7. Feedback, Digital and Intelligent Control

Introduction and Context

Mechatronic systems combine mechanical structures, sensors, actuators, power electronics, embedded computation, and control algorithms. Their performance is therefore determined not only by mechanical design but also by how effectively information is measured, processed, and converted into actuator action. Control engineering provides the framework for achieving accuracy, stability, disturbance rejection, rapid response, safety, and adaptability.

In an open-loop system, the actuator command is generated without measuring the resulting output. Such systems are simple but cannot automatically compensate for disturbances, parameter variations, friction, load changes, or sensor uncertainty. Feedback control closes the loop by comparing a measured output with a desired reference. The resulting error is used to modify the control input. Digital control extends this principle through sampling and computation, while intelligent control incorporates adaptation, prediction, learning, fuzzy reasoning, or optimization.

This chapter develops the principal control architectures used in advanced mechatronics: PID, feedforward, cascade, state-space, optimal, adaptive, robust, and model-predictive control. Particular attention is given to discretisation, stability, controllability, observability, estimation, implementation constraints, and the engineering trade-offs involved in selecting a controller.

Learning Objectives

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

1. Explain the role of feedback in mechatronic systems and distinguish feedback from feedforward control.

2. Design and tune PID, cascade, and feedforward controllers for practical systems.

3. Convert continuous-time models into discrete-time representations and assess sampling effects.

4. Analyse stability using transfer-function and state-space methods.

5. Determine controllability and observability and explain their significance.

6. Formulate state feedback, observer, optimal, robust, adaptive, and model-predictive controllers.

7. Evaluate the influence of sensors, actuators, computation, quantisation, delays, saturation, and noise.

8. Select an appropriate control strategy for an industrial mechatronic application.

Feedback as an Engineering Decision

A basic single-loop feedback system consists of a reference signal, comparator, controller, plant, sensor, and feedback path. For a unity-feedback system with plant transfer function \(G(s)\) and controller \(C(s)\), the closed-loop transfer function is

\[

T(s)=\frac{C(s)G(s)}{1+C(s)G(s)}.

\]

The denominator \(1+C(s)G(s)\) is the characteristic equation. Its roots, or closed-loop poles, determine stability and strongly influence rise time, overshoot, settling time, and oscillation. Feedback generally improves disturbance rejection and reduces sensitivity to plant uncertainty, but it may also amplify measurement noise, introduce instability, and demand greater sensor and computational capability.

The sensitivity function is

\[

S(s)=\frac{1}{1+C(s)G(s)},

\]

while the complementary sensitivity function is \(T(s)=1-S(s)\). High loop gain tends to reduce sensitivity to plant disturbances and parameter variations over the frequency range in which the loop gain is large. However, high gain cannot be increased indefinitely because of unmodelled dynamics, actuator limitations, noise, and delay.

Important terminology includes set point, process variable, manipulated variable, disturbance, tracking error, loop gain, bandwidth, phase margin, gain margin, steady-state error, and actuator saturation. These terms connect mathematical performance with physical operation.

PID Control: The Industrial Baseline

The continuous PID law is

\[

u(t)=K_p e(t)+K_i\int_0^t e(\tau)d\tau+K_d\frac{de(t)}{dt},

\]

where \(e(t)=r(t)-y(t)\). The proportional term responds to present error, the integral term removes persistent offset, and the derivative term anticipates error variation. In transfer-function form,

\[

C(s)=K_p+\frac{K_i}{s}+K_d s.

\]

Proportional control is simple and responsive, but a finite proportional gain commonly leaves steady-state error. Integral action eliminates offset for constant references and disturbances, yet can cause sluggishness, oscillation, and integrator wind-up when the actuator saturates. Derivative action can improve damping and transient response, but ideal differentiation amplifies high-frequency noise.

A practical derivative term is filtered:

\[

C_D(s)=\frac{K_d s}{1+s/N},

\]

where \(N\) is a filter factor. Derivative action is usually applied to the measured output rather than directly to the error when set-point steps would otherwise produce a derivative kick.

