Anda di halaman 1dari 7

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/299569076

Optimal Design of Digital Anticipatory


Servomechanisms with Good Noise Immunity

Conference Paper · July 2016


DOI: 10.1109/NORBERT.2016.7547455

CITATION READS

1 567

1 author:

Hugh Lachlan Kennedy


Technical Knockout Systems Pty. Ltd.
59 PUBLICATIONS 174 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Real-Time (RT) Video Moving Target Indication (VMTI) for an airborne platform using Graphics Processing
Units (GPUs) View project

All content following this page was uploaded by Hugh Lachlan Kennedy on 02 April 2016.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
To appear in Proc. 2nd IEEE Conf. on Norbert Wiener in the 21st Century, Melb. Australia, July 13-15, 2016.

Optimal Design of Digital Anticipatory


Servomechanisms with Good Noise Immunity
Hugh L. Kennedy
School of Engineering
University of South Australia
Hugh.Kennedy@unisa.edu.au

Abstract—A frequency-domain approach to the design of approach usually does result in satisfactory controls, for
feedback controls for SISO systems is presented. The noise gain is adequate sampling rates [11]; however, if rigorous design
minimized, subject to a phase margin constraint. The controller is methodologies are used to carefully balance conflicting
expressed as a repeated-pole expansion, which reduces the requirements and specifications, an optimal design in
optimization to a quadratic problem and a linear system of continuous time may be far from optimal in discrete time.
equations. PID+filter controls are produced as a limiting case.
𝑑(𝑛)
Keywords—Cybernetics; Digital control; Optimal control 𝑟(𝑛) 𝑒(𝑛) 𝑢(𝑛) 𝑦(𝑛)
+ 𝐻𝑒 (𝑧) 𝐻𝑢 (𝑧) +
I. INTRODUCTION
In Chapter IV of [1], on feedback control systems, Wiener 𝑣(𝑛) 𝜀(𝑛)
−𝐻𝑦 (𝑧) +
envisages three servomechanism configurations. The input-
shaping structures described in [2],[3], where a “compensator” Nomenclature:
is cascaded with a simple feedback loop containing the 𝑋(𝑧) = 𝒵{𝑥(𝑛)}, 𝑥(𝑛) = 𝒵 −1 {𝑋(𝑧)} and
“effector” (i.e. the plant), are modern-day examples of the first 𝐻𝑥𝑦 (𝑧) = 𝑌(𝑧)⁄𝑋(𝑧), i.e. the transfer fcn. for 𝑥(𝑛) → 𝑦(𝑛).
configuration. The second servomechanism configuration uses a Some definitions for analysis/design:
“predictor” or an “anticipator” inside the feedback loop for a 𝐿(𝑧) = 𝐻𝑦 (𝑧)𝐻𝑒 (𝑧)𝐻𝑢 (𝑧) (Loop transfer function).
phase lead, to compensate for the “lagging characteristics” of the 𝑆(𝑧) = 1⁄[1 + 𝐿(𝑧)] (Sensitivity function).
effector. A feedback structure incorporating online system 𝑇(𝑧) = 1 − 𝑆(𝑧) = 𝐿(𝑧)⁄[1 + 𝐿(𝑧)] (Complementary sensitivity).
identification is described in the third configuration. On the Some specific relationships of interest:
second approach, Wiener notes: “The conditions of stability and 𝐻𝑑𝑦 (𝑧) = 𝑆(𝑧), 𝐻𝜀𝑦 (𝑧) = −𝑇(𝑧).
effectiveness of anticipatory feedbacks need a more thorough For 𝐻𝑦 (𝑧) = 1: 𝐻𝑟𝑒 (𝑧) = 𝑆(𝑧), 𝐻𝑟𝑦 (𝑧) = 𝑇(𝑧).
discussion than they have yet received”. In Chapter III of [1], on 𝐻𝑟𝑢 (𝑧) = 𝐻𝑒 (𝑧)⁄𝑆(𝑧) , 𝐻𝜀𝑢 (𝑧) = − 𝐻𝑦 (𝑧)𝐻𝑒 (𝑧)⁄𝑆(𝑧).
estimation and communication, Wiener also notes: “The theory
Fig. 1. Control-loop block-diagram with basic definitions. The continuous-
of discrete time series is simpler in many respects than the theory
time signal 𝑣(𝑡) is sampled at times 𝑛𝑇𝑠 using an analogue-to-digital-to
of the continuous series”; and later on, that: “The filters for converter (A2D) with a perfect anti-aliasing filter. The resulting discrete-time
discrete time series are usually not so much physically signal is simply denoted here using 𝑣(𝑛). The continuous-time plant 𝐻𝑢 (𝑠) is
constructible devices to be used with an electric circuit as connected to 𝐻𝑒 (𝑧) via a digital-to-analogue converter (D2A) with a zero-order
mathematical procedures to enable statisticians to obtain the best hold. The A2D and D2A are perfectly synchronized; furthermore, the onset
results with statistically impure data”. times of all inputs are assumed to be an integer multiple of 𝑇𝑠 .

