12. System Identification and Adaptive Control

Introduction and Context

Classical control design usually begins with a mathematical model of the plant. In many industrial systems, however, the governing equations are incomplete, parameters vary with operating conditions, sensors introduce noise, and unmeasured dynamics influence performance. System identification addresses this difficulty by constructing mathematical models from measured input–output data. Adaptive control then uses such models, either explicitly or implicitly, to modify controller parameters as the plant or operating environment changes.

These subjects extend the analysis introduced in earlier control-system chapters. Instead of assuming that the plant model is known and fixed, the engineer treats modelling and control as interacting tasks. A useful identified model must be accurate in the frequency and operating ranges relevant to control; an adaptive controller must respond to uncertainty without becoming unstable or excessively sensitive to noise. The central engineering problem is therefore not simply estimating parameters, but ensuring that estimation, control, excitation, computation, and safety operate together.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Explain the purpose, assumptions, and workflow of system identification.

2. Formulate least-squares and recursive least-squares estimators.

3. Distinguish prediction-error, subspace, and frequency-domain identification methods.

4. Select excitation signals and assess model adequacy using validation data.

5. Explain gain scheduling, self-tuning regulators, and model-reference adaptive control.

6. Interpret persistent excitation, parameter convergence, and closed-loop stability.

7. Identify practical, ethical, and safety risks in adaptive control applications.

From Experimental Data to a Control-Oriented Model

System identification begins with an experiment. Inputs and outputs are measured at sampling instants, and a model is selected whose parameters are estimated from the data. A discrete-time linear model may be written as

y(k)=G(q,θ)u(k)+H(q,θ)e(k),

where q is the forward-shift operator, θ is a parameter vector, e(k) is an unmeasured disturbance, G represents plant dynamics, and H represents disturbance dynamics. Common model classes include finite impulse-response models, autoregressive models with exogenous input (ARX), autoregressive moving-average models with exogenous input (ARMAX), output-error models, and Box–Jenkins models.

The model structure determines what can be represented. An ARX model, for example, is computationally convenient but combines process and disturbance dynamics in a particular way. An output-error model focuses on deterministic plant dynamics and may be preferable when measurement noise is substantial. A Box–Jenkins model gives separate polynomial descriptions for the plant and noise processes, but requires more estimation effort.

A sound identification study normally includes:

• definition of the modelling purpose and operating region;
• selection of measured variables, sampling period, and model structure;
• experiment design and safe excitation of the plant;
• preprocessing, including synchronization, filtering, detrending, and removal of invalid records;
• parameter estimation;
• validation on data not used for estimation; and
• assessment of uncertainty and suitability for control.

The model is not judged only by its fit to the estimation data. A model that reproduces one data record but gives poor predictions under a different input or operating point is of limited value. Validation must consider residual correlation, prediction accuracy, stability, physical plausibility, and the control objective.

Least-Squares Estimation

Consider the linear regression model

y(k)=φ^T(k)θ+e(k),

where φ(k) is a known regressor vector and θ is unknown. For N observations, define Φ=[φ(1) … φ(N)]^T and Y=[y(1) … y(N)]^T. Ordinary least squares minimizes

J(θ)=\|Y-Φθ\|^2.

If Φ^TΦ is nonsingular, the estimate is

θ̂=(Φ^TΦ)^{-1}Φ^TY.

For an ARX model, the regressor may contain previous outputs and inputs, for example

φ(k)=[-y(k-1),-y(k-2),u(k-1),u(k-2)]^T.

Least squares is attractive because it is simple, fast, and analytically transparent. Under assumptions such as correct model structure, sufficiently informative data, and disturbances that are uncorrelated with the regressors, the estimate has desirable statistical properties. These assumptions are often violated in closed-loop experiments. If feedback makes the input correlated with measurement noise, ordinary least squares can be biased. Instrumental-variable methods, prediction-error methods, or carefully designed experiments may then be required.

