Conventional and Modern Control Foundations

Introduction

Intelligent control systems are not designed in isolation from classical control theory. Proportional–integral–derivative (PID) control, root-locus analysis, frequency-response methods, state feedback, observers, robust control, and optimal control provide the benchmarks against which intelligent methods must be judged. A neural controller, fuzzy controller, or reinforcement-learning policy is useful only when it improves a measurable property such as tracking, disturbance rejection, robustness, energy use, or safety.

This chapter develops the foundations required to design and evaluate such controllers. The discussion begins with feedback concepts and transfer-function models, then proceeds to time-domain and frequency-domain analysis. State-space methods introduce controllability, observability, pole placement, and observer design. The final sections connect these foundations to robust and optimal control and to an illustrative mechatronic application.

Learning Objectives

By the end of this chapter, students should be able to:

1. Model a mechatronic plant using transfer functions and state-space equations.
2. Explain the roles of feedback, stability, tracking, disturbance rejection, and sensitivity.
3. Design and evaluate PID controllers using time- and frequency-domain criteria.
4. Determine controllability and observability and apply state-feedback pole placement.
5. Explain observer-based control, robustness, and linear-quadratic regulation.
6. Select suitable benchmarks for evaluating intelligent-control solutions.

1. The Feedback-Control Framework

A control system regulates a plant by measuring its output and using the measurement to calculate an actuator command. In a standard negative-feedback loop, the reference signal r(t) is compared with the measured output yₘ(t). The error is

e(t) = r(t) − y_m(t)

A controller C(s) processes the error, while a plant P(s) represents the physical system. If a sensor and actuator are included explicitly, their transfer functions may be denoted H(s) and A(s). For a unity-feedback loop, the closed-loop transfer function is

Here s is the Laplace variable, R(s) is the reference transform, and Y(s) is the output transform. The formula assumes zero initial conditions, linear time-invariant components, and negative feedback. The loop transfer function is L(s) = C(s)P(s), and the characteristic equation is 1 + L(s) = 0. Its roots are the closed-loop poles.

For a disturbance d entering at the plant input, the disturbance-to-output transfer function in a unity-feedback system is commonly

Thus, increasing loop gain can reduce the effect of certain low-frequency disturbances, although excessive gain may reduce stability margins and amplify measurement noise.

Stability

A continuous-time linear system is asymptotically stable when every closed-loop pole has a strictly negative real part. Poles on the imaginary axis require special care: repeated poles there imply instability, while simple poles may produce marginal stability under ideal conditions. A discrete-time system is asymptotically stable when every pole lies strictly inside the unit circle, meaning |zᵢ| < 1.

Stability is necessary but not sufficient for acceptable performance. A stable motor-position system may still be too slow, oscillatory, inaccurate, noisy, or energy intensive.

Performance measures

Common time-domain measures include:

• Rise time: time for the response to move between specified percentages, often 10% and 90%, of its final value.
• Peak time: time at which the first maximum occurs.
• Settling time: time after which the response remains within a specified tolerance band, commonly ±2% or ±5%.
• Overshoot: the maximum excursion above the final value, usually expressed as a percentage.
• Steady-state error: the limiting difference between reference and output.

The chosen definitions must be stated because software packages and engineering standards may use different percentage bands.

2. Mathematical Models for Mechatronic Plants

Transfer-function representation

For a linear time-invariant system with zero initial conditions, the transfer function is

where U(s) and Y(s) are the Laplace transforms of input u(t) and output y(t). A transfer function is appropriate when the system is linearized around an operating point and the input–output relationship is the main concern.

For a mass–spring–damper system,

where m is mass in kilograms, b is viscous damping in N·s/m, k is stiffness in N/m, x is displacement in metres, and F is force in newtons. With zero initial conditions,

The model is valid for a linear spring, viscous damping, constant parameters, and small enough motion that nonlinear effects can be neglected.

State-space representation

A continuous-time state-space model is

The state vector x contains variables that, together with the input, determine future behavior. A is the state matrix, B the input matrix, C the output matrix, and D the direct-feedthrough matrix. Their dimensions must be compatible: for n states, m inputs, and p outputs, A is n×n, B is n×m, C is p×n, and D is p×m.