Wiener did not live to witness the dawn of the digital age and A. Frequency-domain vs. state-space controller design
experience the extent to which computers, running the very The emphasis here on low-order digital controls is less to do
algorithms he foresaw, have infiltrated our everyday lives. Our with constraints imposed by computing resources and more to
understanding of simple anticipatory compensators, such as do with plant model uncertainty. When a system is only partially
lag/lead networks and proportional-derivative (PD) controls is understood, the safest approach is to model partially-known
also much improved [3][4][5][6][7]; however as noted in [7], low-frequency dynamics using a low-order approximation and
some unsolved problems remain, such as the minimization of to aim for a high degree of robustness to ensure that unknown
high-frequency sensor noise in digital implementations (see Fig. high-frequency dynamics do not yield instability or poorly
1), which is particularly problematic when a large phase lead is damped oscillations in the closed-loop system. Reasonable plant
applied for high performance or robust stability. Relative to the models can usually be constructed using limited empirical data
continuous-time case [3][8][9][10], this aspect of discrete-time or approximate ab-initio calculations; and in some cases,
controller design has not received the attention that it deserves, intuition may be adequate. As noted by Weiner (and Horowitz
which is somewhat surprising, given the proliferation of digital [12]), feedback is not required if the characteristics of the plant
devices and systems in recent times. This may be because many are constant and known with certainty (in the absence of
control engineers are satisfied with the discretization of disturbances) and an open-loop strategy with a compensator
continuous-time designs. Indeed, when done appropriately, this cascaded with the effector is sufficient, indeed it may be
To appear in Proc. 2nd IEEE Conf. on Norbert Wiener in the 21st Century, Melb. Australia, July 13-15, 2016.

preferable in some cases because it avoids the main “cost of problem to be simply solved via the inverse of a small square
feedback”, which is primarily the aforementioned problem of matrix. Furthermore, as recommended in [14], the method
measurement noise amplification [12]. Thus imperfect models reduces to a point-response matching method (PRM) as a
should not be feared in sufficiently robust servomechanism limiting case, which yields discrete proportional-derivative-
designs because the resulting tracking errors are removed over integral (PID), PI & PD controls with an integrated low-pass
time via feedback. filter (+F). The proposed approach provides sufficient flexibility
to balance the various trade-offs associated with controller
Following the pioneering works of Bode and Horowitz, and
designs in a rigorous way, with sufficient simplicity to permit
the more recent discussions in [13][14], a frequency-domain solutions to be found without difficulty, thus it may be used by
design approach, using the loop transfer function 𝐿(𝑧) , is practitioners who do not have time to establish and customize
adopted here (see Fig. 1). Set against a backdrop of ever- dedicated solver apparatus.
growing enthusiasm for the “modern” state-space design
techniques, and decreasing interest in “classical” frequency- II. FORMULATION
domain techniques, Horowitz questions the use of the so-called
modern methods, particularly for single-input/single-output When a low-order lag/lead or PD+F controller is expected to
(SISO) plants [12]. Horowitz raises the issue of an over-reliance be sufficient to achieve the required design objectives, the
of state-space techniques on the model (this issue is also raised discrete-time (pulsed) transfer-function of a one degree-of-
in [14]). Even if robustness issues are carefully addressed, using freedom (1-DOF) controller 𝐻𝑐 (𝑧), is expressed in the complex
more recently-developed techniques to find ℋ∞ solutions, the 𝑧-domain as
model is effectively counted twice because it is used firstly to 𝑧
𝑘
estimate the plant’s internal states (via an observer) then again 𝐻𝑐 (𝑧) = ∑𝐾𝑘=1 𝑐𝑘 (𝑧−𝑝 ) (1a)
𝑓
to determine the appropriate plant input from the states. In
contrast, the so-called classical methods, focus on what is which may be factored in causal form as
directly measurable and manipulable, using analogue or digital,
𝐻𝑐 (𝑧) = 𝐻IIR (𝑧)𝐻FIR (𝑧) where (1b)
infinite- or finite-impulse response (IIR or FIR), filter
realizations. 𝐾
1
𝐻IIR (𝑧) = ( ) and (1c)
1−𝑝𝑓 𝑧 −1
B. ℋ2-norm vs. ℋ∞-norm controller design
Wiener (along with Hopf and Kalman) pioneered the 𝐻FIR (𝑧) = ∑𝐾−1
𝑘=0 ℎ𝑘 𝑧
−𝑘
. (1d)
development of ℋ2 -norm problems, that seek to minimise This basis set was chosen because of its relationship with the
mean-squared errors, in Linear Quadratic Gaussian (LQG) discrete Laguerre polynomials, which are known for their utility
optimal control. This approach was well suited to the new and in the analysis and representation of low-frequency phenomena
challenging aerospace problems of that era, which involved
over short time scales [6]. The FIR filter coefficients ℎ𝑘 and the
inertial plants with well-modelled ballistic properties, perturbed
expansion coefficients 𝑐𝑘 are related via a partial-fraction
by random disturbances that are approximately white Gaussian expansion of (1b) in non-causal form. Its frequency response
noise [15]. Since then, ℋ∞ -norm problems, that seek to 𝐻𝑐 (𝜔), is evaluated using
minimize worst-case absolute errors, have proven to be more
useful because they yield robust controls that are better able to 𝐻𝑐 (𝜔) = 𝐻𝑐 (𝑧)|𝑧=𝑒 𝑖𝜔 = 𝐻𝑐 (𝑒 𝑖𝜔 ) (2)
accommodate the high degree of model uncertainty that is
associated with complex terrestrial systems [15][16]. where 𝑖 is the imaginary unit (𝑖 = √−1) and 𝜔 is the angular
frequency (radians per sample). To simplify notation, frequency
A popular form of the ℋ∞ (SISO) feedback control problem responses 𝐻(𝑒 𝑖𝜔 ) are abbreviated here using 𝐻(𝜔). The noise
is stated as follows: Optimize robust performance (which gain (NG𝐿 ) of the loop transfer function 𝐿(𝑧) = 𝐻𝑐 (𝑧)𝐻𝑢 (𝑧),
implies robust stability and nominal performance) by finding a weighted using 𝑊𝐿 (𝜔), is
controller that minimizes ‖𝑈(𝜔)‖∞ and ensures that 𝑈(𝜔) < 1
+𝜋
for all 𝜔, where 𝑈(𝜔) = |𝑊𝑆 (𝜔)𝑆(𝜔)| + |𝑊𝑇 (𝜔)𝑇(𝜔)| with 1
NG𝐿 = 2𝜋 ∫−𝜋 |𝑊𝐿 (𝜔)𝐿(𝜔)|2 𝑑𝜔 = 1
‖𝑊𝐿 (𝜔)𝐿(𝜔)‖2 . (3)
2𝜋
𝑊𝑆 and 𝑊𝑇 being arbitrary weighting functions [17][18].
Unfortunately, this problem is not easy to solve, although it may Using (1a) in (3) yields
be re-stated in simpler ways to facilitate the discovery of NG𝐿 = 𝒄T 𝑺𝒄 where (4a)
approximate solutions or the production of simple PID controls
[16][17][19]. 𝒄 is a 𝐾 × 1 vector containing the coefficients 𝑐𝑘 and
C. Summary of the proposed approach 𝑺 is a 𝐾 × 𝐾 positive semi-definite matrix with elements
The controller design problem described in this paper 1 +𝜋
𝑆𝑘2,𝑘1 = 2𝜋 ∫−𝜋 𝑊𝐿 (𝜔)𝛷𝑘∗ 2 (𝜔)𝛷𝑘1 (𝜔)𝑑𝜔. (4b)
involves the minimization of a frequency-domain ℋ2-norm (of
the weighted loop-transfer function) for good noise immunity; (The [∙]∗ operator represents complex conjugation and [∙]T is a
subject to an equality constraint (on the phase margin) for transpose operation.) Evaluation of this integral, and the
reasonably robust stability and satisfactory transient-response controller tuning process, are simplified somewhat if the
performance. The main technical contribution of this paper is the weighting function is a binary piecewise constant
expression of the solution as a repeated pole expansion in the
complex 𝑧 domain, which allows a quadratic optimization
To appear in Proc. 2nd IEEE Conf. on Norbert Wiener in the 21st Century, Melb. Australia, July 13-15, 2016.

