Control of Variable Geometry Turbocharged Diesel Engines for Reduced Emissions1

Department of Mechanical and Environmental Engineering, UC, Santa Barbara, CA 93106 Ford Research Laboratory, Ford Motor Company, Dearborn, MI 48121 + Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI 48109
z

A. G. Stefanopoulouz , I. Kolmanovsky , J. S. Freudenberg+

Abstract
An emission control problem for an automotive direct injected compression ignition (diesel) engine equipped with exhaust gas recirculation (EGR) and a variable geometry turbocharger (VGT) is considered. The objective is to operate the engine to meet driver's torque demand and minimize NOx emissions while at the same time avoiding visible smoke generation. It is demonstrated that the steady-state optimization of engine emissions results in operating points where EGR and VGT actuators are in e ect redundant in their e ect on the variables that most directly a ect the emissions. A multivariable feedback controller is proposed which accounts for this actuator redundancy. Furthermore, it coordinates the two actuators to fully utilize their joint e ect on engine emission performance. Experimental results con rm good response properties of the proposed controller. Keywords: Multivariable Control, Emission, Automotive

Variable Geometry Turbine Wt

N tc

Compressor W c1

EGR W 21 Wf Exhaust Manifold 2 P2 F2

Engine
W 1e

1 P1 F1

Intake Manifold

1 Introduction
In this paper we consider an automotive control problem for a variable geometry turbocharged (VGT) compression ignition direct injection (CIDI or also referred to as diesel) engine with an external exhaust gas recirculation system. The engine is shown schematically in Figure 1. The turbine converts energy of the exhaust gas into mechanical energy of the rotating turboshaft that, in turn, drives the compressor. The compressor increases the density of air supplied to the engine this larger mass of air can be burnt with a larger quantity of fuel thereby resulting in a larger torque output as compared to (non-turbocharged) naturally aspirated engines 20]. The power generated by the turbine depends on the pressure and temperature of the exhaust gas and on the mass ow rate of the exhaust gas through the turbine. A turbine with variable geometry uses inlet guide vanes (located on the turbine stator) to modify its e ective ow area and provide a better match between the turbocharger and the CIDI engine 15]. Through changes in inlet guide vane positions, the power transfer to the turbine and, hence, to the compressor can be modi ed and the

Figure 1: Schematic representation of the diesel (CIDI)

engine with the relevant ows, gas property states and rotational speeds.

air ow to the intake manifold can be controlled. For example, if the EGR valve is closed, closing the vanes results in a smaller turbine e ective ow area, higher exhaust manifold pressure, increased power generation by the turbine and increased ow of air from compressor into the intake manifold. Opening the vanes produces the opposite e ect. To reduce the emissions of harmful oxides of nitrogen (NOx ) a portion of the exhaust gas can be diverted back to the intake manifold to dilute the air supplied by the compressor. This process is referred to as exhaust gas recirculation (EGR). It is, typically, accomplished with an EGR valve that connects the intake manifold and the exhaust manifold (See Figure 1). In the cylinders the recirculated combustion products act as inert gas thus lowering the combustion temperature and, hence, decreasing the formation of NOx 11]. The experimental data in Figure 2 illustrate the trend. A high level of burned gas fraction, F1 , on the other hand, can lead to low in-cylinder air-to-fuel ratio (AFR), and consequently, unacceptable smoke generation as Figure 2

1 Support is provided by the National Science Foundation under contracts NSF ECS-94-14822 (Freudenberg) and ECS-9733293 (Stefanopoulou) matching funds to these grants were provided by Ford Motor Company.

demonstrates. First, burned gas fraction from the EGR valve displaces fresh air from the intake manifold and consequently decreases the fresh air mass ows into the cylinders. Secondly, a fraction of the exhaust gas that can be used by the turbine is diverted through the EGR valve to the intake manifold, reducing the turbine power and consequently the ow delivered to the intake manifold through the compressor. In this work, we show that the need to prevent smoke generation will limit our ability to reduce NOx emissions. The experimental data for NOx and smoke generation versus AFR are shown in Figure 3. Note that increase in the in-cylinder air-to-fuel ratio at constant fueling level and intake manifold pressure requires simultaneous reduction in the burned gas fraction. The decrease in burned gas fraction, consequently cause increased NOx emission. The typical air-to-fuel ratio versus NOx emission curves shown in the literature are for constant levels of burned gas fraction. In contrast to the Figure 3 shown here, these typical gures show NOx emission decreasing when air-to-fuel ratio is increasing.
Wf=5-6 kg/hr,N=2480-2520 rpm for different MAPs 9 Mon Dec 21 03:42:25 EST 1998
p1=140 kPa p1=160 kPa p1=180 kPa

Wf=4-5 kg/hr, N=2480-2520 rpm for different MAPs 7 Mon Dec 21 03:29:19 EST 1998
p1=120 kPa p1=140 kPa p1=160 kPa

NOx (g/kWhr) and smoke (Bosch Units)

NOx

Smoke

0 20

25

30

35 40 AFR [afr] (none)

45

50

55

Figure 3: Oxides of nitrogen and smoke generation ver-

sus in-cylinder air-to-fuel ratio. Bosch 4 is considred visible smoke, Bosch 2 is considered a safe bound for avoiding visible smoke. Increasing AFR without maintaining constant dilution in the cylinders causes increased NOx generation.

7 NOx (g/kWhr) and smoke (Bosch Units)

NOx

EGR and V GT actuators do not have the authority to manipulate AFR and F1 independently, i.e. the actuators become essentially redundant. The latter fact manifests itself in that the DC gain matrix of the plant linearization at the optimal operating point is almost rank de cient. To address these issues, we analyze the directional properties of the plant using the singular value decomposition 8] of the plant DC gain matrix. We consequently propose a combination of nonlinear feedforward and gain-scheduled linear feedback controller architecture that accounts for this actuator redundancy. Experimental results con rm good response properties of the proposed control scheme. Finally, in \Concluding Remarks", we point out that the problem of actuator redundancy caused by the necessity to operate at optimized operating points may occur more frequently in automotive applications in the future due to tightening emission regulations and performance requirements. For a comprehensive review of the state of the art in diesel engine control, including VGT control, see 9] and references therein. Here we only comment on the papers that are most relevant to the present investigation. Reference 4] reports on a control system development e ort for a heavyduty diesel engine with EGR valve and VGT. At tip-ins the control system closes the EGR valve and uses a PID feedback on boost pressure to rapidly increase air supply to the intake manifold with VGT. In conditions near steady-state operation, the control system fully opens up the EGR valve and regulates the VGT vane position open loop to the setting required to achieve the desired EGR rate. This strategy would be applicable to the vehicles that operate mostly in steady state conditions, such as heavy commercial vehicles. The approach of treating fast transient operation and close to steady-state operation di erently is motivated to a large extent by the multiple objectives of the control design such as acceleration performance (point-wise in time con-

Smoke
1

0.05

0.1

0.15 0.2 F1 [f1] (none)

0.25

0.3

0.35

Figure 2: Oxides of nitrogen and smoke generation versus

burned gas fraction.