For the mass–spring–damper system, choose x₁ = x and x₂ = ẋ. Then

with y = [1 0]x when displacement is measured. State-space models are especially valuable for multivariable systems, internal-state analysis, observer design, and modern control.

3. PID Control and Classical Design

The ideal PID controller is

where Kₚ is proportional gain, Kᵢ is integral gain, and K_d is derivative gain. In the time domain,

The proportional term responds to present error, the integral term accumulates past error, and the derivative term anticipates error changes. With consistent SI units, the gains have units determined by the chosen input and output units; they are not universally dimensionless.

For a plant with a nonzero constant gain and no unstable pole, integral action can eliminate steady-state error to a constant reference under suitable stability conditions. However, integral action may cause overshoot, sluggish recovery, or windup. Derivative action can improve damping but is sensitive to measurement noise.

A practical derivative term uses filtering:

where N is a positive filter parameter in rad/s. This is proper and therefore more implementable than an ideal differentiator. In digital implementation, sampling time, numerical integration, quantization, and actuator saturation must be considered.

Integral windup and anti-windup

When the actuator saturates, the integral term may continue accumulating error even though further controller output cannot be applied. Once the error changes sign, the stored integral action can produce prolonged overshoot. Anti-windup methods include conditional integration and back-calculation. In back-calculation, the difference between saturated and unsaturated commands is fed back to the integrator through a tuning gain. The exact implementation must respect actuator limits and sample timing.

Classical design tools

The root locus displays how closed-loop poles move as a real gain varies. It supports qualitative selection of gain for desired damping and speed. Frequency-response methods use the magnitude and phase of L(jω), where ω is angular frequency in rad/s.

Important measures are:

• Gain crossover frequency: frequency at which |L(jω)| = 1.
• Phase margin: additional phase lag required at gain crossover to reach −180°.
• Phase crossover frequency: frequency at which phase is −180°.
• Gain margin: factor by which loop gain can increase before reaching instability, subject to the usual single-loop interpretation.

Margins are indicators, not complete guarantees of performance, especially for nonlinear, time-varying, or poorly modeled systems.

Worked example: PI control of a first-order plant

Consider

and a PI controller

For unity feedback, the characteristic polynomial is

Choose Kₚ = 3 and Kᵢ = 4. The polynomial becomes s² + 5s + 4 = (s + 1)(s + 4). Both poles are negative, so the closed loop is asymptotically stable. For a unit-step reference, the integral term gives zero steady-state error if the closed loop remains stable. The poles also indicate two real decay modes, with the slower mode associated with a time constant of approximately 1 s. This conclusion assumes an unsaturated actuator, an accurate linear model, and ideal sensing.

4. State Feedback

State feedback uses measured or estimated states to assign closed-loop dynamics:

where K is the feedback-gain matrix and N is a reference precompensator. Substitution gives

Thus, the eigenvalues of A − BK are the closed-loop poles. State feedback does not automatically eliminate tracking error for arbitrary plants and disturbances; integral augmentation or a suitable precompensator may be required.

Controllability

The pair (A, B) is controllable if every state can be driven to the origin in finite time by an appropriate input. For a single-input n-state system, the controllability matrix is

The system is controllable if rank(𝒞) = n. Pole placement through state feedback requires controllability of the modes being assigned. Numerical rank tests should use a tolerance because floating-point calculations can make an exactly rank-deficient matrix appear full rank.

Pole placement

For a controllable system, K can be selected so that A − BK has desired poles. In practice, reliable numerical algorithms such as Schur-based methods are preferred to direct symbolic formulas. Desired poles should reflect actuator capability, sensor noise, sampling rate, and unmodeled dynamics. Placing poles excessively far left may require unrealistic control effort and amplify model errors.

5. Observers and Observer-Based Control

A state observer reconstructs unavailable states from measured inputs and outputs. A full-order Luenberger observer is

where x̂ is the estimated state and L is the observer gain. Defining estimation error eₓ = x − x̂ gives

The observer error converges asymptotically when A − LC is stable. The pair (A, C) must be observable, with observability matrix

and rank(𝒪) = n for complete observability.