The covariance matrix

P=(Φ^TΦ)^{-1}

indicates the information content of the data. Small eigenvalues of Φ^TΦ indicate weak excitation and poor parameter distinguishability. Parameter estimates may then vary greatly in response to noise, even when the output fit appears acceptable.

Recursive Least Squares

For online estimation, data need not be stored in a growing batch. Recursive least squares updates the parameter estimate as each sample arrives:

K(k)=P(k-1)φ(k)[λ+φ^T(k)P(k-1)φ(k)]^{-1},

θ̂(k)=θ̂(k-1)+K(k)[y(k)-φ^T(k)θ̂(k-1)],

P(k)=λ^{-1}[I-K(k)φ^T(k)]P(k-1).

Here K(k) is the gain and λ is a forgetting factor, usually chosen slightly below or equal to one. With λ=1, old data are retained indefinitely. With λ<1, recent data receive greater weight, allowing adaptation to slowly varying parameters. Excessive forgetting, however, increases noise sensitivity and may produce parameter fluctuations or loss of numerical conditioning. Covariance resetting, projection, parameter bounds, and square-root implementations are practical measures for maintaining robust operation.

Prediction-Error and Subspace Identification

Prediction-error methods select parameters by minimizing a criterion based on one-step-ahead prediction errors:

V_N(θ)=\frac{1}{N}\sum_{k=1}^{N}[y(k)-ŷ(k|k-1,θ)]^2.

The predictor is generated using the candidate model and measured data. Nonlinear optimization is commonly used, so initialization, local minima, parameter scaling, and constraints matter. Prediction-error methods can estimate richer model structures than basic least squares and can explicitly represent disturbance dynamics. Their results are often more statistically efficient when the assumed model and noise description are appropriate.

Subspace identification methods estimate state-space models directly from blocks of measured input and output data. A state-space representation is

x(k+1)=Ax(k)+Bu(k)+w(k),

y(k)=Cx(k)+Du(k)+v(k).

Using projections or singular-value decompositions, subspace methods identify the system order and state sequences before estimating matrices A, B, C, and D. They are particularly useful for multivariable systems, where transfer-function parameterizations become cumbersome. Singular values provide an indication of effective model order: a marked decrease after a particular singular value suggests that additional states contribute little to the measured behaviour. Nevertheless, order selection remains an engineering decision and should be checked through validation and physical interpretation.

Frequency-domain identification provides another route. Sinusoidal or broadband excitation is applied, and spectral estimates of the frequency response are formed. This approach is valuable when a frequency-response model is sufficient or when closed-loop operation makes time-domain assumptions difficult. Leakage, windowing, nonlinear distortion, noise, and insufficient frequency resolution must be considered.

Experiment Design, Excitation, and Validation

Identification requires informative data. A constant input cannot reveal dynamic parameters; a slowly varying input may fail to excite fast modes; and a narrow-band signal identifies behaviour only near selected frequencies. Persistent excitation describes a sufficient richness of the regressor or input for parameter distinguishability. In informal terms, an input is persistently exciting of an adequate order when it contains enough independent variation over time to identify the chosen model parameters.

Useful excitation signals include steps, pseudorandom binary sequences, multisines, chirps, and filtered random signals. The choice depends on actuator limits, safety constraints, expected bandwidth, and whether the plant can be operated in open or closed loop. For a sensitive industrial process, a small multisine superimposed on the normal operating signal may be safer than a large open-loop experiment.

Data should be divided into estimation and validation sets. A validation record should preferably involve an input pattern or operating condition different from that used for estimation. Important checks include:

• one-step-ahead and multi-step prediction errors;
• residual whiteness and lack of correlation between residuals and past inputs;
• reproduction of step, transient, and frequency-response behaviour;
• stability and physically plausible poles and gains; and
• robustness of the model under alternative data selections.

A high fit percentage does not by itself prove model validity. Over-parameterized models can fit noise, while under-parameterized models can hide important dynamics. For control, an apparently modest output fit may still be adequate if the model captures the dominant dynamics in the controller bandwidth.

