gain scheduling

59
Gain scheduling Robust Design and Automated Tuning of Automotive Controllers Ir. G.J.L. Naus 2009.106 26 th of October, 2009 WP3 - robust, low-complexity controller design Supervisors: Prof.dr.ir. M. Steinbuch (TU/e) Dr.ir. M.J.G. v.d. Molengraft (TU/e) Dr.ir. R. Huisman (DAF) Ir. J. Ploeg (TNO) TNO Science and Industry, Helmond (TNO) Business Unit Automotive Department of Advanced Chassis and Transport Systems DAF Trucks N.V., Eindhoven (DAF) Product Development Technical Analysis group University of Technology Eindhoven (TU/e) Department of Mechanical Engineering Division Dynamical Systems Design Control System Technology group

Upload: sabiju-balakrishnan

Post on 20-Jul-2016

9 views

Category:

Documents


1 download

DESCRIPTION

gain scheduling and advanced automotive controls.

TRANSCRIPT

Page 1: Gain Scheduling

Gain schedulingRobust Design and Automated

Tuning of Automotive Controllers

Ir. G.J.L. Naus

2009.10626th of October, 2009

WP3 - robust, low-complexity controller design

Supervisors:Prof.dr.ir. M. Steinbuch (TU/e)Dr.ir. M.J.G. v.d. Molengraft (TU/e)Dr.ir. R. Huisman (DAF)Ir. J. Ploeg (TNO)

TNO Science and Industry, Helmond (TNO)Business Unit AutomotiveDepartment of Advanced Chassis and Transport Systems

DAF Trucks N.V., Eindhoven (DAF)Product DevelopmentTechnical Analysis group

University of Technology Eindhoven (TU/e)Department of Mechanical EngineeringDivision Dynamical Systems DesignControl System Technology group

Page 2: Gain Scheduling

2

Page 3: Gain Scheduling

Contents

1 Introduction 11.1 Background and goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Gain scheduling 32.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1.1 Classical gain scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.1.2 LPV and LFT synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.1.3 Fuzzy gain scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.2 Gain scheduling vs Robust control . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Parameter-dependent plant models 73.1 linearization-based scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83.2 Off-equilibrium linearizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113.3 Quasi-LPV approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123.4 linear, parameter-dependent plant . . . . . . . . . . . . . . . . . . . . . . . . . . . 133.5 Linear fractional transformation description . . . . . . . . . . . . . . . . . . . . . . 143.6 Conclusions / recap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Classical gain scheduling 174.1 LTI controller design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.1.1 Direct gain scheduling controller design . . . . . . . . . . . . . . . . . . . . 184.2 Gain scheduling controller design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.2.1 Linearization scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194.2.2 Interpolation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214.2.3 Velocity-based scheduling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4.3 Hidden coupling terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244.4 Stability properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5 LPV controller synthesis 275.1 General LPV controller synthesis setup . . . . . . . . . . . . . . . . . . . . . . . . . 28

5.1.1 Stability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295.1.2 Performance analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305.1.3 General synthesis problem formulation . . . . . . . . . . . . . . . . . . . . . 31

5.2 Lyapunov-based LPV control synthesis . . . . . . . . . . . . . . . . . . . . . . . . . 325.2.1 Gridding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325.2.2 Polytopic approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

5.3 Scaled small-gain or LFT synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 335.3.1 LFT problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335.3.2 Controller synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345.3.3 Pros and cons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.4 Mixed LPV-LFT approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

3

Page 4: Gain Scheduling

4 CONTENTS

5.4.1 Extended Kalman-Yakubovich-Popov (KYP) lemma . . . . . . . . . . . . . 365.4.2 LFT Lyapunov functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

6 Fuzzy gain-scheduling 396.1 Fuzzy modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

6.1.1 Weighting or scaling functions . . . . . . . . . . . . . . . . . . . . . . . . . 406.1.2 Approximation accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

6.2 Fuzzy gain scheduling control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436.2.1 Parallel Distributed Control . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

7 Conclusions and future work 457.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457.2 Future work (tbd) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

Bibliography 52

A Longitudinal vehicle control 53

B Engine - idle governor 55

Page 5: Gain Scheduling

Chapter 1

Introduction

In this literature survey, an overview of gain-scheduling syntheses is put together. Many differentnotions can be viewed as Gain Scheduling (GS). For example switching or blending of gain values ofcontrollers or models or switching or blending of complete controllers or model dynamics accordingto different operating conditions, or according to preset times. As the terms switching and blendingalready indicate, GS may either involve continuous or discrete scheduling of controllers or modeldynamics.

In general, gain-scheduling encompasses the attenuation of nonlinear dynamics over a range ofoperations, the attenuation of environmental time-variations or the attenuation of parameter vari-ations and uncertainties. Classical gain scheduling involves offline linearization of nonlinear systemdynamics at multiple operating conditions, which are parameterized by a so-called scheduling vari-able θ, and the design of corresponding linear controllers at each point. Next, online scheduling ofcontroller gains is established, based on θ = θ(t) to reflect the nearby operating condition. The in-tention is to extend the reach of a single linearization-based control design over an entire operatingenvelope. In general, the scheduling variable is time-varying and may either be an internal plantvariable (or a function of internal plant variables) called endogenous variables, or an externallyprescribed exogenous variable.

Consequently, difficulties regarding stability and performance requirements may arise whenthe LTI controller designs are implemented for θ = θ(t). More recent LPV and LFT approacheshence take into account the time-varying nature of θ = θ(t) in the controller design. Guaranteedglobal stability and performance requirements can thus be given a priori. This survey encompassesclassical gain-scheduling syntheses as well as more recent LPV and LFT approaches.

1.1 Background and goals

Background

This report is based on the conclusions of (Naus 2007b, Naus 2007a), which discuss the controllerdesign and tuning at Integrated Safety, Business Unit Automotive, TNO Helmond, and DAFTrucks N.V. respectively. Essentially, the main problems regarding controller design at DAF andTNO comprise i) a lack of proper system identification and corresponding modeling and ii) thelack of appropriate performance specifications. In some cases, application of system identification,subsequent modeling and corresponding local controller design is applied. However, schedulingof the resulting controller gains or switching between various controllers is commonly applied.This is done by experience and insight in the system. Stability issues as well as performance androbustness specifications are not considered. Hence, insight in gain scheduling and related controlsyntheses is demanded.

At DAF, gain scheduling is adopted in many practical applications. In all cases, the schedulingis employed in an ad hoc manner, not considering stability issues or performance and robustnessspecifications. Besides employing gain scheduling as a suitable control synthesis for nonlinear

1

Page 6: Gain Scheduling

2 CHAPTER 1. INTRODUCTION

systems, gain scheduling is also adopted for closed-loop performance improvements of linear plantsby scheduling the controller gains on the basis of e.g. a tracking error. At TNO, discrete schedulingor switching as well as smooth scheduling is employed in an ad hoc fashion. The tuning and designof the controller (gains) corresponding to specific working conditions involves much trial-and-error.Furthermore, recent application of Model Predictive Control (MPC) in an explicit fashion (Naus,van den Bleek, Ploeg, Scheepers, van de Molengraft & Steinbuch 2008) yields a gain-scheduledcontroller. The interpretation of this controller and corresponding tuning require more insight ingain scheduling and related control syntheses.

Goals

The main goals of this report involve two issues.

• The report provides an overview of gain scheduling and related controller design syntheses.Available gain scheduling techniques, and guidelines for appropriate application of thesetechniques are discussed.

• Based on the conclusions regarding the application of gain scheduling, goals for furtherresearch have to be defined within the project robust design and automated tuning of auto-motive controllers. Appropriate syntheses and research directions to elaborate on have to bedefined, focusing on generic design and tuning methodologies for applications at DAF andTNO.

1.2 Outline

To start with, an overview of gain scheduling techniques is given (Chapter 2). Parameter-dependent plant models, which serve as a basis for gain scheduling controller design, are discussedin Chapter 3. Next, each of the main classes of gain scheduling techniques is discussed in detail;classical gain scheduling (Chapter 4), LPV and LFT control synthesis (Chapter 5) and fuzzy gainscheduling (Chapter 6). In Chapter ??, application examples, focusing on the automotive industryare discussed. Finally, conclusions and future work regarding the applicability of gain schedulingtechniques in the remainder of this project are given (Chapter 7). The future work will be deter-mined in cooperation with DAF and TNO. Throughout this report, two example case studies areused; the problem of designing a longitudinal vehicle speed controller, i.e. cruise controller, andan engine example involving the design of a low idle governor. These case studies are defined inAppendix A and B respectively.

Page 7: Gain Scheduling

Chapter 2

Gain scheduling

2.1 Overview

Different gain scheduling approaches can be distinguished, which may be classified in differentways. To start with, gain scheduling methods can be divided into i) methods decomposing non-linear design problems into linear sub-problems and ii) methods decomposing nonlinear designproblems into simpler nonlinear sub-problems or affine sub-problems. (Leith & Leithead 2000)give an overview of the main theoretical results and design procedures relating to (continuous)gain-scheduling control in the sense of decomposing nonlinear design problems into linear sub-problems. Focus will lie on the former one, as linear system and corresponding control theory iswell developed.

Furthermore, gain scheduling methods decompose into i) continuous gain-scheduling methodsand ii) discrete, i.e. hybrid or switched gain-scheduling methods. Discrete or hybrid in this senseinvolves the switching of a system or controller between dynamical regimes. E.g. friction modelsthat have a clear distinction between stick and slip phases, backlash in gears, dead zones in cogwheels; phenomena like saturation, hysteresis, sensor and actuator failures. Hybrid systems inthe sense of combining digital controllers with physical processes is not meant here (Schutter &Heemels 2004).

Finally, three main approaches can be distinguished regarding gain-scheduling; i) so-calleddivide and conquer techniques, or the classical gain-scheduling approach, more recent LPV andLFT syntheses and thirdly, fuzzy techniques. A brief overview of each of these approaches is givennext.

2.1.1 Classical gain scheduling

Generally, classical gain scheduling involves the decomposition of the design of a nonlinear con-troller into the design of a number of linear controllers. Hence, well-established linear controldesign methods may be applied without restriction to a certain method, as opposed to nonlinearmethods. No guarantees on the robustness, performance or even nominal stability of the closed-loop gain-scheduled design can be given however. Classical gain scheduling comprises four mainsteps (Rugh & Shamma 2000).

Step 1 A family of LTI approximations of a nonlinear plant at constant operating points (equilib-ria), parameterized by constant values of convenient plant variables or exogenous signals θ iscomputed using linearization-based scheduling. Jacobian linearization of the nonlinear plantabout a manifold of constant equilibria, constant operating points or setpoints is employed.The linearizations have to correspond to zero error. This yields a parameterized family oflinearized plants. Subsequent implementation of the controller requires θ to be a measurablevariable. Other syntheses to derive a parameter-dependent model are

3

Page 8: Gain Scheduling

4 CHAPTER 2. GAIN SCHEDULING

• off-equilibrium or velocity-based linearizations, e.g. when zero error equilibrium pointsor working conditions are not present

• a so-called quasi-LPV approach in which the plant dynamics are rewritten to distinguishnonlinearities as time-varying parameters that are used as scheduling variables

• direct LPV modeling, based on a linear plant incorporating time-varying parameters,i.e. when no nonlinear plant is involved; this also includes black-box or data-basedmodeling methods

Step 2 LTI controllers corresponding to the previously derived set of local LTI models are de-signed, to achieve specified performance and stability at each operating point. Hence, the LTIcontrollers in combination with the corresponding LTI models have desired properties. Theresulting set of controllers is also parameterized by θ. Although θ actually is time-varying,i.e. θ = θ(t), classical design methods are based on fixed or frozen scheduling parametervalues θ. To enable subsequent scheduling, interpolation of e.g. controller coefficients, theset of LTI controllers requires fixed-structure controller designs. Two exceptions are

• in case direct derivation of a Linear Parameter Varying (LPV) controller for a corre-sponding LPV plant model is possible, subsequent scheduling, interpolation becomessuperfluous.

• when discrete or hybrid scheduling instead of continuous scheduling is demanded, theset of controller designs not necessarily need to be fixed-structured.

Step 3 Implementation of the family of LTI controllers such that the controller coefficients arescheduled according to the current value of the scheduling variable, e.g. by controller gaininterpolation or scheduling. At this point, θ = θ(t) is implemented. At each operating point,the scheduled controller has to linearize to the corresponding linear controller design as wellas provides a constant control value yielding zero error at these points. As mentioned inStep 2, in case of direct scheduling, this step becomes superfluous. Furthermore, in caseof discrete scheduling, the implementation of the LTI controllers involves the design of ascheduled selection procedure that is applied to the set of LTI controllers, rather than thedesign of a family of scheduled controllers. The presence of hidden coupling terms is animportant aspect, which yields various additional requirements to the scheduling procedure.

Step 4 Typically, local performance assessment can be performed analytically, whereas assess-ment of global performance and robustness has to be established by extensive simulations.Nonlocal performance of the gain scheduled controller is evaluated and checked by simula-tions.

2.1.2 LPV and LFT synthesis

LPV and LFT syntheses are based on LPV and LFT plant representations respectively. Bothmethods yield direct synthesis of a controller utilizing (l2-)norm based methods, with guaranteeson the robustness, performance and nominal stability of the overall gain-scheduled design (Shamma& Athans 1990)). LPV and LFT syntheses essentially involve only two main steps.

Step 1 The first step corresponds to the classical approach. A family of LTI approximations of anonlinear plant at constant operating points (equilibria), parameterized by constant values ofconvenient plant variables or exogenous signals θ is computed. Subsequent implementation ofthe controller requires θ = θ(t) to be a measurable variable. Besides the already mentionedmethods, which all arrive at Linear Parameter-Varying (LPV) models, in specific cases aLFT description is possible. The LFT description serves as a basis for subsequent LFTcontroller synthesis.

Page 9: Gain Scheduling

2.2. GAIN SCHEDULING VS ROBUST CONTROL 5

Step 2 LPV and LFT control synthesis directly yield a gain-scheduled controller. Stability andperformance specifications can be guaranteed a priori as the time-varying parameter θ(t) in-stead of its corresponding frozen-value θ is addressed in the design process. Only continuousgain-scheduling is considered.

2.1.3 Fuzzy gain scheduling

Fuzzy gain scheduling should overcome the disadvantage of classical gain scheduling regarding therestriction of stability and performance analysis to local rather than global closed-loop behavior.The corresponding fuzzy modeling considers the transient dynamics of the original nonlinear modelinstead of local linearizations only. Fuzzy gain scheduling techniques may involve classical gainscheduling alike as well as LPV techniques. Focusing on the former one, four main steps have tobe considered.

Step 1 Analogous to classical gain scheduling, sets of local LTI models and corresponding LTIcontrollers have to be designed. Focus lies on the regions of the envelope of operatingconditions for which these controllers assure stability and desired performance of the resulting(local) closed-loop systems.

Step 2 To arrive at a fuzzy model, so-called weighting or scaling functions are designed, cor-responding to the before mentioned regions. Utilizing these weighting functions, the localmodels are blended. Specifying a specific approximation accuracy of the fuzzy model withrespect to the original nonlinear model, yields the required number of local models.

Step 3 The set of local controllers is blended analogous to the set of local models. The sameweighting functions are utilized. The blending yields scheduling of the controller outputsrather than scheduling of the controllers or controller coefficients. Hence, members of thecorresponding parameterized set of LTI controllers do not necessarily need to have fixedstructure and dimension.

Step 4 Stability and performance are established by extensive simulations, analogous to classicalgain scheduling. However, in the case of fuzzy controller design, global as well as localspecifications have to be derived from simulations, as the characteristic dynamics of thefuzzy model can not be related to the dynamics of the set of local models.

2.2 Gain scheduling vs Robust control

An important issue in linear controller design involves the robustness and performance of theclosed-loop system in the presence of uncertainty. System uncertainty essentially is of two types; i)dynamical uncertainty, which corresponds to neglected plant dynamics (high-frequency behaviour,nonlinearities, etc.), and ii) constant parametric uncertainty, which results either from inaccurateknowledge of the value of physical parameters or from variations of this value during operation.RobustH∞ theory may be adopted for the former one and µ-synthesis, Lyapunov-based techniquesfor the latter one (Apkarian & Gahinet 1995).

LPV plants can be viewed as i) Linear Time-Invariant (LTI) plants subject to a time-varyingparametric uncertainty θ(t) or ii) as models of linear time-varying plants, either derived as a non-linear model or a set of linear models describing a nonlinear plant. When θ(t) can be measured, thelatter idea enables exploiting the available measurements of θ(t) in an appropriate control strategy.Assuming that the uncertainties are constant or only slowly varying, the former problem yieldsLTI robust control techniques (Apkarian, Gahinet & Becker 1995). Especially in case of largeparameter variations, a single robust controller can be very conservative and plant stabilization bya single LTI controller may not even be feasible. Assuming that the varying parameter θ(t) can bemeasured during operation, gain-scheduling may provide less conservative solutions. Essentially,no a priori knowledge about θ(t) is required but for its range of operation and its rate of variationθ(t).

Page 10: Gain Scheduling

6 CHAPTER 2. GAIN SCHEDULING

Considering gain-scheduled controllers, the controllers are a function of the scheduling variableθ(t). A gain-scheduled controller thus can adjust itself according to the changes in the dynamicsof the plant. As a result, a gain-scheduled controller usually is less conservative than a robustone. Combinations, i.e. robust gain-scheduled controllers exist as well. The parameter θ(t) isassumed to be divided into θ(t) = [θm(t) θu(t)]T where θm(t) is measured in real-time and θu(t)is uncertain. The gain-scheduling part schedules based on θm(t). The controller now adapts tochanges in the plant dynamics due to θm(t) and is robust for changes in θu(t) (Bianchi, Mantz &Christiansen 2007).

The other way to deal with θ(t) is to consider it as a time-varying uncertainty. In this sense, e.g.(Zhou, Khargonekar, Stoustrup & Niemann 1992) present the synthesis of a strongly robust H∞performance criterion, i.e. robust performance for systems with time varying uncertainties. Morerecent LPV and LFT control synthesis techniques elaborate on this idea using i) parameter-varyingLyapunov functions and ii) scaled small-gain theorems.

