There’s a recurring conversation in control engineering that goes something like: “the model doesn’t match the plant.” What follows depends on the engineer. The good ones know that this is the central problem of the field, not an exception to it — every real controller is built on top of a model someone obtained from data, and the gap between the model and the plant is where every interesting failure mode lives.
System identification is the practical art of closing that gap honestly. It’s a craft as much as a body of theory: you wiggle an input, measure the output, look at the response, and slowly converge on a model that captures enough of the plant’s behavior to design a controller against. This primer is a tour through that craft.
What it covers
Ten chapters, about forty-five minutes to read plus tinkering time, written for control / motor / mechatronics engineers.
§1–2 — Foundations and workflow, essential terminology. The mental model. What “a model” means and what it doesn’t. The setup before you collect a single data point.
§3 — Non-parametric methods. Step responses, frequency sweeps, impulse responses. The first tools you reach for when you don’t yet know what model class fits the plant.
§4–5 — First- and second-order models. The two model classes that handle 80% of process control. How to extract gain, time constant, and damping from a step response — by hand, in five minutes.
§6 — Parametric methods. ARX, ARMAX, OE, BJ. The model classes that actually fit real noise structures, and the recursive-least-squares machinery underneath them.
§7 — Input design and excitation. What signal to inject. Why PRBS beats steps for parameter estimation. Spectral content vs amplitude tradeoffs.
§8 — Validation and model choice. Cross-validation, residual analysis, model order selection. The unglamorous part that decides whether you have a model or a fantasy.
§9–10 — Pitfalls and rules of thumb. Sampling rate gotchas. Closed-loop identification bias. Drift, aliasing, and the small handful of mistakes that cost the most time.
Scope is linear SISO, time-invariant. Tools assumed: step tests, PRBS, MATLAB/Simulink, pen and paper.
Read it
← Back to Autonomy