In this paper, we rst describe a procedure for generating optimal setpoints for F1 and AFR as a function of fueling rate (Wf ) and engine speed (N ). Speci cally, we use a nonlinear engine model (see Appendix A) to select desired steady-state operating points where the value of F1 is maximized subject to the constraint that the AFR must remain above an appropriate level that ensures that no visible smoke is generated from the engine. Our next step is to develop a control scheme with the objective of regulating the performance variables to the optimized setpoints. The performance variables are not measured directly and the control scheme relies on the measurements of compressor mass air ow and absolute pressure in the intake manifold. Two issues arise in the development of this control scheme. Firstly, the fact that we have optimized the performance variables implies that the desired operating points tend to lie on the boundary of the set of feasible engine operating points. As a consequence, relatively small errors in the strategy that generates the setpoints (e.g. due to interpolation of values in a look-up table) will tend to generate infeasible setpoints. Secondly, at the desired setpoints, the

straints on smoke) and emission performance (constraints on the averaged performance over a driving cycle). See also a related discussion in 3, 13]. Reference 2] also considers a heavy-duty engine with EGR valve and VGT that are open-loop controlled to their desired settings determined based on the values of the engine speed and fueling rate. Reference 17] reports on the development of a multivariable PI feedback controller for controlling EGR and VGT based on measurements of intake manifold pressure and compressor mass air ow. The feedback controller is developed using model-based linearization, a xed structure H1 optimization-based design and gain-scheduling. The set-points for the controller in 17] are developed for minimum fuel consumption subject to air-to-fuel ratio constraint to prevent visible smoke. Here the set-points are developed so that the aggressive objectives of NOx emission reduction are incorporated. As will be demonstrated in this paper the need to sustain the operation of the engine at these NOx emission friendly set-points lead to a di erent multivariable controller architecture than in 17]. A nonlinear control design methodology based on the method of Control Lyapunov functions and domination redesign is applied to control of EGR valve and VGT in 10]. As in 17], in 10] the important interplay between set-point selection and controller architecture remain unexplored.

N External Signals Control Signals Wf EGR VGT Wc1 p1 Engine Model

Rotational Dynamics

T load

Tq F1 AFR Engine Performance Variables Simplified Emissions Model NOx Smoke

System Performance Variables

Measurements

Figure 4: Signal de nition for the CIDI control problem.

EGR and V GT control in reducing AFR excursions. Consider, temporarily, xed V GT and EGR and an in-

2 Performance Objectives
The system performance objective is to reduce NOx emissions and avoid visible smoke generation while maintaining fast engine torque response. Torque response in CIDI engines depends mostly on the fueling rate, Wf (kg/hr), and engine speed, N (rpm), during lean operation. We assume a simple fueling strategy which is a static map that determines the appropriate fueling rate value based on driver's pedal position and engine speed, N . Thus, the fueling rate, Wf , and engine speed, N , are treated as external inputs to the engine (w) and not as control signals. Figure 4 shows the input-output signal de nition for the CIDI engine. As we previously discussed, the reduction of NOx emissions is achieved with high dilution of the air charge that corresponds to large values of F1 , the burnt gas fraction in the intake manifold. Visible smoke can be avoided by keeping the in-cylinder air-to-fuel ratio, AFR, su ciently lean. Thus, we employ the engine variables, AFR and F1 as performance variables that we want to optimize at each operating point (see Figure 4). Conventional sensors of mass air ow and manifold absolute pressure are used to provide information about the intake gas process. Speci cally, we measure the intake manifold pressure (p1 , using a MAP sensor), and the air ow through the compressor (Wc1 , using a MAF sensor) as depicted in Figure 4. The fuel injection system in CIDI engines enables essentially instantaneous control of the fueling rate. Typically, rapid increases in fueling rate are limited to protect the driveline and to prevent smoke. In this work we are not addressing the design of the fuel limiter and we allow instantaneous increase of fueling rate in order to investigate the ability of

stantaneous increase in fueling rate. Utilizing the energy stored in the resulting exhaust gas the turbocharger will spin up to a new steady-state operating point and eventually increase the amount of air to the engine. Thus, changes in fueling rate cause a disturbance in AFR which is partially rejected through the natural feedback that the turbocharger establishes between the engine exhaust and intake processes. It takes, however, a certain amount of time for the turbocharger to spin up and for the intake manifold to ll to a new level of pressure and air mass, thus, reaching a new equilibrium. To summarize the input-output properties of the plant, changes in fueling rate a ect rapidly the combustion process and the exhaust manifold. Furthermore, changes in fueling rate a ect the intake manifold states primarily through the turbocharger, and secondly, through the open EGR valve. In point of fact, during fueling rate changes the performance variables, AFR and F1 , and the measured variables, Wc1 and p1 , converge to a new equilibrium. This equilibrium and the transient trajectory, however, might not be the optimal with respect to the emission requirements and the time required to achieve convergence might be excessively long.

3 Control Problem Formulation

The control design objective is to coordinate EGR and V GT to (i) speed up the engine open loop dynamics in order to reduce transient AFR excursions, and (ii) to regulate AFR and F1 to a new optimum equilibrium. The resulting control problem is de ned as follows:

x _ y z u w

= f (x u w) = Wc1 p1 T = hy (x) = F1 AFR T = hz (x w) = EGR V GT T = Wf N T

State Equations Measured Outputs Performance Variables Control Inputs External Inputs. (3.1)

In the next section, we calculate for a given w the desired steady state point z (w), the corresponding control signal u (w) = Su (w), and measurements y (w) = Sy (w). We refer to Sy (w) as the setpoint map that determines the desired

values for the measured outputs. The controller is designed to ensure that in steady-state or whenever w is changing slowly z stays as close as possible to the desired z (w). The controller consists of a feedforward part ufw = u = Su (w) and a feedback part C (s). The feedback portion is employed to compensate for uncertainties, see Figure 5. We develop a family of linear controllers for EGR and V GT actuators, parameterized by w0 , that regulate z (t) to z (w) as w varies in a small neighborhood of a nominal operating point dened by w0 . These linear controllers are then scheduled to obtain a controller de ned over the entire operating region. Figure 5 shows the controller structure.
w w Sy y*
Engine

moved in the set-point generation. These manipulations, however, hold only for linear plants and the uncertainties introduced by nonlinear high order terms or other modeling errors will be treated as input disturbances to the closed loop (dI , and dOy respectively). It will be shown that the linear controller C (s) ensures robust feedback properties against input and output disturbances.
w P wz Pz dI w DSy-PyDSu-Pwy

doy

z y

C(s) -

Py

C(s)

Figure 7: Simpli ed linear robust control problem.