Page 11: Gain Scheduling

Chapter 3

Parameter-dependent plantmodels

In general, gain scheduling is considered as a controller synthesis to design controllers for a nonlin-ear plant. The general approach assumed in most literature is to start with a whitebox nonlinearmodel Σnl, describing the nonlinear plant dynamics.

Σnl :

x(t) = f(x(t), u(t), w(t))z(t) = g(x(t), u(t), w(t))y = h(x(t), w(t))

(3.1)

where x(t) is the state of the system, u(t) is the input, z(t) denotes an error signal to be controlled,y(t) denotes the measured output, available to the controller, e.g. penalized variables, trackingcommands, state variables, and w(t) represents external inputs such as reference commands, dis-turbances and noise. When such a model is not available however, a set of linear point-designblack-box models may be used instead. The point-designs should cover the operating range of thenonlinear plant dynamics.

Conceptually, gain-scheduling is based on a linear, time-invariant, parameter-dependent plant.A linear approximation Σl of the nonlinear model Σnl or the above mentioned set of linear point-designs can be used. The linear approximation is parameterized by the scheduling variable θ fora corresponding operating envelope θ ∈ Θ ⊂ Rnθ , where nθ the length of the parameter vectorθ. This yields a family of linear plant models Σl(θ), describing the nonlinear plant dynamics(furtheron, the subscript l, denoting a linear model will be omitted). The corresponding family oflinear parameter-varying plant models Σ(θ) = Σl(θ) is given by

Σ(θ) :

x = A(θ)x+B1(θ)w +B2(θ)uz = C1(θ)x+D11(θ)w +D12(θ)uy = C2(θ)x+D21(θ)w

θ ∈ Θ (3.2)

Throughout the controller synthesis, θ is regarded as a constant variable. However, when imple-menting the resulting gain scheduled controller, θ = θ(t) is used instead.

The model Σ(θ) (3.2) represents a family of linear parameter-varying plant models rather thana single dynamic system. A true Linear Parameter Varying (LPV) model is defined as

Σ(θ(t)) :

x = A(θ(t))x+B1(θ(t))w +B2(θ(t))uz = C1(θ(t))x+D11(θ(t))w +D12(θ(t))uy = C2(θ(t))x+D21(θ(t))w

θ(t) ∈ Θ (3.3)

A LPV model comprises linear, parameter-dependent dynamics. The model matrices are fixedfunctions of some vector of varying parameters θ(t) (Shamma & Athans 1992). As these parametersare time-varying, the model becomes time-varying. Consider a parameterized family of Jacobian

7

Page 12: Gain Scheduling

8 CHAPTER 3. PARAMETER-DEPENDENT PLANT MODELS

linearizations, i.e. local linear plant models. The synthesis of such a family is called linearization-based scheduling (Shamma & Athans 1992). Correspondingly implementing θ = θ(t), in practiceyields an LPV model. Possible variations in the state, input and / or output transformationsof the family members however, are neglected. This may yield closed-loop performance or evenstability problems when implementing the corresponding controller. Different approaches enablingthe design of true LPV models are exact linearization or quasi-LPV techniques (Shamma & Athans1992, Rugh 1991), velocity-based scheduling techniques (Leith & Leithead 2000), or parameter-dependencies arising in a linear plant (Shahruz & Behtash 1990).

In general, three types of linear, parameter-dependent models can be distinguished

• Linear Parameter Varying (LPV) models Σ(θ(t))

• Linear Fractional Transformation (LFT) models ΣLFT (θ(t))

• Polytopic Linear Modeling (PLM) or fuzzy models ΣPLM (θ(t))

Correspondingly, three main methods to derive a linear parameter-dependent model of a nonlin-ear plant can be distinguished. In next sections, the above mentioned methods to derive LPVmodels and subsequently LFT models are presented. The fuzzy modeling is discussed separatelyin Chapter 6.

3.1 linearization-based scheduling

The classical approach, using Jacobian linearization of the nonlinear model about a manifold ofconstant equilibria, constant operating points or setpoints, is called linearization-based scheduling.When a corresponding scheduling variable θ is chosen appropriately to parameterize the set oflinear models, a parameterized family of linearized models representing the original nonlinearmodel results.

Linearization theory

Consider the nonlinear plant dynamics, described by the nonlinear model Σnl (3.1). Define theoperating envelope of the plant Rx. An equilibrium or constant operating point (xo, uo, wo) ∈Ro ⊂ Rx is defined by f(xo, uo, wo) = 0, where f(∗) as defined by (3.1). An equilibrium familyand the corresponding error and output equilibrium families are defined as

Ro :

0 = f(xo, uo, wo)zo = g(xo, uo, wo)yo = h(xo, wo)

(3.4)

Assuming f(x,w, u) is continuously differentiable at (xo, uo, wo) ∈ Ro, the nonlinear model isapproximated by

Σo :

x = fo +∇xf |Ro(x− xo) +∇wf |Ro(w − wo) +∇uf |Ro(u− uo) + rf (x,w, u)z = go +∇xg|Ro(x− xo) +∇wg|Ro(w − wo) +∇ug|Ro(u− uo) + rg(x,w, u)y = ho +∇xh|Ro(x− xo) +∇wh|Ro(w − wo) + rh(x,w, u)

(3.5)

where fo = f(xo, wo, uo), go = g(xo, wo, uo) and ho = h(xo, wo) respectively. The residual termsrf,g,h(x,w, u) satisfy

lim(x,w,u)→(xo,wo,uo)

rf,g,h(x,w, u)√|x− xo|2 + |w − wo|2 + |u− uo|2

= 0 (3.6)

and are assumed to be zero for simplicity.

Page 13: Gain Scheduling

3.1. LINEARIZATION-BASED SCHEDULING 9

Application of a coordinate transformation

x = xo + x, w = wo + w, u = uo + u, y = yo + y, z = zo + z (3.7)

and definition of the time-invariant system matrices

A = ∇xf |Ro B1 = ∇wf |Ro B2 = ∇uf |Ro

C1 = ∇xg|Ro D11 = ∇wg|Ro C12 = ∇ug|Ro

C2 = ∇xh|Ro D21 = ∇wh|Ro

(3.8)

yields the linearization family Σ, locally describing the nonlinear model (3.1)

Σ :

˙x = Ax+B1w +B2uz = C1x+D11w +D12uy = C2x+D21w

(3.9)

Scheduling variable θ

Define a scheduling variable θ ∈ Θ, parameterizing the envelope of working conditions Rx → Θand analogously Ro → Θo with Θo ⊂ Θ. This yields Ro = Ro(θ), where θ ∈ Θo. An equilibriumfamily and the corresponding error and output equilibrium families for any θ ∈ Θo are then definedby

Ro(θ) :

0 = f(xo(θ), uo(θ), wo(θ))zo(θ) = g(xo(θ), uo(θ), wo(θ))yo(θ) = h(xo(θ), wo(θ))

θ ∈ Θo (3.10)

Corresponding to the parameterized equilibrium values (xo(θ), wo(θ), uo(θ), zo(θ), yo(θ)) fol-lowing from (3.10), a linear parameter-dependent linearization family exists, locally describing thenonlinear model (3.1).

Σ(θ) :

˙x = A(θ)x+B1(θ)w +B2(θ)uz = C1(θ)x+D11(θ)w +D12(θ)uy = C2(θ)x+D21(θ)w

θ ∈ Θo (3.11)

Due to the parameterization of the equilibrium values, the system matrices are now parameterizedas well (compare to (3.8)).

As discussed in the introduction, in general, (3.11) describes the nonlinear plant only locally.Hence, Σ(θ), θ ∈ Θo is a set or family of point-designs rather than a true LPV model, see Figure3.1. A LPV model by definition, is valid for θ.

θ ∈ Θo

Θ ⊂ Rnθ

Figure 3.1: Operating envelope Θ ⊂ Rnθ and the set of equilibrium values or constant operatingpoints Θo, parameterized by θ.

Page 14: Gain Scheduling

10 CHAPTER 3. PARAMETER-DEPENDENT PLANT MODELS

In specific cases, the set of point-designs immediately yields an LPV model Σ(θ(t)), θ(t) ∈ Θ.In practice, this is commonly established by implementation of θ = θ(t), yielding

Σ(θ(t)) :

˙x = A(θ(t))x+B1(θ(t))w +B2(θ(t))uz = C1(θ(t))x+D11(θ(t))w +D12(θ(t))uy = C2(θ(t))x+D21(θ(t))w

θ(t) ∈ Θ (3.12)

Possible variations in the state, input and output transformations hence are neglected, which mayresult in closed-loop performance or even stability problems. In case subsequent classical gainscheduling controller design is applied, the set of linearized models Σ(θ), θ ∈ Θo may be usedas a starting point. In this sense, a set of linear, point-design black-box models resulting frome.g. frequency-response or step-response measurements, yields the same result. In case subsequentLPV control synthesis is applied, an LPV model is required.

A parameterized family of linearized models resulting from linearization-based scheduling or anumber of black-box point-designs are only locally valid. In case an LPV model is based on sucha set of linearized models, the accuracy of the resulting linear parameter-dependent model withrespect to the original nonlinear model or plant is unknown and generally follows from extensivesimulations.

Example 3.1 (Longitudinal vehicle control)Consider the nonlinear model (A.5) of the longitudinal vehicle control example (seeA). Linearizing (A.5) around the equilibrium operating point (xe, ue, we) yields

Re :

0 = − 1

MC ′dx2e − g(sinw(1)

e + Cr cosw(1)e ) + 1

M ue

ze = w(2)e − xe

ye = xe

(3.13)

The corresponding equilibrium families, and error and output equilibrium families aregiven by

Re :

xd,e = ye = xeze = 0

we =[w

(1)e , w

(2)e

]T= [αe, xe]

T

ue = C ′dx2e + g(sinw(1)

e + Cr cosw(1)e )

(3.14)

Define a scheduling variable θ = [α, x]T = [w(1), y]T ∈ Θ, where Θ represents theparameterized envelope of operating conditions. This yields

Re(θ) :

xd,e(θ) = ye(θ) := θ(2)

ze(θ) = ze = 0we(θ) = [αe(θ), xe(θ)]

T := θ

ue(θ) = C ′dx2e + g(sinw(1)

e + Cr cosw(1)e )

:= C ′dθ(2)2 + g(sin θ(1) + Cr cos θ(1))

θ ∈ Θe (3.15)

where Θe ⊂ Θ defines the parameterized equilibrium operating points of the nonlinearsystem (A.5). The corresponding parameterized set of linear (LPV) models Σ(θ) isdefined by

Σ(θ) :

˙x(t) = A(θ)x(t) +B1(θ)w(t) +B2u(t)z(t) = C1x(t) +D11w(t) +D12u(t)y(t) = C2x(t) +D21w(t)

θ ∈ Θe (3.16)

where the deviation variables

x(t) = x(t)− xe(θ), w(t) = w(t)− we(θ), u(t) = u(t)− ue(θ)z(t) = z(t)− ze(θ), y(t) = y(t)− ye(θ)

(3.17)

Page 15: Gain Scheduling

3.2. OFF-EQUILIBRIUM LINEARIZATIONS 11

and the parameterized system matrices

A(θ) = − 2MC ′dθ

(2) B1 =[g(Cr sin θ(1) − cos θ(1)) 0

]B2 = 1

MC1 = −1 D11 = [0 1] D12 = 0 C2 = 1 D21 = 0

(3.18)

3.2 Off-equilibrium linearizations

A disadvantage of classical linearization-based scheduling is the restriction to equilibrium-pointmodeling and corresponding controller design. Using so-called velocity-based or off-equilibriumlinearizations, (Leith & Leithead 1998b) enable linearization at every operating point. Rewritingthe nonlinear dynamics (3.1) such that the nonlinearities are incorporated in θ(x,w, u) explicitlyyields

Σnl(θ) :

x = Ax+B1w +B2u+ f(θ)z = C1x+D11w +D12u+ g(θ)y = C2x+D21w + h(θ)

θ = θ(t) ∈ Θ (3.19)

where A, B, C and D are constant matrices and f(θ), g(θ) and h(θ) are nonlinear functionsincorporating the model nonlinearities via θ(x,w, u), with ∂θ

∂x , ∂θ∂x and ∂θ

∂u constant. Consideringan operating point (xo, wo, uo) at which θo = θ(xo, wo, uo), the solution to the correspondingvelocity-based linearization becomes

Σ(θo) :

xo = ζoζo = (A+ ∂f

∂t (θo))ζo + (B1 + ∂f∂t (θo))wo + (B2 + ∂f

∂t (θo))uozo = (C1 + ∂g

∂t (θo))ζo + (D11 + ∂g∂t (θo))wo + (D12 + ∂g

∂t (θo))uoyo = (C2 + ∂h

∂t (θo))ζo + (D21 + ∂h∂t (θo))wo

(3.20)

which is linear. (Leith & Leithead 1998b) show that (3.20) approximates the solution to the nonlin-ear model (3.19) and thus (3.1) locally to an arbitrary operating point (xo, wo, uo). Consequently,there is a velocity-based linearization associated with every operating point of a nonlinear sys-tem and the solutions may be pieced together. The resulting velocity-based linearization family,parameterized by θ, which captures the nonlinearities of the model, globally approximates thesolution to the nonlinear model to an arbitrary degree of accuracy (Leith & Leithead 2000). As norestriction to equilibrium operating points is present, linear approximation of transient dynamicsand operating points far from equilibrium operating points is enabled. Hence, the velocity-basedlinearization family Σ(θ) is an alternative representation of the nonlinear model (3.1).

Σ(θ) :

x = ζ

ζ = (A+ ∂f∂t (θ))ζ + (B + ∂f

∂t (θ))w + (B2 + ∂f∂t (θ))u

z = (C1 + ∂g∂t (θ))ζ + (D11 + ∂g

∂t (θ))w + (D12 + ∂g∂t (θ))u

y = (C2 + ∂h∂t (θ))ζ + (D21 + ∂h

∂t (θ))w

θ = θ(t) ∈ Θ (3.21)

The velocity-based linearization family (3.21) may be adopted as a suitable basis for subsequentgain scheduling controller design (Leith & Leithead 1998c). Classical gain scheduling (Leith &Leithead 1998c), LPV synthesis (Leith & Leithead 2000) or fuzzy-alike blending techniques (Leith,Tsourdos, White & Leithead 2000) may be employed.

The parameterized velocity-based linearization family (3.21) represents the entire dynamics ofthe nonlinear model (3.1) without loss of information. Hence, the velocity-based linearization fam-ily can be regarded as an alternative representation of the nonlinear model, globally approximatingthe solution to the nonlinear model to an arbitrary degree of accuracy. In practice however, a smallset of velocity-based linearizations will be used. The corresponding parameterization θ ∈ Θo ⊂ Θwhere Θ represents the total envelope of working conditions parameterized by θ, and correspondinginterpolation functions are chosen such that a sufficiently accurate approximation of the nonlineardynamics over the envelope of working conditions is ensured. The use of interpolation functionsis comparable to fuzzy or Takagi-Sugeno modeling. However, in this case the dynamics of the

Page 16: Gain Scheduling

12 CHAPTER 3. PARAMETER-DEPENDENT PLANT MODELS

interpolated models are directly related to the dynamic characteristics of the set of local models,which is not the case with fuzzy modeling (Leith & Leithead 1999).

Due to the direct relation between the family of velocity-based linearizations and the originalnonlinear model, stability of the nonlinear model (3.1) can be related to stability conditionsconcerning the velocity-based linearizations (Leith & Leithead 2000, Leith & Leithead 1998b).Provided that the members of the velocity-based linearization family corresponding to a certainnonlinear model are uniformly stable and the rate of evolution of the nonlinear model is sufficientlyslow, the nonlinear model is BIBO stable for some restricted class of inputs and initial conditions.For systems where the slow variation condition is automatically satisfied, the class of allowableinputs and initial conditions is unrestricted and the stability analysis is global.

3.3 Quasi-LPV approach

Quasi-LPV scheduling tries to overcome the general shortcomings of classical gain-schedulingregarding local validity of the resulting linearized model via transformation of a nonlinear modelto an LPV form. The nonlinear terms are hidden by including them in the scheduling variable. Asthis involves a transformation rather than a linearization, the resulting LPV model exactly equalsthe original nonlinear model. LPV analysis and synthesis tools can be used, taking into accountthe presence of parameter variations. However, it is difficult to exploit the endogenous relation ofthe scheduling parameter y, which is the part of the state that accounts for the nonlinearities.

Rewriting the nonlinear model (3.1) correspondingly yields[yx′

]= f(y) +A′(y)

[yx′

]+B′1(y)w

′ +B′2(y)u′ (3.22)

where all nonlinearities are incorporated in f(y) and B1, B2 are assumed to be linear beforehand.Moreover, y is regarded as the scheduling variable. The corresponding equilibrium family is thusparameterized by y, yielding (x′o(y), w

′o(y), u

′o(y)) and is given by

0 = f(y) +A′(y)[

yx′o(y)

]+B′1(y)w

′o(y) +B′2(y)u

′o(y) (3.23)

yielding

f(y) +[A′(11)(y)A′(21)(y)

]y = −

[A′(12)(y)A′(22)(y)

]x′o(y)−B′1(y)w

′o(y)−B′2(y)u

′o(y) (3.24)

where A′(ij) denotes the ijth element of A′. Substitution of (3.24) in (3.22) yields[yx′

]= −

[A′(12)(y)A′(22)(y)

]x′o(y)−B′1(y)w

′o(y)−B′2(y)u

′o(y) + . . .

