6.5 Convolution & Transfer Functions ended with a transfer function: a ratio of polynomials in whose poles say how a system rings after a kick. Every system there was left alone to respond. This lesson closes the loop. We measure the output, compare it with what we wanted, and push the input to make the difference smaller. It is the idea behind a thermostat, a cruise control, the fan of a wind tunnel and the yaw damper of 12.6 Lateral-Directional Modes, where feeding the yaw rate back to the rudder pulled the Dutch roll pole left. Here the idea is stripped to its simplest case: a plant with one pole, a controller with one gain. Even that case shows what feedback gives (a faster pole, a smaller error) and what it charges (noise, delay, instability).
Transcript
- A cruise control that ignores the speedometer is open loop. It applies a calibrated command and never looks at the result. At t equals 10 the speed is 0.6321, at 30 it is 0.9502, and a hill at 30 seconds takes it down to 0.0473, still falling, because nothing pushes back. Closed loop adds a sensor. The speed is subtracted from the reference r to give the error e, the controller K turns it into a command u, the plant P responds, and the disturbance d, the hill, enters beside the command. The price is that the output now feeds its own input.
- Put u equals K times r minus y into the plant and divide by 1 plus K. It is again a first-order lag. The pole is the root of tau s plus 1 plus K, so it moves from minus 1 over tau to minus 1 plus K over tau. With tau equal 10, K equals 1, 4, 9 and 19 puts the pole at minus 0.2, minus 0.5, minus 1 and minus 2, and the time constant shrinks from 10 seconds to 5, 2, 1 and 0.5. At K equals 4 the speed is 0.7947 at t equals 10 where open loop managed 0.6321, and the time to 90 percent of the final value is 4.6 seconds against 23.0.
- Let the loop settle. The speed ends at K r plus d over 1 plus K, so with no hill the steady-state error is 1 over 1 plus K. At K equals 1 the speed settles at 0.4988 and the error is 0.5. At K equals 4 it is 0.8000, error 0.2. At 9, 0.9000 and 0.1. At 19, 0.9500 and 0.05. Feedback makes the error small but never zero, because to hold a nonzero output the plant needs a nonzero command, and the controller needs a nonzero error to produce it.
- Now the hill, d equals minus 1 at t equals 30. The disturbance enters through P over 1 plus L, which at zero frequency is 1 over 1 plus K. Open loop has no defence: 0.9502 before the hill, 0.0473 and falling after. At K equals 4 the speed goes from 0.8000 to 0.6000, a dent of 0.2. At K equals 1 it falls from 0.4988 to 0.0012. The dents at K equal 4, 9 and 19 are 0.2000, 0.1000 and 0.0500. The hill that costs open loop everything dents the closed loop by 1 over 1 plus K.
- How much does feedback protect against a plant that is not what we modelled? Define the sensitivity S as 1 over 1 plus L, with T equal L over 1 plus L, so S plus T equals 1. A fractional error in the plant reaches the closed loop multiplied by S. At K equals 4, S is 0.2000 at zero frequency, 0.2774 at 0.1 radians per second, 0.7211 at 0.5, and 0.8989 at 1, while T falls from 0.8000 to 0.3578. Feedback is strong at low frequency and weakens as the plant's lag removes loop gain: good where L is large, neutral where it is small.
- The sensor is not ideal. Add noise of rms 0.1 metres per second. The transfer function from noise to output is minus T, so the loop that follows the reference follows the noise too, and the command carries K times it. At K equals 4 the command jitters by 0.3996 and the speed by 0.0078. At K equals 19 the command jitters by 1.8786, against a sensor of 0.1, while the speed, a low-pass filter, barely moves, 0.0197. The cost lands on the actuator. At K equals 19 the opening command is 19, nineteen times the open-loop value, and a real throttle saturates long before that.
- A measurement delay of theta seconds multiplies the loop gain by e to the minus s theta. Stability is lost when the loop gain reaches 1 at a phase of minus 180 degrees, which needs arctan of tau omega plus omega theta equal pi. For a delay of 1 second that happens at 1.632 radians per second, giving a critical gain of 16.351. A delay of 0.5 seconds allows 32.056, and 2 seconds only 8.502. Running the loop with a delay of 1, the largest speed over the last 10 seconds is 0.8943 at K equals 14.5, ringing and decaying, 1.4339 at 16, barely decaying, 78.25 at 18 and 6496.8 at 20. Even at K equals 4 the sensitivity peaks at 1.396, so disturbances near that frequency are amplified.
