Automatic Control Theory: The Classical Design Workflow

Automatic control theory is not a catalogue of formulas for making a graph look smooth. It is a chain of claims: a model represents a physical operating regime; a controller changes causal paths through that model; mathematical tests predict stability and performance; and experiments establish whether those predictions survive uncertainty, constraints, sampling, and hardware.

Each analysis view answers a different question. Verification can send the design back to the model, objective, or architecture.

Start with a control contract #

A useful control problem names more than a setpoint. Write down:

• the controlled output $y$ and reference $r$, with units and allowable range;
• the manipulated input $u$, its magnitude and rate limits, and the actuator's failure state;
• disturbances $d$ and measurement noise $n$, including their important frequency or time scales;
• the operating region and states or outputs that must never be entered;
• transient, steady-state, robustness, and control-effort requirements;
• the available sensors, sample timing, computation delay, and mode transitions.

The controller is successful only if the complete contract is met. A small tracking error with a saturated actuator, fragile stability margin, unsafe internal state, or intolerable noise amplification is not successful control.

One loop, several exact maps #

For controller $C$, plant $P$, and unity negative feedback, let $L=PC$. Define

$$S=\frac{1}{1+L},\qquad T=\frac{L}{1+L},\qquad S+T=1.$$

These two functions already show why one performance number is insufficient. Reference-to-output is $T$; reference-to-error is $S$; an input disturbance reaches the output through $PS$; measurement noise often reaches the output through $-T$ and the control signal through $-CS$. Raising loop gain in one frequency band reduces some paths and enlarges or relocates others.

The classical workflow #

1. Model the plant. Apply conservation laws or identify dynamics from data. State units, sign conventions, delays, constraints, and the validity region.
2. Choose an operating point and linearize. A transfer function is usually a local small-signal model, not a global description of the device.
3. Expose the dynamic structure. Factor poles, zeros, gain, integrators, delays, and nonminimum-phase behavior. Check properness and hidden modes.
4. Assemble every signal path. Include sensors, actuators, filters, feedforward, disturbances, noise, and the sign of each summing junction.
5. Establish nominal stability. Use poles, Routh–Hurwitz, Nyquist, or an equivalent state-space test. A plot is evidence only when its contour and assumptions are correct.
6. Measure performance and margins. Inspect time response, steady-state error, bandwidth, sensitivity, gain and phase margins, and actuator demand.
7. Synthesize compensation. Place gain, integral action, lead, lag, notch, feedforward, PID, state feedback, or another mechanism where the contract requires it.
8. Reintroduce reality. Test uncertainty, nonlinearities, saturation, rate limits, quantization, sampling, delay, jitter, mode changes, and failure handling.

Each stage can invalidate an earlier assumption. That is normal design iteration, not a reason to skip the failed check.

What each representation reveals #

The views are translations of one causal system under stated assumptions. They should agree. If a time simulation, pole calculation, and Nyquist count disagree, first suspect a sign, loop opening, delay, cancellation, initial-condition, or sampling convention error.

A worked thread: first-order thermal plant #

Suppose heat input $u$ raises temperature deviation $y$ above ambient:

$$C_{th}\dot y=-\frac{1}{R_{th}}y+\eta u+d.$$

The plant is

$$P(s)=\frac{\eta R_{th}}{R_{th}C_{th}s+1}=\frac{K}{\tau s+1}.$$

Proportional control $u=K_p(r-y)$ gives

$$\frac{Y}{R}=\frac{K_pK}{\tau s+1+K_pK}.$$

The closed-loop time constant becomes $\tau/(1+K_pK)$, but the constant-reference error is $1/(1+K_pK)$. Increasing $K_p$ makes the nominal loop faster and more accurate, yet also increases demanded heat and sensitivity to ignored delay. Integral action can remove constant error, but introduces another state, phase lag, and windup risk. This small example contains the whole workflow: model, close the loop, locate poles, calculate error, identify a limitation, add a mechanism, then verify delay and saturation.

Classical does not mean obsolete #

Classical methods remain the clearest tools for many SISO and weakly coupled loops. Bode shaping makes disturbance/noise tradeoffs visible; Nyquist handles unstable open-loop plants and delay; root locus links a tunable gain to pole motion; error constants expose the internal model needed for polynomial references.

Their boundary is equally important. Strongly coupled MIMO plants, state or input constraints, stochastic estimation, hybrid logic, and nonlinear operating envelopes need state-space, robust, predictive, nonlinear, or hybrid methods. Those approaches extend the same control contract; they do not repeal causality, stability, sensitivity, or implementation limits.

Verification hierarchy #

Progress from cheap structural evidence to representative runtime evidence:

1. units, signs, properness, pole-zero structure, and internal stability;
2. closed-loop maps for every reference, disturbance, and noise path;
3. time and frequency checks over the modeled operating points;
4. uncertainty sweeps and worst-case delays;
5. nonlinear simulation with constraints and mode logic;
6. target-code timing and numerical tests;
7. bounded hardware tests, then commissioning under documented safeguards.

Passing a late test does not excuse an unexplained early contradiction. A controller that “works in simulation” without a traceable model and margin is difficult to trust, diagnose, or safely modify.