. . .+[A′(12)(y)A′(22)(y)

]x′ +B′1(y)w

′ +B′2(y)u′ (3.25)

=[A′(12)(y)A′(22)(y)

](x′ − x′o(y)) +B′1(y)(w

′ − w′o(y))−B′2(y)(u′ − u′o(y)) (3.26)

Application of a change of variables

x =[

yx′ − x′o(y)

], w = w′ − wo(y), u = u′ − u′o(y) (3.27)

with

x =[

yx′ − x′o(y)

]=

[y

x′ −∇yx′o(y)y

](3.28)

Page 17: Gain Scheduling

3.4. LINEAR, PARAMETER-DEPENDENT PLANT 13

then yields

x =[

0 A′(12)(y)0 A′(22)(y)

]x+

[0

−∇yx′o(y)y

]+B′1(y)w +B′2(y)u (3.29)

= A(y)x+B1(y)w +B2(y)u (3.30)

where

A(y) =[

0 A′(12)(y)0 A′(22)(y)−∇yx′o(y)A′(12)(y)

](3.31)

B1(y) =

[B′(1)1 (y)

B′(2)1 (y)−∇yx′o(y)B

′(1)1 (y)

](3.32)

B2(y) =

[B′(1)2 (y)

B′(2)2 (y)−∇yx′o(y)B

′(1)2 (y)

](3.33)

If all nonlinearities are incorporated in f(y) and no nonlinearities concerning x′ are present,the transformation is exact and (3.22) equals (3.30), taking into account the change of variables(3.27). If, more generally the nonlinearities aren’t exactly in y, a residual term rf (x) with dynamicsremains, yielding

x = A(y)x+B1(y)w +B2(y)u+ rf (x) (3.34)

where

limx′→x′

o(y)

|rf (x)||x′ − x′o(y)|

= 0 (3.35)

The subsequent controller synthesis is based on all time-varying plants (3.30) where y = y(t).In practice, however, this introduces additional dynamics, which may not be present or evenfeasible in case of e.g. rapidly changing y(t), for the original model (3.22). Hence, the resultingcontroller synthesis will be conservative as it takes into account all y = y(t).

3.4 linear, parameter-dependent plant

When the starting point is a linear plant with one or several time-varying parameters ratherthan a nonlinear plant, appropriate selection of the scheduling variables directly yields an LPVmodel. Physical nonlinear modeling of the process may be omitted in case black-box models ormeasurements are available. Analogous to linearization-based scheduling, based on for example aset of frequency-response or step response measurements, which are parameterized by θ ∈ Θo, afamily of linear, parameter-dependent models Σ(θ), θ ∈ Θ follows directly (e.g. (Nichols, Reichert& Rugh 1993)).

Example 3.2 (Engine system)Given a model for the engine system (see Appendix B)

He(Tc) = e−0.07s G(Tc)1.85s+ 1

Tc ∈ [300 400] K (3.36)

where G(Tc) = g1Tc + g2 is a temperature Tc-dependent gain of the engine as definedin (B.2). Defining θ = Tc shows that the model (3.36) in fact is a linear parametermodel with one varying parameter. Hence, implementation of Tc = θ = θ(t) directlyyields an LPV model.

He(θ(t)) = e−0.07s G(θ(t))1.85s+ 1

θ(t) ∈ Θ (3.37)

where Θ ∈ [300 400] K.

Page 18: Gain Scheduling

14 CHAPTER 3. PARAMETER-DEPENDENT PLANT MODELS

3.5 Linear fractional transformation description

In special cases, the parameter dependency in the LPV models resulting from a family of lineariza-tions or quasi-LPV modeling can be modeled or approximated as a Linear Fractional Transforma-tion (LFT). A LPV model, which may be formulated as a linear time-invariant system enclosed bya feedback loop with time-varying parameters θ(t), is regarded as an LFT model. The particularLFT structure enables specific approaches for the subsequent controller design. Constructing anLFT essentially is the same problem as the realization of a transfer function. Hence, representa-tions may not be unique or ‘minimal’.

LFTs are a powerful and flexible approach to represent uncertainty in models by separatingwhat is known, represented by M , from what is unknown, represented by ∆, in a feedback-likeconnection. In Figure 3.2, a lower LFT is depicted, where

M =[M (11) M (12)

M (21) M (22)

](3.38)

The possible values of the unknown elements ∆ are bounded.

M

z w

r v

Figure 3.2: uncertainty modeling using LFTs

This lower LFT is denoted by FL(M,∆) and yields

z = FL(M,∆)w= (M (11) +M (12)∆(I −M (22)∆)−1M (21))w

(3.39)

Analogously, an upper LFT is defined as

FU (M,∆) = M (22) +M (21)∆(I −M (11)∆)−1M (12) (3.40)

In general, an LFT description enables representation of any polynomial or matrix function of ascalar variable. Hence, if the matrices of the model (3.3) depend rationally on the unknown, i.e.varying parameter θ, a LFT on a corresponding diagonal, θ-dependent ∆(θ) is enabled, where∆ = diag(θ1In, . . . , θpIn) for θ having p components and M being of order n (Doyle, Packard &Zhou 1991, Zhou, Doyle & Glover 1996).

Example 3.3 (Engine system)Consider the engine system He(Tc) = G(Tc)He,n, where G(Tc) = g1Tc + g2 and g1 =8.7 · 10−3, g2 = 2.1. Suppose Tc takes on values 300 ≤ Tc ≤ 400, which yields4.71 ≤ G(Tc) ≤ 5.58. Rewriting this as a LFT yields G(Tc) = 5.145 + 0.57δG(Tc) withδG(Tc) ∈ [−1, 1], i.e.

G(Tc) =([

5.145 0.571 0

], δG(Tc)

)(3.41)

Figure 3.3 shows the corresponding block-diagrams.

The advantages of LFT modeling involve i) the flexibility to represent different system realiza-tions including uncertainties, e.g. state-space or transfer functions and corresponding parametric

Page 19: Gain Scheduling

3.6. CONCLUSIONS / RECAP 15

5.145 0.571 0

δG(Tc)

Fv

r v

He,n G(Tc)Fvωe

He,n

ωe

(a)

(b)

Figure 3.3: Modeling the dependence of G(Tc) on Tc as an uncertainty δG(Tc) using a lower LFT.

and / or model uncertainties and ii) the possibility to apply typical algebraic operations, e.g.cascade, parallel and feedback interconnections (Doyle et al. 1991). Hence, LFTs are very suitablefor a general hierarchical representation of uncertainty and may be used as a suitable basis forsubsequent gain-scheduling (v. Helvoort, Steinbuch, Lambrechts & v.d. Molengraft 2004, Wu &Dong 2006).

In principle, LPV systems in which the state-space matrices are polynomial or rational func-tions of the parameters θ can be transformed into LFT form. However, this reformulation ingeneral is nontrivial and may involve a considerable increase in the order of the system and cor-responding increase in the order of the controller (Leith & Leithead 2000).

3.6 Conclusions / recap

Linearization-based scheduling yields a set or family of locally valid, linear parameter-dependentplant models (3.2) rather than a true LPV model (3.3). This is comparable to the use of a set ofblack-box point-designs instead of a nonlinear model as a starting point. Direct gain scheduling aswell as a quasi-LPV approach directly yield a true LPV model. Velocity-based scheduling allowsfor the implementation of θ = θ(t) and thus the design of a true LPV model as well. SubsequentLPV control synthesis requires a true LPV model as a basis, whereas a family of linear parameter-dependent plant models is sufficient for subsequent classical gain scheduling, which is based onlocal LTI controller designs.

The distinction between LPV and LFT models follows from i) the different available techniquesto analyze the system, ii) the allowable dependence of the state-space on the parameters and iii)the extent to which information about the parameter’s variation, i.e. the rate of variation θ(t),is exploited in the analysis. Regarding LPV models, the scheduling is limited to parameter vari-ations, whereas LFTs enable for all sorts of uncertainties. However, LFTs exhibit computationaldifficulties.

Open question are how to select the scheduling variables θ appropriately and how to determinehow many point-designs are required in case of linearization-based scheduling or black-box identi-fication. In practice, these issues are commonly resolved by insight in the system and experience.

Page 20: Gain Scheduling

16 CHAPTER 3. PARAMETER-DEPENDENT PLANT MODELS

Page 21: Gain Scheduling

Chapter 4

Classical gain scheduling

As discussed in Section 2.1.1, a set of LTI controllers Λ(θ), parameterized by θ has to be de-signed. Subsequent implementation of the controllers requires the design of a family of parameter-dependent controllers Λ(θ(t)) based on this set of controllers. In the best case, a set of parameter-dependent controllers follows immediately, otherwise for example interpolation techniques may beapplied to enable smooth scheduling or the set of controllers may be used for discrete scheduling.

The derivation of a parameterized set of LTI controllers is discussed briefly after which theimplementation of the resulting set of LTI controllers Λ(θ) is discussed. Various methods to derivea family of parameter-dependent controllers, e.g., via interpolation, and corresponding points ofattention, e.g. hidden coupling terms, are discussed.

4.1 LTI controller design

The main advantage of applying classical gain scheduling techniques is the ability to use linearcontroller design synthesis without the restriction to a specific synthesis. Based on the modelingdescribed in the previous chapter, a set of LTI controllers Λ(θ) is designed, which is parameterizedby the same scheduling variable θ as used in the corresponding modeling.

In case a parameterized set of linearized models is available instead of an LPV model or incase an LPV model is available, but different closed-loop dynamics are demanded for differentoperating points, a set of linear controllers has to be designed. A set of operating conditionsθi ∈ Θi ⊂ Θ is determined for which corresponding LTI controllers are designed. In case of aparameterized set of linearized models, typically the same operating conditions Θi = Θo are used.The corresponding set of linearized models or constant operating points of the LPV model enablemodel-based controller design. The resulting parameterized set of controllers hence is only locallyvalid.

The parameterized set of linear controllers is defined by

Λ(θi) :{ξ(t) = Ac(θ)ξ(t) +Bc(θ)z(t)u(t) = Cc(θ)ξ(t) +Dc(θ)z(t)

θi ∈ Θi (4.1)

with input signal z(t), which is derived from the error signal z(t), including if appropriate areference command r(t), e.g. z(t) = r(t)−y(t), the controller output signal u(t) and the controllerstate ξ(t). Corresponding LTI controller design syntheses are not discussed in this report. Forexample standard loop-shaping or pole-placement techniques as well as robustH∞ and µ synthesesmay be employed.

Conceptually, the resulting set of linearized controllers Λ(θi) is used as a basis for a continuouslyscheduled family of parameter-dependent controllers Λ(θ(t)) lateron. In general, this enforces allcontroller designs to have the same structure and dimension. Notable exceptions are i) the appli-cation of discrete scheduling instead of continuous scheduling methods and ii) the application offuzzy controller blending. Regarding discrete gain scheduling, a distinction between the scheduling

17

Page 22: Gain Scheduling

18 CHAPTER 4. CLASSICAL GAIN SCHEDULING

of controller coefficients and the scheduling of complete controllers has to be made. Concerning theformer, fixed controller structure and dimension obviously are demanded as well. Concerning thelatter, a fixed controller structure and dimension is not demanded, however favorable neverthelessto be able to minimize differences in magnitude between controller transitions.

4.1.1 Direct gain scheduling controller design

When a true LPV model is available (instead of a parameterized set of linearized models), andconstant closed-loop behavior for the complete operating range is desirable, a set of parameter-dependent controllers can be determined immediately.

Example 4.1 (Engine system)

Consider the LPV engine model He(Tc) = Y (s)U(s) (3.37) and the problem of designing a

corresponding idle governor controller. The time-delay is not taken into account forthe moment. For control purposes, an error z(t) = r(t)−y(t) is introduced, where r(t)is a certain reference trajectory to be followed by y(t). An LPV controller is designeddirectly using a (linear) parameter-dependent proportional controller

C(θ(t)) =U(s)Z(s)

=kp

G(θ(t))(4.2)

where Z(s) = R(s) − Y (s). Combining (3.37) and (4.2) yields a closed-loop transferfunction that is independent of θ(t)

Y (s)R(s)

=kp

1.85s+ 2(4.3)

where kp is the tuning parameter (for kp = 23.31, suitable closed-loop behavior isachieved).

4.2 Gain scheduling controller design

Based on the parameterized set of linear controllers Λ(θi), θi ∈ Θi, a family of linear, parameter-dependent controllers Λ(θ), θ ∈ Θ has to be derived. I.e. the operating point designs have tobe extended to the entire operating envelope. To enable actual implementation, the schedulingvariable θ is replaced by the measured variable θ(t), yielding θ = θ(t). This yields a nonlinear,scheduling controller of the form

Λ(θ(t)) :{ξ(t) = fc(ξ(t), z(t), w(t), θ(t))u(t) = hc(ξ(t), z(t), w(t), θ(t))

θ(t) ∈ Θ (4.4)

with input signal z(t), which is derived from the error signal z(t), including if appropriate areference command r(t), e.g. z(t) = r(t) − y(t), the controller output signal u(t), the controllerstate ξ(t), external inputs w(t) and the time-varying scheduling variable θ(t).

In general, the controllers are scheduled utilizing the current value of the scheduling variableto interpolate or schedule the controller coefficients, e.g. the controller gains. In case of directscheduling (see Section 4.1.1), a family of parameter-dependent controllers is present alreadyand the implementation θ = θ(t) has to be considered only. Considering gain-scheduling andinterpolation of Λ(θi), θi ∈ Θi, the goal is to design a family of linear, parameter-dependentcontrollers Λ(θ), θ ∈ Θ, such that at each operating point θ ∈ Θi, the scheduled controller i)linearizes to the corresponding linear controller design and ii) provides a constant control valueyielding zero error. The presence of hidden coupling terms will be shown to yield additionalrequirements to the scheduling procedure.

The complexity of the scheduling or interpolation method depends on the structure of the setof LTI controllers. Consider e.g. fixed structured LTI controllers with only one or a few varying

Page 23: Gain Scheduling

4.2. GAIN SCHEDULING CONTROLLER DESIGN 19

gains for changing plant dynamics or operating conditions with respect to a set of LTI controllerswith varying structure, possibly including different states and topology (Reichert 1992). Theformer often is the case when classical linear control techniques are applied to design the set ofLTI controllers. The latter may be the result of application of multivariable state-space designprocedures.

In case of discrete scheduling, the implementation of the LTI controllers involves the designof a scheduled selection procedure that is applied to the set of LTI controllers, rather than thedesign of a family of scheduled controllers. The selection procedure or switching has the advantageof simplicity in defining the controller family. This comes down to the definition of regions forwhich the members of the set of LTI controllers are valid. However, discontinuities (jumps) mayappear in the control output, or in the controller coefficients in case of discrete scheduling of thecoefficients rather than the total controller, which may result in e.g. chattering behavior.

In Figure 4.1 an example of discrete scheduling by switching between 2 controllers is shown.This works only in case Λ0 as well as Λ1 are open loop stable. Whereas this will be the case ingeneral for classical controller designs, this is less obvious for controllers resulting from H∞ orµ-syntheses (Niemann & Stoustrup 1999).

System orPlant P

y(t)u(t)

Λ0

Λ1

θ(t)

Figure 4.1: Gain scheduling by switching between two controllers.

In general, continuous or smooth scheduling is required, which will be the focus of the re-mainder of this chapter. Issues involve how to schedule the controller, i.e. use interpolation,linearization scheduling or other methods, how many and which operating points to use as a basisfor the scheduling and how to assess global stability and performance properties. Firstly, classicalscheduling and various interpolation techniques are presented. Next, hidden coupling terms andclosed-loop stability properties are discussed.

4.2.1 Linearization scheduling

The objective of linearization scheduling is that the equilibrium family of the controller (4.4)matches the equilibrium family of the plant (3.1), such that i) the closed-loop system still can betuned appropriately with respect to performance and robustness demands and ii) the linearizationfamily of the controller equals the designed family of linear controllers.

The former requirement requires the existence of a function ξe(θ) such that

0 = fc(ξe(θ), ze(θ), we(θ))ue(θ) = hc(ξe(θ), ze(θ), we(θ))

θ ∈ Θi (4.5)

The latter requirement yields combination of (4.1) and (4.4)

∇ξfc(ξe(θ), ze(θ), we(θ)) = Ac(θ)∇zfc(ξe(θ), ze(θ), we(θ)) = Bc(θ)∇ξhc(ξe(θ), ze(θ), we(θ)) = Cc(θ)∇zhc(ξe(θ), ze(θ), we(θ)) = Dc(θ)

θ ∈ Θ (4.6)

Page 24: Gain Scheduling

20 CHAPTER 4. CLASSICAL GAIN SCHEDULING

Both conditions are satisfied by the general controller family

ξ(t) = Ac(θ) (ξ(t)− ξe(θ)) +Bc(θ) (z(t)− ze(θ))u(t) = Cc(θ) (ξ(t)− ξe(θ)) +Dc(θ) (z(t)− ze(θ)) + ue(θ)

θ ∈ Θ (4.7)

Implementation of θ = θ(t) in (4.7) then yields the (nonlinear) gain-scheduled controller

ξ(t) = Ac(θ(t)) (ξ(t)− ξe(θ(t))) +Bc(θ(t)) (z(t)− ze(θ(t)))u(t) = Cc(θ(t)) (ξ(t)− ξe(θ(t))) +Dc(θ(t)) (z(t)− ze(θ(t)))

θ(t) ∈ Θ (4.8)

Example 4.2 (Longitudinal vehicle control)Consider the set of parameterized LPV models (3.16). Given a corresponding set ofparameterized LTI PI-controllers

Λ(θ) :

{˙ξ(t) = z(t)u(t) = ki(θ)ξ(t) + kp(θ)z(t)

θ ∈ Θe (4.9)

locally yielding the closed-loop dynamics

Σcl :

[˙x(t)˙ξ(t)

]=

[A(θ)x(t) +B1(θ)w(t) +B2

(ki(θ)ξ(t) + kp(θ)(z(t))

)z(t)

]=

[A(θ)−B2kp(θ) B2ki(θ)

−1 0

] [x(t)ξ(t)

]+

[B1(θ) + [0 B2kp(θ)]

0 1

]w(t)

y(t) =[

1 0] [

x(t)ξ(t)

]

(4.10)

Using linearization scheduling, a nonlinear, gain-scheduled controller Λnl is to be de-signed.

Λnl :{ξ(t) = fc(ξ(t), w(t), y(t))u(t) = gc(ξ(t), w(t), y(t))

(4.11)

The controller equilibrium family ξe(θ), θ ∈ Θe should be such that the equilibriumconditions of the original (nonlinear) plant and the controller match. Furthermore, atequilibrium, the linearized controller should correspond to the corresponding originalpoint-design controller.

