Feedback Control and Networked Control Systems

Introduction and Context

Cyber-physical systems (CPS) combine physical processes, sensors, actuators, computation, and communication networks. Feedback control is the mechanism that connects these elements into an operating system capable of correcting errors and rejecting disturbances. In a conventional control loop, a sensor measures a physical variable, a controller compares the measurement with a desired reference, and an actuator changes the plant. In a networked control system (NCS), these signals may travel through wired or wireless networks, creating delays, packet losses, bandwidth limits, quantization, and cybersecurity risks.

For mechatronics engineers, control design is therefore no longer limited to selecting a controller for a fixed plant. The engineer must evaluate the interaction between dynamics, sampling, computation, communication, and resource constraints. A controller that is stable in a laboratory with direct wiring may become oscillatory or unsafe when implemented over a congested industrial network.

This chapter develops the foundations needed to design and evaluate feedback, state estimation, digital, distributed, and networked control systems.

Learning Objectives

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

1. Explain closed-loop feedback and distinguish tracking, regulation, disturbance rejection, and feedforward control.
2. Analyze stability and performance using transfer functions and state-space models.
3. Design basic proportional-integral-derivative (PID) and discrete-time controllers.
4. Construct and interpret observers and Kalman-filter-based state estimators.
5. Explain distributed and networked control architectures.
6. Evaluate the effects of sampling, delay, packet loss, quantization, and limited communication resources.
7. Select practical mitigation strategies and assess safety, cybersecurity, and governance issues.

1. Feedback Control Foundations

A plant is the physical process to be controlled, such as a motor, robot joint, thermal chamber, or autonomous vehicle. The reference signal, or setpoint, specifies the desired value. The measured output is compared with the reference to produce the error:

Here, r(t) is the reference, y_m(t) is the measured output, and e(t) is the control error. The controller computes a command u(t), which is applied through an actuator to the plant. Measurement noise, disturbances, and actuator limitations influence the resulting output.

For a linear time-invariant single-input, single-output system with plant transfer function G(s) and controller C(s), unity negative feedback gives the closed-loop transfer function:

The loop transfer function is L(s) = C(s)G(s). The denominator 1 + L(s) is the characteristic equation. Closed-loop poles are the roots of this equation; their locations determine stability and dynamic behavior. This expression assumes linear models, zero initial conditions, unity feedback, and a well-defined Laplace-domain representation.

The sensitivity function describes the effect of plant or disturbance perturbations:

The complementary sensitivity function is:

For unity feedback, S(s) + T(s) = 1. High loop gain generally reduces sensitivity to plant disturbances and modeling errors at frequencies where |L(jω)| is large, but excessive gain can amplify measurement noise and reduce stability margins.

Regulation, Tracking, and Disturbance Rejection

A tracking controller makes y(t) follow a changing reference. A regulator maintains a desired operating point despite disturbances. For example, a robot position controller performs tracking, whereas a temperature controller maintaining 80 °C against changing ambient conditions performs regulation.

Feedforward control uses a model or measured disturbance to act before an error appears. Feedback corrects residual error and modeling uncertainty. A robust design often combines both. Feedforward alone cannot correct unmeasured disturbances or model inaccuracies; feedback alone may react too slowly to large predictable disturbances.

Stability and Performance

For continuous-time linear systems, asymptotic stability requires all closed-loop poles to have negative real parts. Poles close to the imaginary axis produce slow responses, while complex poles can produce oscillation. Common performance measures include rise time, settling time, overshoot, steady-state error, bandwidth, gain margin, and phase margin.

For a standard underdamped second-order model,

ω_n is the natural frequency in rad/s and ζ is the damping ratio. For 0 < ζ < 1, the response is underdamped. An often-used approximation for the 2% settling time is:

This approximation applies to a dominant second-order response with a sufficiently small contribution from other poles. It is not a universal formula for higher-order or strongly nonminimum-phase systems.

2. PID Control and Practical Design

The continuous-time PID controller is:

where K_p is proportional gain, K_i is integral gain, and K_d is derivative gain. The proportional term responds to present error, the integral term removes persistent error, and the derivative term anticipates changes in error.

In the Laplace domain, under zero initial conditions:

The ideal derivative term amplifies high-frequency measurement noise. Practical controllers therefore use a filtered derivative, for example:

where τ_f is a positive filter time constant. Integral action can cause windup when the actuator saturates. Anti-windup methods include conditional integration, integrator clamping, and back-calculation.

