System Identification: Problem & Procedure

System identification infers a model of a dynamical system from observed inputs and outputs. It does not discover “the true system.” It constructs a model family, parameters and uncertainty that are adequate for a declared use over an observed operating domain.

This chapter gives the end-to-end decision loop. The mechanics are developed separately in identification experiments and data conditioning, nonparametric impulse/frequency-response estimation, and prediction-error models and validation.

flowchart TD
  Q["Define model use"] --> E["Design excitation and measurements"]
  E --> S["Choose structure"]
  S --> F["Estimate parameters"]
  F --> V["Validate on independent data"]
  V -->|"residual mechanism remains"| E
  V --> M["Publish validity envelope"]

Begin with the use #

A model for one-step prediction may need a noise model. A model for controller design must preserve gain and phase near loop bandwidth. A model for physical diagnosis needs interpretable parameters. A high predictive score does not make one structure suitable for all three.

Define inputs, outputs, disturbances, sampling, operating range, acceptable error and which decisions depend on the result. This also determines what excitation is safe and informative.

Excitation creates identifiability #

Parameters cannot be learned from behavior the experiment never excites. A constant input cannot distinguish several dynamic modes. A signal concentrated at low frequency says little about fast dynamics. Inputs should cover the amplitudes and frequency band relevant to the model's use while respecting constraints and allowing transients to become observable.

Closed-loop data require special care: controller actions are correlated with disturbances and noise, so naive regression can be biased. Record the reference, controller output and measured plant output; use identification methods whose assumptions match the feedback experiment.

Measurement quality is part of the model. Time alignment, anti-alias filtering, sensor dynamics, saturation, missing samples and calibration uncertainty can masquerade as plant dynamics.

The experiment chapter turns these concerns into a concrete protocol: declare the operating envelope, design spectral coverage, record the intersample behavior and timing chain, preserve repeated runs, and split validation by experiment rather than randomly shuffling adjacent samples.

Choose a structure, then estimate #

Common structures form a spectrum:

• white-box: equations and parameters come from mechanism and conservation laws;
• grey-box: physical structure is known, uncertain parameters are estimated;
• black-box linear: impulse/frequency response, transfer-function, ARX-family or state-space models;
• black-box nonlinear: basis functions, trees, neural networks or other flexible regressors;
• fuzzy rule model: local consequent models blended by membership functions, as in Takagi–Sugeno systems.

More flexible structures reduce approximation bias but increase data need, variance and extrapolation risk. For control, a simpler model with the right uncertainty around crossover can be safer than a high-order model that fits irrelevant noise.

Use a nonparametric impulse or frequency-response estimate first when the data support it. It exposes delay, resonances, usable bandwidth, nonlinear contamination and frequencies with weak excitation before a low-order parametric structure forces those features into poles and zeros.

Parameter estimation commonly minimizes a prediction or simulation error,

$$\hat\theta=\arg\min_\theta \sum_{k\in\mathcal E}\ell\!\left(y_k,\hat y_k(\theta)\right),$$

over estimation data $\mathcal E$. The loss function and predictor encode assumptions about noise and intended use. Report parameter covariance or another uncertainty representation; a point estimate hides whether the data identified the parameters at all.

ARX, ARMAX, output-error and Box–Jenkins models differ mainly in how they constrain the plant and disturbance paths. Those constraints change bias, variance, computation and the meaning of the residual; the model-family chapter derives them from one common $G/H$ representation.

Validate mechanisms, not just fit #

Use separate validation data, preferably from different runs and operating conditions. Check:

1. time-domain simulation and prediction error;
2. frequency response in the decision-relevant band;
3. stability, delays and physically meaningful signs/units;
4. residual autocorrelation;
5. residual correlation with past inputs;
6. parameter stability across data subsets;
7. performance near constraints and regime boundaries.

Residuals $e_k=y_k-\hat y_{k|k-1}$ should not contain predictable structure the model was expected to explain. A visually small residual can still be correlated and therefore diagnostically important. Whiteness does not prove the model is physically correct; it only says the tested predictor has not left detectable structure under that experiment.

Nonlinear, fuzzy and neural models #

Flexible nonlinear models belong in the same identification discipline. Split estimation and validation data by runs or time blocks, preserve temporal order, test interpolation separately from extrapolation, and inspect closed-loop behavior before using the model in a controller. A neural model can fit a MISO mapping without yielding a stable controller. A fuzzy rule base can be readable while its membership functions and local models remain poorly identified.

If a learned model will drive action, include input constraints, uncertainty, out-of-distribution detection and a fail-closed operating envelope in the design. Predictive accuracy alone is not a stability proof.

Deliverable #

Publish the model together with data provenance, preprocessing, sample time, input intersample assumption, structure/order, parameter uncertainty, validation cases, residual diagnostics and validity limits. System identification is complete only when another engineer can reproduce both the estimate and the reasons it is safe—or unsafe—to use for the intended decision.