PID tuning may be based on a process model, frequency response, pole placement, or experimental methods. A useful engineering sequence is to establish safe actuator limits, obtain a nominal model, tune proportional action, introduce integral action carefully, and then add filtered derivative action only when improved damping is required. Tuning must be validated under realistic load changes rather than only under a nominal simulation.

Anti-windup is essential in systems with actuator limits. Common methods include conditional integration, integrator clamping, and back-calculation. In back-calculation, the difference between the demanded and actual actuator signal is fed back to reset the integrator. Bumpless transfer is also required when switching between manual and automatic operation.

Feedforward, Cascade, and Ratio Structures

Feedback reacts after an error has appeared. Feedforward control uses a measured disturbance or reference and a model of the plant to act before the disturbance affects the output. If a disturbance enters through \(G_d(s)\) and the control path is \(G_u(s)\), an ideal feedforward controller is approximately

\[

F(s)=-\frac{G_d(s)}{G_u(s)}.

\]

Exact cancellation is rarely possible because models are imperfect, delays may be non-minimum-phase, and disturbances may not be measurable. Consequently, feedforward is normally combined with feedback. Feedforward supplies rapid compensation; feedback corrects residual error.

Cascade control contains an outer loop and one or more inner loops. The inner loop must be faster than the outer loop and should regulate a variable affected directly by disturbances or actuator dynamics. For example, a servo drive may use a position loop around a velocity loop, which surrounds a current loop. The inner current loop rapidly controls motor torque, the velocity loop shapes motion, and the position loop ensures tracking. Cascade control improves disturbance rejection and reduces the burden on the outer controller, but it requires reliable intermediate measurements and appropriate bandwidth separation.

Ratio control maintains a specified relationship between two process variables, while override control selects among controllers to protect equipment or enforce constraints. These architectures are especially important in coordinated motion, energy systems, and process-mechatronic equipment.

Digital Control and Discretisation

A digital controller operates at sampling instants \(kT_s\), where \(T_s\) is the sampling period. The continuous signal is sampled, converted to a numerical value, processed, and converted back through a digital-to-analogue converter or a digital actuator interface. The sampling frequency must be sufficiently high to represent the relevant dynamics. The Nyquist principle requires a sampling frequency greater than twice the highest frequency component, but practical control design normally uses a considerably higher rate than the closed-loop bandwidth.

Sampling introduces delay, quantisation, computation time, and the possibility of aliasing. An anti-aliasing filter is placed before the analogue-to-digital converter. Quantisation produces a finite-resolution measurement and may create limit cycles in high-gain systems.

A continuous state model is

\[

\dot{x}=Ax+Bu, \qquad y=Cx+Du.

\]

With zero-order-hold input, its exact discrete form is

\[

x_{k+1}=A_d x_k+B_d u_k,

\]

where

\[

A_d=e^{AT_s}, \qquad B_d=\int_0^{T_s}e^{A\tau}B\,d\tau.

\]

Approximate methods include forward Euler, backward Euler, and the bilinear or Tustin transformation. Forward Euler is simple but can distort stability and frequency characteristics. Tustin mapping generally gives better frequency correspondence, although pre-warping may be required near an important frequency.

A digital PID controller can be implemented using positional or incremental form. The incremental form is often less vulnerable to numerical overflow and is convenient for embedded systems. Every implementation should include output saturation, anti-windup, sensor filtering, timing supervision, and safe behaviour when measurements or communications fail.

Stability and Performance Analysis

A continuous-time linear system is asymptotically stable when all closed-loop poles have negative real parts. In discrete time, stability requires all poles to lie inside the unit circle. Marginal poles on the boundary require special interpretation and are generally undesirable in practical systems.

The Routh–Hurwitz criterion assesses continuous polynomial stability without explicitly calculating roots. Nyquist analysis examines encirclements of the critical point and is particularly useful when delays and frequency-domain uncertainty are important. Bode plots provide gain margin, phase margin, bandwidth, and resonance information. Adequate margins indicate tolerance to modelling error, although no single margin guarantees robust performance.