A practical tuning procedure is to:

1. Identify an approximate plant model or perform a safe step test.
2. Select a proportional gain that gives adequate responsiveness without excessive oscillation.
3. Add integral action to reduce steady-state error.
4. Add filtered derivative action only when transient prediction is beneficial.
5. Test saturation, disturbances, sensor noise, and parameter variation.
6. Validate the design on the actual controller hardware and network.

A controller must be evaluated against actuator limits. A command outside the permitted range is not physically realizable, even if the mathematical simulation appears satisfactory.

3. State-Space Models and State Estimation

Transfer functions are convenient for single-loop analysis, but mechatronic systems often have multiple inputs and outputs. State-space models represent internal variables explicitly:

Here, x is the state vector, u is the control input, y is the measured output, w]] is process disturbance, and v is measurement noise. Matrices A, B, C, and D describe system dynamics and have dimensions determined by the numbers of states, inputs, and outputs. The matrices G describes how process noise enters.

A system is controllable if its state can be driven to the desired state by appropriate inputs. For an n-state linear system, the controllability matrix is:

The system is controllable if rank(𝒞) = n. It is observable if its internal state can be reconstructed from output measurements. The observability matrix is:

Observability requires rank(𝒪) = n. These rank tests apply to finite-dimensional linear time-invariant models. Numerical conditioning should also be considered in real implementations.

Observers

If all states are not measured, an observer estimates them. A full-order Luenberger observer has the form:

where x̂ is the estimated state and L is the observer gain. The estimation error e_x = x − x̂ evolves according to:

Ignoring disturbances and noise, the observer error converges to zero if A − LC is stable. Observer poles are commonly selected faster than controller poles, but excessively fast observers amplify measurement noise and may demand unrealistic computation or actuator behavior.

Kalman Filtering

The Kalman filter is a model-based estimator for systems with stochastic process and measurement noise. In discrete time:

The noises are commonly modeled as zero-mean, uncorrelated Gaussian processes with covariance matrices Q and R. The prediction step is:

The measurement update is:

The Kalman gain K_k balances model prediction against measurement information. These equations assume the stated linear model and known covariance properties. Nonlinear systems require extensions such as the extended or unscented Kalman filter, each with additional assumptions and possible failure modes.

4. Digital Control

Digital controllers operate at discrete sampling instants. With sampling period T_s seconds, the continuous signal is represented by x_k = x(kT_s). A zero-order hold maintains the actuator command between updates.

For a continuous state model with constant input over one sampling interval:

where, exactly,

The integral expression remains valid even when A is singular. The common approximation A_d ≈ I + AT_s is accurate only when the sampling period is sufficiently small relative to the system dynamics.

A simple discrete PID implementation uses backward differences:

This equation is an implementation form, not a unique digital PID algorithm. The derivative term should generally be filtered, and the integral sum should be protected against windup. Sampling frequency must be high enough to capture relevant dynamics, but unnecessarily high sampling increases computation, communication, and energy demands. The Nyquist criterion states that a sampling frequency must exceed twice the highest frequency component for ideal reconstruction, but control design typically requires additional margin because of phase delay, noise, and nonideal hold behavior.

Quantization introduces finite resolution. If an analog-to-digital converter has full-scale range V_FS and N bits, its ideal step size is approximately:

Actual accuracy also depends on offset, gain error, nonlinearity, noise, and reference quality.

5. Distributed and Networked Control Systems

A distributed control system divides sensing, computation, and actuation among multiple nodes. A networked control system specifically uses a communication network in one or more feedback paths. Typical architectures include:

• Centralized control: one controller receives measurements and sends commands. It is conceptually simple but creates a central bottleneck and single point of failure.
• Decentralized control: local controllers operate largely independently. It can be robust to communication failure but may neglect important coupling.
• Distributed control: local controllers exchange information with neighboring nodes and coordinate their actions.
• Hierarchical control: supervisory optimization operates above faster local loops.
• Cloud or edge-assisted control: computation is placed in edge devices or remote servers, subject to latency, availability, and security constraints.

A useful abstraction for a networked loop is:

where the controller function κ uses current or estimated information. The network may alter the timing and availability of y_k and u_k.

Delay

A fixed communication delay of τ_d seconds contributes phase lag. For a sinusoidal component at angular frequency ω, the delay transfer function is:

and its frequency response has magnitude 1 and phase −ωτ_d. Thus delay does not attenuate an ideal signal, but it reduces phase margin. Variable delay, or jitter, is often more difficult because the controller cannot rely on a fixed timing relationship.

Mitigation strategies include reducing network load, prioritizing control traffic, time synchronization, timestamping, buffering, prediction, and redesigning the controller for the known delay. A predictor can be effective when the model is reliable, but model mismatch can produce large errors.

