How Feedback Control Works

Feedback control works by measuring what a system is doing, comparing that measurement with what the system should be doing, and using the difference to change the input. The mechanism is simple enough to describe in one sentence, but powerful enough to stabilise aircraft, regulate power converters, control industrial reactors, position robotic joints, maintain room temperature, and coordinate many other engineered systems.

The key idea is closed-loop correction. A controller does not merely issue a command and hope the plant obeys. It observes the result of previous commands, detects error, and modifies future commands. This turns measurement into action.

The feedback loop

A standard feedback loop contains five main pieces:

1. Reference: the desired value or trajectory, often written as r(t).
2. Plant: the physical system to be controlled.
3. Sensor: the device or estimator that provides the measured output.
4. Controller: the algorithm, circuit, or mechanism that computes the control action.
5. Actuator: the component that applies the controller command to the plant.

The measured output is commonly written as y(t). The difference between the reference and the measured output is the error:

The controller receives e(t) and produces a command u(t). The actuator converts that command into physical influence: torque, force, voltage, valve opening, heating power, pump speed, fin angle, brake pressure, or another input.

If the output is too low, the error is positive and the controller may increase the input. If the output is too high, the error is negative and the controller may reduce the input. In negative feedback, the loop acts in the direction that reduces error. This sign convention is essential. If the loop acts in the direction that increases error, it becomes positive feedback and can run away or oscillate.

A thermostat as a minimal example

A thermostat illustrates feedback without heavy mathematics. The reference is the desired room temperature. The sensor measures the actual room temperature. The controller compares the two. If the room is colder than desired, it turns heating on. If the room is warm enough, it turns heating off or reduces heating power.

The thermostat does not need to know exactly why the room cooled. A door may have opened, outside temperature may have dropped, or sunlight may have disappeared. Feedback responds to the measured consequence: the room temperature moved away from the setpoint.

More advanced temperature controllers use proportional or PID action instead of simple on-off switching, but the principle is the same: measure, compare, correct, repeat.

The linear closed-loop equation

For a linear time-invariant single-input single-output system, the basic loop can be written with a plant transfer function G(s) and a controller transfer function C(s). The loop transfer function is:

For unity negative feedback, the closed-loop transfer function from reference to output is:

This equation shows why feedback can make the output follow the reference. If |L(s)| is large at a frequency where the loop is stable, then:

The output approximately follows the reference at that frequency. However, this approximation is not free. Large loop gain can also reduce stability margin, amplify noise through some paths, excite neglected dynamics, or demand more actuator effort.

The sensitivity function is:

It describes how certain disturbances and modelling errors are transmitted to the output. Where |L(s)| is large, |S(s)| is small, so feedback reduces sensitivity. Where |L(s)| is small, feedback has little authority.

These equations are not a complete control design, but they reveal the central tradeoff: control is achieved by shaping the loop transfer function.

How feedback rejects disturbances

Consider a motor speed controller. The reference is a desired speed. A load torque disturbance tries to slow the motor. Without feedback, the motor speed drops according to the motor and load characteristics. With feedback, the sensor detects the speed drop, the error increases, and the controller commands more drive torque.

The controller does not need to measure the disturbance directly. It reacts to the output error caused by the disturbance. This is one of the main advantages of feedback over pure feedforward control.

However, disturbance rejection is frequency-dependent. A controller may reject slow load changes well but respond poorly to high-frequency disturbances. The plant, actuator, sensor, and controller all have bandwidth limits. A valve cannot move infinitely fast. A motor drive has current limits. A temperature sensor may have thermal lag. A digital controller samples at finite intervals. Feedback can only correct what it can measure and influence in time.

Why feedback improves accuracy

Feedback improves accuracy by continually correcting residual error. Suppose the plant gain is uncertain. A nominal open-loop command may produce the wrong output because the actual plant differs from the model. In a feedback system, the output error remains visible to the controller. The controller adjusts the command until the error is reduced.

Integral control is especially important for steady-state accuracy. A proportional controller produces an output command proportional to the current error:

If the plant requires a nonzero input to maintain the desired output, proportional control alone may leave a steady-state error. Integral control accumulates error:

As long as error persists, the integral term changes. In many stable loops, this drives steady-state error to zero for step references or step disturbances. The cost is that integral action adds phase lag and can worsen overshoot or stability if tuned aggressively.

Why feedback can destabilise a system

Feedback is corrective only if the correction arrives with the right sign and timing. If delay, phase lag, or unmodelled dynamics make the corrective action arrive too late, the controller may push the system in a direction that reinforces oscillation.

A simple way to see this is through sinusoidal response. At a given frequency, the loop has a gain and a phase shift. If the loop phase approaches -180^\circ while the loop gain is still at or above one, negative feedback effectively becomes positive feedback at that frequency. The result may be sustained oscillation or instability.

This is why control engineers care about gain margin and phase margin. These margins measure how much gain increase or phase lag the loop can tolerate before crossing the boundary of instability. They do not guarantee good performance by themselves, but they are practical indicators of robustness for many classical designs.