The separation principle states that, for an appropriate linear time-invariant system, controller poles and observer poles can be designed independently: the combined controller–observer dynamics have poles comprising those of A − BK and A − LC. This does not remove practical limitations. Fast observer poles can magnify measurement noise and create large transient estimates. A Kalman filter provides a probabilistic alternative when process and measurement noise can reasonably be modeled.

6. Robust Control Foundations

A model is never exact. Uncertainty arises from parameter variation, neglected flexible modes, friction, load changes, sensor error, and actuator dynamics. Robust control asks whether stability and performance persist for a specified family of models.

A useful scalar concept is sensitivity:

and complementary sensitivity:

For a unity-feedback loop, S describes how some plant disturbances and model perturbations influence the error, while T describes how reference signals and measurement noise can pass through the loop, depending on injection location. Since S + T = 1, reducing sensitivity in one frequency range generally affects the other. High low-frequency loop gain often reduces tracking error, whereas high-frequency gain can transmit noise and excite neglected dynamics.

Robust design therefore requires:

• explicit uncertainty descriptions;
• gain and phase margins or more formal robustness measures;
• validation over parameter and operating ranges;
• actuator and sensor constraints;
• tests for nonlinearities, saturation, dead zones, and friction.

H-infinity methods formulate a worst-case gain problem, while μ-analysis addresses structured uncertainty. These methods are powerful but require credible uncertainty models; mathematical sophistication cannot compensate for unrealistic assumptions.

7. Optimal Control

Optimal control selects a control signal by minimizing a performance index. For linear-quadratic regulation (LQR), the continuous-time cost is

where Q is positive semidefinite and weights state deviations, and R is positive definite and weights control effort. For the system ẋ = Ax + Bu, the optimal law is

where P is the stabilizing solution of the continuous-time algebraic Riccati equation

The standard existence result requires, among other conditions, stabilizability of (A, B) and detectability of (Q¹ᐟ², A). LQR offers a systematic trade-off between regulation and effort, but the selected Q and R encode engineering priorities and do not automatically handle actuator saturation or hard safety constraints.

Model predictive control extends optimal control by repeatedly solving a finite-horizon optimization, applying only the first command, and then re-optimizing. It can incorporate constraints, but requires a sufficiently accurate model and adequate computation.

8. Benchmarking Intelligent Controllers

An intelligent controller should be compared with a credible baseline under identical conditions. A benchmark protocol should specify:

1. plant model and operating region;
2. sampling frequency and computational hardware;
3. reference trajectories and disturbance profiles;
4. actuator, velocity, temperature, and safety limits;
5. performance metrics and weighting;
6. training and test data, if learning is used;
7. failure-handling and fallback behavior.

Useful metrics include integral absolute error,

integral squared error,

and control-energy measures such as

where T is the evaluation interval. IAE penalizes persistent errors linearly, ISE penalizes large errors more strongly, and Eᵤ measures effort under the chosen scaling. These metrics have units and should not be added without normalization or meaningful weighting.

Intelligent methods may improve adaptation or nonlinear compensation, but they can introduce instability, poor interpretability, distribution-shift failures, and difficult verification. A conventional controller may therefore serve as a stabilizing inner loop, safety monitor, or fallback controller. Claims of superiority should include statistical repetition where randomness is present, uncertainty analysis, and tests outside the training set.

9. Illustrative Case: Servo-Position Control

Background

Consider a laboratory servo that positions a robotic joint. The motor drives a load through a gearbox. An encoder measures position, the amplifier limits voltage, and friction varies with load. A nominal linear model is used near a selected operating point.

Problem

An initial proportional controller gives acceptable motion for small references but leaves steady-state error under load. Increasing gain reduces error but produces oscillation and actuator saturation. A learning controller is proposed, but its safety and generalization are uncertain.

Analysis

A baseline PI controller can remove constant-load steady-state error, but anti-windup is required because of voltage saturation. A derivative term may improve damping, yet raw encoder differentiation would amplify quantization noise. The plant is also subject to parameter uncertainty in inertia and friction.