- So what did a bare gain buy? A faster pole, minus 1 plus K over tau, an error of 1 over 1 plus K, and a hill or plant error cut by the same factor. What it charged: noise on the command, and a critical gain of 16.351 once there is a 1 second delay. The yaw damper of lesson 12.6 is the same mechanism in state space, with the Dutch roll damping going from 0.2065 to 0.733 and the spiral root from minus 0.00678 to minus 0.173. The next lesson in build order, 13.2, is not yet written. It makes the pole picture precise: step responses, rise time, overshoot, settling time and the final value theorem behind 1 over 1 plus K.
Open loop and closed loop
A cruise control that ignores the speedometer is open loop. You calibrate it once, “this throttle position holds m/s on flat road”, and it applies that throttle forever. It works while the world matches the calibration; a hill or a headwind changes the answer and nothing notices.
A closed-loop (feedback) controller measures the result and acts on the difference. The speedometer reading is subtracted from the wanted speed, and the throttle is changed in proportion to what is left. The parts have standard names:
- the plant : the thing being controlled (the car, a room, an aircraft);
- the controller : the rule that turns error into command (here a pure gain );
- the sensor: what measures the output (here ideal, later noisy or delayed);
- the reference : what we want; the output : what we get;
- the error , and the command ;
- the disturbance : anything we did not command that pushes the plant, such as a hill.
The price of closing the loop is that the output now feeds its own input, so the system can chase its tail. Everything below is about when that helps and when it hurts.
A cruise-control plant
Let be the speed deviation from cruise in m/s. Newton gives with linearised drag . Take kg and N s/m (assumed round numbers, in the right range for a car around m/s, from memory). Dividing by and calling gives the plant
Here is the command expressed as the steady speed it would buy, and the hill in the same units: is a hill that, left alone, takes m/s off the speed. This is the form of 3.4 Linear First-Order Equations & Integrating Factors, and its transfer function is the first-order lag of 6.5 Convolution & Transfer Functions. Open loop, gives : at s and at s. A hill of at s sends the car to the bottom: and falling toward . The hill ate the whole command and nothing pushed back.
Block diagram algebra
A block diagram draws signals as arrows and operations as boxes. Three pieces are enough: a box multiplies a signal by its transfer function, a summing junction adds or subtracts, and a branch point copies a signal to two places. Boxes in series multiply and boxes in parallel add.
For the loop above, the controller sees , where is sensor noise, and the plant sees . Write that out in the domain, with the loop gain (multiply the blocks around the loop once):
Solving for gives the three transfer functions every control course keeps returning to:
The rule behind the algebra is forward over one plus loop: the transfer function around a negative-feedback loop is the forward path divided by one plus the loop gain. Reference and noise share the factor : a loop that follows the reference well follows the sensor noise too.
Proportional feedback moves the pole
Put into the plant equation with no noise:
It is again a first-order lag. The time constant has shrunk from to and the pole has moved from to
The pole is the root of : gives . In state-space language (7.1 Systems & State Space and 7.2 Linear Systems & the Eigenvalue Method) the plant is with , and feeding back gives : the eigenvalue moves by exactly the gain. That is the whole mechanism; lesson 13.6 will do it for matrices.
The lab simulates the loop at with the exact plant step , , s, reference from and a hill at s. Measured values at and s:
| pole | error | hill dent | ||||
|---|---|---|---|---|---|---|
| s | ||||||
| s | ||||||
| s | ||||||
| s |
At the speed is at s, where open loop managed . The time to reach of the final value is s against s open loop.
Steady-state error and disturbance rejection
Let the loop settle. Setting in the last equation gives
With the speed settles at of the reference, so the steady-state error is
Open loop had zero error only because the model was perfect. Closed loop has an error by design, at , at . And the disturbance enters through , which at zero frequency is : the hill that dented open loop by dents the closed loop by , the “hill dent” column. Feedback buys a factor against disturbances, and against plant changes too.
Notice what feedback did not do. It did not remove the error. No finite makes zero. The remedy is to make the controller remember the error (integral action), the subject of 13.3 PID Control. For now, a bare gain trades error for gain, and the cost of gain is the rest of this lesson.
Sensitivity and robustness
How much does feedback protect against a plant that is not what we modelled? Write the reference-to-output transfer function as , the complementary sensitivity, and define the sensitivity
Differentiating with respect to gives , so
A fractional error in the plant arrives at the closed-loop response multiplied by . Open loop passes it at full strength. For the lab’s is
| (rad/s) | ||||
|---|---|---|---|---|
These come from . Feedback is strong at low frequency and weakens as the plant’s lag removes loop gain; beyond about rad/s, well above any speed change a car makes, creeps toward and feedback does nothing. The loop is good where is large and neutral where it is small.