Substituting the equilibrium values (3.15) for the deviation variables in (4.9) yields thecontroller equilibrium ξe(θ)

ξe(θ) =1

ki(θ)ue(θ) θ ∈ Θe (4.12)

=1

ki(θ)

(C ′dθ

(2)2 + g(sin θ(1) + Cr cos θ(1)))

θ ∈ Θe

Adopting the general controller family (4.7) yields

Λ(θ) :{ξ(t) = z(t)− ze(θ)u(t) = ki(θ) (ξ(t)− ξe(θ)) + kp(θ) (z(t)− ze(θ)) + ue(θ)

(4.13)

Again substituting the equilibrium values then yields the LPV controller family

Λ(θ) :{ξ(t) = z(t)u(t) = ki(θ)ξ(t) + kp(θ)z(t)

θ ∈ Θ (4.14)

Page 25: Gain Scheduling

4.2. GAIN SCHEDULING CONTROLLER DESIGN 21

By replacing the parameter θ by its measured value θ(t) = [w(1)(t), y(t)]T , the non-linear gain-scheduled controller results

Λnl :{ξ(t) = z(t)u(t) = ki(w(1)(t), y(t))ξ(t) + kp(w(1)(t), y(t))z(t)

(4.15)

4.2.2 Interpolation methods

Most of the literature on continuous gain scheduling assumes that linearization scheduling isavailable and possible. However, in the case of isolated point-designs, corresponding interpolationof the controller (coefficients) is required. In practice, ad hoc interpolation of local point-designcontrollers is gladly adopted to arrive at a gain-scheduled LPV controller. However, in general noglobal or even local stability and performance guarantees can be given. Hence, more theoreticallyjustified methods are proposed, which guarantee local stability. Practical implementation of thesemethods however, is difficult.

Ad hoc interpolation methods

Although ad hoc interpolation methods lack a theoretical basis, many practical applications arepresent in literature. In general, these examples perform satisfactorily for the application consid-ered. However, as they are based on intuition rather than a theoretical basis, application in othercases often is difficult or even prohibited. The lack of a theoretical basis implies that no perfor-mance guarantees can be given at intermediate operating points and even closed-loop instabilitymay occur. Ad hoc interpolation methods in general require a set of local LTI controller designswith fixed structure as well as dimension as a basis. I.e. when focusing on transfer functions, allcontrollers should have the same number of poles and zeros.

(Nichols et al. 1993) interpolate point-design H∞ controllers by means of interpolating thecorresponding poles, zeros and gains. The number of operating points involved, depends on therate and size of migration of poles, zeros and gains. Although no guarantees can be given regardingglobal stability and performance specifications of the resulting LPV controller, more recent researchby (Paijmans, Symens, v. Brussel & Swevers 2008) is based on the same principle. The mainadvantage of the method is the practical applicability. Stability and performance is validated bymeans of extensive simulations, thereby specifically focusing on intermediate operating points.

(Reichert 1992) design a set of LTI H∞ controllers using Riccati equations. Linear interpola-tion of the Riccati equations yields a family of parameter-dependent controllers. No guaranteesregarding performance, robustness or even stability regarding the nonlinear plant can be given,locally nor globally. All point-design controllers have to be of the same structure and dimension,which is not obvious for a set of H∞ controllers.

Other ad hoc interpolation methods, such as for example presented by (Kellet 1991, Koc,Knittel, de Mathelin & Abba 2002) will not be discussed further at this point. Finally, con-troller blending sometimes is referred to as a special case of an ad hoc interpolation technique(e.g. (Niemann & Stoustrup 1999)). As this can be interpreted as fuzzy scheduling, this will beconsidered separately in Chapter 6.

More theoretically justified methods

The main goal of theoretically justified interpolation techniques for the synthesis of gain schedulingcontrollers with respect to the before presented ad hoc interpolation methods, comprises a stabilitypreserving condition. A true LPV model Σ(θ), θ ∈ Θ of the original plant is assumed to beavailable. Analogous to ad hoc methods, the interpolation of a set of LTI controllers Λ(θi),θi ∈ Θi, which yield different closed-loop behavior for different operating points is considered.When constant closed-loop behavior is required, a gain-scheduled family of LPV controllers oftenfollows immediately from the corresponding LPV model (see Section 4.1.1). The resulting LPVcontroller Λ(θ), θ ∈ Θ is called stability preserving if (Stilwell & Rugh 2000)

Page 26: Gain Scheduling

22 CHAPTER 4. CLASSICAL GAIN SCHEDULING

• the coefficients of Λ(θ) are continuous functions of θ

• Λ(θ) = Λ(θi) for θ = θi

• Λ(θ) stabilizes Σ(θ) for all θ ∈ Θ

The interpolation methods are based on a stability covering condition for Λ(θi), θi ∈ Θi withrespect to Σ(θ), θ ∈ Θ. The stability covering condition demands the existence of at least onecontroller Λ(θi), θi ∈ Θi that stabilizes Σ(θ), for all θ ∈ Θ. I.e. if Θ ⊂

⋃θi∈Θi Uθi

where Uθi⊂ Θ

an open neighborhood containing θi ∈ Θi, such that Λ(θi) stabilizes Σ(θ) for all θ ∈ Uθi , the setof local LTI controllers Λ(θi), θi ∈ Θi is stability covering for Σ(θ), θ ∈ Θ. This is a sufficientcondition on the selection of parameter values Θi for which locally valid LTI controllers Λ(θi),θi ∈ Θi are designed.

θi ∈ Θi

Uθj,k,l

Θ ⊂ Rnθ

θj,k,l

Uθjk

Figure 4.2: Operating envelope Θ ⊂ Rnθ with the set Θi and three example stability regions Uθj,k,l

corresponding to θj,k,l. The intersection or overlapping region Uθjkof Uθj

and Uθkis indicated in

grey.

Define an intersection or overlap of multiple regions Uθi as Uθijk..., e.g. Uθjk

represents theintersection of Uθj

and Uθk(see Figure 4.2). The proposed interpolation method generates corre-

sponding stability preserving controllers Λ(Uθijk...), Uθijk...

∈ UΘ, where UΘ the set comprising alloverlapping regions, that follow from interpolation of the controllers Λ(θi) of the correspondingregions Uθi , θi ∈ Θi.

The resulting gain-scheduled controller Λ(θ) for the complete envelope of operating conditionsθ ∈ Θ then becomes

Λ(θ) :{

Λ(θi), θ ∈ Uθi, θ /∈ UΘ

Λ(Uθijk), θ ∈ Uθijk

∈ UΘθ ∈ Θ (4.16)

The method actually involves a particular form of controller blending. The resulting controlleryield switching between controllers with guaranteed smoothness. Additional slow-variation argu-ments yield local exponential stability of the gain-scheduled nonlinear closed-loop system.

Various approaches are presented in literature. (Shahruz & Behtash 1990) describes a methodto linearly interpolate the controller gains of a state feedback controller resulting from eigenvalueplacement. (Stilwell & Rugh 1999) elaborate on this idea introducing an observer-based statefeedback and linearly interpolating the observer gains. Combining this with a Youla parame-terization subsequently ((Stilwell 1999)), yields a basis for state-space interpolation of a set ofobserver-based point-design controllers (Stilwell & Rugh 1998). Analogously, a transfer functioninterpolation method is proposed. Both methods are presented by (Stilwell & Rugh 2000), who alsodetermine an upper bound on the rate of variation of the scheduling variable. Regarding state-space interpolation, the method proposed by (Stilwell & Rugh 2000) is restricted to full-ordercontrollers. (Pellanda, Apkarian & Alazard 2000) generalize the approach to augmented orderstate-space controllers, targeting at the application of H∞ and µ-synthesis. (Claveau, Chevrel

Page 27: Gain Scheduling

4.2. GAIN SCHEDULING CONTROLLER DESIGN 23

& Knittel 2007) present a practical example of linear interpolation of a special observer-basedrealization.

An analogous approach from a mathematical point of view is provided by (McNichols &Fadali 2003). Interval mathematics are applied to determine regions for which closed-loop polesremain within specified bounds. Correspondingly, the required number of point-designs is deter-mined. The methodology is restricted to proper transfer functions with time-varying parameters(coefficients) where Θ is known and θ is known to be small. Global stability has to be inferredfrom extensive simulations.

4.2.3 Velocity-based scheduling

The velocity-based linearization family (3.21) may be adopted as a suitable basis for subsequentgain scheduling controller design (Leith & Leithead 1998c). Classical gain scheduling (Leith &Leithead 1998c), LPV synthesis (Leith & Leithead 2000) or fuzzy-alike blending techniques (Leithet al. 2000) may be employed.

A velocity-based linearization family by definition has infinitely many constant operating pointsas linearizations are assigned to every operating point instead of equilibrium operating pointsonly. The velocity-based linearization family thus broadens the application range of classicalor divide-and-conquer gain scheduling. In practice, LTI controller design will be applied for agrid of operating points in the total operating envelope. Subsequent scheduling or interpolationtechniques (see Section 4.2.1, 4.2.2) can be adopted to design a parameter-varying controller.

(Leith & Leithead 1999, Leith & Leithead 2000) propose to use a blended multiple modelrepresentation, which has close similarity to Takagi-Sugeno or fuzzy modeling. The velocity-basedlinearization family is approximated by a small number of local models. The difference with respectto fuzzy modeling is that the blended local models directly equal the velocity-based linearizationmodel at the corresponding operating point. The blended model is approximated by the weightedcombination of the local model solutions.

Difficulties arise when practical implementation is considered due to the resulting velocity-based controller, which requires y rather than y as input, see Figure 4.3. The correspondingdifferentiator and integrator may be absorbed in the controller in special cases (Leith & Leithead1998a). The corresponding velocity-based controller is given by

Nonlinearplant

Velocity-based

controller∫d

dt u(t) u(t) y(t)

z(t)w(t)w(t)

y(t)

w(t)

Figure 4.3: Implementation of a velocity-based controller design.

Λ(%) :

ξ = ηη = Ac(%)η +B1c(%)w +B2c(%)yu = C(%)η +D(%)w

(4.17)

Stability and performance of velocity-based gain-scheduled systems is established using generalnonlinear system theory such as small gain or Lyapunov theories. Hence, generally conservativeresults are obtained, since, except in special cases, only sufficient conditions are known for thestability of nonlinear systems. In contrast to conventional gain scheduling approaches, the resultingnonlinear controller is valid throughout the operating envelope of the plant, not just in the vicinityof the equilibrium operating points.

Page 28: Gain Scheduling

24 CHAPTER 4. CLASSICAL GAIN SCHEDULING

4.3 Hidden coupling terms

The fact that the scheduling variable θ(t) is time-varying, whereas the controller design generallyis based on corresponding fixed values θi ∈ Θi introduces so-called hidden coupling terms. Theseterms introduce differences between (4.7) and the corresponding linearization of (4.8). Due to the(endogenous) variable θ(t), additional terms with respect to the linearization of (4.7), which isgiven by (4.1), may appear.

Define g(t) = θ(t). Linearization of (4.7) then yields (the time-dependencies are omitted forreasons of clarity)

˙ξ = Ac(θ)ξ +Bc(θ)z − . . .

. . .−{Ac(θ)

∂ξo(θ)∂θ +Bc(θ)

∂zo(θ)∂θ

{∂g(zo(θ),wo(θ))

∂z z + ∂g(zo(θ),wo(θ))∂w w

}u = Cc(θ)ξ +Dc(θ)z + . . .

. . .+{∂uo(θ)∂θ − Cc(θ)

∂ξo(θ)∂θ −Dc(θ)

∂zo(θ)∂θ

{∂g(zo(θ),wo(θ))

∂z z + ∂g(zo(θ),wo(θ))∂w w

}(4.18)

Comparing (4.18) to (4.1) reveals the additional hidden coupling terms that result due to imple-mentation of θ = θ(t). These terms represent additional feedback terms in case of componentsinvolving z(t) or exogenous disturbances in case of components involving w(t). Hence, local sta-bility as well as local performance of the resulting closed-loop differs from the original local LTIdesigns. Correspondingly, the issue of preventing hidden coupling terms revolves around the no-tion that the linearized closed-loop system at any operating condition, parameterized by fixedvalues of the scheduling variable, should precisely match the interconnection of the correspondinglinearized plant and linear controller design at that specific point.

(Lawrence & Rugh 1995, Nichols et al. 1993, Rugh & Shamma 2000) discuss the phenomenonof hidden coupling terms, presenting sufficient conditions to prohibit the existence of these termsin case of a SISO or decoupled problem. Nevertheless, typically hidden coupling terms can not beprevented in all cases. No hidden coupling terms will be present when, in addition to (4.5) and(4.6), the following conditions are fulfilled (Rugh & Shamma 2000, Nichols et al. 1993)

∇θfc(ξo(θ), zo(θ), wo(θ)) = 0∇θhc(ξo(θ), zo(θ), wo(θ)) = 0 θ ∈ Θ (4.19)

which yield

Ac(θ)∂ξo(θ)∂θ +Bc(θ)

∂zo(θ)∂θ = 0

∂uo(θ)∂θ − Cc(θ)

∂ξo(θ)∂θ −Dc(θ)

∂zo(θ)∂θ = 0

θ ∈ Θ (4.20)

Practical application of (4.19) generally requires insight in the system at hand and comes downto physical interpretation of the corresponding terms. The extent to which hidden coupling termscan be prevented is discussed by (Lawrence & Rugh 1995).

Example 4.3 (Longitudinal vehicle control)Consider the nonlinear, gain-scheduled controller (4.15). Define g(t) = θ(t) = g(w(t), y(t)) =[w(1)(t), y(t)]T . Linearization of (4.15) around an equilibrium point yields a local, pa-

Page 29: Gain Scheduling

4.4. STABILITY PROPERTIES 25

rameterized controller Λ(θ), θ ∈ Θe

Λ(θ) :

˙ξ = z − ∂ze(θ)

∂θ

(∂g(ye(θ),we(θ))

∂y y + ∂g(ye(θ),we(θ))∂w w

)= z= [0 1] w − x

u = ki(θ)ξ + kp(θ)z +(∂ue(θ)∂θ − ki(θ)

∂ξe(θ)∂θ − kp(θ)

∂ze(θ)∂θ

). . .

. . .(∂g(ye(θ),we(θ))

∂y y + ∂g(ye(θ),we(θ))∂w w

)= ki(θ)ξ + kp(θ)z +

(∂ue(θ)∂θ − ki(θ)

(1

ki(θ)∂ue(θ)∂θ + ue(θ) ∂∂θ

1ki(θ)

))(y + [1 0]w(t))

= ki(θ)ξ + kp(θ)z − ki(θ)ue(θ) ∂∂θ1

ki(θ)(y + [1 0]w(t))

= ki(θ)ξ +[−ki(θ)ue(θ) ∂∂θ

1ki(θ)

kp(θ)]w −

(kp(θ) + ki(θ)ue(θ) ∂∂θ

1ki(θ)

)x

(4.21)

Hence, the corresponding closed-loop system, combining (3.16) and (4.21) becomes

Σcl(θ) :

[˙x˙ξ

]=

[−B2

(kp(θ) + ki(θ)ue(θ) ∂∂θ

1ki(θ)

)B2ki(θ)

−1 0

] [x

ξ

]+ . . .

. . .+

[B1(θ) +

[−ki(θ)ue(θ) ∂∂θ

1ki(θ)

kp(θ)]

0 1

]w

y =[

1 0] [

x

ξ

](4.22)

Comparing the linearization result (4.22) to the original, locally designed closed-loopdynamics (4.10) shows additional terms. These terms are the result of substituting θby its corresponding measured values g(t). By doing so, additional (feedback) loopsare created in the closed-loop system, which might influence stability and performanceat the corresponding equilibrium points.

In this case, the additional constraints (4.19) are easily fulfilled by choosing anotherparameterization of the controller. Rewriting (4.9) as

Λ(θ) :

{˙ξ(t) = ki(θ)z(t)u(t) = ξ(t) + kp(θ)z(t)

θ ∈ Θe (4.23)

does not change the controller as it just another non-unique parameterization. How-ever, the controller equilibrium becomes ξe(θ) = ue(θ), as opposed to (4.12). Usingthis equilibrium value, the additional conditions (4.19) indeed are fulfilled. Hence, nohidden coupling terms will be present anymore when implementing θ = θ(t).

4.4 Stability properties

Stability and performance properties of the closed-loop system incorporating the original nonlinearplant or model (3.1) and the parameter-dependent controller (4.8) form an important issue in thedesign and application of gain-scheduled controllers. Regarding classical gain-scheduling synthesisas described in this chapter, difficulties arise due to

• the equilibrium-operating-point-based controller designs, which disregard off-equilibrium op-erating points

• the use of a frozen-valued scheduling variable θ instead of θ(t) in the controller design, whichdisregards transient dynamics

Page 30: Gain Scheduling

26 CHAPTER 4. CLASSICAL GAIN SCHEDULING

Stability analysis of nonlinear systems and hence closed-loop LPV systems is difficult. A classi-cal result by (Desoer 1969) state that LPV systems are stable for sufficiently small θ(t). (Shamma& Atans 1991) extend this result considering robust performance for unmodeled dynamics as well.(Shamma & Athans 1990) derive sufficient conditions for robust stability and performance ofquasi-LPV systems, which again comes down to a sufficiently small θ(t). The conditions also holdfor non-quasi-LPV systems if the nonlinearities in the state variables other than the schedulingvariables are small. The corresponding rule-of-thumb ’the plant nonlinearities should be capturedby the scheduling variables’ is often referred to. However, the conditions typically are conser-vative as they are based on classical Lyapunov and small-gain stability analyses. Furthermore,implementation of θ = θ(t) to derive a nonlinear, parameter-dependent controller may introduceadditional dynamics, so-called hidden terms, as the controller design is based on the frozen-valuedscheduling parameter θ. Hidden terms may influence stability (see Section 4.3).

A priori assessment of guaranteed stability and performance properties of classical gain-schedulingregarding i) non-local or off-equilibrium operating points and ii) fast variations of the schedulingvariable, is impossible. As this is not considered in the controller design process, guaranteed prop-erties of the overall design cannot be established. (Shamma & Athans 1992) present an examplein which the frozen parameter designs are based on stable plant models, but the actual plantmay be unstable. The linear models, therefore, are an inaccurate representation of the true plant.Stability and performance specifications thus are typically inferred from extensive simulations.