Packet Loss and Intermittent Measurements

Packet loss creates missing measurements or commands. Common responses include holding the last value, using a model-based prediction, switching to a safe local controller, or declaring a fault after a specified timeout. Holding the last control command is simple but can be dangerous for unstable plants or actuators that must return to a safe state.

If each packet is independently lost with probability p, the probability that a packet is successfully received is 1 − p. The probability of m consecutive losses under this independent model is:

This calculation does not apply directly to bursty or correlated losses, which require a model such as a Markov chain or measured loss statistics.

Resource Constraints

Networked controllers compete for bandwidth, energy, processor time, and memory. Resource-aware design may use event-triggered sampling, where a transmission occurs only when an error or state change exceeds a threshold. Periodic sampling is predictable and easier to certify; event-triggered sampling can reduce traffic but complicates timing analysis and may create undesirable transmission bursts.

Other techniques include message compression, quantization, scheduling, prioritization, local autonomy, and graceful degradation. Safety-critical commands should not be delayed behind noncritical traffic. Time-sensitive networking can provide bounded latency and synchronized communication when appropriately engineered and configured.

6. Networked Stability and Performance Evaluation

A networked controller should be evaluated at three levels:

1. Physical performance: tracking error, overshoot, energy use, vibration, and settling time.
2. Network performance: latency, jitter, packet-loss rate, throughput, queue occupancy, and synchronization error.
3. Safety and resilience: behavior during faults, attacks, sensor failures, and loss of communication.

The closed-loop characteristic equation must include communication effects. For example, a delayed loop may be represented conceptually by:

This is a transcendental equation with infinitely many characteristic roots, unlike an ordinary finite-order polynomial. Therefore, a controller verified without delay cannot automatically be assumed stable with delay. Approximation methods, frequency-domain analysis, Lyapunov–Krasovskii methods, or numerical delay-system tools may be used depending on the required rigor.

Robustness means maintaining acceptable behavior under uncertainty. A useful practical test is to vary plant parameters, delay, packet-loss patterns, sampling period, and sensor noise in simulation and hardware-in-the-loop testing. Monte Carlo testing can reveal rare combinations of disturbances, although it does not prove safety.

7. Worked Example: Digital Speed Control with Communication Delay

Consider a simplified motor speed plant:

The plant time constant is 0.5 s. Suppose a proportional controller with gain K_p = 4 is used under unity feedback.

The closed-loop transfer function is:

The closed-loop pole is at s = −10 rad/s, so the ideal first-order time constant is 0.1 s. The steady-state output for a unit step reference is:

Thus the steady-state error is 0.2, or 20 percent. Integral action would be needed to remove this error for a constant reference, provided the actuator has sufficient authority and anti-windup protection is included.

Now suppose the controller is sampled every T_s = 0.02 s. The sampling frequency is 50 Hz, while the closed-loop pole magnitude is 10 rad/s, approximately 1.59 Hz. The sampling rate is therefore substantially higher than the dominant closed-loop frequency, although implementation and delay must still be tested.

If the network adds a constant delay of 0.05 s, the phase lag at angular frequency ω = 10 rad/s is:

This delay consumes phase margin and may cause overshoot or instability in a more aggressive controller. The correct engineering response is not simply to assume failure or safety; it is to calculate or measure the actual stability margin, simulate the delayed loop, and test worst-case timing.

8. Illustrative Case: Networked Conveyor and Robotic Sorter

Background

Consider an illustrative industrial cell containing a conveyor, an encoder, a vision sensor, and a robotic arm. The conveyor speed is controlled locally, while object coordinates are sent from a vision computer to the robot over an industrial Ethernet network. A supervisory system also changes production recipes.

Problem

During peak production, network traffic increases. Vision packets arrive with variable latency between 20 and 80 ms, and occasional bursts of three consecutive packets are lost. The robot sometimes reaches a pick location too early or too late. A direct increase in controller gain improves response in a laboratory test but causes oscillatory behavior on the production network.

Analysis

The system contains two different control problems. Conveyor speed is a fast local loop and should not depend on a remote supervisory link. Object localization is a slower, event-based coordination problem. Treating both as one centralized loop unnecessarily exposes the fast loop to network delay.

Measurements are timestamped at acquisition. The robot estimates the object position using conveyor speed and elapsed time. If the conveyor velocity is approximately constant over the prediction interval, the predicted position is:

where p₀ is the measured position at timestamp t₀, and v is estimated conveyor velocity. The approximation must be revised when acceleration is significant.