Su u*

Figure 5: Nonlinear controller structure.

4 Steady-State Optimization
The desired steady-state values for the burnt gas fraction and air-to-fuel ratio, z , are generated using the following procedure. Given w, we search for the operating point u = u (w) so that F1 is maximized subject to (1.) incylinder air-to-fuel ratio, AFR, being greater than AFR and (2.) intake manifold pressure, p1 , being less than p1 . The rst constraint ensures that no visible smoke is generated in steady-state. The lower bound AFR can be a function of fueling rate and engine speed. For simplicity, however, we assume a constant bound AFR = 25. The second constraint protects the engine from damage due to high intake manifold pressure (over-boost). We assume that p1 = 220 kPa. The value of F1 is maximized to achieve NOx reduction. After selecting the desired set-points for the performance variables, z , and its associated inputs, u , we determine the corresponding measured outputs, y , using simulation of the engine model. In Figures 8 to 10 we illustrate the procedure for generating desired setpoints by assuming speci c external inputs, Wf = 6 kg/hr and N = 2500 rpm. We show how to determine the actuator positions, EGR and V GT , that correspond to maximum F1 for AFR AFR. A grid that spans the feasible set for the controller inputs u is selected. The grid is based on the actuator authority, speci cally, EGR and V GT actuator positions vary between 0 and 1, where 0 corresponds to the completely closed position and 1 corresponds to the completely open position as shown in Figure 8. The signi cance of the region enclosed by the small box in Figure 8 is explained in the following section. The set of all steady-state points for the performance variables, z , that correspond to Wf = 6 kg/hr, N = 2500 rpm is shown in Figure 9. Similarly, the set of all steady-state points for the measured outputs, y, that correspond to Wf = 6 kg/hr, N = 2500 rpm are shown in Figure 10. The circles indicate the equilibria that satisfy the constraints, the \x"s those that violate the constraints. Our procedure for setpoint generation indicates that F1 = 0:2262 and AFR = 25:79 is to be chosen for this engine speed and load values. This

To design one of these linear controllers, C (s), consider the linearization of the system at a nominal operating point, w0 , and its corresponding values of the control input u0 = Su (w0 ) and the output y0 = Sy (w0 ). Let w, u, y, and z denote the deviations of the corresponding variables from their nominal values. Let Py (s) and Pz (s) denote the transfer function matrices from the input u to the measured outputs y and performance outputs z , respectively. Similarly, let Pwy (s) and Pwz (s) denote the transfer function matrices from the input w to the measured outputs y and performance outputs z , respectively. Figure 6 shows the resulting block-diagram of the linearized closed loop system. Note that in Figure 6 we have also included the linearization of the set-point map, DSy , and of the feedforward controller, DSu .
P wz z

Pz w w ufb DSy
-

Pwy y

C(s) ufw

Py

DSu

ufw

Figure 6: Block diagram showing the linearized control

problem.

In Figure 7 we further simplify the block diagram of the linearized control problem by standard block diagram algebra. Assuming small signal, the e ects of the fuel disturbance in the output and the feedforward controller signal can be

setpoint corresponds to the largest value of F1 that is consistent with AFR and p1 satisfying the stated limits.
U* Set 1

Y* Set 200 190 180

y*
170 160 p1

0.8

150 140

0.6 VGT
130 120 110

0.4

0.2

100 0.03

0.035

0.04

0.045

u
0

0.05

0.055

0.06

0.065

c1

0.2

0.4 EGR

0.6

0.8

Figure 10: Feasible set of measured outputs for Wf = 6

kg/hr and N = 2500 rpm.

Figure 8: Selected grid of control signals for Wf = 6 kg/hr

and N = 2500 rpm.

Z* Set 40 38 36 34 32 AFR 30 28 26 24 22 20 0.05

feedback controller will shape the transient response and correct for modeling errors and uncertainties. Let us rst examine the control authority available in steady-state to do so. The DC gain matrix of the linearized system at this setpoint, Pz (0), is given by: F1 0:23 ;0:55 EGR (5.1) AFR sc = ;0:90 1:84 V GT where in calculating Pz (0) the outputs were scaled based on their relative importance, F1 AFR ]T = diag( 1:0 0:1 ]) F1 AFR ]T : sc The (scaled) Pz (0) is \almost singular" or \almost rank de cient" in the sense that it has both a large condition number ( 64) and a small singular value ( min 0:04) 8]. Although scaling should not be used to conceal potential sensitivity problems it is an important consideration to the control designer. In this particular case modifying scaling cannot alleviate the rank de ciency problem. It can be shown that scaling the performance variables cannot reduce the condition number of Pz (0) lower than min 33. Furthermore, the best condition number is obtained with an output scale that is unreasonable from physical prespective. This fact demonstates a structural di culty that is scale-independent. It is also important to note here that although the rows of Pz (0) are almost colinear the rows of the state x to performance variables z matrix, Cz , are independent. Similarly, the columns of the input u to state x matrix, B , are independent. In other words, the rank de ciency problem doesn't occur due to actuator or sensor redundancy (B or Cz matrices) but rather in the overall DC gain matrix. Analysis of the plant singular values and the actuator and sensor authority in various frequencies demonstrates similar behavior with the above DC analysis. The rank de ciency arises because the two performance variables cannot be independently controlled by the actuators i.e., the two rows of Pz (0) = Cz A;1 B are almost colinear. The physical reason for this property is originated in the set-point optimization for aggressive emission reduction. The desired operating set points are derived based on

z*

0.05

0.1

0.15 F1

0.2

0.25

0.3

Figure 9: Feasible set of performance variables for Wf = 6

kg/hr and N = 2500 rpm.

The u setpoint that generated F1 = 0:2262 and AFR = 25:79 is EGR = 0:9, V GT = 0:1. The corresponding values of y are p1 = 176:23 kPa and Wc1 = 0:0364 kg/sec. The optimization has to be repeated for external signals on the selected grid on w and the resulting values of y and u are interpolated to derive the set-point map Sy and the feedforward controller Su .