In the vicinity of equilibrium operating conditions, the stability of a nonlinear system maybe related to the stability of the corresponding frozen-input nonlinear system. Following (Rugh1991, Lawrence & Rugh 1995), the deviation of a (nonlinear) system from its equilibrium surfaceremains small if

1. the initial conditions lie in the neighborhood of an equilibrium point, which yields∣∣∣∣∣∣∣∣( xi(θ(t))ξi(θ(t))

)−

(xe(θ(t))ξe(θ(t))

)∣∣∣∣∣∣∣∣ < γ, γ > 0, θ(t) ∈ Θ (4.24)

The inputs and initial conditions are thus confined to lie in the neighborhood of the corre-sponding equilibrium operating points (Leith & Leithead 1998b, Leith & Leithead 1998c).

2. a slow variation requirement is present, which yields∣∣∣∣∣∣θ(t)∣∣∣∣∣∣ < µ, µ > 0 (4.25)

This condition restricts the class of allowable inputs and initial conditions to remain suffi-ciently close to the equilibrium operating conditions.

3. the linearization of the nonlinear system equals the local point-designs, i.e. no hidden cou-pling terms are present, and is stable, i.e. <(λ) ≤ −ε < 0, ∀ θ(t) ∈ Θ

If all three conditions are fulfilled, the deviation from the equilibrium surface remains small and thestability of the nonlinear system may be related to the stability of a set of LTI models representingthe nonlinear system.

More recent controller design syntheses, which stem from classical gain scheduling synthe-sis such as LPV and LFT syntheses incorporate θ(t) in the controller design process. As thetime-varying nature of the scheduling parameter is taken into account in the controller design syn-thesis explicitly, global stability and performance properties can be derived. In general however,these syntheses guarantee stability and performance ∀ θ instead of for θ(t). Hence, conservatismis introduced. Quasi-LPV design suffers the same drawback (see Section 3.3). Typically, thisconservatism is reduced by defining explicitly a range and rate of variation corresponding to thescheduling parameter. Taking these parameters into account in the controller design, decreasesthe conservatism.

Page 31: Gain Scheduling

Chapter 5

LPV controller synthesis

The term Linear Parameter Varying (LPV) control synthesis is commonly used to indicate con-troller synthesis based on either LPV or Linear Fractional Transformation (LFT) models. LPVcontrol synthesis yields parameter-dependent controllers with a priori guaranteed stability andperformance properties. The available real-time information of the parameter variation is used inthe control synthesis. The time-varying nature of the corresponding LPV dynamics is thus incor-porated in the LPV control synthesis as opposed to classical gain-scheduling methods (Shamma& Atans 1991). The a priori stability and performance guarantees prohibit the need for validationof stability and performance characteristics afterwards by means of extensive simulations. LPVcontrol synthesis focuses on an LPV model rather than a nonlinear model of the plant, whichexplains the naming. LPV modeling is discussed in Chapter 3.

The main advantages of LPV control synthesis are i) there exist a solid theoretical foundationguaranteeing a priori stability and performance for all θ(t) given a corresponding range and rateof variation of θ(t), ii) the corresponding controller design is global with respect to the parame-terized operating envelope Θ, whereas classical gain scheduling techniques focus on local systemproperties, and iii) a controller is synthesized directly, rather than its construction from a familyof local linear controllers. The main disadvantages are i) with respect to classical gain schedulingtechniques, the controller synthesis is much more involved, which results in focusing on appro-priate problem formulation rather than the actual controller design, ii) generally, conservatismhas to be introduced to arrive at a feasible and convex problem, and iii) with respect to classicalgain scheduling, allowing for arbitrary linear controller design techniques, a predefined controllerdesign synthesis has to be adopted.

As the latter point already indicates, LPV syntheses constitutes a specific performance evalu-ation framework, whereas classical gain scheduling provides an open framework. Typically, LPVsyntheses employ the induced L2-norm as a performance measure, which is directly related tolinear H∞-techniques. As a result, an LPV control synthesis applied to a time-invariant systemis equivalent to a standard H∞-approach.

In general, LPV control syntheses can be categorized into

• techniques utilizing a Lyapunov-based approach, also known as (genuine) LPV-techniques

• techniques exploiting the specific structure of systems with LFT parameter dependence,utilizing a small-gain approach, which are also determined as LFT approaches

• a combination of the two preceding points, which will be referred to as ‘mixed’ LPV-LFTapproaches

The LPV paradigm was introduced circa 1990 as a framework to analyse gain scheduling (Shamma& Atans 1991, Shamma & Athans 1992) and has seen considerable activity since (Rugh & Shamma2000). By now, application examples are present in literature, e.g., (Wang, Nishimura & Shimogo2007, Dinh, Scorletti, Fromion & Magarotto 2005, Dettori & Scherer 2002). However, a general

27

Page 32: Gain Scheduling

28 CHAPTER 5. LPV CONTROLLER SYNTHESIS

solution for the synthesis of an LPV controller still has not been determined. Until now, researchhas come up with solutions for special extensions rather than genuine generalizations of linearH∞-designs. Correspondingly, in general, conservatism is introduced, which is one of the maindrawbacks of LPV controller synthesis. Current research hence focuses on solving the problem fortrue generalizations (Rieber, Scherer & Allgower 2008, Wu & Dong 2006, Dinh et al. 2005).

In the next sections an overview of the general LPV synthesis problem is given. Next, thedifferent approaches are detailed: LPV, LFT and ‘mixed’ LPV-LFT approaches. Finally, a con-clusion with guidelines for application of each approach in practice is given. Elaboration of allmatrices, algorithms and LMIs is saved for a case study.

5.1 General LPV controller synthesis setup

Consider the LPV model Σ(θ(t)) (3.3) (see Figure 5.1). The parameter values θ(t) are measured inreal-time and time variations are not known in advance. The parameter vector θ(t) may compriseexogenous as well as endogenous variables. Conceptually, θ(t) is regarded as an exogenous variablewhen LPV control synthesis is considered, i.e. θ(t) = θ(w(t)). When θ(t) is an endogenous variable,θ(t) = θ(x(t)), this is called quasi-LPV (see Section 3.3). In a more complete model, θ(t) naturallywill always be endogenous as is depicted in Figure 5.1.

LPVmodel

z(t)

u(t)

θ(t)

w(t)

y(t)

Figure 5.1: LPV model

A general LPV controller is defined as

Λ(θ(t)) :{ξ(t) = Ac(θ(t))ξ(t) +Bc(θ(t))y(t)u = Cc(θ(t))ξ(t) +Dc(θ(t))y(t)

(5.1)

The rate of variation of the parameter vector, θ(t), can be regarded as a scheduling variable aswell, yielding

Λ(θ(t), ψ(t)(t)) :{ξ(t) = Ac(θ(t), θ(t))ξ(t) +Bc(θ(t), θ(t))y(t)u = Cc(θ(t), θ(t))ξ(t) +Dc(θ(t), θ(t))y(t)

(5.2)

In this way, θ(t) can be accounted for explicitly in the controller design. Assume the time-variationof the parameter vector θ(t) = [θ1(t), . . . , θnθ

(t)]T is known to be constrained by

θ(t) ∈ Θ ⊂ Rnθ (5.3)ψ(t) = θ(t) ∈ Ψ ⊂ Rnθ (5.4)

where nθ the length of θ(t). Define correspondingly

φ(t) ∈ {θ(t), ψ(t)} ∈ Θ×Ψ ⊂ R2nθ (5.5)

Consider a closed-loop LPV system combining an LPV model Σ(θ(t)) (3.3) and a correspondingLPV controller Λ(θ(t)) (5.1). The corresponding closed-loop system state vector equals ζ = [x, ξ]T

and the closed-loop system becomes[ζ(t)z(t)

]=

[A(θ(t)) B(θ(t))C(θ(t)) D(θ(t))

] [ζ(t)w(t)

](5.6)

Page 33: Gain Scheduling

5.1. GENERAL LPV CONTROLLER SYNTHESIS SETUP 29

with [A BC D

]=

A+B2DcC2 BCc B1 +B2DcD21

BcC2 Ac BcD21

C1 +D12DcC2 D12Cc D11 +D12DcD21

(5.7)

where the parameter-dependency of the matrices is omitted for clarity. Moreover, it has to beremarked that the closed-loop matrices in (5.6) may be dependent on ψ(t) as well. LPV controllersynthesis is based on stability and performance analyses considering the closed-loop system (5.7).

5.1.1 Stability analysis

Regarding Linear Time-Invariant (LTI) systems, asymptotic stability is well-known to be charac-terized by the set of LMIs

ATX + XA ≺ 0, X � 0 (5.8)

where A the (closed-loop) system matrix. Considering the time-varying-parameter-dependentclosed-loop system (5.7), exponential stability of the LPV system is defined analogous to (5.8) by

A(θ(t))TX + XA(θ(t)) ≺ 0, X � 0, ∀θ(t) ∈ Θ (5.9)

Again, (5.6) may be dependent on ψ(t) as well, changing the stability criterion (5.9) accordingly.The LMI (5.9) guarantees stability for all θ(t) ∈ Θ, not considering the rate of variation ψ(t) of θ(t).Hence, (5.9) guarantees stability for all possible rates of variation ψ(t) ∈ Rnθ , e.g. for infinitelyfast variations, which in practice obviously is not possible and thus introduces conservatism.

Conservatism of (5.9) may be reduced, using a parameter-dependent Lyapunov function X =X (θ(t)) instead. In this case, additional information on ψ(t) may be taken into account as well.However, in general, constant θ(t) = θ is considered, i.e., ψ(t) ∈ ∅, when a parameter-dependentLyapunov function X = X (θ(t)) is adopted. Consider a closed-loop LPV system combining theLPV model Σ(θ(t)) (3.3) and a corresponding LPV controller Λ(θ(t), ψ(t)) (5.2). The correspond-ing closed-loop system becomes[

ζ(t)z(t)

]=

[A(θ(t), ψ(t)) B(θ(t), ψ(t))C(θ(t), ψ(t)) D(θ(t), ψ(t))

] [ζ(t)w(t)

](5.10)

A classical result regarding stability of this time-varying system is that it is stable if i) the cor-responding frozen-time, time-invariant system with system matrix A(θo, 0) is stable and ii) therates of time variation θ(t) are sufficiently slow. This is captured in the following theorem (e.g.(Rugh & Shamma 2000)).

Theorem 5.1 (Lyapunov stability for time-varying-parameter dependent systems)Suppose there exists X (θ(t)) = X T (θ(t)) > 0 such that

X (θ(t))A(θ(t), ψ(t)) +AT (θ(t), ψ(t))X (θ(t)) + ∂X (θ(t), ψ(t)) < 0 (5.11)

for all (θ(t), ψ(t)) ∈ Θ×Ψ, where

∂X (θ(t), ψ(t)) ≡nθ∑i=1

∂X(θ(t))∂θi(t)

ψi(t) (5.12)

Then the closed-loop system (5.10) is exponentially stable for all parameter trajectories (θ(t), ψ(t)) ∈Θ×Ψ, where

V (x(t), θ(t)) = x(t)TX (θ(t))x(t) (5.13)

is a corresponding parameter-dependent quadratic Lyapunov function.

Page 34: Gain Scheduling

30 CHAPTER 5. LPV CONTROLLER SYNTHESIS

Equation (5.11) yields a set of LMIs

[I A(θ(t), ψ(t))T

] [∂X (θ(t), ψ(t)) X (θ(t))

X (θ(t)) 0

] [I

A(θ(t), ψ(t))

]≺ 0 (5.14)

Lyapunov functions are widespread as a tool for stability analysis of nonlinear systems. Stan-dard Lyapunov theory for exponential stability of nonlinear systems is established in (Khalil 1992).Adopting particular classes of Lyapunov functions, e.g. quadratic Lyapunov functions (5.13), in-troduces conservatism. Subsequent control synthesis might hence be conservative as (5.11) is asufficient condition for stability only. Although research to non-quadratic Lyapunov functions ispresent as well, focus is mainly on quadratic Lyapunov functions.

5.1.2 Performance analysis

Performance analysis of the closed-loop system combining an LPV model Σ(θ(t)) (3.3) and acorresponding LPV controller Λ(θ(t), θ(t)) (5.2), is commonly related to the induced norm of theexogenous signals w(t) to the error signals z(t), the so-called performance channel. The energygain over the performance channel w(t) 7→ z(t) is defined by the induced L2-gain as

supw(t) 6=0

||z(t)||2||w(t)||2

(5.15)

where w(t) ∈ L2, L2 = {x : [0,∞) → Rnx | ‖x‖2 <∞}, and

‖x‖2 =

√∫ ∞

0

x(t)Tx(t)dt (5.16)

Good performance of the corresponding closed-loop system is guaranteed if (5.15) is small. Definingz(t) = Tzww(t) and accounting for (5.3) and (5.4), this yields (e.g., (Dettori 2001))

‖Tzw‖∞ = sup(θ,ψ)∈Θ×Ψ

supw(t) 6=0

||z(t)||2||w(t)||2

(5.17)

Introducing a performance value γ, closed-loop performance is guaranteed by

||Tzw||2 < γ (5.18)

Combination of (5.17), (5.16) and (5.18) yields∫ ∞

0

[w(t)z(t)

]T [−γI 0

0 1γ I

] [w(t)z(t)

]dt ≤ ε

∫ ∞

0

w(t)Tw(t)dt (5.19)

More generally, the quadratic performance specification is given by∫ ∞

0

[w(t)z(t)

]T [Qp SpSTp Rp

] [w(t)z(t)

]dt ≤ ε

∫ ∞

0

w(t)Tw(t)dt (5.20)

with Rp ≥ 0.Combining (5.14) and (5.20) guarantees stability as well as quadratic performance of the cor-

responding closed-loop system. This leads to the following theorem (Dettori 2001):

Theorem 5.2 (Closed-loop quadratic performance) Quadratic performance and exponentialstability is achieved if there exists

X (θ(t)) = X T (θ(t)) > 0 (5.21)

Page 35: Gain Scheduling

5.1. GENERAL LPV CONTROLLER SYNTHESIS SETUP 31

such that1 0

A(θ(t)) B(θ(t))0 1

C(θ(t)) D(θ(t))

T

∂X (θ(t), ψ(t)) X (θ(t)) 0 0X (θ(t)) 0 0 0

0 0 Qp Sp0 0 STp Rp

1 0A(θ(t)) B(θ(t))

0 1C(θ(t)) D(θ(t))

< 0

(5.22)

for all θ(t) ∈ Θ and ψ(t) ∈ Ψ, where ∂X (θ(t), ψ(t)) =m∑i=1

∂X (θ(t),ψ(t))∂θ(t) ψ(t) and x = Ax + Bw,

z = Cx+Dw for x(0) = 0.

5.1.3 General synthesis problem formulation

The LPV control synthesis is based on (5.21) and (5.22). The goal is, given a parameter-dependentLPV system Σ(θ(t), ψ(t)) (3.3), to design a parameter-dependent controller

K = {Ac(θ(t), ψ(t))Bc(θ(t), ψ(t))Cc(θ(t), ψ(t))Dc(θ(t), ψ(t))} (5.23)

using the constraints (5.21) and (5.22).Considering the constraints (5.21) and (5.22), three inherent problems are present. Firstly,

(5.22) is nonlinear in the decision variables; for example, A(θ(t)) includes Ac(θ(t)) and hence aproduct between two decision variables X (θ(t)) and A(θ(t)). However, employing a nonlinearchange of variables ((Masubuchi, Ohara & Suda 1995, Scherer, Gahinet & Chilali 1997)), thisproblem can be solved efficiently: Introduce the new set of variables

V (θ(t)) = {X(θ(t))Y (θ(t))K(θ(t), ψ(t))L(θ(t))M(θ(t))N(θ(t))} (5.24)

and apply the nonlinear transformation

{X (θ(t)) |Ac(θ(t), ψ(t))Bc(θ(t))Cc(θ(t))Dc(θ(t))} 7→ V (θ(t)) (5.25)

to the original search variables X(θ(t)), determining the Lyapunov function, and Ac(θ(t), ψ(t)),Bc(θ(t)), Cc(θ(t)) andDc(θ(t)), the feedback controller matrices. This yields a convex optimizationproblem. The original search variables are derived from V (θ(t)).

The second problem is the fact that the parameter-dependent decision variables V (θ(t)) yieldfunctional inequalities, i.e., the dependency of the decision variables on the varying parameterθ(t) is unknown a priori. Standard LMI algorithms do not allow solving functional inequalities.To overcome this problem, several approaches have been proposed in literature. Until now, how-ever, no general solution is present. One might adopt a fixed, a priori chosen structure for thedecision variables, e.g., affine of polytopic; this is proposed in both LPV and LFT approaches.Furthermore, one might assume an LFT parameter-dependency of the controller in combinationwith parameter-independent Lyapunov functions, i.e., the LFT approach, or various variations onthis, i.e., ‘mixed’ LPV-LFT approaches. In all cases, a feasible solution is achieved at the cost ofadditional conservatism.

Finally, the problem is parameterized by (θ(t), ψ(t)) ∈ Θ×Ψ, which implies that it constitutesinfinitely many inequalities. Again, no general solution to this problem is present, although inspecific cases nonconservative solutions are presented. The most commonly adopted approaches,in practice, are to use gridding of the parameter range or to use a polytopic approximation. Inliterature, is often referred to as (genuine) LPV controller synthesis. Exploiting the fact that thesystem has rational dependence on the varying parameters enables a transformation to a finite setof LMIs. This is referred to as the LFT approach. More recently, ‘mixed’ LPV-LFT approacheshave been proposed to transform the infinite set of LMIs to a finite one. All approaches, however,arrive at a set of finite dimensional LMIs at the cost of additional conservatism.

Page 36: Gain Scheduling

32 CHAPTER 5. LPV CONTROLLER SYNTHESIS

5.2 Lyapunov-based LPV control synthesis

When the LPV paradigm was first introduced, research focused on quadratic Lyapunov functions.(Shahruz & Behtash 1990) present the synthesis of a continuous time state feedback controllerusing a common quadratic Lyapunov functions V (x) = xTx. (Becker, Packard, Philbrick &Balas 1993) extend this to output feedback controllers, adding specific L2-norm performancebounds as well, i.e. V (x) = xTXx, X = XT > 0, where x the closed-loop system state. (Apkarian& Gahinet 1995) extend the previous results to discrete time systems. To reduce conservatism,parameter-dependent Lyapunov functions can be adopted.