Adaptive Control Architectures

Adaptive control is a feedback strategy in which controller behaviour changes in response to estimated plant conditions or observed performance. It differs from robust control, which generally maintains one controller designed to tolerate a specified uncertainty set. Adaptive control can achieve improved performance over a wider operating range, but it introduces estimation transients, computational demands, and additional stability questions.

Gain Scheduling

Gain scheduling uses a family of controllers selected or interpolated according to a measurable scheduling variable, such as speed, load, temperature, or operating point. The controller gains are obtained offline from models or experiments. Gain scheduling is often reliable when the scheduling variable is measured accurately and changes slowly. It is not, strictly speaking, adaptive in the strongest sense because the controller does not generally estimate unknown parameters online.

Poor scheduling can cause discontinuities, excessive gain changes, or instability between experimentally characterized operating points. Interpolation should preserve meaningful controller structure, and transitions should be bumpless. The scheduling variable must also be causally and reliably measurable; using a noisy variable directly may create unwanted controller modulation.

Self-Tuning Regulators

A self-tuning regulator contains an online parameter estimator and a controller-design procedure. The estimator identifies an ARX, ARMAX, or state-space model; the design step computes controller parameters from the current estimate; and the resulting controller is applied to the plant. This is known as indirect adaptive control because the model is explicitly estimated.

A typical self-tuning regulator operates as follows:

1. Measure input and output data.

2. Update the plant estimate using recursive estimation.

3. Check parameter bounds, covariance, and model validity.

4. Solve a control-design problem, such as pole placement or minimum-variance control.

5. Apply the new controller only if stability and implementation constraints are satisfied.

The controller should not be redesigned on the basis of every noisy parameter fluctuation. Dead zones, update-rate limits, regularization, and supervisory logic can prevent aggressive changes.

Model-Reference Adaptive Control

In model-reference adaptive control (MRAC), a reference model specifies desired behaviour. The controller contains adjustable parameters, and an adaptation law modifies them so that the plant output follows the model output. If e(t)=y(t)-y_m(t), a Lyapunov-based design may select parameter updates that make a candidate Lyapunov function decrease. A simplified adaptation law has the form

θ̇̂(t)=-Γφ(t)e(t),

where Γ is a positive adaptation-gain matrix and φ(t) is a regressor. The exact sign and form depend on the plant and controller parameterization.

MRAC can provide a direct relationship between design objectives and adaptation. However, parameter convergence is not guaranteed merely because tracking error becomes small. If the reference and plant signals do not sufficiently excite all parameters, different parameter values may produce nearly identical outputs. Thus tracking convergence and parameter convergence are distinct properties.

Convergence, Stability, and Persistent Excitation

Adaptive systems must be evaluated at several levels. Tracking convergence means that the difference between plant output and desired output tends to zero or remains acceptably small. Parameter convergence means that estimated parameters approach their true values. Boundedness means that signals and parameter estimates remain finite. These properties are related but not equivalent.

Persistent excitation is central to parameter convergence. If the regressor φ(t) satisfies, for some positive constants α, β, and time interval T,

αI ≤ ∫_t^{t+T} φ(τ)φ^T(τ)dτ ≤ βI,

then the data contain sufficient directional information over every interval of length T. Without this condition, an adaptive controller may track well while retaining large uncertainty in unexcited parameter directions.

Stability proofs commonly use Lyapunov functions, passivity arguments, small-gain reasoning, or bounded-real conditions. In practical implementations, ideal assumptions are supplemented with robustness modifications: σ-modification, e-modification, projection, normalization, dead zones, leakage, and parameter constraints. These devices can prevent drift caused by noise or unmodelled dynamics, although they may introduce steady-state error or reduce adaptation speed.