The cost: noise
The sensor is not ideal. Add noise of rms m/s, uniform and sampled at Hz (seeded generator, so the run repeats). The transfer function from to is : the loop that follows the reference follows the noise, and the command carries times it. Measured over s as the rms difference between the noisy and the clean run:
| command jitter | speed jitter | |
|---|---|---|
The throttle jitters by about : at and at , against a sensor of . The car, a low-pass filter, barely moves ( and m/s). The cost lands on the actuator, whose limits the sim ignores: at the opening command is , nineteen times the open-loop value, and a real throttle saturates long before that.
The cost: delay and the edge of instability
A measurement delay of seconds multiplies the loop gain by : . A delay adds phase lag proportional to frequency, and nothing in the first-order lag can fight it. Stability is lost when, at some frequency , the loop gain reaches magnitude at a phase of , so that the feedback, meant to be negative, comes back positive. That requires
For s the lab finds rad/s and . Without delay no gain is critical. With s the limit is , and with s only . Running the loop at and watching over the last s: at it is (ringing, decaying), at it is (barely decaying), at it is and at it is .
The sensitivity shows the squeeze before the break. With and the peak of over to rad/s is , at : disturbances near that frequency are amplified, not rejected. A famous result (from memory, the Bode sensitivity integral) says that for stable loops of this kind the area below is paid for by area above it: pushing down at low frequency raises it somewhere else.
Back to the aircraft
12.6 Lateral-Directional Modes built a yaw damper without calling it feedback: the rudder was moved by from a rate gyro. That is proportional feedback from a measured state, and the closed loop matrix was , with (numbers from that lesson’s sim, not this one). The Dutch roll damping went from at to at , and the slow spiral root moved from to : a side effect that fought steady turns. A faster, better-damped mode, bought with a gain, with a cost visible only in what else the feedback touched.
Worked examples
Hill at gain 4. , , , after . Closed-loop equation: . Pole , s. Before the hill ; after it . Dent of the open-loop . The table’s row agrees: and .
Playground: the loop as a simulation
The exercise
Implement the four formulas of the lesson and test them against the loop.
Concept checks
Problems
A room has with min, the temperature deviation in degrees and in degrees of steady temperature. A thermostat applies with . Find the closed-loop pole, the time constant, the steady-state error for a degree step in , and the change in when an open door adds .
Show worked solution
The pole is per minute, so min. The error is degree. The door changes by degree, against open loop.
The sensor multiplies the output by , so . Show that and find for , , and . Compare with and say which speed the controller thinks it has reached.
Show worked solution
From , . At : . With : . The measured speed in the first case is , and the controller sees an error of , the usual formula applied to the reading: it works on the reading, not the speed.
The true plant has static gain instead of : . (a) Find the closed-loop final speed for and compare the fractional change with the open-loop . (b) Compare with the first-order estimate . (c) What keeps the closed-loop change under ?
Show worked solution
(a) against for : a change of . Open loop changes by . (b) , so the estimate is ; the exact value is a bit larger because itself changes as moves. (c) The ratio must be at least : , so .
For s and s verify that rad/s satisfies the phase condition, compute , and explain why a first-order plant with no delay is stable for every .
Show worked solution
, and : the sum is . Then . Without delay the phase of is , which never reaches (its limit is ), so the loop never has a frequency where it turns positive.
What comes next
This lesson used one gain and one pole, and judged the response by looking at plots. The next lesson in the course, 13.2 (13.2 Transfer Functions & Time Response), makes that precise: first- and second-order step responses, rise time, overshoot and settling time, the final value theorem behind the above, and how pole locations map to what a response looks like. A lagging sensor or actuator adds a second pole, the road to the oscillatory instability previewed above.
Key Takeaways
- Open loop applies a calibrated command and ignores the result; closed loop measures the output, forms the error , and acts on it. Plant, controller, sensor, reference, error and disturbance are the parts.
- Block-diagram algebra: boxes in series multiply, a negative loop is forward over one plus loop gain, and with .
- For with gain the pole moves from to and the time constant to ( s to s at , s at ).
- The steady-state error is and a constant disturbance is reduced by the same factor ( at , at ); no finite gain makes the error zero.
- Sensitivity and satisfy ; a fractional plant error reaches the closed loop multiplied by , which is small only where the loop gain is large.
- Costs: sensor noise reaches the command as roughly ( at , at for ), and a measurement delay of s makes the loop unstable above .
- The yaw damper of 12.6 is the same mechanism in state space, ( from to ).
- Next: 13.2 Transfer Functions & Time Response (13.2) turns the first-order pole picture into time-response specifications and the final value theorem.