13.3 PID Control ended on a warning. At the cruise loop with an engine lag was fine at and unstable at , and the edge between them was , found by putting into one polynomial. A simulator or an autopilot needs the general answers: is this loop stable, and if I turn one knob, where do all the poles go? Routh-Hurwitz answers the first without a root finder, and the root locus draws every pole at once. A delay, at the end, will show where both fall short.
Transcript
- A closed loop is stable when every pole sits in the open left half plane: the standard statement, quoted from memory. Marginal is not safe. The poles that matter are the roots of 1 plus L of s, the characteristic equation D plus K times N. Take the cruise PI loop with K p equal 4. At K i equal 0.5 all three poles are in the green half. At K i equal 6 a pair has crossed into the red, at plus 0.0141 plus or minus 0.7291 i.
- The Routh array tests the polynomial itself. Write the coefficients alternately in two rows and build each later row from the two above. The rule, from memory and not proved here: the number of right-half-plane roots equals the number of sign changes in the first column. At K i equal 0.5 the first column is 10, 11, 4.5455, 0.5: no sign change. At K i equal 6 it is 10, 11, minus 0.4545, 6: two sign changes, so two unstable roots, counted without computing them.
- Now recover the edge that 13.3 found by substituting s equal i omega. The first column has no sign change exactly when K i is below 1.1 times 1 plus K p: 5.5 at K p equal 4. At the edge the s to the 1 row is all zeros, and the row above gives 11 s squared plus 5.5, so the oscillation is at plus or minus 0.7071 i, with the third root at minus 1.1. With derivative gain the 11 becomes 11 plus K d: at K p equal 10 the edge is 12.1, and 18.7 with K d equal 6.
- The root locus plots every closed-loop pole as K runs from 0 to infinity. A point is on it when K equals minus D over N is a positive real number, which gives the angle condition: the angle of N minus the angle of D is 180 degrees. Three rules are from memory: n branches, symmetric about the real axis; they start at the poles and end at zeros or infinity; a real point is on the locus if the count of real poles and zeros to its right is odd. Three are derived: asymptotes through the centroid sigma, breakaway points where D N prime minus D prime N is zero, and crossings at the Routh edge.
- Now two cruise loci. Left: proportional gain on the engine lag, poles at minus 0.1 and minus 1. The branches break away at s equal minus 0.55, at K equal 2.025, matching 13.2. The pole sum is fixed, so the pair then climbs the vertical line through minus 0.55. K equal 4 gives damping 0.7778, K equal 19 gives 0.3889, and no gain is unstable. Right: the integral gain at K p equal 4. The centroid is minus 0.3667 and the discriminant is negative, so no breakaway. The complex branches run out at 60 degrees and cross at K i equal 5.5, omega 0.7071.
- A locus by hand: K over s times s plus 1 times s plus 2. The centroid is minus 1. Breakaway: 3 s squared plus 6 s plus 2 equals 0, so s equal minus 0.4226 at K equal 0.3849. Below that all roots are real. Above it a pair bends toward the 60 degree lines and crosses where Routh gives K equal 6, auxiliary polynomial 3 s squared plus 6, omega 1.4142. At K equal 1 the damping is 0.5148.
- The yaw damper of Module 12. The rudder feedback makes the characteristic polynomial linear in K, so the same rules apply to a fourth-order system. At K equal 0.5 the Dutch-roll pair is at minus 1.6647 plus or minus 1.5465 i, damping 0.7326. The pair breaks away at K equal 0.7165, with best damping 0.9776 at K equal 0.7. The branch moving right meets the spiral root at the break-in, K equal 0.9665, and that pair is bound for the zeros at minus 0.0509 plus or minus 0.5797 i: damping 0.5124 at K equal 2, 0.3711 at 3.
- A delay is not rational, so the rules no longer apply. To use polynomial tools, approximate. The exact critical gain from 13.1 is 16.351. The first-order Pade form gives a quadratic stable up to K equal 21.0000, at omega 2.0976: 28 percent above the exact value, an unsafe error. The second-order form gives 16.4866, 0.8 percent above. Pade puts a zero at plus 2, in the right half plane: the signature of a delay. At K equal 14.5 the two forms give damping 0.1846 and 0.0587.
- A loop is stable when every pole is in the left half plane, and Routh counts the bad ones from the coefficients alone. It recovered the integral edge of 5.5 at K p equal 4, and the root locus showed the cruise breakaway at minus 0.55 and the yaw damper breakaway at 0.7165. A delay was the gap: Pade puts the critical gain at 21, against the exact 16.351. Lesson 13.5 is not yet written. It is meant to take the exact route, gain and phase against frequency as a Bode plot, with the margin to instability read off it, for a model or a measured system.