Applied Scenario: Adaptive Speed Control of a Variable-Speed Drive

Background

Consider an illustrative medium-voltage motor drive used in a process line. Motor inertia changes when a mechanical tool is engaged, and load torque varies during production. A fixed speed controller gives acceptable performance at nominal load but produces sluggish recovery under heavy load and overshoot when the tool is removed.

Situation and Problem

A detailed electromechanical model is unavailable because friction and load coupling are uncertain. Speed, torque command, and current are sampled digitally. The drive must maintain speed without violating current, torque, or thermal limits. Direct aggressive experimentation is unacceptable because a large speed excursion could damage the process.

Analysis

A safe identification experiment is conducted around several operating points. A small filtered pseudorandom signal is added to the torque command while the drive remains under a stabilizing baseline controller. Data are detrended and divided into estimation and validation records. Recursive least squares estimates a low-order discrete-time model, while a subspace method is used offline to check whether an additional state is needed. Residual tests show that the dominant variation is associated with effective inertia and load torque rather than a new unstable mode.

The experiments reveal that a single controller is conservative at low load and too aggressive at high load. However, load torque is not measured directly. Motor current and speed acceleration are used as scheduling and estimation information, with current limits included in the controller design.

Approach and Solution

A supervisory self-tuning regulator is implemented. The baseline controller remains active during initialization and whenever the estimator covariance becomes excessive. Recursive least squares uses a forgetting factor slightly below one, parameter projection, and a bounded controller-update rate. The estimated model is used to calculate proportional and integral gains within precomputed safe intervals. A bumpless transfer mechanism limits changes in the control signal.

Persistent excitation is not guaranteed during steady production, so the system does not claim full parameter convergence. Instead, it monitors prediction error and estimator covariance. Small diagnostic excitation is introduced only when operating conditions permit. If prediction error rises sharply, the supervisor freezes adaptation and returns to the robust baseline controller.

Outcome

The adaptive arrangement improves recovery over the identified load range while avoiding large gain changes during normal operation. Its principal success is not an assertion that every physical parameter has been identified; it is the controlled use of an uncertain model with validation, constraints, and fallback logic.

Learning Points

The scenario illustrates four principles. First, identification should be designed around the control objective rather than model complexity alone. Second, closed-loop data require attention to noise correlation and excitation. Third, adaptation must be supervised by safety and validity checks. Finally, good tracking does not prove parameter convergence.

Practical Risks, Limitations, and Professional Responsibilities

Identification can disturb a plant, expose equipment to unsafe inputs, or reveal sensitive operational information. Experiments therefore require risk assessment, actuator saturation analysis, emergency limits, and approval from responsible operations personnel. The identification signal must remain within safe amplitude, rate, and energy limits.

Common technical mistakes include choosing a sampling period that misses important dynamics, identifying a model from insufficiently rich data, ignoring sensor bias, filtering away relevant phase information, fitting and validating on the same record, and allowing parameter estimates to control the plant without bounds. Integral windup and actuator saturation are especially dangerous because the estimator may interpret saturation-induced behaviour as a change in plant dynamics.

Adaptive control also raises governance questions. Operators should be able to determine which controller is active, why adaptation was suspended, and what limits apply. Software changes require version control, traceability, testing, and, in safety-related systems, formal verification appropriate to the application. Data used to construct models should be documented with operating conditions, sensor calibration, preprocessing steps, and known anomalies. An unexplained model change must not be treated as evidence of real plant degradation without investigation.

Digital Tools and Emerging Directions

Modern identification increasingly uses automated pipelines combining data acquisition, signal-quality assessment, model-order selection, uncertainty quantification, and deployment monitoring. Digital twins can provide simulation environments in which identification and adaptive algorithms are tested before field commissioning. Machine-learning models may capture nonlinear behaviour, but purely predictive accuracy does not ensure stability, extrapolation reliability, or interpretability. Hybrid methods that combine physical equations with data-driven residual models are often more suitable for control.