Time-domain specifications include rise time, peak overshoot, settling time, steady-state error, and disturbance-rejection time. A fast response is not automatically a good response: excessive bandwidth may excite structural resonances, amplify sensor noise, or exceed actuator capacity. Control design is therefore a compromise among speed, accuracy, robustness, energy consumption, and safety.

State-Space Control, Controllability, and Observability

Transfer functions describe input–output behaviour, whereas state-space models represent internal dynamic variables. A state-space model is

\[

\dot{x}=Ax+Bu,\qquad y=Cx+Du.

\]

A state-feedback law has the form

\[

u=-Kx+Nr,

\]

where \(K\) is selected to place closed-loop poles or optimise a performance criterion. State feedback can shape multivariable dynamics more systematically than independent PID loops, but it requires the states to be measured or estimated.

The pair \((A,B) is controllable if the controllability matrix

\[

\mathcal{C}=[B\ AB\ A^2B\ \cdots\ A^{n-1}B]

\]

has rank \(n\). Controllability means that the actuator can influence every dynamic mode. The pair \((A,C) is observable if

\[

\mathcal{O}=\begin{bmatrix}C\\CA\\\vdots\\CA^{n-1}\end{bmatrix}

\]

has rank \(n\). Observability means that internal states can be reconstructed from measured outputs. A system may be physically controllable but poorly controllable numerically, or observable in theory but too noisy for reliable estimation.

An observer estimates states using model and measurement information. A Luenberger observer is

\[

\hat{x}_{k+1}=A_d\hat{x}_k+B_du_k+L(y_k-C\hat{x}_k),

\]

where \(L\) is the observer gain. The separation principle permits controller and observer poles to be designed independently for linear systems under suitable assumptions. Excessively fast observer poles, however, amplify noise and create sensitivity to model error.

The Kalman filter provides a statistically optimal state estimate for linear systems with known stochastic models, Gaussian disturbances, and Gaussian measurement noise. Extended and unscented Kalman filters address nonlinear systems approximately. In practice, covariance tuning is a modelling task, not merely a software setting.

Optimal Control

Optimal control chooses inputs by minimising a performance index. The linear–quadratic regulator minimises

\[

J=\sum_{k=0}^{\infty}(x_k^TQx_k+u_k^TRu_k),

\]

where \(Q\) penalises state deviations and \(R\) penalises control effort. The resulting law is \(u=-Kx\), with \(K obtained from the algebraic Riccati equation. Increasing state penalties generally produces more aggressive correction; increasing input penalties produces smoother and less energy-intensive control.

Linear–quadratic–Gaussian control combines LQR with a Kalman filter. Optimal methods provide a principled treatment of competing objectives, but the result depends strongly on model quality, weighting matrices, actuator limits, and state availability. A mathematically optimal controller may be unacceptable if it violates constraints or produces excessive actuator activity.

Adaptive and Robust Control

Adaptive control modifies controller parameters online as the plant or environment changes. In gain scheduling, different controllers are selected according to operating conditions. In model-reference adaptive control, parameters are adjusted so that the plant follows a desired reference model. Self-tuning regulators identify a model online and redesign the controller.

Adaptation creates risks. Poor excitation can make parameter estimates unreliable; noisy measurements can cause parameter drift; unmodelled dynamics may be misinterpreted as parameter changes. Projection, dead zones, parameter bounds, monitoring, and safe fallback controllers are therefore important.

Robust control assumes that the plant belongs to an uncertainty set and seeks acceptable performance for all or most models in that set. H-infinity design, for example, limits the worst-case gain from disturbances to regulated outputs. Structured singular-value methods address structured uncertainty. Robust control is valuable when uncertainty is significant, but it may be conservative and mathematically demanding.

Adaptive and robust control address different problems. Adaptation uses observed behaviour to change the controller, whereas robust control seeks a controller that tolerates uncertainty without changing its structure. Hybrid strategies are increasingly used in systems that require both guaranteed protection and limited online adaptation.