A state-space model with position and velocity states supports state feedback. Since velocity is not directly measured, an observer or filtered velocity estimate is required. Controllability and observability must be checked for the chosen model. Frequency-response tests can identify resonant frequencies that should not be excited by excessive bandwidth.

Approach and solution

A defensible architecture is a conventional stabilizing controller with a bounded intelligent supervisory component. The baseline controller provides the primary command and enforces saturation and rate limits. The intelligent component may schedule gains according to load or compensate for repeatable friction, but its output is bounded. If confidence, sensor validity, or tracking error violates prescribed limits, the system reverts to the verified baseline.

Evaluation should compare PID, state-feedback/observer control, and the augmented controller using identical trajectories, loads, disturbances, and hardware limits. Metrics should include settling time, overshoot, IAE, peak control voltage, energy, tracking under load changes, and recovery after sensor disturbances.

Outcome

The important outcome is not that an intelligent controller is automatically better. A successful design demonstrates improvement without loss of stability, constraint compliance, or explainability. If the intelligent component reduces error but increases saturation events or fails under an unseen load, it has not met the engineering objective.

Learning points

The case illustrates four principles: classical control supplies measurable benchmarks; state estimation is part of implementation rather than an optional theory; robustness must be tested across uncertainty; and intelligent augmentation should be constrained by safety and fallback mechanisms.

10. Design and Implementation Best Practices

• Linearize only within a stated operating region and validate the approximation.
• Keep units consistent and document signal scaling.
• Check closed-loop poles, time response, frequency margins, and actuator demand.
• Include saturation, quantization, delays, computation time, and sensor noise in simulation.
• Use anti-windup whenever integral action and actuator limits coexist.
• Avoid placing observer poles so fast that noise dominates the estimate.
• Treat gain scheduling as a collection of operating-point designs requiring transition analysis.
• Test nominal performance, worst credible parameters, disturbances, and failure modes.
• Preserve a safe fallback and log controller decisions for commissioning and maintenance.
• Distinguish simulation evidence from hardware evidence.

Challenges, Ethics, and Governance

Control engineers are responsible for more than numerical performance. A controller that can move a robot quickly but cannot detect an encoder failure is unsafe. Safety-related systems may require documented hazard analysis, independent limits, traceability, verification, and compliance with applicable machinery, industrial, automotive, or functional-safety standards. The exact standard depends on the application and jurisdiction.

Data-driven controllers introduce additional governance concerns: training data may omit rare but dangerous conditions; adaptive policies may change after deployment; and opaque decisions can complicate incident investigation. Human override, cybersecurity controls, access management, version control, and clearly defined operating envelopes are therefore engineering requirements, not administrative additions.

Chapter Summary

Feedback control compares a desired reference with measured behavior and shapes plant dynamics through a controller. Transfer functions support classical input–output analysis, while state-space models expose internal dynamics and support multivariable design. Stability depends on closed-loop poles, but useful performance also requires acceptable speed, damping, accuracy, robustness, and actuator demand.

PID control remains widely applicable, provided derivative filtering, anti-windup, sampling, and saturation are addressed. State feedback assigns closed-loop poles when the system is controllable; observers reconstruct unavailable states when the system is observable. Robust control explicitly addresses uncertainty, while LQR provides a principled trade-off between state deviation and control effort. Intelligent controllers should be evaluated against these methods using reproducible metrics, constraints, uncertainty tests, and safety mechanisms.

Key Terms

Feedback: Use of measured output to influence a control command.

Reference: Desired value or trajectory for a controlled variable.

Plant: Physical process being controlled.

Closed-loop pole: Root of the characteristic equation determining linear closed-loop modes.

Stability: Property that responses remain bounded and, for asymptotic stability, converge appropriately.

PID controller: Controller combining proportional, integral, and derivative actions.

Integral windup: Excess accumulation of integral action during actuator saturation.

Root locus: Plot of closed-loop pole locations as a gain varies.

Phase margin: Additional phase lag tolerated at gain crossover before the stability boundary.

State: Minimum collection of variables needed to describe future system evolution with input information.

State feedback: Control law based on system-state measurements or estimates.

Controllability: Ability to move the state using available inputs.

Observability: Ability to infer internal states from inputs and outputs.

Observer: Dynamic estimator of unmeasured states.