5.2.1 Gridding

⇒ consider a parameterized subset (p, q) ∈ (Πs × Πs) ⊂ (Π× Π); compute for these fixed points a

solution (Apkarian et al. 1995, Becker & Packard 1994, Wu, Yang, Packard & Becker 1996, Zhou

et al. 1996)

+ widely used in LPV synthesis; often suitable for practical applications

+ rate of variation can be accounted for

− interpolation required for intermediate points

− dense grid / large m yields large LMI

The number of required gridding points increases very fast for an increasing number of param-eters. The more dense the grid, the larger the optimization problem. However, to assure globalproperties, a dense grid is required. Gridding is commonly adopted in LPV control design to arriveat a finite dimensional set of LMIs. It enables to account for bounded rates of variation of theparameters. E.g., (Wu, Packard & Balas 1995, Wu et al. 1996) present a LPV control synthesisyielding a LPV controller with a ψ(t)-dependent controller matrix Dc(θ(t), ψ(t)). Consideringthis specific example, however, measurement of ψ(t) in practice is difficult, i.e. implementation ofψ(t)-dependent controllers is difficult.

5.2.2 Polytopic approximation

⇒ i) assume affine or polytopic parameter dependencies for V (p), i.e., plant and controller, ii) con-

sider a convex polytope (Πo× Πo) ⊃ (Π× Π); compute a solution for extreme points of the polytope

(Gahinet, Apkarian & Becker 1996, Apkarian & Gahinet 1995, Yu & Sideris 1995)

+ no need for gridding

+ solution for functional inequalities

− nonlinear parameter dependencies are not considered / all parameters vary independently(conservative)

− polytope is (generally) a (conservative) overestimation

example: in (5.22), q enters affinely via ∂X (p, q), hence q ∈ Π0, where Π0 the extreme points of Π

Page 37: Gain Scheduling

5.3. SCALED SMALL-GAIN OR LFT SYNTHESIS 33

An affine LPV model is defined as an affine dependency of e.g. the system matrix A(θ(t))on θ(t), i.e. A(θ(t)) = Ao + θ(t)A1. Considering affine LPV models only, implies that nonlinearparameter dependencies are not considered. Hence, all parameters are assumed to vary inde-pendently, which introduces conservatism. Furthermore, if Φ is not a convex polytope, a largerpolytope Φ ⊂ Φ ⊂ Rnθ , which is convex may be considered. However, this introduces parametercombinations, which may not be possible in the original model and hence introduces conservatism.

5.3 Scaled small-gain or LFT synthesis

Exploiting LFT parameter-dependency of LPV systems enables application of a generalized H∞control synthesis using the optimally scaled small-gain theorem. The parameter variations aretemporarily regarded as unknown perturbations (Packard 1994, Apkarian & Gahinet 1995). Mul-tipliers or scalings describing the nature of the unknown, time-varying parameter θ(t) are intro-duced to decrease conservatism with respect to LPV methods using a constant, common quadraticLyapunov function only. In the LFT approach, a constant, common quadratic Lyapunov functionis adopted as well, hence, infinitely fast parameter variations are accounted for.

5.3.1 LFT problem formulation

An LFT model (see Section 3.5) is used as a basis. An LFT model in fact is a special case ofa LPV model, which is transformed using an upper LFT, yielding a constant, known part Mand a corresponding time-varying, unknown part ∆(θ(t)). To enable an LFT representation, thesystem should have rational dependence on the parameter θ and no poles in zero. Analogously,transformation of the LPV controller Λ(θ(t)) (5.1) using a lower LFT, yields a constant, known partK and a corresponding unknown part ∆K(θ(t)), where ∆K(θ(t)) represents a possibly nonlinear,controller scheduling function, see Figure 5.2. In general, ∆K(θ) 6= ∆(θ(t)). For the controllerto be represented in LFT form, the same restrictions hold as for the system. The closed-loopinterconnection of the resulting LFTs is transformed, again using a lower LFT, yielding the closed-loop interconnection T over the channel w 7→ z

T = FL (FU (M,∆(θ(t))), FL(K,∆K(θ(t))) ) (5.26)

Rewriting this yields

T = FU (FL(M,K), ∆ ) (5.27)

where ∆ = diag(∆(θ(t)),∆K(θ(t))). The closed-loop system matrix is defined as follows: ζ(t)zu(t)z(t)

=

A Bu BpCu Du DupCp Dpu Dp

ζ(t)wu(t)w(t)

(5.28)

where w(t) 7→ z(t) the performance channel and zu(t) = ∆wu(t). In Figure 5.2, the correspond-ing interconnection structures are shown. The closed-loop system has repeated linear fractionalparameter dependence.

The use of a constant Lyapunov function and the required rational dependence of the controlleron the scheduling parameters introduces conservatism. However, the resulting problem is a gener-alized robust performance problem for the augmented system M with respect to the uncertainty∆ = ∆(θ(t)), and the theoretical background covering analysis and controller synthesis for theseproblems is extensive. Assume that ∆(θ(t)) is block-diagonal with diagonal terms |∆i(θ(t))| ≤ 1,which is not restrictive. An LTI controller K has to be found that renders the augmented systemM stable such that robust quadratic performance is achieved for the performance channel w 7→ z.After the design, the corresponding parameter variations are incorporated in the controller designagain, yielding an LPV controller, also with LFT parameter-dependence. For θ(t) = θ, a genuinerobust control problem results.

Page 38: Gain Scheduling

34 CHAPTER 5. LPV CONTROLLER SYNTHESIS

(b)

K

∆K

rK vK

M

z w

r v

y u

KrKvK

Mz w

r v

y u

(a)

Figure 5.2: LPV LFT interconnections, (a) represents FL (FU (M,∆(θ(t))), FL(K,∆K(θ(t))) ),(b) represents FU (FL(M,K), ∆ ).

5.3.2 Controller synthesis

Analogous to the previously presented LPV controller synthesis, the scaled small-gain controlsynthesis is based on the robust performance analysis of the closed-loop system. Consider theclosed-loop LFT interconnection system which is represented by T (5.27), (5.28). The corre-sponding analysis involves three main conditions regarding the closed-loop model T , comprising(Aangenent 2008)

• well-posedness, i.e. for all initial conditions corresponding to the augmented system M andall inputs w(t) to the system, the system should have a unique solution.

• exponential stability of the closed-loop system, which is analyzed via the small-gain theory,which states that a closed-loop system is stable provided that the loop-gain is less than unity

• performance; for example quadratic performance, i.e. asymptotic convergence of the per-formance channel w(t) 7→ z(t), or l2-performance, i.e., the loop-gain over the performancechannel w(t) 7→ z(t) is smaller than γ

Combining all three conditions yields a set of LMIs, which is directly derived from the generalLPV synthesis constraints (5.21) and (5.22), e.g., (Dettori 2001).

Theorem 5.3 (LFT system analysis) If there exists a symmetric matrix X and scalingsQ = QT , S and R = RT such that

X > 0, (5.29)

and I 0 0A Bp Bu0 I 0Cp Dp Dpu0 0 ICu Dup Du

T 0 X 0 0 0 0X 0 0 0 0 00 0 Qp Sp 0 00 0 STp Rp 0 00 0 0 0 Q S0 0 0 0 ST R

I 0 0A Bp Bu0 I 0Cp Dp Dpu

0 0 ICu Dup Du

< 0, (5.30)

and [∆(θ(t))

I

]T [Q SST R

] [∆(θ(t))

I

]> 0, ∀θ(t) ∈ Θ, (5.31)

Page 39: Gain Scheduling

5.3. SCALED SMALL-GAIN OR LFT SYNTHESIS 35

then the system[xzp

]=

[A(θ(t)) B(θ(t))C(θ(t)) D(θ(t))

] [xwp

](5.32)

is uniformly exponentially stable and achieves quadratic performance over the channel w(t) 7→ z(t).

Analogously, for[Qp SpSTp Rp

]=

[−γI 0

0 1γ I

](5.33)

the L2-performance criterion results, achieving an L2-gain over the performance channel w(t) 7→z(t), smaller than γ. In order to arrive at this result, specific scalings or so-called multipliers

P =[Qp SpSTp Rp

](5.34)

are employed. This explains the naming of scaled small-gain control synthesis. Solving the LMIsyield the corresponding Lyapunov functions and the multipliers from which the controller can bereconstructed.

Again, a nonlinear transformation is required to transform (5.30) into a convex LMI, see Section5.1.3. However, (5.31) is nonconvex as well, which is a result of the use of multipliers. Additionalconditions on the multipliers, relaxations, are required to arrive at a standard LMI problem, whichcan be solved using existing algorithms (Scherer 1999, Scherer 2001). Several approaches have beenproposed in literature:

• Assuming the matrix ∆(θ(t)) to be diagonal, a corresponding diagonal structure of themultiplier P yields (5.31) automatically true, see, e.g., (Apkarian et al. 1995):

Qp = diag(Q1, . . . , Qm), Sp = 0, Pp = Q (5.35)

However, conservatism is introduced as less freedom is left to solve (5.30).

• Following the previous item, however, taking S 6= 0 but block diagonal and skew-symmetricyields the same results; (5.31) is automatically true at the cost of introducing conservatismin solving (5.30), see, e.g., (Helmersson 1995, Scorletti & Ghaoui 1995).

• Finally, full-block scalings may be applied; either with additional inertia or (partial) concav-ity constraints on the multipliers, e.g., Q < 0, R > 0 (Dettori 2001, Scherer et al. 1997), orwith general full-block multipliers, also known as the full-block S-procedure (Scherer 2001).In the latter case, the final synthesis problem comprises many parameters, which makes it a(very) complex problem. The former approach is complex as well, and, besides, still imposesstructural requirements on P , introducing conservatism.

5.3.3 Pros and cons

The main advantages of the LFT approach are the possibility to use the optimally scaled small-gaintheorem, which is well-established in literature, and the possibility to arrive at a finite dimensionalset of LMIs using multipliers. However, the general full-block multiplier approach is complex,whereas relaxations on the multipliers introduce (much) conservatism. Furthermore, the use of aconstant, common quadratic Lyapunov function, implying that infinitely fast parameter variationsare accounted for, introduces conservatism, especially in the case of large parameter variations.Finally, the controller inherits the LFT parameter dependency, which might be conservative, andpossible realness of the parameter θ(t) is not exploited, introducing conservatism as well.

Page 40: Gain Scheduling

36 CHAPTER 5. LPV CONTROLLER SYNTHESIS

5.4 Mixed LPV-LFT approaches

5.4.1 Extended Kalman-Yakubovich-Popov (KYP) lemma

• KYP lemma: equivalence of infinite dimensional FDIs and finite dimensional LMIs (Rantzer1996)

⇒ replace infinite dimensional (LFT) LMI by single LMI via introduction of additional multiplier

(comparable to full-block S-procedure (Scherer 2005))

+ parameter-dependent Lyapunov function X (p) (q ∈ ∅)

− no general synthesis method (Iwasaki & Hara 2005, Iwasaki, Meinsma & Fu 2000)

Specific extension

• one parameter, rational decision variables (Dinh et al. 2005)

5.4.2 LFT Lyapunov functions

⇒ replace infinite dimensional (LPV) LMI by single LMI via quadratic LFT Lyapunov functions

and full-block multipliers

+ implicit model to account for q bounds (analysis (Helmersson 1996, Iwasaki & Shibata 2001))

+ general parameter-dependent controller (Wu & Dong 2006)

− quadratic LFT Lyapunov functions: difficult to choose (no general procedure)

5.5 Conclusions

As already mentioned in the introduction, current LPV control syntheses comprise solutions tospecial extensions rather than genuine generalizations of linear H∞-designs. In all approaches,conservativeness has to be introduced to arrive at a standard LMI problem. Hence, currentresearch focuses on the definition of an appropriate set of LMIs, which can be solved by standardLMI algorithms, introducing the least possible conservatism.

The main advantages and disadvantages of LPV control synthesis with respect to classical gainscheduling are discussed in the introduction of this chapter. Focusing on the class of systems theLPV control synthesis may be applied to, a true LPV model rather than, for example, a familyof linearizations is used as a basis. Furthermore, the open control framework enabled by classicalgain scheduling, allows accounting for possible modeling errors intuitively. LPV control synthesis,on the other hand, constitutes a predefined control synthesis relying on the LPV model extensively.The preceding LPV modeling hence has to be very adequate.

Whereas classical gain scheduling requires extensive simulation-based testing as in general,stability and performance can only be guaranteed locally, this is not required for LPV controlsyntheses. However, in practice, extensive testing is still critical as LPV control syntheses reliesextensively on the preceding LPV modeling.

Page 41: Gain Scheduling

5.5. CONCLUSIONS 37

LPV control synthesis

Focusing on gridding, which is a commonly employed technique when solving an LPV controlproblem, a parameterized family of controllers is generated rather than a true LPV controller.The family of controllers has to be interpolated to arrive at an LPV controller, thus neglectingtransient behavior. In this case, classical gain scheduling techniques suffer the same disadvantage,however exhibit the advantage of employing an open control design framework, which may yield lessstringent modeling demands. Especially in the case of many parameters, the problem size mayexplode when gridding is adopted. Nevertheless, LPV control synthesis enables to incorporateparameter trajectories θ(t), as well as ψ(t), and design the controllers in one shot. Consideringaffine or polytopic parameter-dependent systems, gridding is not required. The correspondingLPV controller synthesis problem becomes relatively easy. For truely affine or polytopic systems,the only conservatism lies in the corresponding structural restrictions on the resulting controller.

LFT control synthesis

In standard small-gain control synthesis, robustness to parameter variations is ensured. How-ever, at the cost of overestimating admissible parameter trajectories due to the use of a constantLyapunov function. This yields conservative results. The use of multipliers, i.e., scaled small-gain control synthesis, decreases conservativeness, although at the cost of numerical difficulties.Multiplier relaxations are introduced to resolve these difficulties, however, again at the cost ofconservatism.

Mixed LPV-LFT syntheses

More recent approaches target genuine generalizations of the LPV and / or LFT approaches.Two examples are the use of an extended KYP-lemma and the use of Lyapunov functions withLFT parameter dependency. Until now, all results still correspond to special extensions ratherthan genuine generalizations (Scherer 2005). Nevertheless, for a rational dependence of the sys-tem matrices on θ(t), (Wu & Dong 2006) propose a quite general solution using LFT Lyapunovfunctions. Furthermore, for one varying parameter, i.e., θ(t) ∈ R1, and rational decision vari-ables, (Dinh et al. 2005) present a solution with the leas possible conservatism using the extendedKYP-lemma.

LPV vs LFT

Focusing on application of LPV control synthesis in practice, some guidelines on when to useeither LPV or LFT techniques are, see Figure 5.5.

• In case the parameter vector θ(t) constitutes only one parameter and relevant rate boundsfor ψ(t) are known, LPV techniques seem most suitable. Parameter-dependent Lyapunovfunctions are a suitable way of taking into account ψ(t).

• In case θ(t) constitutes many quickly varying parameters, or if the system exhibits a nicedependence on θ(t) allowing for a LFT representation, LFT techniques seem most suitable.The LFT representation may be restrictive as parameter-rate-of-bounds are not explored.

• More recent ‘mixed’ LPV-LFT syntheses are still evolving. No general control synthesis ispresent yet, although some solutions to specific extensions are present.

In general, the LMI problems resulting from LFT control synthesis are less computational de-manding than the corresponding LMI problems resulting from LPV control synthesis. Includingrobustness for unmeasured uncertainties is the easiest using an LFT approach. The LPV approachis most suitable for introducing practical controller modifications, e.g., adding an anti-windup.

Page 42: Gain Scheduling

38 CHAPTER 5. LPV CONTROLLER SYNTHESIS

affi

ne

rati

onal

dependence

non-r

ati

onal

# parametersmany

few

rate of variation q

zero (constant p)

bounded

infinitely fastLFT /

LPVLFT

LPV

LPV

LPV∗elseLPV

Xq ∈ Rm: LFT,

else X∗

oneq ∈ Rm: LFT,

else LPV∗∗∗∗p ∈ R1: KYP,

else LPV

Page 43: Gain Scheduling

Chapter 6

Fuzzy gain-scheduling

The restriction of classical gain scheduling techniques to near-operating point system behavior,not taking into account transient system dynamics, and additionally imposing a slow variation re-quirement, has lead to the development of fuzzy gain scheduling techniques. Fuzzy gain schedulingoriginates from the idea to relax these restrictions, while remaining closely related to the classicalgain scheduling approach (Leith & Leithead 2000).

Literature and research to fuzzy modeling and controller design methods is abundant andextensive, see e.g. (Guerra & Vermeiren 2001). It is not the intention to give an overview offuzzy modeling at this point. In this chapter, a brief overview of fuzzy methods leading to gainscheduling alike techniques is given. The basis of fuzzy gain scheduling is a fuzzy model or so-called blended multiple model representation, which is discussed in Section 6.1. The correspondingcontroller design is discussed in Section 6.2.

6.1 Fuzzy modeling

A blended multiple model representation or fuzzy modeling involves the blending of a set of modelsdescribing some nonlinear dynamics via point-designs. Examples of multiple model representationsare Takagi-Sugeno (fuzzy) models (e.g. (Sugeno & Kang 1988)), local model networks (Shorten,Murray-Smith, Bjorgan & Gollee 1988) or Polytopic Linear Models (PLMs) (Schulte 2005, Angelis2000). Targeting at subsequent application of gain scheduling alike techniques, i.e. to retain thepossibility of applying linear controller design techniques, commonly linear local models are used,e.g. (Gawthrop 1995, Tanaka & Sano 1994).

Consider a set of Nm local, indexed models Σ(θi), where i = 1, . . . , Nm, and θi ∈ Θi ⊂ Θparameterizing e.g. the system’s equilibrium operating points or a set of constant operating pointscovering the nonlinear, changing system dynamics. The local models are thus parameterized byθ = θi (compare to (3.11)). The equilibrium or constant operating points corresponding to theindexed models are defined by Ro(θi) = (xo(θi), wo(θi), uo(θi), zo(θi), yo(θi)).