0, 0 ≤ |𝜔| < 𝜔𝑐 ensures the transferal of the desired properties to the various
𝑊𝐿 (𝜔) = { . (5) transfer functions of the closed-loop system (see Fig. 1). Stating
1, 𝜔𝑐 ≤ |𝜔| ≤ 𝜋
the optimization problem in this way (in addition to making it
These inner/outer domains are the signal/noise bands, easy to solve) allows a reasonable degree of performance and
respectively. The 𝜔𝑐 parameter defines the “available robustness to be attained, for a stable closed-loop system. This
bandwidth” of the loop [13]. Frequencies above this threshold is achieved by cutting low-frequency sensitivity (in the near-dc
are either: corrupted by sensor noise, beyond the capabilities of region of the signal band) to meet the phase-margin specification
the actuator, and/or suspected to contain unknown (possibly and tracking-error requirements, while ensuring that high-
non-linear) plant dynamics. This high frequency region is frequency sensitivity (in the noise band) is as flat and as close to
therefore assumed to contain no useful/meaningful information unity as possible. Using 𝜔g ≪ 𝜔c ensures that no attempt is
for feedback control purposes, thus input signals in the noise made to flatten the sensitivity peak at low-to-mid frequencies (at
band should ideally pass through the closed-loop system as if it the upper end of the signal band), which would be in defiance of
were an open-loop system (i.e. unity sensitivity).
Bode’s integral theorem. However, peak sensitivity may be
Substitution of (5) into (4b), and utilizing real symmetry and reduced for improved stability, at the expense of a longer
imaginary anti-symmetry around 𝜔 = 0, yields transient response, by either increasing sensitivity at the low end
1 𝜋
of the signal band, and/or by increasing sensitivity over the
𝑆𝑘2,𝑘1 = 𝜋 Re [∫𝜔 𝛹𝑘∗2 (𝜔)𝛹𝑘1 (𝜔)𝑑𝜔] with (6a) entire high-frequency noise band, in accordance with Bode’s
𝑐
integral theorem. These considerations highlight the fact that
𝑘
𝑒 𝑖𝜔 controller design is essentially a zero-sum ‘game’ [13].
𝛹𝑘 (𝜔) = ( ) 𝐻𝑢 (𝑒 𝑖𝜔 ). (6b)
𝑒 𝑖𝜔 −𝑝
B. Extra constraints and limiting cases
(The Re[∙] operator extracts the real part of the complex
Additional equality constraints on 𝐿(𝜔) may be applied if
argument.) The non-zero lower limit of 𝜔𝑐 ensures that any
desired/required. For instance: constraining 𝐿(𝜔)|𝜔=0 to equal
singularities at 𝜔 = 0, due to plant (or controller) integrators, a finite real constant (in the absence of integrators); or using
are avoided. It also means that rapid changes in the integrand
due to severe low-pass filtering (using 𝑝𝑓 → 1 in the controller) Re [ lim 𝐿(𝜔)] = 0 , may help to promote stable solutions for
𝜔→0
are tolerated. In the absence of plant integrators, a ℎ𝐼 ⁄(1 − 𝑧 −1 ) some plants.
term may be added to the summation in (1a), for a PID+F-type Use of 𝐾 = 2 reduces (9a) to (9b), for a PD+F controller,
controller, where ℎ𝐼 is a fixed coefficient, or an unknown to be designed using a one-point PRM
determined.
1 2
Specification of the desired phase margin 𝜙𝑚 at 𝜔g , where 𝐻𝑐 (𝑧) = (1−𝑝𝑧 −1 ) (ℎ0 + ℎ1 𝑧 −1 ). (10a)
𝜔g is the gain crossover frequency, such that |𝐿(𝜔g )| = 1 and
Similarly, if 𝐾 = 1 and a discrete integrator (i.e. an
∠𝐿(𝜔g ) = 𝜋 + 𝜙𝑚 , means that accumulator) is appended to the sum in (1a), a PI+F controller
𝐿(−𝜔g ) = 𝑒 −𝑖(𝜋+𝜙𝑚 ) and 𝐿(+𝜔g ) = 𝑒 +𝑖(𝜋+𝜙𝑚) . (7) is produced
1 ℎ
Using (1a) in (7) yields 𝐻𝑐 (𝑧) = (1−𝑝𝑧 −1 ) ℎ0 + 1−𝑧𝐼−1. (10b)
𝒍 = 𝑱𝒄 where (8a) If 𝐾 = 2 is instead used to derive the PI+F controller along with
−𝑖(𝜋+𝜙𝑚 ) an additional 𝐿(𝜔)|𝜔=𝜋 = 0 constraint, for a Nyquist null to
𝒍 = [𝑒 +𝑖(𝜋+𝜙 ) ] and (8b) reduce high-frequency noise, the optimization reduces to a two-
𝑒 𝑚
point PRM (at 𝜔 = ±𝜔g and 𝜔 = 𝜋) and a PID+F controller
𝑱 is a 2 × 𝐾 matrix, with the first and second rows containing results
the elements 𝛹𝑘 (−𝜔g ) and 𝛹𝑘 (+𝜔g ) respectively, where 2
1 ℎ
𝛹𝑘 (𝜔) is defined in (6b). 𝐻𝑐 (𝑧) = (1−𝑝𝑧 −1 ) (ℎ0 + ℎ1 𝑧 −1 ) + 1−𝑧𝐼−1. (10c)
Minimization of NG𝐿 in (4) subject to the phase-margin If backwards differences are used to derive the discrete D and I
equality constraint in (8), with 𝐾 > 2 , is solved using the operators, and if the two 1st-order low-pass filters are cascaded
method of Lagrange multipliers 𝝀, i.e. with both the P and D terms, instead of the more common form
−𝟏 with a single low-pass filter cascaded with the D term, (10c) may
𝒄 𝑺 𝑱T
[ ]=( ) [𝟎] or (9a) be expressed in a parallel PID+F form
𝝀 𝑱 𝟎 𝒍
1 2 1−𝑧 −1 𝑇𝑠
𝒄 = 𝑱−𝟏 𝒍 when 𝐾 = 2. (9b) 𝐻𝑃𝐼𝐷 (𝑧) = (1−𝑝𝑧 −1 ) (𝐾𝑃 + 𝐾𝐷 𝑇𝑠
) + 𝐾𝐼
1−𝑧 −1
(10d)
A. Shaping the sensitivity function where
SISO design methods that focus directly on 𝑇(𝑧) , e.g.
𝐾𝑃 = ℎ0 + ℎ1, 𝐾𝐷 = −𝑇𝑠 ℎ1 and 𝐾𝐼 = ℎ𝐼 ⁄𝑇𝑠 . (10e)
polynomial methods, tend to emphasize the end (i.e.
performance) rather than the means by which it is achieved
[11][20]. Even though 𝐿(𝜔) is the object of the optimization
used here, the simple relationship between 𝐿(𝜔) and 𝑆(𝜔)
To appear in Proc. 2nd IEEE Conf. on Norbert Wiener in the 21st Century, Melb. Australia, July 13-15, 2016.