Model-Predictive Control and Intelligent Methods

Model-predictive control (MPC) uses a model to predict future outputs over a finite horizon. At each sampling instant, it solves an optimisation problem subject to constraints, applies only the first control move, and repeats the process after receiving new measurements. A typical objective is

\[

J=\sum_{i=1}^{N_p}\|y_{k+i}-r_{k+i}\|_Q^2+\sum_{i=0}^{N_c-1}\|\Delta u_{k+i}\|_R^2.

\]

MPC naturally handles actuator limits, rate constraints, multiple variables, and future reference changes. Its limitations include computational demand, dependence on the prediction model, tuning complexity, and sensitivity to poorly represented disturbances. Modern embedded processors have made MPC increasingly feasible for motion systems, energy management, thermal control, and autonomous machines.

Fuzzy control represents knowledge through linguistic rules such as “if position error is large and velocity is small, apply a strong correction.” Neural and reinforcement-learning controllers can approximate nonlinear policies or learn from data. These methods may be useful where accurate first-principles models are unavailable, but verification, explainability, data quality, and safety remain major concerns. In safety-critical mechatronics, learning components are commonly supervised by conventional constraints, monitors, or fallback controllers rather than being given unrestricted authority.

Applied Case: Precision Motion Stage

Background

Consider an illustrative precision linear stage used to position a cutting tool. The stage consists of a ball screw, servo motor, encoder, drive amplifier, and embedded controller. The required position accuracy is high, while the load varies as the tool engages different materials.

Problem

Initial operation with a single position PID loop produces acceptable low-speed accuracy but exhibits vibration near a structural resonance at higher speed. When the tool load changes, tracking error increases. During emergency stopping, actuator saturation causes integral wind-up and delayed recovery.

Analysis

The plant contains nested electrical, mechanical, and structural dynamics. A single outer loop cannot reject motor torque disturbances rapidly without exciting the resonance. Encoder quantisation becomes significant near the target position. The system is controllable through motor torque, but the structural mode is only weakly observable from the available encoder location.

Approach

A cascaded architecture is adopted. The innermost current loop regulates motor torque, the velocity loop is tuned with a higher bandwidth than the position loop, and the position controller uses proportional–integral action with a filtered velocity feedback term. A measured acceleration signal is used cautiously for notch filtering and vibration suppression. Feedforward terms based on desired velocity and acceleration reduce tracking error during programmed motion. Anti-windup is implemented in the position loop, and the controller enters a controlled safe state when encoder communication is lost.

A discrete model is identified from frequency-response tests. The sampling frequency is selected well above the position-loop bandwidth, and computation delay is included in the model. An observer estimates velocity where direct differentiation of encoder position would amplify noise. Robustness is checked by varying inertia, friction, delay, and resonance frequency in simulation and hardware tests.

Outcome

The cascaded design improves disturbance rejection because torque and velocity disturbances are corrected before they accumulate into position error. Feedforward reduces predictable tracking error, while feedback retains robustness to load variation. Filtering and bandwidth separation reduce resonance excitation. The final design is accepted only after tests of saturation, sensor faults, communication delay, emergency stop, and restart behaviour.

Learning Points

The case illustrates that controller selection must follow plant structure. A sophisticated algorithm cannot compensate for inadequate sensing, poor sampling, or unmodelled resonance. Feedforward and feedback are complementary, cascade loops must be coordinated by bandwidth, and practical safety functions are part of control design rather than later additions.

Implementation Risks and Professional Practice

Common mistakes include tuning a controller without identifying actuator limits, ignoring transport delay, differentiating noisy signals directly, selecting a sampling rate from processor speed alone, and validating only with nominal simulations. Other failures arise from incorrect sensor polarity, unit inconsistency, hidden saturation, poorly scaled state variables, and inadequate reset behaviour after faults.

Control engineers should document assumptions, operating ranges, sensor calibration, model uncertainty, software versions, parameter changes, and test evidence. Safety-related functions require hazard analysis, independent protective mechanisms, access control, and traceable verification. Cybersecurity is also relevant because networked controllers can be affected by unauthorised parameter changes, timing disruption, or falsified measurements.