5 Rank De cient Plant

Suppose we wish to design a linear controller to regulate z in the neighborhood of the operating point corresponding to F1 = 0:2262, AFR = 25:79, N = 2500, Wf = 6. Note here that the feedforward control input, Su will rapidly regulate the system close to the desired operating point. The linear

maximization of the burned gas fraction for xed air-to-fuel ratio. When the burnt gas fraction increases the fresh air charge is displaced from the intake manifold and the incylinder air-to-fuel ratio, AFR, decreases. Let us consider the de nition of in-cylinder air-to-fuel ratio (1.6) where W1e is the engine intake mass ow rate, F1 is the burned gas fraction and Wf is the fueling rate. The maximization of F1 for xed AFR is achieved for VGT and EGR positions that maximize W1e . Thus, at least locally, VGT and EGR cannot a ect W1e . Denote by the attained maximum value of W1e at the desired point. Then (5.2) becomes AFR = ; F1 (5.3) which results in completely coupled outputs. We note that the DC gain matrix Py (0) is relatively well-conditioned, which re ects the fact that EGR and VGT have independent authority to regulate the measured outputs from equivalent deviations from the equilibrium. The fact that the actuators are locally almost redundant is also illustrated in Figures 8 to 10. Consider how the performance variables change as the actuators move in a small region about the nominal setpoint given in Figure 8. The size of this region in u, is chosen so that the nonlinear plant transformations from u to y and z can be approximated with the linear plant transformations Py (0) and Pz (0). Mapping this set of control commands into the performance and measured outputs gives us the regions shown in Figures 9 and 10. Note that the region in Figures 9 is essentially a line in AFR and F1 coordinates. Were the actuators completely redundant, in the sense that they had identical e ect on performance variables, this skinny region would reduce to a line similar to the one shown in Figure 9. On the other hand, the region in the y coordinates indicates independent control of the measurements in steady-state by using EGR and V GT . The consequence of the rank de cient Pz (0) can now be recognized. Small errors in the steady-state scheduling scheme that correspond to z outside the skinny set will result in y set points outside the feasible set of Wc1 and p1 . A controller that enforces y(t) to converge to y using high gain in both EGR and V GT actuators may cause saturation of both actuators or may result in large actuator e orts and sensitivity. In light of this di culty a possible solution is to generate a di erent set-point strategy that results in engine operation away from \optimal" but \di cult" setpoints. However, this solution would degrade emission performance of the engine and may be undesirable due to very stringent emission regulations that are currently being put in place.

AFR = (1 ; F1 ) W1e W
f

(5.2)

numerical interpolations, there is a large probability for infeasible set-points z . Any high gain feedback controller (in both actuators) will attempt to track both feasible and infeasible commands using large control signals and causing frequent actuator saturation. Instead as we demonstrate below, a feedback controller based on a single integrator design (high gain in a selected direction) can achieve good performance with small control signals. We refer to this controller as a \directional" compensator. In the following subsections we design the directional compensator with a single integrator to coordinate the actuators to achieve maximum authority and to reduce the impact of strategy errors by not attempting to track infeasible setpoints. To do that we rst assume that we can measure the performance variables z . We design the controller based on the directional properties of the plant Pz using the singular value decomposition of the plant DC gain matrix. This analysis identi es the e ective range of the plant, which determines the set of setpoints that may be tracked without incurring excessively large control signals, and the e ective rowspace of the plant, which determines how the actuators should be coordinated to achieve maximum authority. We then translate our feedback controller to lter the measured y commands that correspond to infeasible z commands and analyze the performance and robustness properties of the resulting feedback loop. Finally, we gain schedule the feedback controller to span the entire operating range of the engine.

6.1 Feedback on Performance Variables

6 Feedback Controller
A feedback controller that allows robust engine operation at the optimal F1 AFR]T = z is presented below. The optimum region is narrow and close to the boundary of the feasible region of operation. Because the set-point generation may entail large errors due to model uncertainty and

Consider the singular value decomposition (SVD) of Pz (0): T T Pz (0) = 1 u1 v1 + 2 u2 v2 (6.1) where 1 2 are the largest and smallest singular values of Pz (0), and (u1 u2 ) and (v1 v2 ) are the corresponding sets of left and right singular vectors, respectively. Suppose temporarily that Pz (0) were rank de cient, so that 2 = 0. Then it is not possible to track an arbitrary setpoint command the set of commands that can be tracked is given by the range, or columnspace, of Pz (0), R(Pz (0)) = R(u1 ). For our problem, Pz (0) is \almost" rank de cient, in the sense that 0 < 2 1. It follows that certain commands can be tracked only by using large control inputs. To see this, consider a command z , together with the corresponding steady state control signal u. Using the SVD, we have that uss = (1= 1 ) v1 uT z + (1= 2 ) v2 uT z : (6.2) 1 2 Hence if z = u2 for a scalar , then k uk = (1= 2 ) k z k k z k, where we denote by k k the vector 2;norm. It follows that relatively small commands may result in large control signals. On the other hand, commands that satisfy z 2 R(u1 ) may be tracked without using large control signals we say that such commands lie in the \e ective range" of the plant. A controller with an integrator for each input-output pair will thus achieve perfect steady state tracking of all commands at the expense of large control signals, which are undesirable because they may result in actuator saturation. Although saturation isn't necessarily bad, it is di cult to analyze for systems with multiple actuators and strong interactions, and it often has the potential

to degrade performance in an unpredictable way. Similar comments apply to controllers whose DC gain is nite but very large. To avoid potential di culties with large control signals, we propose a controller of the form C (s) = Gz KC (s)Hz , where Gz 2 R2 1 and Hz 2 R1 2 are constant unit vectors and KC (s) is a stabilizing controller with integral action (Figure 11). To simplify the controller implementation and gain-scheduling a simple PI controller for KC has been chosen. We choose Hz and Gz based on the following considerations: (i) Hz enforces tracking of commands that lie in the \effective range" of the plant Pz (0): Hz = uT . 1 (ii) Gz minimizes the control e ort that is needed to track feasible commands: Gz = v1
z*
dI
Hz KC(s) Gz

dOz

Pz(s)

It is particularly interesting to investigate the robustness properties of the resulting 2I2O feedback controller. The robustness and performance properties of the 2I2O feedback system assuming measured performance variables provide a benchmark of how well we can do given appropriate sensors for the rank de cient plant. It is expected that these properties will degrade in the case when the measured variables are di erent from the performance variables. We introduce the notation for sensitivity and complementary sensitivity functions. SOz (s) = (I + Pz (s)C (s));1 TOz (s) = I ; SOz (s) SIz (s) = (I + C (s)Pz (s));1 TIz (s) = I ; SIz (s) (6.4) The closed loop system behavior in tracking, disturbance rejection, noise attenuation and output uncertainties is determined by the output sensitivity and complementary sensitivity functions. For example, if we assume step inputs for all exogenous inputs in Figure 11 (dOz (t) = dOz 1(t) etc.), the steady state error is given by

ess = SOz (0) z ; SOz (0)dOz ; SOz (0)Pz (0)dIz ; SOz (0)n:
n

Figure 11: Linear feedback controller based on measuring

Similarly, the controller response and closed loop stability robustness against input uncertainties is determined by the input sensitivity and complementary sensitivity functions:

the performance variables.