Scalar weighting or validity functions αi(θα) are defined, where θα represents the dependenceof the weighting functions on the working conditions, i.e. θα = θα(x,w, u). Hence, by definitionθ differs from θα. Assuming all models have the same states, inputs and outputs, the nonlineardynamics (3.1) are approximated by a weighted combination of the models Σ(θi), θi ∈ Θi, yieldinga nonlinear fuzzy model

Σnl(θα) :

x(t) =Nm∑i=1

αi(θα){A(θi)xi +B1(θi)wi +B2(θi)ui

}z(t) =

Nm∑i=1

αi(θα){C1(θi)xi +D11(θi)wi +D12(θi)ui

}y(t) =

Nm∑i=1

αi(θα){C2(θi)xi +D21(θi)wi

} (6.1)

39

Page 44: Gain Scheduling

40 CHAPTER 6. FUZZY GAIN-SCHEDULING

The indexing of the states ˙xi(t), inputs wi(t), ui(t), and outputs zi(t), yi(t) of the local mod-els results from the model specific equilibrium values Ro(θi). Rewriting (6.1) according to thecoordinate transformation (3.7) yields

Σnl(θα) :

x(t) =Nm∑i=1

αi(θα){A(θi)x+B1(θi)w +B2(θi)u+ fo(θi)

}z(t) =

Nm∑i=1

αi(θα){C1(θi)x+D11(θi)w +D12(θi)u+ go(θi)

}y(t) =

Nm∑i=1

αi(θα){C2(θi)x+D21(θi)w + ho(θi)

} (6.2)

where

fo(θi) = −A(θi)xo(θi)−B1(θi)wo(θi)−B2(θi)uo(θi)go(θi) = zo(θi)− C1(θi)xo(θi)−D11(θi)wo(θi)−D12(θi)uo(θi)ho(θi) = yo(θi)− C2(θi)xo(θi)−D21(θi)wo(θi)

(6.3)

which is a time-variable, weighted combination of linear and / or affine state-space systems. Gen-erally, appropriate modeling assures the latter two terms go(θi) and ho(θi) to equal zero. Besideslinear modeling, which is considered throughout this report, literature on fuzzy modeling in gen-eral considers affine modeling as well. In case affine instead of linear local models are used, anadditional term appears in fo(θi) due to f(xo(θi), wo(θi), uo(θi)) 6= 0, see e.g. (Schulte 2005).

The main disadvantage of fuzzy modeling comprises the fact that the nonlinear, blended modeldynamics (6.2) can not be related to the original set of local point-design models, i.e. to theoriginal nonlinear plant dynamics. This is shown when considering the corresponding indexedlinearizations. Assuming the set of local models contains models of the form (3.11), the indexedlinearizations of (6.2) are given by

Σi(θα) :

˙x = (Ai +∇xαi(θα))x+ (Bi1 +∇wαi(θα))w + (Bi2 +∇uαi(θα))uz = (Ci1 +∇xαi(θα))x+ (Di

11 +∇wαi(θα))w + (Di12 +∇uαi(θα))u

y = (Ci2 +∇xαi(θα))x+ (Di21 +∇wαi(θα))w

(6.4)

where Ai = A(θi), Bi1 = B1(θi), etc. With respect to the original set of local models (3.11),the indexed linearizations (6.4) contain cross-terms involving derivatives of the weighting func-tion αi(θα). Regarding subsequent controller design, the time-varying, operating-point-dependentnature of αi(θα) thus has to be taken into account or blending should be minimal. Moreover, anal-ogous to classical gain scheduling, a slow-varying condition has to be fulfilled, however now withrespect to αi(θα) Furthermore, off-equilibrium points need not be stable, nor can performance beguaranteed (Shorten et al. 1988).

6.1.1 Weighting or scaling functions

The blending, utilizing the weighting or scaling functions αi(θα) is essential to the fuzzy gainscheduling approach. Typically, the weighting functions are normalized such that

∑Nm

i=1 αi(θα) = 1.

Furthermore, the functions are parameterized such that the local models switch over into each othersmoothly. Two example sets of scaling functions αi(θα) are shown in Figure 6.1. The design ofthese functions will not be discussed further to this point.

A distinction can be made between the case of local models describing the original, nonlinearplant dynamics i) at distinct operating points and ii) at operating regions of the operating envelopeof the nonlinear plant dynamics rather than distinct points. In the latter case, each local model isvalid in an extended operating region and blending correspondingly is confined to small transitionsat the boundaries between these operating regions. Blending thus plays a minor role with respectto the former case in which blending should provide smooth interpolation between the local models.

Page 45: Gain Scheduling

6.1. FUZZY MODELING 41

θα(x, w, u)0

1α4α3α2α1 α5

θα(x, w, u)0

1α4α3α2α1 α5

Figure 6.1: Two example distributions of the scaling functions αi(θα), i = 1, . . . , 5.

6.1.2 Approximation accuracy

One of the main advantages of fuzzy modeling as a basis for fuzzy gain scheduling involves thefact that the original nonlinear model may be approximated arbitrarily close by a fuzzy model.Moreover, the fuzzy model is a global approximation of the nonlinear model, thus taking intoaccount transient dynamics as well, contrary to a linearisation-based parameter-dependent set oflocal models as adopted in classical gain scheduling. In (Schulte 2005) an upper bound on thenumber of models Nm that is sufficient to construct a fuzzy model with predefined accuracy withrespect to the original nonlinear plant model is derived. These results are based on (Angelis 2000).The upper bound depends on an upper bound of the error, the size of the operating space and theestimation of the largest nonlinearity of the plant model.

Theorem (approximation accuracy)

(Schulte 2005) Consider the dynamics x = f(x,w, u) of the nonlinear model (3.1). Let f(x,w, u) =F [x,w, u]T + fnl(x,w, u) be the corresponding partitioning of the linear and nonlinear dynamicsand let ε ∈ R+ be an error bound. Determine the distance between the original nonlinear dynamicsf(x,w, u) and a corresponding nonlinear fuzzy model approximation f(x,w, u) as the supremumof the Euclidian error norm

d(f, f) := sup(x,w,u)

||f(x,w, u)− f(x,w, u)||2 (6.5)

Then the equation d(f, f) ≤ ε is satisfied, i.e. the original nonlinear model is approximated withε-accuracy by the corresponding fuzzy approximation, if Nm local models are employed, with

Nm =nnl∏i=1

ceil(

βi

2√

√λHnnl

√n

)(6.6)

where nnl = dim(fnl(x,w, u)) the number of nonlinear terms, βi the range of the correspondingnonlinear term i = 1, . . . , nN , n = dim(F ) the order of the model and λH = max

Gb,j(λHj

) the

maximum eigenvalue of all eigenvalues corresponding to the Hessian matrices that are derivedfrom the nonlinear terms.

Example 6.1 (Longitudinal vehicle control)Consider the nonlinear model (A.5), focusing on the dynamics x = f(x,w, u), where

Page 46: Gain Scheduling

42 CHAPTER 6. FUZZY GAIN-SCHEDULING

x ∈ Rnx the state, u ∈ Rnu the input and w ∈ Rnw as defined in (A.4). Rewriting this,explicitly discriminating between the linear and nonlinear parts, yields

x(t) =[0, 0, 0,

1M

]︸ ︷︷ ︸

F

x(t)w(1)(t)w(2)(t)u(t)

+− 1MC ′dx(t)

2 − g(sinw(1)(t) + Cr cosw(1)(t)︸ ︷︷ ︸fnl=[fnl,1, fnl,2]T

(6.7)

where the order of the system dynamics nx = 1 and the number of nonlinear termsnnl = 2. The bounds for the working space G = Rnx ×Rnw ×Rnu are determined via

0 < x(t) ≤ xmax, 0 ≤ w(1)(t) ≤ π4 ,

0 < w(2)(t) ≤ xmax, 0 < u(t) ≤ umax(6.8)

where xmax the maximum allowable velocity of the vehicle and umax the maximumdrive force. This yields a working envelope Gb. Using (6.8), the working ranges of theparameters, βi = Gb,i ⊂ Rnl,i, i = 1, 2 corresponding to the nonlinear terms fnl,i, aregiven by β1 = xmax, β2 = π/4. The Hessian matrices from the terms fnl,i, i = 1, 2 arecalculated using

Hi = 2.dfnl,ix, i = 1, 2 (6.9)

which yields the corresponding eigenvalues λHi

λH1 = − 2MC ′d

λH2 = g(sinw(1) + Cr cosw(1))(6.10)

Using the parameter values for a general passenger car, the maximal eigenvalue thenis

λH = max(x,w,u)∈Gb

{λHi} , i = 1, 2 (6.11)

= max(x,w,u)∈Gb

{− 2MC ′d, g(sinw

(1) + Cr cosw(1))}

(6.12)

= max{−5.1 · 10−4, 7.0

}(6.13)

Demanding an ε-accuracy of ε = 10 and using a maximal vehicle velocity xmax =150/3.6 m s−1, requires the use of Nm local models as a basis for the nonlinear fuzzymodel, with

Nm =nnl∏i=1

ceil(

βi

2√

√λHnnl

√nx

)(6.14)

= ceil(xmax

2√

20

√7.0 · 2

)· ceil

(π/4

2√

20

√7.0 · 2

)(6.15)

= 18 · 1 = 18 (6.16)

Assuming that the models with system matrices as defined in (3.18) are locally valid,equally spaced triangular functions are defined in the domain (x,w(1)) ∈ Gb,1 × Gb,2.The fuzzy model approximating the original nonlinear model (A.5) then is definedby (6.2) and (eq:TSmodelRes). Implementation of the nonlinear fuzzy model usingNm = 16 linear models, yields the results shown in Figure 6.2.

Page 47: Gain Scheduling

6.2. FUZZY GAIN SCHEDULING CONTROL 43

0 200 400 600−0.4

−0.2

0

0.2

0.4

f nl a

nd f fu

zzy

0 200 400 6000

10

20

30

40

x nl a

nd x

fuzz

y

0 200 400 600−5

0

5

10x 10

−3

erro

r f nl

−f fu

zzy

time [s]0 200 400 600

−0.4

−0.2

0

0.2

0.4

erro

r x nl

−x fu

zzy

time [s]

Figure 6.2: Comparison of simulation results corresponding to the original nonlinear model (A.5)(solid black) and the approximated model using Nm = 16 equally distributed LTI models (dottedgrey).

6.2 Fuzzy gain scheduling control

Consider the fuzzy model Σnl(θα) (6.2), based on a set of local, indexed models Σ(θi), where θi ∈Θi ⊂ Θ parameterizing e.g. the equilibrium operating points Θi = Θe or a set of constant operatingpoints Θi = Θo covering the nonlinear, changing system dynamics appropriately. Utilizing Σ(θi)as a basis, either blended, local controller design techniques, or LMI-based LPV methods maybe applied to arrive at a parameter-dependent controller. As no direct relationship between thedynamic characteristics of the fuzzy model and the characteristics of the original set of localmodels is present, the analysis and design of corresponding controllers with guaranteed stabilityand performance specifications is difficult. In the remainder of this section, the Parallel DistributedControl (PDC) synthesis (Guerra & Vermeiren 2001), which is an often applied specific form ofblended, local controller designs, is briefly explained. LMI-based LPV methods, such as describedin e.g. (Tanaka & Sano 1994, Guerra & Vermeiren 2001, Angelis 2000), will not be discussedfurther.

6.2.1 Parallel Distributed Control

Parallel Distributed Control (PDC) synthesis involves the blending of a set of LTI controllers Λ(θi),corresponding to Σ(θi). The same weighting or scaling functions αi(θα) as for the modeling areutilized, yielding a nonlinear gain-scheduled controller immediately. The number of local controllerdesigns equals the number of models, which is determined on the basis of the desired approximationaccuracy (see Section 6.1.2). PDC is strongly related to classical gain scheduling controller designtechniques. However, the fixed-structure LTI controller design requirement to enable appropriateinterpolation or scheduling when employing classical gain scheduling techniques, may be relaxedsignificantly. The controller outputs are blended (scheduled) rather than scheduling e.g. thecontroller coefficients.

Page 48: Gain Scheduling

44 CHAPTER 6. FUZZY GAIN-SCHEDULING

The PDC control law comprises a feedback and a feedforward alike part. The feedback partΛfb(θα) directly follows from the local controller designs Λ(θi).

Λfb(θα) :=Nm∑i=1

αi(θα)Λ(θi) (6.17)

Focusing on the nonlinear fuzzy model Σnl(θα) (6.2), the feedback part Λfb(θα) only holds for theparameterized, linear part. I.e. the corresponding offsets (6.3) are not compensated. Accordingly,a feedforward alike part, Λff (θα) is determined. As this compensator will depend on θα(x, u, w),which may consist of measured states, it is no feedforward in its strict sense. Consider the casewhen go(thetai) = ho(thetai) = 0. The corresponding feedforward compensator then becomes

Λff (θα) := −{B2(θα)TB2(θα)

}−1B2(θα)T

Nm∑i=1

αi(θα)fo(θi) (6.18)

where

B(θα) :=Nm∑i=1

αi(θα)B2(θi) (6.19)

The resulting gain-scheduled controller then equals Λ(θα) = Λfb(θα) + Λff (θα), where θα =θα(x, u, w) is the scheduling vector.

Page 49: Gain Scheduling

Chapter 7

Conclusions and future work

7.1 Conclusions

Modeling and scheduling variable

Most literature on gain scheduling focuses on the design of a scheduled controller rather than thecorresponding modeling. Commonly, it is assumed that a parameterized set of local models or evena LPV model is present. In that case, this report discusses various techniques and guidelines how todesign a corresponding scheduled controller. However, focusing on continuous gain scheduling of aparameterized set of local models, several issues involving the design of this set have not explicitlybeen considered; How to select the number of models to design, how to select the correspondingoperating points and how to select the scheduling variables θ? As said, in general these issues arenot considered in gain scheduling.

In this sense, fuzzy gain scheduling, more specifically the theorem considering the approxima-tion accuracy of a parameterized set of local models with respect to the original nonlinear model,might provide a good starting point. In general, research on fuzzy modeling is abundant, contraryto research to the design of a parameterized set of LTI models. A required number of LTI modelsis determined. The corresponding operating points are assumed to be equally distributed overthe envelope of operating conditions. If these points do not represent constant operating points,equilibrium operating conditions have to be considered instead. Otherwise, velocity-based gainscheduling might resolve the problem.

Typically, determination of the scheduling variable θ is based on the parameterized set ofLTI models. This requires insight in the system as well as experience. An often used rule-of-thumb states that the scheduling variable should incorporate the main nonlinearities of thesystem. Changing operating conditions typically are a result of nonlinearities. Furthermore, theparameterized set of LTI models or the LPV model should reflect the nonlinear system behaviorfor the complete envelope of operating conditions. Hence, the scheduling variable should capturethe system’s nonlinearities. Quasi-LPV techniques are based on this idea an may thus provide asuitable solution. Besides that, the scheduling variable θ obviously should be measurable or atleast be observable to enable actual implementation of the scheduled controller.

Application of gain scheduling

All gain scheduling techniques involve firstly parameter-dependent modeling and subsequentlyparameter-dependent, i.e. scheduled controller design. Analysis of the modeling involves stabilityanalysis as well as analysis of the approximation accuracy. The controller design requires stabilityand performance analysis. Taking in mind that gain scheduling is employed in an ad hoc manner atDAF as well as at TNO (see 1), focus lies on the stability and performance analysis correspondingto the controller design. I.e. the main question is not whether a system can be approximatedsufficiently close with a scheduling model based on a set of linear models, but whether gain

45

Page 50: Gain Scheduling

46 CHAPTER 7. CONCLUSIONS AND FUTURE WORK

scheduling of the controller gains is required to assure stability and performance specifications oreven enable tightening of these specifications over the entire operating envelope.

The before mentioned fuzzy theorem regarding approximation accuracy of the parameterizedset of LTI models, can be interpreted as a measure of the nonlinearity of the system. Hence, ismight be used as a measure of the need to apply a gain scheduling controller design. Obviously, thisdoes not hold in case gain scheduling is adopted to improve closed-loop performance by employinga nonlinear controller to a linear system, e.g. by scheduling the controller parameters accordingto the error rather than a system nonlinearity. In general, this is an open issue in control theory(Aangenent 2008).

In general, application of gain scheduling in case of a system with slowly varying dynamicsis assumed to be beneficial regarding the resulting closed-loop performance. When a measurablescheduling variable, reflecting the changing dynamics appropriately, can be determined, applicationof gain scheduling instead of robust control should be considered. Typically, nonlinear systems areconsidered, whereas robust control focuses on linear system representations with immeasurableparameter variations or disturbances.

Gain scheduling methods

Classical gain scheduling and various related gain scheduling techniques, LPV and LFT synthesisand fuzzy gain scheduling are discussed in this report.

The main advantage of classical gain scheduling is that it inherits the benefits of linear controllerdesign methods, including intuitive classical design tools and time as well as frequency domainperformance specifications. PID control is the most used control strategy in industrial applicationsdue to its relatively simple and intuitive design, hence, this is a major advantage with respectto other nonlinear controller design syntheses. The approach thus enables the design of lowcomputational effort controllers. Conceptually, gain scheduling involves an intuitive ’simplification’of the problem into parallel decompositions of the total system.

The major drawback of classical gain scheduling involves the lack of guaranteed global ro-bustness, performance and especially stability. The robustness, performance and even nominalstability properties of the global gain scheduled controller are not addressed explicitly in the de-sign process. Such properties are rather inferred from extensive simulations. Only in the case ofslowly varying parameters, these designs can guarantee stability. The key issues are i) the factthat transitions among operating conditions are not addressed in the design and ii) shortcomingsof the linearizations. Hence, extensive offline testing is required to establish global stability andperformance guarantees. An often used rule-of-thumb states that the scheduling variable shouldonly vary slowly to prohibit the introduction of additional dynamics.

Generally, the scheduling is based intuitively on physical variables in the plant. In case ofthe quasi-LPV however, a coordinate transformation has to be applied. As this approach isconservative, this might lead to infeasible controller design problem, i.e. no feasible controllermay be available. Quasi-LPV techniques on the other hand, enable global stability as well asperformance and robustness analysis. Furthermore, in practice, classical gain scheduling often isapplied in an ad hoc manner, which is not suited for more difficult problems.

LPV and LFT synthesis require a true LPV model as a basis. In general however, gainscheduling may be employed in the absence of an analytical model, e.g. on the basis of a collectionof plant linearizations. Consequently, controller design based on whitebox as well as blackboxand even data-based ’modeling’ is possible. If the possibility of fast parameter variations is notaddressed in the design process, guaranteed properties of the overall gain-scheduled design cannotbe established. The main advantage of LPV and LFT control synthesis is that they do accountfor parameter variations in the controller design, which results in a priori guarantees regardingstability and performance specifications.

