13.5 Frequency Response & Bode Plots described a loop by how it treats a sinusoid, for one input, one output and a controller whose shape was fixed in advance. It never said how to choose the dynamics, and the yaw damper of 12.6 Lateral-Directional Modes fed back exactly one state of an aircraft that has many. Here we feed back all of them, , and ask what that buys: the eigenvalues of 7.2 Linear Systems & the Eigenvalue Method become design knobs. The price, as always, is paid in control effort, and the lab prints it.
Transcript
- Every linear system in the course fits one template: x dot equals A x plus B u, and y equals C x. The state is the smallest set of numbers that fixes the future. The cruise plant is two states, speed y and engine a, with eigenvalues minus 0.1 and minus 1. The double integrator, a position and a velocity driven by an acceleration, has both poles at the origin. The short period is the w q block of its matrix, eigenvalues minus 2.6649 plus or minus 3.6921 i. The lesson prints no elevator column for it, so the lab uses a stand-in, B equal 0 and minus 4: illustrative, not the aircraft's own.
- Feed back every state: u equals minus K x plus N r. Substitute, and the closed loop is x dot equals A minus B K times x plus B N r. Its eigenvalues are the poles, so placing poles means choosing K. For the cruise plant the characteristic polynomial is s squared plus 1.1 plus k 2, times s, plus 0.1 times 1 plus k 1 plus k 2. Match it to a damped pair at zeta 0.7: omega n equal 1 gives K equal 8.7 and 0.3, poles minus 0.7 plus or minus 0.7141 i. Omega n 0.5 gives k 2 equal minus 0.4: slower than the plant, so placement removes damping if you ask.
- Can every pole be placed? Only if the controllability matrix, B, A B and so on, has rank n. The cruise plant gives determinant minus 0.1 and the double integrator minus 1: both controllable. If a left eigenvector w satisfies w B equal 0, then w times A minus B K equals lambda w for every K: that pole is stuck. The lab's diagonal minus 1, minus 2 with B equal 1, 0 has rank 1 of 2, and the pole at minus 2 cannot move. Otherwise Ackermann's formula, from memory, gives K in one line: on the cruise plant the last row of the inverse is 10, 0, and K is 8.7, 0.3 again.
- Placement is exact, and the design can still be bad. After a unit step in r the first command is N r, which for the cruise plant is 10 omega n squared. The peak commands at zeta 0.7 are 3.0642 at omega n 0.5, then 10, 40 and 160 at 1, 2 and 4, while the settling time falls from 5.985 seconds at 1 to 3.000 and 1.500. Doubling the speed quadruples the command and only halves the settling time. The double integrator is the same: 4 at 2, 16 at 4. And the sensor noise gain on y rises from 8.7 to 154.5, a factor of 17.8.
- Now tracking. The feedforward gain N is minus 1 over C times A minus B K inverse times B, and for the cruise plant it is 10 omega n squared. At omega n 1 the step settles at 1 with 4.60 percent overshoot in 5.985 seconds. But N comes from the model. Give the car 20 percent less authority and the output ends at 0.97561. At omega n 0.5 the miss is 0.90909, and at omega n 4 only 0.99844: more loop gain, smaller miss, but never zero.
- Integral augmentation adds a state z with z dot equal r minus y and places one more pole. On the cruise plant with a third pole at minus 3: K equal 47.7 and 3.3, k I equal minus 30. Overshoot 4.27 percent, settling 6.33 seconds. With the same 20 percent error the output still ends at exactly 1.0000, because the command ends at 1.25 to make up the lost authority. And this is PID: proportional 51, derivative 33, integral 30, with characteristic polynomial 10 s cubed plus 44 s squared plus 52 s plus 30: the three poles we asked for. That is 13.3 again.
- Saturation changes the loop. At omega n 2 the feedforward design asks for 40, a limit of 10 gives it 10, and it still settles in 3.54 seconds with 3.32 percent overshoot: there is nothing to wind up. The integral design, K equal 118.3 and 4.7, k I minus 120, has 3.12 percent overshoot and settles in 3.285 seconds unlimited. With the limit of 10 it is 16.75 percent and 3.975 seconds: windup, with no anti-windup in the lab. And feedback does not move zeros. The short period stand-in overshoots 58.6 percent at omega n 6, and 128.0 at omega n 10.
- A Module 12 example: the lateral model of 12.6, four states, rudder input, controllability rank 4. The yaw damper moved the Dutch roll to minus 1.6647 plus or minus 1.5465 i but dragged the spiral to minus 0.1731. Ask instead for a Dutch roll at omega n 2.5, zeta 0.7 and keep roll and spiral: Ackermann gives a yaw-rate gain of 0.49943, close to 12.6's 0.5, and the eigenvalues come out minus 9.3790, minus 1.7500 plus or minus 1.7854 i, and minus 0.0068. But the spiral stays slow: only 0.0950 of the reference in 15 seconds. Move roll and spiral too and the rudder command is 12.15 radians per unit of yaw rate: about 35 degrees at a reference of 0.05, with 572 percent overshoot.