uss = SIz (0)dIz + s!0(C (s)SOz (s))( z ; dOz ; n): lim

The sensitivity and complementary sensitivity functions of the \sd" plant shown in Figure 12, Ssd (s) = 1=(1 + Hz Pz (s)Gz ), and Tsd (s) = Hz Pz (s)Gz =(1 + Hz Pz (s)Gz ), respectively, are directly shaped based on the PI controller gains. Stability of the \sd" plant implies that Ssd (0) = 0 and Tsd (0) = 1. Although Ssd (0) and Tsd (0) are important for the stability and performance of the feedback system they are only indirectly related to the 2I2O stability robustness properties. Recall here that by using the directional compensator Gz KC Hz we cannot track all commands, thus the output sensitivity and complementary sensitivity functions for the 2I2O plant in Figure 11 at DC must satisfy kSOz (0)k 1 and kTOz (0)k 1. It is Timportant, to note that the choices of Gz = v1 and Hz = u1 achieve the minimum value (see Appendix B) and result in:

Recall here that because Pz (0) is almost singular there exist many control signals u that will result in the similar steady-state z . Having Hz and Gz we form the single-input single-output \squared-down" plant Hz Pz (s)Gz , which we denote by \square down" plant, or \sd" (shown in Figure 12). The \sd" plant has non-zero DC gain and we can tune a PI controller to achieve steady-state tracking of the commands that lie in the e ective range of the plant at DC. Apart from steady state behavior, the PI controller can be tuned to achieve fast response taking into account bandwidth contraints imposed in the 2I2O con guration by the available actuators and sensors. The speci cations on the transient characteristics have to re ect constraints in the measured variables and the performance variables simultaneously and it is the subject of current research 21].
esd
KC(s)

usd
HzPz(s)Gz

kSOz (0)k = kTOz (0)k = kSIz (0)k = kTIz (0)k = 1: Thus, the particular choices for Hz and Gz ensure reason-

Figure 12: Single input, single output (\square-down")

plant.

able stability robustness against output and input plant uncertainties. We complete now the analysis of the performance and robustness characteristics for the system when the measured outputs are di erent from the performance outputs which is the desired implementation con guration.

With the current choice of Hz = uT and Gz = v1 the steady 1 state control signal, uss , to a command z = 1u1 + 2 u2 for some scalar 1 , and 2 is uss = Gz Hz z =(Hz Pz (0)Gz ) = 1 v1 = 1 (6.3) which is the smallest control signal that achieves at steady state tracking of the feasible command component. Indeed, the steady state output is zss = Pz (0) u = 1 u1 .

6.2 Feedback on Measured Variables

The directional compensator of the previous subsection needs to be modi ed for implementation based on the measured variables y instead of the performance variables z . The DC gain matrix of the input to measurements Py (0) is well conditioned for the majority of the operating points. The reference inputs are y = Py (0)Pz (0);1 z . Let the T singular controller be C (s) = Gy KC (s)Hy with Hy and Gy 2 1 and KC (s) a PI controller. constant matrices in R

We choose Hy so that the controller attempts to track only those y-commands that correspond to z -commands that lie in the e ective range of Pz (0): Hy = Hz Pz (0)Py;1 (0)=kHz Pz (0)Py;1 (0)k (6.5) T as above. where Hz = u1 The transfer function mapping y-commands to the control signal in Figure 13 is given by C (s) (I + Py (s)C (s));1 , and has DC gain of magnitude kGy Hy k=jHy Py (0)Gy j = kGyTHz Pz (0)Py;1 (0)k=jHz Pz (0)Gy j. Because Hz Pz (0)Gy = 1 v1 Gy , it is easy to verify that choosing Gy = v1 (i.e., Gy = Gz independently of the sensor set) minimizes the size of the steady state control signal. Indeed, a reference command y = 1 Py (0)Pz (0);1 u1 + 2 Py (0)Pz (0);1 u2 (6.6)
|

Modeling errors in the interaction between fuel and the performance variables Pwz can be expressed as an output multiplicative uncertainty and assessed through the e ects of an output disturbance dOz (t) = dOz 1(t) to the Pz (s) plant. Obviously the output disturbance dOz (t) does not a ect the measured variables and cannot be rejected. Finally, to assess the importance of modeling errors in the Pwy term and sensor uncertainties we analyze the e ects of step output disturbance dOy (t) = dOy 1(t) to y. These are determined by yss = SOy (0)dOy , where

will generate the steady state control signal uss = Gy Hy y =(Hy Py (0)Gy ) = v1uT Pz (0)Py (0);1 y =(Hy Py (0)Gy ) (6.7) 1 = 1v1 = 1 : Moreover, the combination of Hy and Gy ensures that the smallest control signal will be used to achieve steady state tracking of the component of y that corresponds to the e ective range of the performance variables. To see that consider the control signal from Eq. 6.7, then the measured output is yss = Py (0) uss = 1 Py (0)v1 = 1 = 1 Py (0)Pz (0);1 u1 which is the feasible component of the reference input. For illustration in Figure 8 we show the line on which the actuator signals lie as speci ed by Gz = v1 . Similarly, the line in Fig. 9 shows the direction Pz (0)Gz and the line in Fig. 10 shows the direction Py (0)Gz . Note also that at DC the single-input single-output system (\sd" plant) that emerges from the Py (0) based feedback is equivalent to the \sd" plant that is based on measuring the performance variables directly, i.e., Hy Py (0)Gy = Hz Pz (0)Gz . The input and output robustness properties of the resulting feedback loop determine the closed loop behavior for (i) actuator uncertainties or errors in the feedforward strategy u (w) = DSu w (de ned in Sec. 3, Fig. 6.) (ii) sensor uncertainties or modeling errors of the fueling disturbance Pwy w in the measured variables and Pwz w in the performance variables. In particular the e ect of EGR and VGT actuator uncertainties can be investigated based on the e ect of an input disturbance dI (t) = dI 1(t) as shown in the Figure 11. The steady state control signal is determined by uss = SIy (0)dI where SIy (0) = I ; Gy Hy Py (0)=(Hy Py (0)Gy ) = I ; Gz Hz Pz (0)=(Hz Pz (0)Gz ) (6.8) = SIz (0) Thus, kSIy (0)k = kSIz (0)k and the input robustness properties of the feedback system are preserved when y is measured instead of z . This is due to the topology of the input disturbance in the feedback loop.

feasible component

{z

(6.9) and, in general, kSOy (0)k kSOz (0)k depending on the anT gle yz = 6 (Hy Py (0)Gy ). In particular, it is shown in 6] that kSOy (0)k = 1=cos( yz ). The physical interpretation is clear: to have good output robustness and steady state tracking of yss , the range of the achievable steady state measured outputs, R(Py (0)Gy ), has to correspond to the T range of the feasible component of the y which is R(Hy ) (see also Eq 6.6). Therefore, the output feedback properties will be preserved if the angle