Current research emphasizes nonlinear and distributed-parameter identification, Bayesian uncertainty representation, safe exploration, reinforcement-learning-assisted adaptation, and networked control under communication constraints. These developments do not remove the need for excitation analysis, stability guarantees, validation, and engineering safeguards. In high-consequence applications, a simpler model with known uncertainty may be preferable to a highly accurate but opaque model.

Best-Practice Workflow

A defensible project follows a sequence such as the following:

1. Define the operating envelope, control objective, and acceptable risk.

2. Select sensors, sampling rate, excitation, and candidate model classes.

3. Conduct a safe experiment with synchronized and calibrated data.

4. Estimate several plausible models rather than relying on one algorithm.

5. Validate using independent records and residual analysis.

6. Quantify uncertainty and test sensitivity to data preprocessing.

7. Design the adaptive mechanism with parameter bounds and fallback control.

8. Test saturation, disturbances, initialization, sensor failure, and loss of excitation.

9. Commission gradually, monitor performance, and preserve an audit trail.

Chapter Summary

System identification constructs dynamic models from measured data for analysis, prediction, and control. Least-squares methods provide efficient parameter estimation for linear regressions, while recursive least squares supports online adaptation through covariance updating and forgetting. Prediction-error methods optimize one-step prediction performance and can represent richer disturbance structures. Subspace methods estimate state-space models and are particularly useful for multivariable systems.

Reliable identification depends on experiment design, persistent excitation, appropriate sampling, independent validation, and careful residual analysis. Adaptive control may be implemented through gain scheduling, self-tuning regulation, or model-reference adaptation. Tracking convergence, parameter convergence, and boundedness are different requirements; persistent excitation is especially important for parameter convergence. Practical adaptive systems require constraints, supervisory logic, safe fallback operation, and documentation. The most effective designs combine mathematical guarantees with engineering judgement and operational governance.

Key Terms

System identification: Estimation of a mathematical dynamic model from measured data.

ARX model: A model using autoregressive output terms and exogenous input terms.

Prediction-error method: Identification method that minimizes errors between measured and model-predicted outputs.

Subspace identification: State-space estimation using matrix projections and numerical decompositions of data.

Regressor: Vector of measured or constructed variables multiplying unknown parameters.

Recursive least squares: Online least-squares estimation updated as new samples arrive.

Forgetting factor: Weighting factor that reduces the influence of older data.

Persistent excitation: Sufficiently rich input or regressor variation to distinguish parameters.

Residual: Difference between measured output and model prediction.

Model validation: Assessment of model adequacy using data or tests not used for estimation.

Gain scheduling: Selection or interpolation of controller parameters according to an operating variable.

Self-tuning regulator: Adaptive controller that estimates a model and redesigns control parameters online.

Model-reference adaptive control: Control in which adjustable parameters force plant behaviour toward a reference model.

Tracking convergence: Tendency of the controlled output to approach the desired output.

Parameter convergence: Tendency of estimates to approach the true parameter values.

Projection: Restriction of adaptive estimates to a prescribed feasible parameter set.

Dead zone: Region in which adaptation is suspended when error is sufficiently small.

Bumpless transfer: Change between controllers that avoids an abrupt control-signal discontinuity.

Covariance matrix: Matrix representing parameter-estimation uncertainty or information quality.

Robust fallback controller: A prevalidated controller used when adaptation becomes unreliable.

Review Questions

Short Answer Questions

1. What is the purpose of separating estimation and validation data?

2. State the principal update equations of recursive least squares and explain the role of the forgetting factor.

3. Distinguish tracking convergence from parameter convergence.

4. Why can ordinary least squares be biased in a closed-loop experiment?

5. What is the function of a reference model in MRAC?

Long Answer Questions

1. Explain the complete workflow of a system-identification study, from experiment design through model validation.

2. Compare ARX, prediction-error, and subspace identification methods with respect to structure, computation, and application.

3. Develop and discuss the operation of a self-tuning regulator, including estimation, controller redesign, and supervisory safeguards.