What stability means
The standard statements (from memory of the textbook theorems): a transfer function with common factors cancelled is BIBO stable, every bounded input giving a bounded output, exactly when all its poles lie in the open left half plane ; and is asymptotically stable exactly when every eigenvalue of has (7.2 Linear Systems & the Eigenvalue Method). A simple pair on the imaginary axis rings forever, and a bounded input at that frequency gives an unbounded output, the resonance of 5.6 Forced Oscillations & Resonance: marginal is not safe.
In a loop the poles that matter are the roots of . Write ; the characteristic equation is
For the PID loop of 13.3 it was . Test the whole loop, not a cancelled : a pole cancelled against a zero is hidden from but not from the plant.
The Routh-Hurwitz test
A polynomial with whose roots all have negative real part has every coefficient positive (products of and factors). A missing or negative coefficient proves instability; the converse holds only up to degree two.
The array. Write the coefficients alternately in two rows and build each new row from the two above:
Missing entries are zeros, and a row may be scaled by a positive number. The rule (from memory, not proved here): the number of right-half-plane roots equals the number of sign changes in the first column.
Worked example 1: the cruise PI loop. For the first column is . No sign change needs and
which is 13.3’s edge with no root finding. With a derivative gain the becomes and the edge : at it is for and for (both reported by the lab’s cruise PI and PID plants, at ). Derivative damping buys room for more integral gain.
The lab’s plant for the table is (, gain ):
| entry | changes | roots | of pair | |
|---|---|---|---|---|
| zero row | ||||
At the column has two changes: two right-half-plane roots, counted without computing them.
Special case 1: a zero pivot. Replace a zero first entry (with a nonzero rest of the row) by , continue, and read signs as . For the rows are , , , then negative, then : two changes, two unstable roots. (In the lab type 0 for num and 1 2 1 2 1 for den: roots , , .)
Special case 2: a row of zeros means roots in symmetric positions. Form the auxiliary polynomial from the row above; its roots are roots of the original. Replace the zero row by the coefficients of its derivative and continue. At the zero row is , the row above is , and gives , leaving . A zero row with no sign changes is the edge: marginal.
Worked example 2: a gain range. For , has rows , , , : stable for . At the auxiliary polynomial gives ; the lab agrees, and at gives , ().
The root locus
The root locus plots all roots of as runs from to . A point is on it when is a positive real number: the magnitude gives the gain there, and the angle condition picks the points. Let have degree with poles and degree with zeros .
Rules. The first three are from memory; the rest are derived.
- There are branches, symmetric about the real axis.
- They start at the poles (); end at the zeros and go to infinity.
- A real point is on the locus if the number of real poles and zeros to its right is odd.
- Asymptotes. For large , times a constant, with (compare the coefficients). The locus needs , negative real, so the branches leave along lines through the centroid at angles .
- Breakaway and break-in. Where branches meet on the axis, is stationary: . Real roots with are the candidates.
- Imaginary-axis crossings are the Routh edge: from the zero row, from the auxiliary polynomial.
By Vieta, when the coefficient of is independent of , so the sum of the closed-loop poles is constant. The rules give a sketch; the lab solves on a grid of for the numbers (the leading coefficient of rescales but not the shape).
Locus 1, by hand: . Segments and ; centroid , angles . Breakaway: gives (on a segment) or (not), with . Crossing at , . The lab prints the same. Below all roots are real; above it two become a pair that bends toward the lines and crosses at .
Locus 2: the cruise plant with engine lag, . has poles , and : segment , centroid , asymptotes vertical. The pole sum is fixed at , so after leaving the axis the pair stays on , which is why 13.2 found at every gain. Breakaway: at , matching 13.2’s discriminant (the lab prints , ). At the roots are (, ); at , (, ). There is no crossing: the lab reports it stable over its whole range.
Locus 3: PI with at . Poles and , no zeros: segment , centroid , angles , . Rule 5: has discriminant , so no breakaway. The complex branches run out along , into the right half plane, and must cross: at , . The table above is this locus sampled. Everything the integral gain does to the pair is paid out of damping.