C. 1-DOF and 2-DOF configurations In order of decreasing priority, the closed-loop design
The controls defined in (1a) and (10) may all be simply objectives are: 1) robust stability (i.e. robustness), as quantified
implemented in a 1-DOF configuration using 𝐻𝑒 (𝑧) = 𝐻𝑐 (𝑧) by the gain margin and the modulus margin; 2) immunity to
and 𝐻𝑦 (𝑧) = 1 (see Fig. 1). However, the poles of the various sensor noise amplification, as quantified by ‖𝑊𝐿 (𝜔)𝑇(𝜔)‖2; 3)
transient-response and steady-state response performance, as
closed-loop transfer functions between any two points around
the loop are uniquely determined by 𝐿(𝑧), i.e. the roots of 1 + might be quantified using time-domain ℋ2 -norms involving
various inputs, e.g. 𝑟(𝑡) or 𝑑(𝑡), and signals of interest in the
𝐿(𝑧). Thus the 𝐻𝑦 (𝑧)𝐻𝑒 (𝑧) product in 𝐿(𝑧) may be arranged to
loop, e.g. 𝑢(𝑡) or 𝑦(𝑡). Unfortunately, for a given sampling rate,
change the zeros of the closed-loop transfer functions, without an improvement in one of the areas is usually achieved at the
changing their poles. Therefore, the generic controller structure expense of the other two; however, it will be shown that controls
in (1a) or the PD+F controller in (10a), i.e. without an integrator,
designed using the proposed method may be readily configured
may be factored to fine tune the responses of these transfer to achieve a satisfactory compromise. The various controller
functions, according to design priorities. configs. (A-G) are described in Table I; the corresponding
A 2-DOF configuration with 𝐻𝑦 (𝑧) = 𝐻FIR (𝑧)⁄𝐾dc and closed-loop metrics are defined and presented in Table II.
𝐻𝑒 (𝑧) = 𝐾dc 𝐻IIR(𝑧), where 𝐾dc = 𝐻FIR(𝜔)|𝜔=0 is particularly The responses of various discrete-time transfer functions
useful. The high-frequency gain of 𝐻𝑟𝑢 (𝜔) is decreased by around the loop, as a function of normalized frequency 𝑓 =
placing the low-pass IIR filter (with a phase lag) in the forward 𝜔⁄2𝜋 (cycles per sample), are also plotted in Fig. 2 - Fig. 5 for
path and the high-pass FIR filter (with a phase-lead) in the selected controller configurations. Note that for 𝜙𝑚 = 60°, an
backward path. This arrangement avoids actuator (hyper-) equilateral triangle is formed in the Nyquist diagram, with
activity, which may be a problem when the sampling frequency |𝐿(𝜔g ) + 1| = |𝐿(𝜔g )| = |1|, thus |𝐿(𝜔g )| = |𝑆(𝜔g )| = |𝑇(𝜔g )| =
(𝐹𝑠 = 1⁄𝑇𝑠 ) is very high and a large phase lead is applied (see
0 dB in Fig. 2, Fig. 3 & Fig. 5. Magnitude responses of 𝐻𝜀𝑦 (𝜔)
[21] for an alternative mixed ℋ1, ℋ2, & ℋ∞ formulation of this
problem). This 2-DOF arrangement also increases the high- and 𝐻𝜀𝑢 (𝜔) are plotted in Fig. 3. Note that |𝐻𝜀𝑦 (𝜔)| = 𝑇(𝜔) for
frequency gain of 𝐻𝑑𝑒 (𝜔); however, this is inconsequential if any 𝐻𝑦 (𝑧) and |𝐻𝜀𝑦 (𝜔)| = |𝐻𝑟𝑦 (𝜔)| if 𝐻𝑦 (𝑧) = 1 . Unlike
machine precision is sufficient, because 𝑒(𝑛) not a physical 𝐻𝜀𝑢 (𝜔), which is of practical interest, 𝐻𝜀𝑦 (𝜔) is independent of
signal, it is just a number in a computer. Other transfer functions the plant gain 𝐾𝑢 , which was arbitrarily selected here; thus
around the loop, with 𝐻𝑦 (𝑧)𝐻𝑒 (𝑧) products, e.g. 𝐻𝑟𝑦 (𝑧) and 𝐻𝜀𝑦 (𝜔) is used to generate the immunity metric.
𝐻𝜀𝑢 (𝑧), or with unity factors, e.g. 𝐻𝑑𝑦 (𝑧) or 𝐻𝑟𝑒 (𝑧), in the
