+1
= eξ (s) + γ j −1 × ξ(s) ADRC closed-loop system.
q(s) q(s) This is due to the fact that the closed-loop characteristic
j =2
n   polynomial coefficients uniquely determine all the coefficients
n j (s)
+ γ j −1 ξ(s) (159) in the flat filter controller. We, therefore, have
q(s)
j =2
p(s) = s 2n + k2n−1 s 2n−1 + · · · + kn+1 s n+1 + · · · + k0
with = (s n + γn−1 s n−1 + γ1 s + γ0 )
n j (s) = s(s j −2 + λn−1 s j −3 + · · · + λn− j +2 ) (160) ×(s n + λn−1 s n−1 + · · · + λ1 s + λ0 ). (167)
n n+1 = s(s n−1
+ λn−1 s n−2
+ · · · + λ2 s + λ1 ). (161) Clearly, given a ROESO-based ADRC controller design,
The disturbance-driven closed-loop system output, y1 , there exists a unique stable flat filter controller that has exactly
evolves in accordance with the following dynamics: the same set of fundamental transfer functions (sensitivity,
  complementary sensitivity, and open-loop transfer functions).
m(s) On the other hand, given an FF controller design, there
e y (s) = ξ(s) (162)
r (s) exists nonunique equivalent ROESO-based ADRC controllers.
where Thus, the equivalence is only one-way. The resulting, proper,
disturbance-driven closed-loop transfer function will exhibit
m(s) = s[s n−1 + (λn−1 + γn−1 )s n−2 enhanced disturbance attenuation and good low-frequency ref-
+(λn−2 + γn−1 λn−1 + γn−3 )s n−3 erence trajectory tracking features with high-frequency output
measurement noise rejection characteristics.
+ · · · + (λ1 + γn−1 λ2 + γn−2 λ3 + · · · + γ1 λn )]
R EFERENCES
r (s) = p(s)q(s) = (s + γn−1 s
n n−1
+ · · · + γ1 s + γ0 ) [1] K. J. Åström and T. Hägglund, “The future of PID control,” Control
Eng. Pract., vol. 9, no. 11, pp. 1163–1175, Nov. 2001.
× (s + λn−1 s
n n−1
+ · · · + λ1 s + λ0 ). [2] K. J. Astrom and T. Hagglund, PID Controllers: Theory, Design and
Tuning. Raleigh, NC, USA: Instrument Society of America, 1995.
Clearly, the denominator r (s) is of the form [3] R. Garrido and J. L. Luna, “On the equivalence between PD+DOB and
PID controllers applied to servo drives,” IFAC-PapersOnLine, vol. 51,
r (s) = s 2n + k2n−1 s 2n−1 + · · · + k1 s + k0 (163) no. 4, pp. 95–100, Dec. 2018.
[4] J. Han, “From PID to active disturbance rejection control,” IEEE Trans.
Ind. Electron., vol. 56, no. 3, pp. 900–906, Mar. 2009.
while the numerator is of the form [5] V. I. Utkin, Sliding Modes in Control Optimization. Berlin, Germany:
Springer, 1992.
m(s) = s n + k2n−1 s n−1 + · · · + kn+1 s. (164) [6] C. Edwards and S. Spurgeon, Sliding Mode Control: Theory and
Applications. London, U.K.: Taylor & Francis, 1998.
The characteristic polynomial of the closed-loop system [7] H. Sira-Ramírez, E. W. Zurita-Bustamante, M. A. Aguilar-Orduña, and
factors into the product of the ROESO characteristic poly- E. Hernández-Flores, Sliding Mode Control Devoid of State Measure-
nomial [i.e., n = (n − 1) + 1] and the nth-order charac- ments. Berlin, Germany: Springer, 2018, ch. 6, pp. 73–102.
[8] W. Xue and Y. Huang, “On frequency-domain analysis of ADRC
teristic polynomial of the closed-loop system, obtained by for uncertain system,” in Proc. Amer. Control Conf., Jun. 2013,
straightforward pole placement on the nth-order pure inte- pp. 6637–6642.
gration plant system, as if all the tracking error phase vari- [9] Y. Huang, W. Xue, Z. Gao, H. Sira-Ramirez, D. Wu, and M. Sun, “Active
disturbance rejection control: Methodology, practice and analysis,” in
ables had been available for feedback. All this is, evidently, Proc. 33rd Chin. Control Conf. (CCC), Jul. 2014, pp. 1–5.
in accordance with the observer controller design separation [10] R. Madoński and P. Herman, “Survey on methods of increasing the
principle for state feedback through an observer in linear efficiency of extended state disturbance observers,” ISA Trans., vol. 56,
pp. 18–27, May 2015.
systems. [11] H. Sira-Ramirez, A. Luviano-Juárez, M. Ramírez-Neria, and
An output tracking error stabilization feedback controller, E. W. Zurita-Bustamante, Active Disturbance Rejection Control
for an nth-order pure integration system, y (n) = u + ξ(t), was of Dynamic Systems: A Flatness Based Approach. London, U.K.:
Butterworth-Heinemann, 2017.
found to be characterized, in the frequency domain, by [12] E. W. Zurita-Bustamante, H. Sira-Ramirez, J. Linares-Flores,
  O. D. Ramírez-Cárdenas, M. A. Contreras-Ordaz, and
kn s n + · · · + k1 s + k0
eu = − e y (s) (165) J. F. Guerrero-Castellanos, “On the active disturbance rejection
s(s n−1 + k2n−1 s n−2 + · · · + kn+1 ) control of the permanent magnet synchronous motor,” in Proc. IEEE
Power Energy Conf. Illinois (PECI), Feb. 2018, pp. 1–6.
with the filter gains chosen to guarantee a Hurwitz closed-loop [13] M. Fliess, R. Marquez, E. Delaleau, and H. Sira-Ramirez, “Correcteurs
characteristic polynomial. The disturbance-driven closed-loop proportionnels-intégraux généralisés,” ESAIM, Control, Optim. Calculus
Variat., vol. 7, no. 1, pp. 23–41, 2002.
system is described by [14] H. Sira-Ramírez, A. Luviano-Juárez, M. Ramírez-Neria, and
  R. Garrido-Moctezuma, “Flat filtering: A classical approach to robust
n(s)
e y (s) = ξ(s) (166) control of nonlinear systems,” in Proc. Amer. Control Conf., Jul. 2016,
d(s) pp. 3844–3849.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