A sound implementation sequence is: define performance and safety requirements; identify the plant and uncertainty; choose sensors and actuators; select architecture; design continuous and discrete controllers; simulate nonlinear and constrained behaviour; test hardware-in-the-loop; commission with conservative limits; and monitor performance in service. Controller performance should be evaluated using both tracking and disturbance tests, including worst-case operating conditions.

Current Developments

Contemporary mechatronic control increasingly combines physics-based models with data-driven estimation. Digital twins support commissioning and predictive maintenance, while edge computing enables more complex optimisation near the machine. Distributed control systems coordinate multiple actuators through industrial networks, making synchronisation and communication delay important design variables. Learning-based tuning, safe reinforcement learning, disturbance observers, and economic MPC are active development areas.

The central trend is not replacement of classical control by artificial intelligence. Rather, practical systems combine dependable feedback loops with higher-level estimation, optimisation, adaptation, and diagnostics. The most successful designs preserve interpretable safety boundaries while using computational intelligence where it provides measurable value.

Best-Practice Design Principles

• Begin with a physically meaningful model and state the uncertainty explicitly.
• Use feedback for correction and feedforward for predictable disturbances.
• Separate cascade-loop bandwidths and verify sensor quality at every level.
• Design sampling, filtering, computation delay, and quantisation as part of the controller.
• Check controllability, observability, actuator authority, and sensor placement before advanced design.
• Include saturation handling, anti-windup, fault detection, safe fallback, and bumpless transfer.
• Validate robustness through parameter variation, disturbances, delays, and hardware testing.
• Prefer the simplest controller that meets performance, safety, and maintainability requirements.

Chapter Summary

Feedback control compares measured output with desired behaviour and uses the error to regulate a plant. PID remains widely used because its proportional, integral, and derivative actions provide an effective balance of simplicity and performance, but practical implementations require filtering, anti-windup, and attention to saturation. Feedforward improves response to measurable disturbances, while cascade control uses nested loops to isolate fast dynamics and improve disturbance rejection.

Digital control introduces sampling, quantisation, delay, and discretisation issues. State-space methods support multivariable control, pole placement, observers, Kalman filtering, and systematic tests of controllability and observability. Optimal control formalises trade-offs between performance and effort. Adaptive control changes parameters in response to plant variation, whereas robust control protects performance against uncertainty. MPC explicitly handles predictions and constraints, while fuzzy, neural, and learning methods can address nonlinear or poorly modelled behaviour but require strong verification and safety measures.

Effective mechatronic control is ultimately a systems-engineering activity. Algorithm choice must be consistent with sensors, actuators, computation, safety requirements, uncertainty, and the physical dynamics of the machine.

Key Terms

Feedback control: Control that uses measured output to modify the input.

Feedforward control: Control based on a reference or measured disturbance before an error develops.

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

Integral wind-up: Excessive accumulation of integral action during actuator saturation.

Cascade control: Nested control architecture with inner and outer loops.

Sampling period: Time interval between successive digital measurements.

Aliasing: Misrepresentation of high-frequency signals caused by insufficient sampling.

Quantisation: Conversion of continuous measurement values into finite digital levels.

Closed-loop stability: Condition in which system responses remain bounded and converge appropriately.

Controllability: Ability of actuators to influence all dynamic states.

Observability: Ability to reconstruct internal states from measured outputs.

State feedback: Control law based on measured or estimated system states.

Observer: Dynamic estimator that reconstructs unmeasured states.

Kalman filter: State estimator based on a stochastic model and noise statistics.

Optimal control: Control obtained by minimising a specified performance index.

Adaptive control: Control that modifies its parameters as plant behaviour changes.

Robust control: Control designed to maintain performance despite model uncertainty.

Model-predictive control: Receding-horizon optimisation using a predictive plant model.

Bandwidth: Frequency range over which a control system responds effectively.