- Pole placement answers where the poles should be, and leaves the harder question open. Omega n equal 1 was a guess, and the obvious lateral placement asked for 12.15 radians of rudder. Lesson 13.7, on LQR and state estimation, is not yet written. It is meant to choose K by optimisation, paying for state error and control effort, and to add an observer, a model run alongside the plant that estimates the states you cannot measure.
State-space form
Every linear time-invariant system the course has met fits one template:
The state is the smallest set of numbers that, together with the future input, fixes the future. is the matrix of 7.1 Systems & State Space, says where the input pushes, and says what the sensor sees. The transfer function is , and its poles are eigenvalues of .
The cruise plant, two states. The car of 13.1 Feedback & Block Diagrams with the engine lag of 13.2 has and . Take :
The eigenvalues are and , the two poles of . A PID controller of 13.3 sees only (and its derivative). State feedback also sees the engine state .
Two more plants appear in the lab. The double integrator (, , a position and a velocity driven by an acceleration) is the cleanest test plant: both open-loop poles sit at the origin. The short period of 12.5 Trim, Linearization & Longitudinal Modes is the block of its matrix, rows and columns two and three:
with eigenvalues (the lab’s readout; the matrix has because the speed and pitch states couple in slightly). That lesson prints no elevator column for this block, so the lab uses a stand-in, per radian, of plausible sign and size: the elevator numbers below are illustrative, not the 12.4 aircraft’s.
Full-state feedback moves the eigenvalues
Choose , where is a row of gains and the reference. Substitute:
The input has disappeared and a new matrix has taken the place of . Its eigenvalues are the closed-loop poles, so placing the poles means choosing so that has the eigenvalues we want.
Worked example 1: the cruise plant by matching coefficients. With ,
We want the polynomial of a damped pair, . Matching, and . For , : and . The lab reports with closed-loop eigenvalues .
Two gains, two poles anywhere; proportional control on alone (13.2) had one gain and a locus. With the formula gives : a negative gain, because the target poles are slower than the plant’s own. Placement will happily remove damping if you ask.
Controllability: when placement is possible
For one input, the controllability matrix is . The input pushes the state along ; the dynamics rotate that push to , then , and so on. If these directions span the whole state space, the input can reach every state. The rule (a theorem, stated here with the necessity argument only): every closed-loop pole can be placed anywhere (conjugate pairs together) if and only if has rank .
For the cruise plant has determinant : controllable. For the double integrator has determinant .
Why failure means a stuck pole. Suppose a row vector satisfies and : a left eigenvector the input is blind to. Then for any ,
so is still an eigenvalue of . The lab’s last plant shows it: and . The controllability matrix has rank of , the lab refuses to place poles, and the pole at cannot move.
Pole placement by Ackermann’s formula
Matching coefficients is fine for and gets painful beyond that. Ackermann’s formula (from memory, not derived here) does it for one input in one line. If is the desired characteristic polynomial,
exists exactly when the system is controllable. Here is the desired polynomial evaluated at the plant matrix.
Worked example 2: Ackermann on the cruise plant. , so its last row is . For , : , and
The same answer, as it must be. The double integrator is faster: matching with gives , which the lab confirms: at and at ().
. The formula is the same with matrices. For a chain of three integrators () the closed loop is , so the gains are the coefficients of in reverse: poles at give . The exercise checks this case.
Tracking a reference: feedforward and integral action
Regulating to zero is not the job; making follow is. Two ways.
Feedforward gain . At steady state and we want , so
For the cruise plant the closed-loop transfer function is , since the plant’s numerator is untouched by , so . At , and settles at with overshoot ( gives ) and settling in s. The weakness: comes from the model. Give the car less authority (the lab’s model-error slider, real input gain with ) and the feedforward design ends at instead of . At the miss is and at it is : more loop gain, smaller miss, the sensitivity story of 13.1 Feedback & Block Diagrams again.
Integral augmentation. Add a state with and feed it back: . The augmented plant has states,
and we place poles with the same Ackermann formula. At steady state forces whatever the plant gain is. Cruise plant, , and a third pole at (the lab’s third-pole slider): , (negative because integrates the error). The step has overshoot, settles in s and peaks at a command of . With the same model error the output still ends at exactly , the command ends at because it must make up the lost authority, the overshoot rises to and the settling to s.
This is PID. On this plant , so : a proportional term , a derivative term and the integral term . Substituting , , into 13.3’s characteristic polynomial gives , which is : the three poles we asked for. (Here the reference enters only through the integrator.) So PID is state feedback for this plant, three gains placing three poles. It stops being one when the plant has more states than a PID has gains. The short period has two states, the lateral model four, and a full-state design uses all of them.
What pole placement does not tell you
The method is exact and the design can still be bad. Each warning has a lab number.