SIRA-RAMÍREZ et al.: FLAT FILTERS, PID, ADRC, AND SMC 15

[15] E. W. Zurita-Bustamante, A. Luviano-Juárez, and H. Sira-Ramírez, Eric William Zurita-Bustamante received the B.S.
“On the robust flat-filtering control of MIMO nonlinear systems: The degree in electronics engineering and the M.S.
PMSM experimental case study,” in Proc. Annu. Amer. Control Conf. degree from the Universidad Tecnologica de la Mix-
(ACC), Jun. 2018, pp. 1–4. teca, Huajuapan, Mexico, in 2007 and 2010, respec-
[16] R. Kelly, “A simple set-point robot controller by using only posi- tively. He is currently pursuing the Ph.D. degree
tion measurements,” in Proc. 12th IFAC World Congr., vol. 6, 1993, with the Center of Research and Advanced Studies
pp. 173–176. (CINVESTAV), Mexico City, Mexico.
[17] H. Sira-Ramírez, A. Hernández-Méndez, J. Linares-Flores, and His current research interests include automatic
A. Luviano-Juárez, “Robust flat filtering DSP based control of the control, nonlinear systems, power electronics, and
boost converter,” Control Theory Technol., vol. 14, no. 3, pp. 224–236, mechatronic systems.
Aug. 2016.

Hebertt Sira-Ramírez received the Ph.D. degree in

electrical engineering from the Massachusetts Insti- Congzhi Huang (SM’14) received the B.S. degree
tute of Technology, Cambridge, MA, USA, in 1977. in automation and the Ph.D. degree in control theory
He is currently with the Departamento de Inge- and applications from the School of Control and
niería Elíctrica, Sección de Mecatrónica, Center Computer Engineering, North China Electric Power
of Research and Advanced Studies (CINVESTAV), University, Beijing, China.
Mexico City, Mexico. His current research interests He is currently an Associate Professor with
include theoretical and practical aspects of feedback North China Electric Power University. His current
regulation of nonlinear dynamic systems with a research interests include networked control sys-
special emphasis on sliding mode control, algebraic tems, disturbance rejection control theory, and their
techniques, active disturbance rejection control, and applications in electric power systems.
their applications.