numerator, are of course not impacted. Even when plant TABLE I. CONTROLLER CONFIGURATION AND DESIGN RATIONALE
dynamics are slow, it will be shown that oversampling may be A 𝑇𝑠 = 0.01 , 𝐾 = 2, 𝜙𝑚 = 60° @ 𝜔g = 2𝜋𝑇𝑠 for good performance,
employed as a strategy to reduce sensor noise, with an i.e. tracking of reference inputs 𝑟(𝑛) and rejection of disturbance
appropriately configured controller: using a narrow bandwidth inputs 𝑑(𝑛). The type-I plant ensures a zero steady-state error for step
with 𝜔g → 0, to lengthen the transient response of 𝐻𝑟𝑦 (𝑧); and inputs, if the closed-loop system is stable. For good robustness and
moderate immunity: 𝜔c = 2𝜋 × 0.1 (𝜔c > 𝜔g ) and 𝑝𝑓 = 1⁄8 . In a
a very low-pass filter with 𝑝 → 1, to decrease the bandwidth of
1-DOF config.
𝐻IIR (𝜔). B As above but with 𝑝𝑓 = 1⁄2 for improved immunity, at the expense
of robustness and performance.
III. ILLUSTRATION OF THE PRINCIPAL DESIGN CONCEPTS C As above but with 𝐾 = 3 for further immunity improvement.
During the World War II period, Wiener immersed himself Increasing 𝑝𝑓 further is counterproductive because the lag introduced
in the problem of anticipatory anti-aircraft fire-control filters, in by 𝐻IIR (𝑧) is too large to be overcome by the lead due to 𝐻FIR (𝑧).
man-in-the-loop configurations, realized using analogue D As above but with 𝐾 = 4 for further immunity improvement.
electronics. This effort led to the famous filter that now bears his E Oversampling. As above but with 𝑇𝑠 = 0.001 and 𝑝𝑓 = 3⁄4 for
improved robustness, performance and immunity.
name. The example considered in this section, is an idealized F As above but in a 2-DOF config. to decrease the high-frequency gain
abstraction, of just one element of that problem. It is used to of 𝐻𝑟𝑢 (𝜔) . ‖𝑊𝐿 (𝜔)𝐻𝑟𝑢 (𝜔)‖2 decreases from 8.81 dB to
illustrate the characteristics of the proposed controller design −41.04 dB dB. This reduces |𝑢(𝑛)|, or the so-called derivative
method. Like Wiener’s filter, an ℋ2-norm problem is posed and “kick” for a step input [11]. ‖𝑊𝐿 (𝜔)𝐻𝑟𝑦 (𝜔)‖2 correspondingly
solved via the frequency domain; however, a man-at-the-loop decreases from −97.16 dB to −135.47 dB , which results in
system is considered here for digital implementation. The increased transient-response damping.
following second-order plant, is used as a crude model of turret G As for Config. B, but designed for a zero steady-state tracking-error
dynamics: 𝐻𝑢 (𝑠) = 𝐾𝑢 𝑒 −𝑠𝜏𝑢 ⁄[𝑠(𝑠 + 𝑝𝑢 )]. It is a reasonable model for ramp reference inputs. An integrator is added and 𝜔c = 2𝜋𝑇𝑠 ×
0.1 is used (𝜔c < 𝜔g ). These changes increase the gain of 𝐿(𝜔) in
for inertial systems, with a force input and a position output, that
the near-dc region of the signal band.
are subject to dissipative forces, e.g. drag or friction, as specified H Oversampling. As above but with 𝑇𝑠 = 0.001 and 𝑝𝑓 = 3⁄4 for
by the 𝑝𝑢 parameter. It is an example of a lag system with a improved robustness, performance and immunity. See Fig. 5.
delay of 𝜏𝑢 seconds that requires lead compensation. Values of
𝐾𝑢 = 10, 𝑝𝑢 = 0.01 and 𝜏𝑢 = 0.01 were used here.
To appear in Proc. 2nd IEEE Conf. on Norbert Wiener in the 21st Century, Melb. Australia, July 13-15, 2016.
TABLE II. CLOSED-LOOP DESIGN METRICS FOR THE VARIOUS CONTROLLER CONFIGURATIONS