SOy (0) = I ; Py (0)Gy Hy =(Hy Py (0)Gy ) = I ; Py (0)Gz Hz Pz (0)Py (0);1 =(Hz Pz (0)Gz )

cos

T yz = cos6 (Hy

Py (0)Gy ) = cos6 (Py (0);T v1 Py (0)v1 )

is small which is ensured if the DC gain matrix Py (0) is well conditioned. In the few operating points that an ill-conditioned Py (0) matrix is obtained through the linearization we check explicitly the angle yz to assess the potential degradation. This angle remained small over the entire optimized operating region of the engine. It is intuitively clear that the cases where the angle yz will be large are the cases where the measurements do not correlate well with the performance variables.
dOz
Pz(s) Hy

y*

dI
Hz KC(s) Gz

dOy

Py(s)

Py(0)

-1

Pz(0)

Figure 13: Linear feedback controller.

We note here that di erential sensitivity calculations can be used to analyze the sensitivity of the closed loop due to errors or uncertainty in the e ective range of the plant Pz (0). Having small SIy and SOy is a necessary condition for reducing the sensitivity of the closed loop performance and stability robustness against these uncertainties, but further work is needed to determine the exact relations. This is going to be the subject of future work.

Measurements 0.06 1

Control Signals 0.4

Performance
160 140 120 Wc1

Measurements 120 100

Control Signals 50

Performance

0.05 EGR
c1

0.8

0.3

40 80 EGR 60 40 20 20 0 10 20 30 0 0 10 20 30 10 0 10 20 30 AFR F1 40 20 0 0.2 0.1 0 0 10 20 Seconds 30 0 30

0.04

0.6

F1

0.2

100

0.03

0.4

0.1

80 60

0.02

10

20

30

0.2

10

20

30

10

20

30

220 210 200 VG p 190 180 170 160 150 0 10 20 Seconds 30

1

0.4 0.3 0.2 0.1 0 0.1

35

200 180 160

100 80 60

0.5 0.4 0.3

30 AFR

25

140 120 100

20

10

20 Seconds

30

15

10

20 Seconds

30

10 20 Seconds

30

VGT

p1

10 20 Seconds

30

Figure 14: Simulation response of the two integrator

(dashed) and the single integrator (solid) controllers. The setpoints are indicated by the dotted line.

Figure 15: Experimental controller response to fuel steps.

7 Simulation and Experimental Results

The di erences between a two integrator controller and the single integrator controller, C (s) = Gz Kc(s)Hy , can be demonstrated by simulating the linearized model. The controller with two integrators was designed so that asymptotic tracking of the set-points by the measured outputs is enforced, and by applying the LQG control design methodology to the augmented plant. The PI controller for the SISO plant Hy Py Gz was tuned to achieve satisfactory response without violating bandwidth constraints imposed by the cut-o frequency of the actuator. The responses of the two controllers to fuel steps between the levels 6 5:5 5:0 7:0 7:5 8:0 8:5 7:0 6:0 kg/hr at 2500 rpm are shown in Figure 14. We observe that the single integrator controller requires less actuator position deviations from their nominal positions, and while strict tracking of the measured outputs (Wc1 and p1 ) is not enforced, the deterioration in the values of performance outputs (F1 and AFR) is very small. For example, at t = 10 seconds the two integrator controller closes V GT beyond its physical limits (limits were not enforced in this linear simulation) while the single integrator controller closes V GT only to about 0:02. This results in an error of about 15 kPa in p1 tracking which translates only to very small performance output errors. Smaller actuator e ort inherent to the operation of the single integrator controller makes actuator saturation and loss of linear performance associated with actuator saturation and antiwindup less frequent. The experimental test of the controller was done in a Visteon's engine dynomometer in Dunton, UK using dSPACE rapid prototyping environment. The matrices Gz and Hy were determined for each operating point in the same grid of engine speed and fueling rate values as was used to generate the set-points. The gain-scheduling of the controller was accomplished by linearly interpolating the values of the entries of the matrices Gz and Hy for all engine speed and

fueling rate values. The desired steady-state positions for EGR and VGT were used as a feedforward and the controller Kc was designed as a PI controller where its two gains (proportional and integral) were tuned on the engine. We found that with xed PI controller gains across engine operating range we are able to obtain good transient responses. Consequently, only the matrices Gz and Hy were gain-scheduled based on the plant DC gains with a clearly de ned procedure that we outlined above. To handle possible actuator saturation we used a conventional antiwindup compensation that would limit the contribution of the integral term of Kc to a certain value and would stop updating the integrators when saturation is encountered. The setpoint map generates the desired measured outputs shown with dotted line in the rst column of plots in Figure 15. The controller generated the desired positions for EGR and V GT actuators that would be passed as set-points to the inner loop controllers for each of the actuators that regulate them to the speci ed settings. A load acceptance type test was used to evaluate the controller performance where the fueling rate was varied in stepwise fashion between levels 2:0 5:0 6:0 5:0 2:0 kg/hr. The measured outputs and the control signals responses to these fuel steps are shown in Figure 15. The engine speed was regulated by the dynocontroller to a constant value of N = 2000 rpm. The performance outputs AFR and F1 were not measured directly but estimated by post-processing experimental data. Because the experimental setup did not have an operational fueling rate limiter at the time of the testing during the transients the air-to-fuel ratio is seen to drop below the threshold selected to avoid visible smoke (shown by the dashed line). Since the fueling rate limiter was not active the quality of achievable load acceptance performance can be judged indirectly, by the speed of the air-to-fuel ratio recovery in response to a fuel step. Fast recovery seen in Figure 15 correlates with good transient performance and is indicative of small turbo-lag.

8 Concluding Remarks
In this paper we treated an emission reduction control problem for an advanced technology diesel engine. We demonstrated that at the optimal operating points the performance variables that most directly a ect the emissions become closely dependent. Speci cally, the performance variables cannot be controlled independently using EGR and V GT actuators. This plant singularity at the optimal operating points does not allow the application of integral control design (or high gain in both input-output pairs) and leads to a di cult tracking problem. We designed nonlinear feedforward and a gain scheduled multivariable controller that enables engine operation at these optimal operating points by coordinating EGR and V GT actuators at DC to achieve maximum authority. Our controller design directly addresses the issues arrising from errors in the set-point generation and the plant singularity by using the singular value decomposition of the plant DC gain matrix. From development perspective, it is important to deal directly with the problem arising from rank de cient plants. The problem is bound to appear more and more in automotive applications where innovative actuators are often introduced to allow optimization of certain performance variables. Frequently, as a result of the optimization the actuators end up operating very close to the feasible boundary which in turn causes a rank de cient problem.

