A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
System identification determines a mathematical model of a dynamic system from measured input-output data, bridging the gap between theoretical control design (which assumes a known plant model) and real-world implementation (where the plant is initially unknown). The step response method applies a step input to the plant and fits the output to a first-order-plus-dead-time (FOPDT) model G(s) = K·e−Ls/(τs + 1) by extracting the steady-state gain K, the time constant τ (time to reach 63.2% of final value), and the apparent dead time L from the response curve. Frequency response identification applies sinusoidal inputs at multiple frequencies and measures the steady-state amplitude ratio and phase shift to construct an empirical Bode plot, from which a transfer function can be fitted. More advanced methods include least-squares parameter estimation, which minimizes the sum of squared prediction errors between the model output and measured data, and subspace identification methods that estimate state-space models directly from input-output sequences. The quality of an identified model depends critically on the excitation signal: it must be sufficiently rich (containing enough frequency content) to excite all the dynamics of interest, and the data must be collected under conditions representative of the intended operating regime.
Collect step response data from a simulated plant with known parameters, then apply the graphical FOPDT fitting method and compare your estimated K, τ, and L against the true values. Next, add measurement noise and repeat to see how noise degrades the estimates. Progress to frequency response identification by sweeping sinusoids and constructing a Bode plot from the measured data, fitting a transfer function model, and comparing with the known plant.
Every control design technique you have learned — PID tuning, pole placement, root locus — begins with a transfer function G(s) that describes the plant. In practice, that transfer function is rarely handed to you. System identification is the process of building that model from data: you apply a known input, measure the output, and infer the mathematical relationship between them. It is the experimental counterpart to theoretical modeling.
The simplest and most widely used approach is the step response method. You apply a step change to the plant input and observe how the output responds over time. For many industrial processes, the response resembles an S-curve: the output delays briefly, then rises and approaches a new steady state. This shape matches the first-order-plus-dead-time (FOPDT) model G(s) = Ke−Ls/(τs + 1). You extract three parameters graphically: the steady-state gain K = (change in output)/(change in input), the time constant τ (the time to reach 63.2% of the final value after the initial delay), and the apparent dead time L (the lag before the output begins to respond). These three numbers fully specify the FOPDT model and directly feed into IMC-PID or Ziegler–Nichols tuning rules you learned in transfer function analysis.
Frequency response identification gives a richer model at the cost of more experimental effort. Instead of one step, you apply sinusoidal inputs at many different frequencies and measure the steady-state amplitude ratio and phase shift at each frequency. Each measurement gives you one point on the Bode plot. Connecting these empirical points reveals the plant's bandwidth, any resonances, and the high-frequency roll-off — features invisible in a single step test. Once you have an empirical Bode plot, you can fit a transfer function to it or use it directly in frequency-domain design.
Both methods depend critically on the excitation signal — the input you inject into the plant. The input must contain frequency content in the bands where you need model accuracy. A step test excites all frequencies in theory, but the high-frequency energy decays rapidly, making identification of fast dynamics noisy. A sinusoidal sweep is ideal but slow. Pseudo-random binary sequences (PRBS) strike a practical balance: they are easy to generate, safe to apply to industrial plants, and spread energy broadly across frequencies. The key insight — which connects to your prior understanding of transfer functions — is that a system cannot reveal dynamics it was never forced to exhibit. You must excite what you want to identify.
The model you identify is always an approximation valid near the operating point where data was collected. Nonlinear plants may require multiple FOPDT models across different operating regions, or a nonlinear identification framework entirely. The apparent dead time extracted from a step test may also reflect high-order dynamics rather than true delay — a cluster of slow poles creates an S-shaped response that mimics delay, but responds differently to certain controllers. Checking whether your identified model accurately predicts the system's response to a new test input (not the data used for fitting) — called model validation — is the essential final step before using the model for control design.