Config. ‖𝑾𝑳 (𝝎)𝑳(𝝎)‖𝟐 ‖𝑾𝑳 (𝝎)𝑻(𝝎)‖𝟐 𝒒𝒄 a 𝑲𝒎 b 𝑹𝒎 b 𝒚𝒑 c 𝒕𝒑 c 𝒕𝐬 c


(dB) (dB) (samples) (s) (s)
A -31.31 -30.89 -38.87 13.01 0.88 1.22 0.52 1.23
B -37.69 -37.52 -54.03 9.11 0.8 1.19 0.52 1.51
C -41.67 -41.7 -82.78 5.99 0.74 1.15 0.5 1.98
D -52.46 -52.49 -178.06 4.71 0.69 1.1 0.43 2.8
E -97.16 -97.16 -391.53 12.23 0.87 1.22 0.53 1.25
F -97.16 -97.16 -391.53 12.23 0.87 1.02 1.25 0.9
G 0.32 -13.38 -55.16 9.01 0.8 1.2 0.53 1.62
H 2.2 -22.98 -355.42 17.44 0.91 1.23 0.53 1.2
a. Group delay of 𝐻 (𝜔) at the dc limit, for lag/lead components (i.e. PD+F terms only; I terms in configs G & H are ignored in this calculation). A low-frequency phase-lead is applied in all cases.
𝑐

b. Stability margins: For 𝜙 = 60°; 𝐾 is the gain margin (a linear factor); 𝑅 is the modulus margin. 𝑅 = 1⁄‖𝑆 (𝜔)‖ , i.e. the shortest distance from the -1 point to the 𝐿(𝜔) curve in a Nyquist diagram.
𝑚 𝑚 𝑚 𝑚 ∞

c. Performance parameters: Transient response of 𝐻 (𝑧), for a unit step input starting at 𝑡 = 0; 𝑦 is the peak output at 𝑡 ; 𝑡 is the settling time. Configs. G & H are tuned for ramp reference tracking.
𝑟𝑦 𝑝 p s

Fig. 4. Frequency response of controller 𝐻𝑐 (𝑧) = 𝐻𝑦 (𝑧)𝐻𝑒 (𝑧) . High mid-


Fig. 2. Magnitude response of 𝐿(𝜔) and 𝑆(𝜔). Magnitudes at 𝜔g and 𝜔c are
frequency gain is essential to provide the required low-frequency phase lead
marked using circles and squares respectively. For step-input tracking, using
for the desired level of performance and robustness. For low-order controls
𝜔g ≪ 𝜔c permits the formation of a taller sensitivity peak, to offset cuts at low
(𝐾 = 2) this elevated gain extends over the entire high-frequency noise band
frequencies.
(lines A & B). Using high-order noise-minimized controls (𝐾 > 2) attenuates
this superfluous high-frequency gain in different ways (lines C & D).