9 Acknowledgment
We acknowledge Michiel van Niewstadt, Paul Moraal (Ford Forschugszentrum Aachen, Germany), Paul Wood, Mike Criddle and Mick Campbell (Visteon, Dunton, UK) for facilitating the experimental validation of the controller deisgn. We also thank Je Cook, Mrdjan Jankovic (Ford Motor Co., USA), and Rick Middleton (Univ. of Newcastle, Australia) for useful discussions.

1] Amstutz, A., and Del Re, L.R., \EGO sensor based robust output control of EGR in diesel engines," IEEE Transactions on Control System Technology, vol. 3, no. 1, 1995. 2] Baert, R.S.G., Beckman, D.E., Verbeek, R., \New EGR technology retains HD diesel economy with 21st Century Emissions," SAE paper 960848. 3] Buratti, A. Carlo, E. Lanfranco, and A. Pisoni, \DI Diesel engine with variable goemetry turbocharger: a model-based boost pressure control strategy," Proc. 1st ICC on Control and Diagnostics in Automotive Applications, pp. 119-129, Genova, 1996. 4] Dekker, H.J., and Sturm, W.L., \Simulation and control of a HD diesel engine equipped with new EGR technology," SAE paper 960871. 5] J. S. Freudenberg and R. Middleton, \Design Rules for Multivariable Systems," Proc. Conf. on Decision and Control, Japan, 1997. 6] J. S. Freudenberg and R. Middleton, \Properties of

References

Single Input, Two Output Feedback Systems," Proc. 1998 Amer. Contr. Conf., Philadelphia, June 1998. 7] J. S. Freudenberg and R. Middleton, \Feedback Systems with an Almost Rank Defecient Plant," submitted Proc. 1999 American Control Conference. 8] Golub, G.H., and van Loan, C.F., Matrix Computations, The Johns Hopkins University Press. 9] Guzzella, L., and Amstutz, A., \Control of diesel engines," IEEE Control Systems Magazine, vol.18, No. 2, pp.53-71, 1998. 10] Jankovic, M. and M., and Kolmanovsky, I., \Robust nonlinear controller for turbocharged diesel engines," Proceedings of 1998 American Control Conference, pp. 13891394, 1998. 11] Heywood, J.B., Internal Combustion Engine Fundamentals, McGraw-Hill, Inc., 1988. 12] Kao, M., and Moskwa, J.J., \Turbocharged diesel engine modeling for nonlinear engine control and estimation," ASME Journal of Dynamic Systems, Measurement and Control, Vol. 117, pp. ?, 1995. 13] Kolmanovsky, I., Moraal, P., van Nieuwstadt, M., and Stefanopoulou, A., \Issues in modelling and control of intake ow in variable geometry turbocharged engines," Proceedings of 18th IFIP Conference on System Modelling and Optimization, Detroit, July 1997, to appear in System Modelling and Optimization, Addison-Wesley Longman. 14] Kolmanovsky, I., Moraal, P., van Nieuwstadt, M.J., Wood, P., and Criddle, M., \Modelling and identi cation of a 2.0 liter turbocharged diesel engine - Part II: VNT, steady-state and dynamic," Ford Motor Company Technical Report SRR-1998-0087. 15] Moody, J.F., \Variable geometry turbocharging with electronic control," SAE paper No. 860107, 1986. 16] M. Morari and E. Za riou, Robust Process Control, Prentice Hall, 1990. 17] van Nieuwstadt M., Moraal, P., Kolmanovsky, I., and Stefanopoulou, A., \A comparison of SISO and MIMO designs for EGR-VNT control of a light duty Diesel engine," Proceedings of IFAC Workshop on Advanced in Automotive Control, Mohican State Park, Ohio, February 1998, pp. 191196. 18] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control, John Wiley & Sons, 1996 19] A.G. Stefanopoulou, I. Kolmanovsky, and J.S. Freudenberg, \Control of Variable Geometry Turbocharged Diesel Engine for Reduced Emissions," Proc. 1998 Amer. Contr. Conf., pp. 1383-1388, Philadelphia, June 1998. 20] Watson, N., and Janota, M.S., Turbocharging the Internal Combustion Engine, Wiley Interscience, New York, 1982. 21] A. Woodyatt, J. S. Freudenberg, R. H. Middleton, \A design tradeo for single input two output feedback systems," Proc. 1997 Allerton Conference , Monticello, IL, 1997.

A Engine Model
The engine model used in this study is a mean-value controloriented model of a high speed turbocharged diesel engine

with the variable geometry turbocharger and the exhaust gas recirculation. The model is based on the standard representations for the gas lling dynamics in the intake manifold and the exhaust manifold, turbocharger dynamics, actuator dynamics and sensor dynamics. The speci c subsystem representations and modelling assumptions are well-known see e.g. 1, 12, 13]. Here we follow 13] that provides careful treatment of the burnt gas fraction dynamics and VGT model, not available from other sources. We start by de ning the actuator dynamics. Stable inner loop controllers for the pneumatic EGR valve and VGT vane actuators are assumed and their closed loop behavior is incorporated in the engine model. The closed loop dynamics are represented by a rst order system which introduces two additional states: EGR valve position ( egr ), and VGT actuator position ( vgt ).
d dt egr = d dt vgt =

Similarly, three states are introduced to represent the gas lling dynamics in the exhaust manifold: density ( 2 , kg/m3 ), pressure (p2 , kPa) and burned gas fraction (F2). The evolution of the gas properties in the exhaust manifold is represented by the following equations:

d 1 dt 2 = V2 (We2 ; W2t ; W21 ) d We2 (Fe2 ;F2 ) (1.3) 2 V2 dt F2 = d p2 = R (We2 Te2 ; W2t T2 ; W21 T2 ) V2 dt where We2 = W1e + Wf is the mass ow rate out of the
engine cylinders into the exhaust manifold (mean-value averaged over a cycle), Fe2 is the burned gas fraction in this ow and Te2 is its temperature, Wf is the fueling rate, W2t is the mass ow rate through the turbine and V2 is the volume of the exhaust manifold.