The yaw damper of Module 12
In 12.6 Lateral-Directional Modes the damper changes one column of the matrix: with from 12.6’s lab cell. A rank-one change makes the characteristic polynomial linear in , . Numerically, rounded to five places (the lab uses these),
The open-loop poles are roll , Dutch roll , spiral ; the zeros of are (next to the roll pole, so roll barely reacts) and a lightly damped pair . With there is one asymptote at . At the lab gives , , (), 12.6’s table. The locus explains it:
- Breakaway at , : the Dutch-roll pair splits (12.6 said “about ”). The pair’s best damping is just before, at .
- Break-in at , : the branch moving right meets the spiral root moving left, and at they are a pair .
- That pair is bound for the zeros. At it is (), at (), ending where . Past , more gain trades the fast Dutch roll for a slow, poorly damped oscillation near rad/s.
The loop stays stable over the slider range. Reading the damping off the locus, is near .
Delay: a transfer function that is not a polynomial
A delay multiplies by , which is not rational, so has infinitely many roots and the rules above do not apply. 13.1 used the phase condition: for with s the loop is critical when , at and (recomputed: , ).
To use polynomial tools, approximate. The first-order Pade form (from memory) is , which matches the series through ( against ). Then , stable for , with , at the edge: the lab gives at rad/s, above the exact , an unsafe error. The second-order form (from memory) gives at rad/s, above.
Two lessons. The Pade form puts a zero at , in the right half plane (a non-minimum-phase zero): the signature of a delay, since the locus must end on it and the gain is capped. And the right structure can still miss the number that matters: at the first form gives (), the second () with ; 13.1 saw ringing that decays at and barely decays at .
More worked examples
Counting unstable roots. For the rows are , , , : the column has two changes. Check: , roots , .
Concept checks
Playground and exercise
The first cell finds loci numerically with a root finder and bisects for the edge gain, the second asks you to write the Routh test itself.
Problems
The cruise plant with the engine lag is under PID control with and . Find the largest stable and the frequency of oscillation there.
Show worked solution
The polynomial is . The array is , , , , so . At the auxiliary polynomial is , rad/s. Check: , so the third root is . In general the edge is and does not depend on .
Sketch the locus of : real-axis segments, asymptotes and centroid, and the breakaway point (as a root of a cubic). Then use Routh to show the loop is stable for every .
Show worked solution
Poles , zero : , . Real-axis points with an odd count of poles and zeros to their right: and . Two branches go to infinity: angles and , centroid , so the asymptotes run along the imaginary axis. The third branch ends on the zero at . Breakaway: , so , whose roots are , and . Only is on the locus (on ), at . For the other roots . Stability: has array , , , . The first column is positive for all . The pair approaches the imaginary axis from the left as grows (centroid ) without ever crossing it, and the sum of the poles stays , so the third root tends to . A numerical check: at the roots are and .
Apply the Routh test to . What happens in row , and what do you conclude about the roots?
Show worked solution
Rows: , , , and then for the row, which is entirely zero. The auxiliary polynomial from the row is , so are roots. Replacing the zero row by the derivative gives the column with no sign change, so the other two roots are in the left half plane. Indeed has roots : marginally stable, with an undamped oscillation at .
The plant with a delay s is under proportional gain . Use the first-order Pade form to find the critical gain and its frequency. The exact phase condition is . Which is larger?
Show worked solution
. The coefficient vanishes at , where , so rad/s. The exact condition is solved numerically by , giving . So the Pade value is the larger, about too generous: it would call stable a loop that is not, the same direction of error as the against of the lesson.
What comes next
Everything here needed a polynomial, a model. The delay was the gap: Pade moved the critical gain from to against the exact , while the exact answer came from a phase condition on the frequency response that needed no polynomial. Lesson 13.5 (13.5 Frequency Response & Bode Plots) takes that route: gain and phase against frequency as a Bode plot, with the margin to instability read off it, for a model or a measured system.
Key Takeaways
- A loop is stable when every root of is in the open left half plane; the imaginary axis is marginal, not safe.
- Routh counts right-half-plane roots as first-column sign changes (from memory, unproved). A zero pivot takes ; a zero row uses the auxiliary polynomial, and at an edge its roots are the oscillation.
- The cruise PI loop is stable for : at , at , with .
- Root locus rules: branches, odd-count real segments, asymptotes at through , breakaway where , crossings from Routh. For the pole sum is constant.
- Cruise plant with lag: breakaway at , vertical locus. PI at : crossing at , .
- Yaw damper: breakaway ( peak at ), break-in , then a slower pair heads for the zeros at .
- Pade (1,1) puts the delay’s critical gain at , Pade (2,2) at , the exact value is .