Robustness: Persistence of stability or performance despite uncertainty.

Sensitivity: Transfer relationship describing response to selected disturbances or model changes.

LQR: Linear-quadratic regulator minimizing a quadratic state-and-effort cost.

Benchmark: Reference method or test protocol used for fair comparison.

Fallback controller: Verified controller used when an advanced controller is unavailable or unsafe.

Review Questions

Short Answer Questions

1. What is the difference between open-loop and closed-loop control?
2. Why can integral action eliminate steady-state error for a constant reference?
3. State the rank condition for controllability of an n-state system.
4. What practical problem does derivative filtering address?
5. Why should an intelligent controller be compared with a conventional baseline?

Long Answer Questions

1. Explain the roles of proportional, integral, and derivative actions, including their main limitations.
2. Derive the closed-loop characteristic polynomial for a unity-feedback plant and controller, and explain how its roots affect stability.
3. Compare transfer-function and state-space representations for a multivariable mechatronic system.
4. Explain observer design, estimation-error dynamics, and the separation principle.
5. Discuss how robustness, actuator constraints, and safety governance affect intelligent-control deployment.

Analytical / Application Questions

1. For P(s) = 1/(s + 3) and C(s) = 2 + 4/s, derive the characteristic polynomial and determine whether the continuous-time closed loop is asymptotically stable.
2. A servo has excellent nominal tracking but becomes oscillatory when inertia doubles. Identify likely causes and propose a validation and redesign procedure using robustness and frequency-domain analysis.
3. Design a fair experiment comparing PID, LQR with an observer, and a learning-based controller for a robotic joint. Specify test conditions, metrics, constraints, and fallback behavior.

Multiple Choice Questions with Answers

Question 1

For a unity-feedback system, which equation gives the closed-loop transfer function?

A. C(s)P(s)

B. C(s) + P(s)

C. C(s)P(s)/[1 + C(s)P(s)]

D. P(s)/C(s)

Correct Answer: C. C(s)P(s)/[1 + C(s)P(s)]

Question 2

For a continuous-time linear system, asymptotic stability requires all closed-loop poles to have:

A. Positive real parts

B. Strictly negative real parts

C. Zero imaginary parts

D. Magnitude greater than one

Correct Answer: B. Strictly negative real parts

Question 3

Which PID term primarily accumulates past tracking error?

A. Proportional

B. Integral

C. Derivative

D. Feedforward

Correct Answer: B. Integral

Question 4

For an n-state single-input system, controllability is tested using a matrix containing:

A. Only C

B. B, AB, A²B, through Aⁿ⁻¹B

C. Only A and C

D. B, BA, BA², through BAⁿ⁻¹

Correct Answer: B. B, AB, A²B, through Aⁿ⁻¹B

Question 5

The estimation-error dynamics of a Luenberger observer are governed by:

A. A + LC

B. A − BK

C. A − LC

D. B − CA

Correct Answer: C. A − LC

Question 6

What is the main purpose of anti-windup protection?

A. To increase sensor noise

B. To prevent excessive integral accumulation during actuator saturation

C. To remove all plant uncertainty

D. To make the plant open loop

Correct Answer: B. To prevent excessive integral accumulation during actuator saturation

Question 7

In LQR, the matrix R primarily weights:

A. State estimation error

B. Reference amplitude

C. Control effort

D. Sensor bandwidth

Correct Answer: C. Control effort

Question 8

If the loop transfer function is L(s), the sensitivity function is:

A. 1 − L(s)

B. 1/[1 + L(s)]

C. L(s)²

D. 1/L(s)

Correct Answer: B. 1/[1 + L(s)]

Question 9

For P(s) = 1/(s + 3) and C(s) = 2 + 4/s, the characteristic polynomial is:

A. s² + 3s + 2

B. s² + 5s + 4

C. s² + 7s + 4

D. s² + 3s + 4

Correct Answer: C. s² + 7s + 4

Question 10

Why may excessively fast observer poles be undesirable?

A. They always make the plant uncontrollable

B. They can amplify measurement noise

C. They eliminate feedback

D. They guarantee actuator saturation

Correct Answer: B. They can amplify measurement noise