Approach and Solution

The engineering team adopts the following architecture:

1. A local drive controller closes the high-speed conveyor loop.
2. Vision messages contain object position, velocity estimate, timestamp, confidence, and sequence number.
3. The robot predicts object position using time-stamped data and rejects stale messages.
4. A watchdog switches the robot to a safe state when communication exceeds a specified timeout.
5. Critical control traffic receives priority over recipe and diagnostic traffic.
6. The system is tested with measured delay distributions, burst losses, sensor faults, and emergency-stop conditions.
7. Authentication, network segmentation, access control, and event logging are applied to reduce cyber risk.

Outcome

The architecture prevents the supervisory network from destabilizing the conveyor loop. Timestamping allows the robot to compensate for ordinary latency, while confidence thresholds prevent low-quality vision data from commanding a pick. The safe-state policy handles prolonged communication loss rather than indefinitely holding an old command.

Learning Points

The case demonstrates that networked control is an architectural problem as well as a controller-tuning problem. Local autonomy, explicit timing information, bounded delays, fault handling, and security controls are more reliable than simply increasing feedback gain.

9. Safety, Cybersecurity, and Governance

A CPS can cause physical harm when software, communication, or control logic fails. Safety engineering should define hazards, safe states, operating limits, fault-detection requirements, and independent protective functions. A communication failure must be treated according to the plant hazard: stopping immediately may be safe for one machine but harmful for another if it causes a suspended load to drop.

Cybersecurity and control performance are interdependent. An attacker who delays, reorders, suppresses, or falsifies measurements can create a control failure without directly changing the controller software. Good practice includes least-privilege access, authenticated messages where appropriate, network segmentation, secure updates, credential management, logging, anomaly detection, and recovery procedures. Security mechanisms must be assessed for their latency and computational overhead.

Engineers should preserve traceability: requirements, model assumptions, calibration data, software versions, test results, and change approvals should be documented. Artificial intelligence may assist with anomaly detection, predictive maintenance, or system identification, but an AI model should not be assumed safe merely because it performs well on historical data. Distribution shift, sensor faults, adversarial inputs, and lack of explainability require independent safeguards and human oversight for consequential decisions.

10. Best Practices and Contemporary Directions

Effective feedback and networked control design commonly follows these principles:

• Model the plant, sensors, actuators, computation, and network together.
• Separate fast safety-critical loops from slower supervisory communication.
• Specify timing bounds, not only average latency.
• Include packet loss, jitter, saturation, quantization, and sensor faults in testing.
• Use timestamps, sequence numbers, watchdogs, and explicit stale-data policies.
• Design observers with realistic noise and model uncertainty.
• Verify controllability and observability before attempting advanced control.
• Use hardware-in-the-loop and fault-injection tests before deployment.
• Apply defense-in-depth cybersecurity and maintain secure recovery procedures.
• Monitor both physical performance and network quality after commissioning.

Contemporary developments include time-sensitive networking, wireless industrial control, edge computing, digital twins, distributed model predictive control, resilient estimation, and learning-enabled control. These technologies can improve flexibility and efficiency, but they also increase verification complexity. A sophisticated method is valuable only when its assumptions, failure modes, computational demand, and safety implications are understood.

Chapter Summary

Feedback control compares desired and measured behavior and uses the resulting error to regulate a physical process. Closed-loop stability depends on the characteristic equation, while performance involves speed, accuracy, robustness, noise sensitivity, and actuator limits. PID control remains widely useful, but derivative filtering and anti-windup are essential in practical systems.

State-space models support multivariable analysis, controllability, observability, and state estimation. Luenberger observers and Kalman filters reconstruct unavailable states, with performance determined by model quality and noise assumptions. Digital control introduces sampling, zero-order hold, quantization, computation, and discrete-time implementation issues.

Networked and distributed control systems add delay, jitter, packet loss, bandwidth constraints, and cybersecurity threats. These effects must be included in stability and performance analysis. Local autonomy, timestamping, prediction, prioritization, watchdogs, safe-state behavior, and robust testing are central design practices. A successful CPS design integrates control theory with communication engineering, software assurance, safety engineering, and governance.

Key Terms

Feedback control: Control that uses measured output to correct error.

Reference: Desired value or trajectory for a controlled variable.

Tracking: Following a time-varying reference.

Regulation: Maintaining a desired condition despite disturbances.

Feedforward control: Control action based on a model or measured disturbance before error appears.

Sensitivity function: S(s) = 1/[1+L(s)], describing closed-loop sensitivity to disturbances and uncertainty.

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

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

State-space model: Representation using internal states, inputs, and outputs.

Controllability: Ability to move system states through suitable inputs.