The actuation signals, EGR and V GT , are generated by the control system, and the time constants, e and v , are approximately equal to 0.07 sec and 0.60 sec respectively. Their scaled range is between 0 and 1, where the position of 0 is completely open and 1 is completely closed. The gas properties in the intake manifold are represented by three states: density ( 1 , kg/m3 ), pressure (p1 , kPa) and burned gas fraction (F1) de ned as the ratio of density of burned gas to the total density of air and burned gas in the intake manifold. The term \burned gas" refers to the products of combustion (other than air) that are in the intake manifold due to EGR. Note that since diesel engines operate lean the ow through the EGR valve can contain as much as fty percent of air. Hence, the use of the burned gas fraction is indispensible to keep track of the EGR. The equations for the gas lling dynamics are derived on the basis of the total and burned gas mass balances, ideal gas law and energy balance under the adiabatic assumption. For the intake manifold:

1 (EGR ; egr ) e 1 (V GT ; vgt) v

(1.1)

The primary source of nonlinearities in the model are the dependencies of the mass ow rates on the model states and model inputs. Speci cally, the total engine intake mass ow rate is given by

W1e =

where Vd is the engine displacement volume and the volumetric e ciency ( vol ) is determined from regressing engine experimental data as (1.5) vol = vol (p1 p2 Ne): The air-to-fuel ratio is then calculated by

Ne vol 1 Vd 120

(1.4)

AFR = (1 ; F1 ) W1e : W
f

(1.6)

ow mass ow rate of gas from the intake manifold into the engine (mean value averaged over an cycle), T1 is the temperature of the gas in the intake manifold and T2 is temperature of the gas in the exhaust manifold. The ideal gas law is used to relate the temperature values to the pressure and density values: Ti = pi =(R i ) i = 1 2: The thermodynamic constants used throughout are the di erence of speci c heats at constant pressure and constant volume (R, kJ/(kg K)) and the ratio of these speci c heats ( ). The dependence of these variables on the burnt gas fraction has been neglected.

d 1 dt 1 = V1 (Wc1 + W21 ; W1e ) d Wegr (F2 ;F1 );Wc F1 (1.2) 1 V1 dt F1 = d p1 = R (Wc1 Tc1 + W21 T2 ; W1e T1 ) : V1 dt Here V1 is the intake manifold volume, Wc1 is the mass ow rate through the compressor and Tc1 is the temperature of this ow, W21 is the EGR mass ow rate, W1e is the total

The mass ow rate of EGR (W21 ) is calculated based on an approximation of the ow through an ori ce, see 11]: W21 = Aegr ( egr )d(p1 p2 T1 T2 ) (1.7) where egr is the normalized EGR valve position, Aegr is the e ective ow area function for the EGR valve that is determined by regressing the steady-state experimental data and d is the standard ori ce ow function: ( pp2 ( p1 ) if p1 p2 RT d(p1 p2 T1 T2 ) = pp1 2 ( p2 ) if p1 > p2 , p2 RT1 p1 where (x) =
8 < :
1 2

2 (x ;1
;1

2 +1

2( 2

+1 ;1) +1

;x

if x rc ) if x > rc

2 where rc = +1 is the critical pressure ratio. A similar equation is used to represent the mass ow rate through the variable geometry turbine (W2t ), where the turbine mass ow rate is a function of the normalized vane positions vgt , exhaust and intake manifold pressures and exhaust manifold temperature. The dependence of the turbine mass ow rate on the turbocharger speed Ntc was

determined to be weak and neglected. The engine temperature rise map (Te2 ) is calculated from a regression of experimental data and has the form Te2 = Te2 (Wf W1e Ne F1 ): (1.8) The turbocharger rotational dynamics are represented with one state, the turbocharger rotational speed (Ntc rpm). The state equation is derived from the power balance on the turbocharger shaft: where Itc is the moment of inertia of the rotating parts, Pt is the power generated by the turbine and Pc is the power consumed by the compressor. The turbine power and the compressor power, Pt and Pc respectively, are calculated from regressions obtained from steady-state data provided by the turbocharger manufacturer. In particular, a \turbine map", ft , is used to determine the pressure ratio between po the upstream and the downstream turbine pressures, p2 , and the e ciency, t , based on the turbine ow, W2t , and the turbocharger speed, Ntc:

together with our choices of Gz = v1 and Hz = uT , we get 1 from Eq 2.1:

SOz (0) = I ; Pz (0)Gz (Hz Pz (0)Gz );1 Hz = I ; u1 uT 1 which results in kSOz (0)k = 1.
A comprehensive treatment of the feedback properties of a 2I2O rank de cient plant using a directional controller can be found in 7]. Using the notions of the alignment angle of the plant Pz with the two controller directions R(Gz KC Hz ) = R(Gz ) and N(Gz KC Hz ) = N(Hz ) it can be shown that the particular choices for Gz and Hz can minimize the input sensitivity and complementary sensitivity functions of the TITO plant with respect to the sensitivity and complementary sensitivity functions of the \sd" plant. This results follows directly from 5, 7] where it is shown that kSOz (0)k = kTOz (0)k = cos6 (Pz1Gz HzT ) and kSIz (0)k = kTIz (0)k = cos6 (Gz 1(Hz Pz )T ) (2.3) For the particular choices of Gz = v1 and Hz = uT , it can be 1 T shown that cos6 (Pz Gz Hz ) = 1 and cos6 (Hz Pz GT ) = 1. z This follows directly from
T T T j T( 1 2v ) = kuu1kk( 1 uuv1 T++ 2 uu2 2 Tv1 Tj k = T 1 1 v1 2 v2 )v1 1 j T z Gz j cos6 (Pz Gz Hz ) = kHHz PPz Gz k z kk
1 1

d (ItcNtc ) dt Ntc = Pt ; Pc

(1.9)

W2t = ft po Ntc p2 t
;1 Pt = W2t cp t m T2 1 ; ( po )( ) p2

(1.10) (1.11)

= 1:

(2.4)

similarly, using data from the \compressor map" we can obtain the compressor characteristics.

Wc1 = fc p1 Ntc po c Pc = Wc1 cp 1 To ( p1 )( po c

;1 )

(1.12) (1.13)

Similar derivation holds for cos6 (Hz Pz GT ) = 1. Eq. 2.3 z provides a clear relation between the sensitivity and complementary sensitivity of the \sd" plant that can be directly manipulated with the PI controller and the TITO feedback properties.

;1 :

The \compressor map" regressions also provide the temperature of the compressor ow, Tc1 . The model that was used for simulations and analysis was developed as an S-function in Simulink and validated for a high speed 2 liter diesel engine 14].

B Calculations
One can calculate SOz directly from Figure 11. SOz (s) = = I ; Pz (s)Gz Hz KC (I + Pz (s)Gz Hz KC );1 = I ; Pz (s)Gz (1 + Hz KC Pz (s)Gz );1 Hz KC H C ( )G = I ; Pz (s)Gz 1 + z KKPzPs(s)z (Hz KC Pz (s)Gz );1 Hz KC H G
|

(2.1) Using the DC value of the complementary sensitivity of the \sd" plant TOsd (0) = s!0 1 + z KC PzPs)s) Gz = 1 lim HH K ( ( G (2.2) z C z z

z C z {z TOsd (s)