The role of the controller

The controller shapes the loop response. In a PID controller:

Each term has a distinct effect:

• Proportional action responds immediately to present error.
• Integral action accumulates past error and can eliminate steady-state offset.
• Derivative action responds to the rate of change and can improve damping.

Derivative action is usually filtered because differentiating noisy measurements can produce large command noise. In many industrial controllers, the derivative term is applied to the measured output rather than the error to avoid a derivative kick when the setpoint changes.

PID control is common because it is simple, interpretable, and effective for many processes. It is not universal. Systems with strong constraints, multiple interacting variables, unstable dynamics, flexible modes, long delays, or severe nonlinearities may require lead-lag compensation, state feedback, observers, model predictive control, robust control, or nonlinear methods.

Sensor and actuator realities

The ideal feedback diagram hides several practical limits.

A sensor measures an output with finite accuracy, bandwidth, resolution, and delay. If the sensor is noisy, high controller gains may convert measurement noise into actuator motion. If the sensor is slow, the controller reacts to old information. If the sensor is placed far from the physical variable of interest, transport delay or structural dynamics may reduce usable bandwidth.

An actuator also has limits. It can saturate, move only so fast, heat up, wear out, or respond with its own dynamics. Saturation breaks the assumptions of linear control. When an actuator saturates, increasing the controller command no longer increases the actual plant input. If integral action continues to accumulate during saturation, the controller may overshoot badly after the actuator comes out of saturation. Anti-windup protection is used to reduce this problem.

The loop must therefore be designed around the real sensor and actuator, not an ideal block diagram.

Feedback and noise

Feedback uses measurements, so it must deal with measurement noise. Strong feedback at low frequencies is often desirable because it improves tracking and disturbance rejection. Strong feedback at high frequencies is often undesirable because sensors may be noisy and the plant model may be inaccurate.

The complementary sensitivity function,

describes how some measurement noise paths can affect the output in a unity-feedback loop. Designers often seek high loop gain at low frequency and low loop gain at high frequency. This is one reason Bode plots are so useful: they show how gain and phase vary with frequency.

Filtering can reduce noise, but filtering adds phase lag. Excessive filtering can make the controller less stable or slower to respond. The filter is part of the control system and must be included in the analysis.

Feedback and feedforward together

Feedback corrects error after it appears in the measured output. Feedforward acts before the error appears, using a model or measured disturbance. The two methods are complementary.

For example, in a temperature control system, a feedforward term may increase heating power when a known cold inflow begins. Feedback then corrects the remaining error caused by model mismatch, heat losses, or sensor uncertainty.

In motion control, feedforward may compute the torque needed for a desired acceleration, while feedback corrects position and velocity errors. In process control, feedforward may compensate for measured changes in inlet flow or composition, while feedback maintains the controlled output.

Good feedforward reduces the burden on feedback. Good feedback makes the system tolerant of feedforward errors.

Discrete-time feedback

Most modern controllers are digital. A sensor is sampled, the controller computes a command, and the actuator holds or updates that command. Digital feedback introduces new issues:

• sampling must be fast enough relative to the closed-loop bandwidth;
• anti-aliasing may be needed before sampling;
• computation and communication add delay;
• numerical precision and scaling matter;
• discrete-time poles must remain inside the unit circle;
• the continuous design may need careful discretisation.

A rough engineering rule is to sample significantly faster than the desired closed-loop bandwidth. The exact requirement depends on phase margin, delay, plant dynamics, noise, and implementation details. Sampling is not only a software choice; it changes the loop.

Commissioning and loop tuning evidence

Feedback loops should be commissioned with evidence, not only adjusted until the response looks acceptable. Useful checks include controller sign, sensor scaling, actuator direction, saturation limits, manual-to-auto transfer, setpoint steps, disturbance response, alarm behavior, and recovery after power or communication loss.

Loop tuning should state the operating condition used for the test. A controller tuned at low load may oscillate at high load. A loop tuned with a clean sensor may become noisy after installation. A valve, motor, pump, heater, or fin may have deadband, rate limits, backlash, or thermal constraints that are invisible in a simple transfer-function sketch.

Deployed feedback evidence should preserve gains, filters, sampling time, operating mode, known constraints, and reason for change. Without that record, later adjustments can remove stability margin while trying to solve a local performance complaint.

What makes feedback successful

A successful feedback controller usually has these properties:

• the controlled output is measurable or accurately estimable;
• the actuator has enough authority and bandwidth;
• the feedback sign is correct;
• the closed-loop system has adequate stability margin;
• low-frequency loop gain is high enough for tracking and disturbance rejection;
• high-frequency loop gain is low enough to avoid noise amplification and unmodelled dynamics;
• saturation, rate limits, delay, and sensor filtering are included in the design;
• the controller is tested against plausible uncertainty and faults.

Feedback control is powerful because it turns error into correction. It is risky when the correction is delayed, mis-signed, too aggressive, or based on poor measurements. The engineering task is to keep the corrective mechanism strong where it helps and restrained where it would create instability, noise, or unsafe actuator demands.