Observability: Ability to reconstruct internal states from outputs.

Observer: Dynamic estimator of unmeasured system states.

Kalman filter: Recursive stochastic state estimator using model and measurement covariance.

Sampling period: Time between successive digital control updates.

Quantization: Representation of a continuous-valued signal using discrete levels.

Networked control system: Control system whose feedback path uses a communication network.

Jitter: Variation in communication or computation timing.

Packet loss: Failure of a transmitted data packet to reach its destination or be accepted.

Event-triggered control: Communication or control updates initiated when a specified condition occurs.

Watchdog: Mechanism that detects missing or late operation and initiates a defined response.

Review Questions

Short Answer Questions

1. What is the difference between tracking and regulation?
2. Why can integral action cause windup?
3. State the rank conditions for controllability and observability of an n-state linear system.
4. What effect does a pure communication delay have on the phase of a sinusoidal signal?
5. Why are timestamps useful in networked control systems?

Long Answer Questions

1. Explain the operation of a PID controller and discuss derivative filtering and anti-windup.
2. Compare centralized, decentralized, distributed, and hierarchical control architectures.
3. Derive the discrete-time state model under a zero-order hold and explain the meaning of A_d and B_d.
4. Explain the prediction and measurement-update stages of a discrete Kalman filter.
5. Discuss how packet loss, jitter, quantization, and actuator saturation can jointly affect closed-loop performance.

Analytical / Application Questions

1. A plant is controlled by a proportional controller with K_p = 9. For unity feedback and a constant unit reference, calculate the ideal steady-state output and error. Explain how integral action would change the result.
2. A network introduces a delay of 30 ms. Calculate the phase lag at ω = 20 rad/s and discuss why the result matters for stability margins.
3. Design a fault-response policy for a remotely controlled robotic actuator that experiences intermittent packet loss. Justify your choices for buffering, prediction, timeout, safe state, and event logging.

Multiple Choice Questions with Answers

Question 1

For unity negative feedback, which expression is the closed-loop transfer function?

A. C(s)G(s)/(1+C(s)G(s))

B. C(s)+G(s)

C. 1/(C(s)G(s))

D. C(s)G(s)/(1−C(s)G(s))

Correct Answer: A. C(s)G(s)/(1+C(s)G(s))

Question 2

What is the primary purpose of integral action in a PID controller?

A. To filter all measurement noise

B. To remove persistent steady-state error

C. To reduce actuator saturation automatically

D. To increase sampling frequency

Correct Answer: B. To remove persistent steady-state error

Question 3

A linear n-state system is controllable when which condition holds?

A. The output matrix is diagonal

B. The plant has no disturbances

C. The controllability matrix has rank n

D. All inputs are quantized

Correct Answer: C. The controllability matrix has rank n

Question 4

What is the principal effect of a pure communication delay on an ideal frequency response?

A. It increases magnitude and adds positive phase

B. It leaves magnitude unchanged and adds negative phase

C. It eliminates all measurement noise

D. It makes the plant memoryless

Correct Answer: B. It leaves magnitude unchanged and adds negative phase

Question 5

If each packet is independently lost with probability 0.1, what is the probability of three consecutive losses?

A. 0.001

B. 0.01

C. 0.1

D. 0.3

Correct Answer: A. 0.001

Question 6

Which Kalman-filter quantity determines the relative weighting of a new measurement and the model prediction?

A. The sampling period only

B. The Kalman gain

C. The actuator saturation limit

D. The derivative time constant

Correct Answer: B. The Kalman gain

Question 7

For an ideal N-bit converter with full-scale range V_FS, the approximate quantization step is:

A. NV_FS

B. V_FS/N

C. V_FS/2^N

D. 2^N/V_FS

Correct Answer: C. V_FS/2^N

Question 8

Which architecture keeps a fast conveyor-speed loop least dependent on supervisory network delay?

A. Closing the speed loop entirely through a remote cloud

B. Using a local speed controller and a slower supervisory link

C. Increasing network packet size without changing control location

D. Removing feedback from the drive

Correct Answer: B. Using a local speed controller and a slower supervisory link

Question 9

For the worked motor example with K_p = 4, what is the ideal unit-step steady-state error?

A. 0

B. 0.04

C. 0.2

D. 0.8

Correct Answer: C. 0.2

Question 10

Why is a watchdog used in a networked actuator system?

A. To increase sensor resolution

B. To detect missing or late communication and initiate a defined response

C. To replace all feedback sensors

D. To guarantee zero packet loss

Correct Answer: B. To detect missing or late communication and initiate a defined response