Phase margin: Frequency-domain measure of additional phase lag tolerated before instability.

Review Questions

Short Answer Questions

1. Why is feedback generally more robust to plant uncertainty than open-loop control?

2. What are the principal functions of the proportional, integral, and derivative terms in a PID controller?

3. Why is anti-windup required in a practical PID implementation?

4. Define controllability and observability.

5. What is the purpose of an anti-aliasing filter in a digital control system?

Long Answer Questions

1. Explain the operation, tuning considerations, and practical limitations of a PID controller.

2. Compare feedback, feedforward, and cascade control using suitable mechatronic examples.

3. Derive the discrete state-space model of a continuous system under zero-order-hold sampling and discuss sampling effects.

4. Explain state feedback, observer design, controllability, and observability.

5. Compare optimal, adaptive, robust, and model-predictive control with respect to assumptions, strengths, and limitations.

Analytical / Application Questions

1. A motor-position system has large load variation and actuator saturation. Propose a control architecture and justify the roles of cascade control, feedforward, and anti-windup.

2. A sampled-data controller becomes oscillatory after its sampling period is increased. Analyse possible causes involving discrete poles, delay, aliasing, and bandwidth.

3. Design a validation plan for an MPC or adaptive controller used in a safety-relevant robotic positioning system, including modelling, fault testing, fallback control, and performance evidence.

22. Multiple Choice Questions with Answers

Question 1

Which action in a PID controller is primarily responsible for eliminating steady-state error caused by a constant disturbance?

A. Proportional action

B. Integral action

C. Derivative action

D. Filtering action

Correct Answer: B. Integral action

Question 2

What is the principal purpose of feedforward control?

A. To amplify sensor noise

B. To correct predictable disturbances before they produce an error

C. To replace all feedback measurements

D. To guarantee stability without a plant model

Correct Answer: B. To correct predictable disturbances before they produce an error

Question 3

For a discrete-time linear system, asymptotic stability requires the closed-loop poles to be located

A. In the right half of the s-plane

B. On the imaginary axis

C. Inside the unit circle

D. Outside the unit circle

Correct Answer: C. Inside the unit circle

Question 4

A system is controllable when

A. Every state can be influenced by the available inputs

B. Every state is directly measured by a sensor

C. The output is independent of the input

D. The system has no disturbances

Correct Answer: A. Every state can be influenced by the available inputs

Question 5

Why is derivative action usually filtered in an industrial PID controller?

A. To remove integral wind-up

B. To reduce amplification of high-frequency measurement noise

C. To increase actuator saturation

D. To eliminate the need for feedback

Correct Answer: B. To reduce amplification of high-frequency measurement noise

Question 6

In a cascade control system, the inner loop should generally be

A. Slower than the outer loop

B. Faster than the outer loop

C. Independent of all sensors

D. Implemented without feedback

Correct Answer: B. Faster than the outer loop

Question 7

What does an observer estimate?

A. Only the actuator rating

B. Unmeasured internal system states

C. The sampling frequency alone

D. The controller’s software cost

Correct Answer: B. Unmeasured internal system states

Question 8

Which feature is a distinctive strength of model-predictive control?

A. It cannot include constraints

B. It operates without a model

C. It explicitly predicts future behaviour and handles constraints

D. It requires no computation during operation

Correct Answer: C. It explicitly predicts future behaviour and handles constraints

Question 9

A high observer bandwidth may be undesirable because it can

A. Eliminate controllability

B. Amplify measurement noise and model error

C. Prevent state estimation entirely

D. Remove actuator saturation

Correct Answer: B. Amplify measurement noise and model error

Question 10

Which statement best distinguishes robust control from adaptive control?

A. Robust control changes parameters continuously, whereas adaptive control never changes them

B. Robust control tolerates specified uncertainty, whereas adaptive control modifies parameters using observed behaviour

C. Robust control is only for open-loop systems

D. Adaptive control cannot be used with feedback

Correct Answer: B. Robust control tolerates specified uncertainty, whereas adaptive control modifies parameters using observed behaviour