4. Explain persistent excitation mathematically and discuss its relationship to parameter convergence.

5. Discuss the practical, safety, ethical, and governance issues involved in deploying adaptive control in an industrial plant.

Analytical / Application Questions

1. An estimator gives excellent one-step prediction on its estimation data but poor multi-step prediction on a new operating condition. Diagnose possible causes and propose a validation and redesign strategy.

2. Design a safe identification experiment for a temperature process with slow dynamics, actuator saturation, and significant measurement noise. Justify the excitation, sampling, model class, and validation procedure.

3. A self-tuning controller begins changing gains rapidly after the actuator saturates. Explain the likely mechanism and design corrective measures involving estimation, constraints, and supervisory control.

22. Multiple Choice Questions with Answers

Question 1

Which condition most directly supports parameter convergence in an adaptive estimator?

A. Constant input with low measurement noise

B. Persistent excitation of the regressor

C. A high controller gain

D. Validation using the estimation data

Correct Answer: B. Persistent excitation of the regressor

Question 2

Why is a separate validation data set important in system identification?

A. It guarantees zero measurement noise

B. It increases the number of estimated parameters automatically

C. It tests whether the model generalizes beyond the data used for fitting

D. It eliminates the need for residual analysis

Correct Answer: C. It tests whether the model generalizes beyond the data used for fitting

Question 3

In recursive least squares, a forgetting factor below one primarily causes the estimator to:

A. Ignore all new data

B. Give greater relative weight to recent data

C. Guarantee parameter convergence without excitation

D. Remove actuator saturation

Correct Answer: B. Give greater relative weight to recent data

Question 4

Which identification approach is especially suited to multivariable state-space modelling using matrix decompositions?

A. Subspace identification

B. Single-step testing only

C. Static gain estimation

D. Manual gain scheduling

Correct Answer: A. Subspace identification

Question 5

A model has excellent fit on estimation data but correlated residuals with past inputs. What is the most appropriate interpretation?

A. The model is necessarily physically exact

B. The residual test suggests unmodelled dynamics or inadequate disturbance treatment

C. Persistent excitation is automatically proven

D. The controller should be made infinitely fast

Correct Answer: B. The residual test suggests unmodelled dynamics or inadequate disturbance treatment

Question 6

Which statement correctly distinguishes gain scheduling from self-tuning control?

A. Gain scheduling always estimates parameters online

B. Self-tuning control never uses a plant model

C. Gain scheduling selects controller settings using an operating variable, whereas self-tuning control estimates and redesigns online

D. They are mathematically identical in every implementation

Correct Answer: C. Gain scheduling selects controller settings using an operating variable, whereas self-tuning control estimates and redesigns online

Question 7

In MRAC, small tracking error does not necessarily imply parameter convergence because:

A. The reference model has no output

B. Unexcited parameter directions may remain indistinguishable

C. The plant must always be unstable

D. Least squares cannot be used in control

Correct Answer: B. Unexcited parameter directions may remain indistinguishable

Question 8

Which measure is most appropriate when adaptive parameters drift because of noise and unmodelled dynamics?

A. Remove all validation tests

B. Use projection, leakage, dead zones, or related robustness modifications

C. Increase excitation without considering safety

D. Disable actuator limits

Correct Answer: B. Use projection, leakage, dead zones, or related robustness modifications

Question 9

In the illustrative variable-speed drive case, why was a robust baseline controller retained?

A. To prevent any use of measured data

B. To provide safe fallback when estimation or adaptation became unreliable

C. To ensure parameter convergence during constant operation

D. To replace all model validation

Correct Answer: B. To provide safe fallback when estimation or adaptation became unreliable

Question 10

What is a major danger of identifying a plant from data collected during actuator saturation?

A. Saturation can be interpreted as a change in plant dynamics and bias the estimate

B. Saturation always improves excitation

C. It guarantees a minimum-phase model

D. It removes the need for parameter bounds

Correct Answer: A. Saturation can be interpreted as a change in plant dynamics and bias the estimate