Effort grows like . After a step in the first command is . For the cruise plant , and the lab’s peak commands at are (; the peak is because the negative lets the command overshoot), , and at , and . Doubling the speed quadruples the command, and settling time falls only from s to s and s. For the double integrator it is the same: at and at . A throttle saturates long before.
Saturation changes the loop. With the command limited to at , the feedforward design requests , receives , and still settles ( s, overshoot): there is no integrator to wind up. The integral design at (third pole at , , ) has overshoot and s settling unlimited, and with the same limit of the lab shows and s: windup, the lesson of 13.3, with no anti-windup in the lab.
Gain on noise. A sensor error in reaches the actuator multiplied by . The cruise gain on is at and at , a factor of . Faster poles are noisier poles.
Unmodelled dynamics. The cruise model has no sensor lag, rate limit or delay; poles far to the left make the loop fast compared with exactly those neglected effects, and the delay margin of 13.5 Frequency Response & Bode Plots shrinks. Placement guarantees the poles of the model only.
Zeros do not move. Feedback changes the poles and leaves the zeros alone. The short-period output has a zero at whatever is. Place the pair at , and the step overshoots by ( at ), far from the of a plain second-order pair. The peak elevator also scales as ( and rad per unit of reference: a ratio of , the stand-in making the units illustrative). Poles set the modes; zeros decide how much of each a reference excites.
A Module 12 example: placing the Dutch roll
The lateral model of 12.6 is a system in with the rudder column per radian. The lab uses those printed five-digit entries, so its eigenvalues are 12.6’s: roll , Dutch roll , spiral . The controllability matrix has rank : the rudder can move all four poles in principle.
The yaw damper of 12.6, with s, moved the Dutch roll to but dragged the spiral to . Full-state feedback can move one mode and leave the others. Ask for a Dutch roll at , , i.e. , and keep roll and spiral where they are (the lab’s checkbox). Ackermann on the gives
in rad of rudder per , and the closed-loop eigenvalues come out , and . The yaw-rate gain is very close to 12.6’s s, and the other gains are small. The unit-reference command is only rad. But the output (yaw rate) is slow: with the spiral kept at , the step reaches only of the reference in s. A mode you leave alone stays as slow as you left it.
Move the roll and the spiral too, to and with the same Dutch roll, and the lab’s gains become with a rudder command of rad per unit of yaw-rate reference. At a reference of rad/s that is rad, about of rudder, and the response overshoots by . Placement did what it was told; the aircraft cannot do it. These two experiments are why the next lesson exists.
Placement against classical design
A root locus moves poles along curves with one gain; placement picks the destination and finds the gains. PID has three gains in a fixed structure; state feedback has one gain per state but needs the states. The margins of 13.5 remain the test of the finished loop.
Concept checks
Playground and exercise
The first cell applies Ackermann’s formula at four speeds, the second simulates the step and measures the peak command (with SciPy’s place_poles as a cross-check), and the exercise has you write Ackermann’s formula yourself.
Problems
For with , find that puts the poles at and , and say what and that is.
Show worked solution
The closed loop is , to match . So . The product of the roots is , , and gives : overdamped, with no overshoot.
For the cruise plant, find and that put the closed-loop poles at and .
Show worked solution
From worked example 1 the closed-loop polynomial is . Match . Then gives , and gives . The closed-loop transfer function is with DC gain , so . Check: has trace and determinant .
For decide whether and give a controllable pair. For the one that fails, find the pole that cannot move.
Show worked solution
: is parallel to , so has rank . A left eigenvector with must be , and : the pole at cannot move (the second state is untouched by the input and obeys ). : and , controllable. The input enters the second state, which feeds the first through the coupling.
What comes next
Pole placement answers “where should the poles be?” and leaves the harder question open. For the cruise plant, was a guess; for the lateral aircraft the “obvious” placement asked for rad of rudder. The next lesson, 13.7 on LQR and state estimation (13.7 LQR & State Estimation), chooses by optimisation instead: pay a cost for state error and a cost for control effort, and let the compromise decide where the poles land. It also answers where the states come from: an observer is a model run alongside the plant that estimates the states you cannot measure, so the full-state feedback of this lesson becomes usable.
Key Takeaways
- State-space form , ; the cruise plant is two states with eigenvalues and .
- With the closed loop is , so the poles are its eigenvalues. On the cruise plant, places (, ).
- Placement needs to have rank ; a left eigenvector blind to the input fixes its eigenvalue for every .
- Ackermann’s formula (from memory) places poles for one input; for matching coefficients is just as good.
- tracks only if the model is exact ( with a gain error); integral augmentation tracks regardless (), and on the cruise plant it is a PID with , , .
- Effort and noise gain grow fast: the cruise step command is (, , at ), and zeros do not move (short period: overshoot at ).
- On the lateral aircraft, moving only the Dutch roll to needs , close to 12.6’s yaw damper; dragging every pole costs rad of rudder per unit of yaw rate.