The main drawback of LPV and LFT control synthesis involves conservativeness, which hasto be introduced to enable solving the resulting LMIs. As a result of that, current LPV and LFTsyntheses comprise specific extensions of robust control techniques rather than true generalizations.However, current and future research still provides and will provide less conservative solutions.

Page 51: Gain Scheduling

7.2. FUTURE WORK (TBD) 47

Furthermore, LPV and LFT control syntheses are relatively involved, i.e. designing a SISO LPVcontroller for a LPV model of order 4 has to be regarded as computationally complex already.

Fuzzy gain scheduling is only considered briefly in this report. Focus lies on fuzzy modelingand controller design following classical gain scheduling methods. The main drawback of fuzzygain scheduling involves the lack of a relation between the dynamic characteristics of the originalnonlinear model and the fuzzy model. Even locally, the dynamics of the fuzzy model can not berelated to the original nonlinear model. Whereas some stability and performance specifications canbe derived in special cases for classical gain scheduling and even a priori stability and performancecan be guaranteed when LPV or LFT control synthesis is adopted, this is thus not the case forfuzzy gain scheduling.

7.2 Future work (tbd)

Page 52: Gain Scheduling

48 CHAPTER 7. CONCLUSIONS AND FUTURE WORK

Page 53: Gain Scheduling

Bibliography

Aangenent, W. (2008), Nonlinear control of linear motion systems, PhD thesis, Eindhoven Uni-versity of Technology. (to be published).

Angelis, G. (2000), System analysis, modelling and control with polytopic linear models, PhDthesis, Eindhoven University of Technology.

Apkarian, P. & Gahinet, P. (1995), ‘A convex characterization of gain-scheduled H∞ controllers’,IEEE Trans. Auto. Control 40(5), 853–864.

Apkarian, P., Gahinet, P. & Becker, G. (1995), ‘Self-scheduled H∞ control of linear parameter-varying systems: a design example’, Automatica 31(9), 1251–1261.

Becker, G. & Packard, A. (1994), ‘Robust performance of linear parametrically varying systemsusing parametrically-dependent linear feedback’, Syst. Control Lett. 23(23), 205–215.

Becker, G., Packard, D., Philbrick, D. & Balas, G. (1993), ‘Control of parameterically-dependentlinear systems, a single quadratic Lyapunov approach’.

Bianchi, F., Mantz, R. & Christiansen, C. (2007), ‘Multivariable PID control with set-pointweighting via BMI optimisation’, Automatica 44, 472–478.

Claveau, F., Chevrel, P. & Knittel, K. (2007), ‘A 2DOF gain-scheduled controller design method-ology for a multi-motor web transport system’, Control Engineering Practice . CorrectedProof.

Desoer, C. (1969), ‘Slowly varying system x = a(t)x’, IEEE Transactions on Automatic Control14(6), 780–781.

Dettori, M. (2001), LMI techniques for control with application to a compact disc player mecha-nism, PhD thesis, Delft University of Technology.

Dettori, M. & Scherer, C. (2002), ‘Mimo control design for a compact disc player with multiplenorm specifications’, IEEE Trans. Control Syst. Tech. 10(5), 635–645.

Dinh, M., Scorletti, G., Fromion, V. & Magarotto, E. (2005), ‘Parameter dependent h∞ controlby finite dimensional LMI optimization: application to trade-off dependent control’, Int. J.Robust Nonlin. Control 15(9), 383–406.

Doyle, J., Packard, A. & Zhou, K. (1991), ‘Review of LFTs, LMIs, and µ’, Proc. of the 30th Conf.on Decision and Control pp. 1227–1232.

Gahinet, P., Apkarian, P. & Becker, G. (1996), ‘Affine parameter-dependent lyapunov functionsand real parametric uncertainty’, IEEE Trans. Autom. Control 41(3), 436–442.

Gawthrop, P. (1995), ‘Continuous-time local state local model networks’, Proc. of the IEEE con-ference on Systems, Man and Cybernetics pp. 852–857.

49

Page 54: Gain Scheduling

50 BIBLIOGRAPHY

Guerra, T. & Vermeiren, L. (2001), ‘Control laws for Takagi-Sugeno fuzzy models’, Fuzzy Setsand Systems 120, 95–108.

Helmersson, A. (1995), Methods for robust gain-scheduling, PhD thesis, Linkoping University,Sweden.

Helmersson, A. (1996), Pearameter dependent Lyapunov functions based on linear fractional trans-formation, in ‘Proc. IFAC World Congress’.

Iwasaki, T. & Hara, S. (2005), ‘Generalized KYP lemma: unified frequency domain inequalitieswith design applications’, IEEE Trans. Automat. Control. 50(1), 41–59.

Iwasaki, T., Meinsma, G. & Fu, M. (2000), ‘Generalized S-procedure and finite frequency KYPlemma’, Math. Prob. Eng. 6, 305–320.

Iwasaki, T. & Shibata, G. (2001), ‘LPV system analysis via quadratic separator for uncertainimplicit systems’, IEEE Trans. Automat. Control 46(8), 1195–1208.

Kellet, M. (1991), ‘Continuous scheduling of H∞ controllers for a MS760 paris aircraft’, Robustcontrol design using H∞ and related methods pp. 197–219.

Khalil, H. (1992), Nonlinear systems, MacMillan, New York.

Koc, H., Knittel, D., de Mathelin, M. & Abba, G. (2002), ‘Modeling and robust control for windingsystems for elastic webs’, IEEE Transactions on Control Technology 10(2), 197–208.

Lawrence, D. & Rugh, W. (1995), ‘Gain scheduling dynamic linear controllers for a nonlinearplant’, Automatica 31(3), 381–390.

Leith, D. & Leithead, W. (1998a), ‘Appropriate realisation of MIMO gain-scheduled controllers’,Int. Journal of Control 70(1), 13–50.

Leith, D. & Leithead, W. (1998b), ‘Gain-scheduled and nonlinear systems: dynamic analysis byvelocity-based linearization families’, Int. Journal of Control 70(2), 289–317.

Leith, D. & Leithead, W. (1998c), ‘Gain-scheduled controller design: an analytic framework di-rectly incorporating non-equilibrium plant dynamics’, Int. Journal of Control 70(2), 249–269.

Leith, D. & Leithead, W. (1999), ‘Analytic framework for blended multiple model systems usinglinear local models’, Int. Journal of Control 72(7), 605–619.

Leith, D. & Leithead, W. (2000), ‘Survey of gain-scheduling analysis and design’, InternationalJournal of Control 73(11), 1001–1025.

Leith, D., Tsourdos, A., White, B. & Leithead, W. (2000), ‘Velocity-based gain-scheduled lateralauto-pilot for an agile missile’, IFAC Journal of Control Engineering Practice 9(10), 1079–1093.

Masubuchi, I., Ohara, A. & Suda, N. (1995), LMI-based output feedback controller design using aconvex parameterization of full-order controllers, in ‘Proc. 1995 Am. Control Conf.’, Vol. 5,Seattle, WA, USA, pp. 3473–3477.

McNichols, K. & Fadali, M. (2003), ‘Selecting operating points for discrete-time gain scheduling’,Computers and Electrical Engineering 29, 289–301.

Naus, G. (2007a), Evaluation of controller design and tuning at DAF Trucks N.V., Internal report51051/07-071, DAF Trucks N.V.

Naus, G. (2007b), Evaluation of controller design and tuning at Integrated Safety, Business UnitAutomotive, TNO Helmond, Internal report, TNO Automotive.

Page 55: Gain Scheduling

BIBLIOGRAPHY 51

Naus, G. & Huisman, R. (2007), Tuning AS-Tronic manoeuvring mode with Multi-Pulse fuelinjection, Internal report 51051/07-034, DAF Trucks N.V.

Naus, G., van den Bleek, R., Ploeg, J., Scheepers, B., van de Molengraft, M. & Steinbuch, M.(2008), Explicit MPC design and performance evaluation of an ACC Stop-&-Go, in ‘2008American Control Conference’, Seattle, WA, United States, pp. 224–229.

Nichols, R., Reichert, R. & Rugh, W. (1993), ‘Gain scheduling for H∞ controllers: a flight controlexample’, IEEE Transactions on Control Systems Technology 1(2), 69–79.

Niemann, H. & Stoustrup, J. (1999), ‘Gain scheduling using the Youla parameterization’, Proc.of the 38th Conf. on Decision and Control pp. 2306–2311.

Packard, A. (1994), ‘Gain scheduling via linear fractional transformations’, Systems and ControlLetters 22, 79–92.

Paijmans, B., Symens, W., v. Brussel, H. & Swevers, J. (2008), ‘Identification of interpolatingaffine LPV models for mechatronic systems with one varying parameter’, European Journalof Control accepted for publication.

Pellanda, P., Apkarian, P. & Alazard, D. (2000), ‘Gain-scheduling through continuation ofobserver-based realizations - applications to H∞ and µ controllers’, Proc. of the 39th IEEEConf. on Decision and Control 3, 2787–2792.

Rantzer, A. (1996), ‘On the Kalman-Yakubovich-Popov lemma’, Syst. Control Lett. 28(1), 7–10.

Reichert, R. (1992), ‘Dynamic scheduling of modern-robust-control autopilot designs for missiles’,IEEE Control Systems pp. 35–42.

Rieber, J., Scherer, C. & Allgower, F. (2008), ‘Robust l2 performance analysis for linear systemswith parametric uncertainties’, Int. J. Control 81(5), 851–864.

Rugh, W. (1991), ‘Analytical framework for gain scheduling’, IEEE Control Systems Magazine11(1), 79–84.

Rugh, W. & Shamma, J. (2000), ‘Research on gain scheduling’, Automatica 36, 1401–1425.

Scherer, C. (1999), Robust mixed control and LPV control with full block scalings, Siam, Philadel-phia.

Scherer, C. (2001), ‘LPV control and full block multipliers’, Automatica 37, 361–375.

Scherer, C. (2005), ‘Relaxations for robust linear matrix inequality problems with verification forexactness’, SIAM J. Matrix Anal. A. 27, 365–395.

Scherer, C., Gahinet, P. & Chilali, M. (1997), ‘Multiobjective output-feedback control via LMIoptimization’, IEEE Trans. Automat. Control 42, 896–911.

Schulte, H. (2005), ‘Approximate modeling of a class of nonlinear oscillators using Takagi-Sugenofuzzy systems and its application to control design’, Proc. of the 44th IEEE Conf. on Decisionand Control and the European Control Conference pp. 3387–3392.

Schutter, B. D. & Heemels, W. (2004), ‘Modeling and control of hybrid systems’, Lecture notes ofthe DISC course .

Scorletti, G. & Ghaoui, L. E. (1995), Improved linear matrix inequality conditions for gain-scheduling, in ‘Proc. 34th IEEE Conf. Dec. Control’, New Orleans, LA, pp. 3626–3631.

Shahruz, S. & Behtash, S. (1990), ‘Design of controllers for Linear Parameter-Varying systemsby the gain scheduling technique’, Proc. of the 29th Conf. on Decision and Control pp. 2490–2491.

Page 56: Gain Scheduling

52 BIBLIOGRAPHY

Shamma, J. & Atans, M. (1991), ‘Guaranteed properties of gain scheduled control for LinearParameter-Varying plants’, Automatica 27(3), 559–564.

Shamma, J. & Athans, M. (1990), ‘Analysis of gain scheduled control for nonlinear plants’, IEEETransactions on automatic control 35(8), 898–907.

Shamma, J. & Athans, M. (1992), ‘Gain scheduling: potential hazards and possible remedies’,IEEE Control Systems Magazine 12(3), 101–107.

Shorten, R., Murray-Smith, R., Bjorgan, R. & Gollee, H. (1988), ‘On the interpretation of localmodels in blended multiple model structures’, International Journal of Control 72(7), 620–628.

Stilwell, D. (1999), ‘J −Q interpolation for gain scheduled controllers’, Proc. of the 38th Conf.on Decision and Control pp. 749–754.

Stilwell, D. & Rugh, W. (1998), ‘Interpolation methods for gain scheduling’, Proc. of the 37thIEEE Conf. on Decision and Control pp. 3003–3008.

Stilwell, D. & Rugh, W. (1999), ‘Interpolation of observer state feedback controllers for gainscheduling’, IEEE Transactions on Automatic Control 44(6), 1225–1229.

Stilwell, D. & Rugh, W. (2000), ‘Stability preserving interpolation methods for the synthesis ofgain scheduling controllers’, Automatica 36, 665–671.

Sugeno, M. & Kang, G. (1988), ‘Structure identification of fuzzy model’, Fuzzy sets and systems28, 15–33.

Tanaka, K. & Sano, M. (1994), ‘A robust stabilization problem of fuzzy control systems and itsapplication to backing up control of a truck-trailer’, IEEE Transactions on Fuzzy Systems2(2), 119–134.

v. Helvoort, J., Steinbuch, M., Lambrechts, P. & v.d. Molengraft, M. (2004), ‘Analytical andexperimental modelling for gain-scheduling of a double scara robot’, 3rd IFAC Symposiumon Mechatronic Systems, Sydney, Australia pp. 61–66.

Wang, D., Nishimura, H. & Shimogo, T. (2007), ‘Active control of shock by gain scheduling’,Journal of Sound and Vibration 308, 647–659.

Wu, F. & Dong, K. (2006), ‘Gain-scheduling control of LFT systems using parameter-dependentLyapunov functions’, Automatica 42, 39–50.

Wu, F., Packard, A. & Balas, G. (1995), ‘LPV control design for pitch-axis missile autopilots’,Proc. of the 34th Conf. on Decision and Control pp. 188–193.

Wu, F., Yang, X., Packard, A. & Becker, G. (1996), ‘Induced l2-norm control for LPV systemswith bounded parameter variations rates’, Int. J. Robust Nonlin. Control 6(9–10), 983–998.

Yu, J. & Sideris, A. (1995), H∞ control with parametric Lyapunov functions, in ‘Proc. 34th Conf.Dec. & Control’.

Zhou, K., Doyle, J. & Glover, K. (1996), Robust and optimal control, Prentice Hall, New Yersey.

Zhou, K., Khargonekar, P., Stoustrup, J. & Niemann, H. (1992), ‘Robust stability and performanceof uncertain systems in state space’, Proc. of the 31st conference on decision and controlpp. 662–667.

Page 57: Gain Scheduling

Appendix A

Longitudinal vehicle control

Consider a vehicle with mass M , which is subject to the following forces (see Figure A.1)

Fair(t) = −12ρCdAdq(t)2

= −C ′dq(t)2 (A.1)Fslope(t) = −Mg sinα(t) (A.2)Froll(t) = −CrMg cosα(t) (A.3)

where Fair(t), Froll(t) and Fslope(t) are the air drag, the rolling resistance and road slope resultantforce respectively, and ρ, Cd, Ad, g and Cr appropriate vehicle and environmental characteristics.The slope α(t) is regarded as an external disturbance and the vehicle speed q(t) is regarded as thesystem output. Corresponding parameter values for a passenger car are given in Table A.1.

q Fair

FrollFrollα

Fslope

Figure A.1: A vehicle and its corresponding environmental disturbances.

The control goal is to achieve a desired vehicle speed qd(t), which is established by control ofthe throttle and brakes, yielding a resulting control force Ft/b(t). Define the external input vectorw(t)

w(t) =[w(1)(t), w(2)(t)

]T= [α(t), xd(t)]

T (A.4)

The nonlinear model for longitudinal vehicle control is hence given by

Σnl :

x(t) = − 1MC ′dx(t)

2 − g(sinw(1)(t) + Cr cosw(1)(t)) + 1M u(t)

z(t) = w(2)(t)− x(t)y(t) = x(t)

(A.5)

53

Page 58: Gain Scheduling

54 APPENDIX A. LONGITUDINAL VEHICLE CONTROL

Table A.1: General passenger car characteristic parameter valuesQuantity Symbol Value Unitmass M 1400 kgfrontal area Ad 2 m2

air drag coefficient Cd 0.3 −rolling resistance Cr 0.015 −wheel radius wr 0.3 mfinal drive ratio fr 4.0 −gear ratio gr 3.4-2.1-1.4-1.0-0.77 −idle speed ωi 73.3 rad s−1

air density ρ 1.2 kg m−3

gravity g 9.81 m s−2

where u(t) = Ft/b the control input, x(t) = q(t) the state of the model and correspondinglyxd(t) = qd(t) the desired vehicle velocity.

Page 59: Gain Scheduling

Appendix B

Engine - idle governor

The engine is regarded as a system with fuel value Fv in mg as input and the rotational speedωe in rpm as output (see Figure B.1). Depending on the present load, the operating point ofthe engine and the temperature of the engine, the plant He exhibits different characteristics. Ina gain-scheduling context only the temperature dependency is taken into account for the sake ofsimplicity. The temperature of the coolant is measured and adopted as an approximation for thetemperature of the engine, i.e. Te ≈ Tc. Furthermore, the temperature dependence is regarded asan exogenous variable.

Engine= He(Te)

ωe rpmFv mg

Te K

Figure B.1: The engine system He

Based on step-response as well as frequency-response measurements, a nominal LTI model He,n

is determined for a heated engine (Naus & Huisman 2007).

He,n(Tc,n) = e−0.07s 2501.85s+ 1

for Tc,n = 355 K (B.1)

Identification of the temperature dependency of the engine response is based on testbenchmeasurements (Naus & Huisman 2007). These measurements indicate that only the system gainis temperature-dependent. The higher the engine temperature Te, the higher the viscosity of thelubricant and thus the less friction will be present. The effect of expanding parts is assumed tobe minor with respect to this phenomenon. Assuming the differences between the testbench andthe experimental measurement results can be related to a different static gain only, the resultingtemperature-dependent engine gain equals

G(Tc) = 2.2Tc(t)− 529.8, Tc ∈ [300 400] K (B.2)

where gaine,tb(t) in rpm/% and Tc the temperature of the cooling water in Kelvin. The %indicates the percentage of the maximal possible injection and corresponds to about 3 mg perpercent.

The resulting model for the engine system is given by

He(Tc) = e−0.07s G(Tc)1.85s+ 1

Tc ∈ [300 400] K (B.3)

where G(Tc) = g1Tc + g2 is the temperature-dependent gain of the engine as defined in (B.2).

55