Fig. 3. Magnitude response of 𝐻𝜀𝑦 (𝜔) (solid lines) and 𝐻𝜀𝑢 (𝜔) (dotted lines).
Inset shows the low-frequency region around 𝑇(𝜔g ) = 1. Immunity (to noise,
in the mid-to-high frequency range) is improved by increasing 𝑝𝑓 & 𝐾. Using Fig. 5. Magnitude response of 𝐿(𝜔), 𝑆(𝜔) & 𝑇(𝜔). Magnitudes at 𝜔g and
𝐾 = 4 places a complex pair of closed-loop zeros on the unit circle in the noise 𝜔c are marked using circles and squares respectively. For ramp-input tracking,
band (line D). using 𝜔c ≪ 𝜔g allows wider and deeper sensitivity cuts at low frequencies, to
be offset by the formation of a broader peak which degrades immunity, rather
than a taller peak, which would degrade robustness.
To appear in Proc. 2nd IEEE Conf. on Norbert Wiener in the 21st Century, Melb. Australia, July 13-15, 2016.

IV. COMPARISON, DISCUSSION AND CONCLUSION and loop shaping,” IEEE Control Syst. Mag., vol. 27, no. 1, pp. 44-54,
Feb. 2007.
Given the elegance of the PRM concept [22], it is surprising [5] R. Zanasi and S. Cuoghi, “Direct method for digital lead-lag design:
that so many instantiations result in complicated expressions analytical and graphical solutions,” in Proc. IEEE Conf. Autom. Sci. Eng.,
involving transcendental and/or irrational functions that do not pp. 804-809, 24-27 Aug. 2011.
necessarily have simple closed-form solutions [23][24][25][26]. [6] H.L. Kennedy, “An Adaptive Digital Filter for Noise-Attenuation in
This is due to the use of analytic/geometric reasoning that begins Sampled Control Systems,” Int. J. Adapt. Control Signal Process., 2015
with a model of a particular mathematical form (to appear).
[25][26][27][28]. The first-order lag-system with dead time is a [7] D.Y.Abramovitch, “A unified framework for analog and digital PID
controllers,” in Proc. IEEE Conf. Control Appl., pp.1492-1497, 21-23
common choice due to its simplicity and ubiquity, i.e. 𝐻𝑢 (𝑠) = Sep. 2015.
𝐾𝑢 𝑒 −𝑠𝜏𝑢 ⁄(𝑠 + 𝑝𝑢 ) [27][28]; with unstable forms the focus of
[8] T.B. Sekara and M.R. Matausek, “Optimization of PID Controller Based
more recent research [26]. For this plant (with 𝐾𝑢 = 10.0, 𝑝𝑢 = on Maximization of the Proportional Gain Under Constraints on
1.0, 𝜏𝑢 = 0.01and 𝑇𝑠 = 0.01) the proposed method (with an Robustness and Sensitivity to Measurement Noise,” IEEE Trans. Autom.
integrator) was compared other PRMs; namely [23][27][28] and Control, vol. 54, no. 1, pp. 184-189, Jan. 2009.
the “modified Zeigler-Nichols” method in [22]. All methods [9] A.D. Micić and M.R. Mataušek, “Optimization of PID controller with
gave reasonable results; however, the following points are worth higher-order noise filter,” J. Process Control, vol. 24, no. 5, pp. 694-700,
May 2014.
noting: 1) The other PRMs do not include a filter for increased
immunity [22]-[28]; however, all design processes could [10] B. Kristiansson and B. Lennartson, “Robust tuning of PI and PID
controllers: using derivative action despite sensor noise,” IEEE Control
conceivably operate on a modified plant 𝐻 ̅𝑢 (𝑠) = 𝐹(𝑠)𝐻𝑢 (𝑠), Syst. Mag., vol. 26, no. 1, pp.55-69, Feb. 2006.
where 𝐹(𝑠) is a low-pass unit that is realized inside the [11] Katsuhiko Ogata, Discrete-Time Control Systems, 2nd Ed., Englewood
controller. 2) All of the other PRMs cited above are designed in Cliffs, N.J. Prentice Hall, 1995.
continuous time [22]-[28]. After discretization, the achieved [12] I. Horowitz, “Survey of Quantitative Feedback Theory (QFT),” Int. J.
phase margin is only approximately equal to the specified phase Control, vol. 53, no. 2, pp. 255–291, 1991.
margin (60° in all cases); however, the discrepancy decreases as [13] G. Stein, “Respect the unstable,” IEEE Control Syst. Mag., vol. 23, no. 4,
𝑇𝑠 decreases. 3) The internal-model-control method in [28] pp. 12-25, Aug. 2003.
sacrifices robustness and immunity to produce a high- [14] D.Y. Abramovitch, “Trying to keep it real: 25 Years of trying to get the
performance controller with a very fast response; however, due stuff I learned in grad school to work on mechatronic systems,” in Proc.
IEEE Conf. Control Appl., pp. 223-250, 21-23 Sep. 2015.
its relationship with the Smith predictor it is able to produce
[15] M. Green, D.J.N Limebeer, Linear robust control, Englewood Cliffs, N.J.,
stable controllers in cases where 𝜏𝑢 ⁄𝑇𝑠 ≫ 1. 4) The proposed Prentice Hall, 1995.
method is not restricted to plants of a particular form. Other [16] M. Morari and E. Zafiriou, Robust process control, Englewood Cliffs,
(second-order) plants were also examined, e.g. the double N.J., Prentice Hall, 1989.
integrator, a low-frequency oscillator, an unstable inverted [17] J.C. Doyle, B.A. Francis and A.R. Tannenbaum, Feedback Control
pendulum model, and a stable plant with repeated real poles. In Theory, Dover Publications, Mineola, N.Y., 2009.
all cases, for all combinations of the 𝑇𝑠 = 0.01 & 𝑇𝑠 = 0.01 and [18] H. Kwakernaak, “Minimax frequency domain performance and
𝑝𝑢 = 10.0 & 𝑝𝑢 = 0.01 parameters, stable controls were robustness optimization of linear feedback systems,” IEEE Trans. Autom.
attained after careful analysis and tuning. 5) Like the proposed Control, vol. 30, no. 10, pp. 994-1004, Oct. 1985.
method, none of the other PRMs guarantee stability; however, [19] M. Grimble, Robust industrial control systems : optimal design approach
this is the price that is paid for simplicity. Imposing stability for polynomial systems, Hoboken, NJ, Wiley, 2006.
constraints on the solution greatly complicates the problem [26]; [20] A.S. Hauksdottir and S.P. Sigurosson, “The matching coefficients PID
controller,” in Proc. Am. Control Conf., pp.710-715, Jun. /Jul. 2011.
however, multiple iterations are not required to find a stable
[21] P.-O. Larsson and T. Hagglund, “Control signal constraints and filter
causal controller. order selection for PI and PID controllers,” in Proc. Am. Control Conf.,
In summary, it would appear that the proposed method, pp. 4994-4999, Jun./Jul. 2011.
inspired by pioneering work of Norbert Wiener, is the perfect [22] K. Astrom and T. Hagglund, “PID Controllers: Theory, Design, and
Tuning”, 2nd Ed., Instrum. Soc. Am., Research Triangle Park, NC., 1995.
compromise between the flexibility and rigour of optimal ℋ2
[23] Qing-Guo Wang, Ho-Wang Fung, Yu Zhang, “PID tuning with exact gain
methods on the one hand, and the simplicity and practicality of and phase margins,” ISA Trans., vol. 38, no. 3, Jul. pp. 243-249, 1999.
PRMs on the other. Optimal continuous-time design techniques
[24] L. Ntogramatzidis and A. Ferrante, “Exact tuning of PID controllers in
usually focus on the robustness/performance trade-off; however, control feedback design,” IET Control Theory Appl., vol. 5, no. 4, pp.
the issue of (noise) immunity cannot be ignored in discrete-time 565-578, Mar. 2011.
systems. The proposed method specifically addresses this issue. [25] Wuhua Hu, Gaoxi Xiao, Xiumin Li, “An analytical method for PID
controller tuning with specified gain and phase margins for integral plus
REFERENCES time delay processes,” ISA Trans., vol. 50, no. 2, pp. 268-276, Apr. 2011.
[1] N. Wiener, Cybernetics: Or Control and Communication in the Animal [26] Q.B. Jin, Q. Liu, B. Huang, “New results on the robust stability of PID
and the Machine. Paris, MIT Press 1948. controllers with gain and phase margins for UFOPTD processes,” ISA
[2] J.R. Huey and W. Singhose, “Design of proportional-derivative feedback Trans., 2015 (to appear).
and input shaping for control of inertia plants,” IET Control Theory Appl., [27] Weng Khuen Ho, Chang Chieh Hang, Lisheng S. Cao, “Tuning of PID
vol. 6, no. 3, pp. 357-364, Feb. 2012. controllers based on gain and phase margin specifications,” Automatica,
[3] T. Hägglund, A unified discussion on signal filtering in PID control, vol. 31, no. 3, pp. 497-502, Mar. 1995.
Control Eng. Pract., vol. 21, no. 8, pp. 994-1006, Aug. 2013. [28] Ibrahim Kaya, “Tuning PI controllers for stable processes with
[4] W.C. Messner, M.D. Bedillion, Lu Xia and D.C. Karns, “Lead and lag specifications on gain and phase margins,” ISA Trans., vol. 43, no. 2, pp.
compensators with complex poles and zeros design formulas for modeling 297-304, Apr. 2004.

View publication stats

Anda mungkin juga menyukai