Fuzzy Logic and Knowledge-Based Control

Introduction

Conventional control design usually depends on a mathematical model and precise numerical measurements. Many mechatronic systems, however, are difficult to model accurately because of nonlinearities, friction, changing loads, uncertain parameters, sensor noise, or qualitative operating knowledge. An experienced operator may describe a suitable action as “increase the motor voltage slightly when the position error is small but still increasing” rather than by a differential equation. Fuzzy control provides a systematic way to represent and use such knowledge.

Fuzzy logic does not treat uncertainty as randomness. Instead, it represents gradual concepts such as slow, near, or high temperature using degrees of membership between 0 and 1. A fuzzy controller maps measured variables to control actions through linguistic rules, while retaining a connection with conventional feedback control.

This chapter develops the design of fuzzy sets, membership functions, rule bases, inference mechanisms, and defuzzification procedures. It compares Mamdani and Takagi–Sugeno controllers, examines interpretability and rule consistency, and applies the methods to a motor-positioning problem.

Learning Objectives

After studying this chapter, you should be able to:

1. Explain fuzzy sets, linguistic variables, membership functions, and fuzzy rules.
2. Design membership functions from engineering ranges and expert knowledge.
3. Construct and simplify a fuzzy rule base.
4. Apply Mamdani inference and common defuzzification methods.
5. Describe Takagi–Sugeno inference and calculate its output.
6. Implement normalization, scaling, saturation, and rule evaluation in a controller.
7. Evaluate interpretability, consistency, stability-related concerns, and limitations.
8. Select an appropriate fuzzy-control structure for a mechatronic application.

1. From Crisp Control to Fuzzy Control

A crisp set assigns an element either to a set or not to it. Its membership is conventionally 0 or 1. A fuzzy set allows partial membership. For a universe of discourse X, a fuzzy set A is described by a membership function:

Here, x is a value of the variable and μ_A(x) is the degree to which x belongs to concept A. The value 0 means no membership, 1 means full membership, and intermediate values represent gradual compatibility with the concept. Membership is not a probability: μ_A(x) = 0.7 does not mean that an event has a 70% chance of occurring.

For example, the linguistic variable position error may have the terms Negative Large (NL), Negative Small (NS), Zero (ZE), Positive Small (PS), and Positive Large (PL). A measured error can belong simultaneously to PS and PL, with different membership grades. This overlap is essential for smooth control.

A fuzzy controller normally contains:

• Fuzzification: conversion of numerical inputs into membership grades.
• Knowledge base: definitions of fuzzy sets and control rules.
• Inference engine: calculation of rule activation and output fuzzy sets or values.
• Defuzzification: conversion of the inferred result into a numerical control signal.
• Scaling and interface logic: conversion between physical units and the normalized controller domain.

The resulting controller is still a feedback controller. If r(t) is the reference and y(t) is the measured output, the error is

When error dynamics are relevant, a commonly used second input is the error rate

where T_s is the sampling period. This backward difference assumes sampled measurements and is an approximation to de/dt. Its units are error units per second.

2. Membership Functions and Fuzzy Sets

2.1 Design requirements

A membership-function design should reflect:

• the physical operating range;
• sensor resolution and noise;
• actuator limits;
• meaningful linguistic distinctions;
• desired smoothness of the control surface; and
• the amount and quality of available expert knowledge.

The membership functions should normally cover the admissible domain. Excessive gaps produce weak or zero rule activation, whereas excessive overlap can make linguistic distinctions unclear. The number of terms is a design compromise: three to seven terms per variable is common in practical controllers.

2.2 Common membership functions

A triangular membership function with left foot a, peak b, and right foot c is

for a < b < c. At x = b, μ(x) = 1. The function is piecewise linear and computationally inexpensive.

A trapezoidal function has a plateau of full membership. With feet a and d and plateau limits b and c, where a ≤ b ≤ c ≤ d,

with the rising or falling fraction omitted when the corresponding interval has zero width. Trapezoids are useful when a range of values is equally representative of a linguistic term.

A Gaussian membership function is

where m is the centre and σ > 0 controls spread. Gaussian functions are smooth and useful when a smooth control surface is important, but they require exponential evaluation and may be less transparent to operators than triangular functions.

A membership function should be dimensionless, although its argument x has physical units. Parameters such as a, b, c, m, and σ must use the same units as x.

2.3 Example of fuzzification

Suppose normalized position error e ranges from −1 to 1. Define a triangular Positive Small set with parameters (0, 0.5, 1). For e = 0.25,

Thus the error has membership 0.5 in PS. If it also has membership 0.2 in Zero, both terms may contribute to the control action. This overlapping representation avoids abrupt switching at a category boundary.

3. Rule Bases and Knowledge Representation

A fuzzy rule is commonly written:

IF e is PS AND Δe is NS THEN u is PM.

The antecedent describes a condition; the consequent describes the recommended action. Here e is error, Δe is error change, and u is actuator command. The linguistic labels must be defined for each variable.

A two-input rule base can be represented as a matrix. For example, rows may denote e and columns Δe. A rule table is not merely a collection of labels: each entry expresses a control policy. For a position controller, positive error generally requires positive actuator action, but the error trend affects how aggressively the controller should act. If positive error is already decreasing rapidly, a smaller command may prevent overshoot.

3.1 Rule activation

For a rule with antecedent memberships α₁, α₂, …, the AND operation is often implemented using the minimum or product operator:

or

The minimum operator has a direct linguistic interpretation: the rule is only as strong as its least satisfied condition. The product operator penalizes partial membership more strongly. The choice is a design convention, not a universal law.

For OR, common choices are maximum or a probabilistic sum. The maximum operator is

and the minimum operator for intersection is

These are the standard max–min fuzzy operations.

3.2 Rule-base quality

A useful rule base should have:

• coverage: relevant input combinations activate at least one rule;
• consistency: similar antecedents do not demand contradictory outputs without justification;
• completeness: important operating regions are represented;
• minimal redundancy: duplicate rules do not unnecessarily increase complexity;
• smoothness: small input changes should not cause unreasonable output jumps; and
• physical plausibility: rules should respect actuator direction, safety limits, and process behaviour.

Two rules are contradictory when their antecedents substantially overlap but their consequents recommend incompatible actions. Contradictions are not always errors; they may indicate competing objectives or a need for additional context. They should nevertheless be identified and documented.

A practical consistency audit compares neighbouring cells of a rule table. For a position loop, moving from negative to positive error should generally not cause an unexplained reversal in command. Monotonicity is not mandatory in every controller, but violations should be justified by dynamics such as damping, backlash, or gravity compensation.

4. Mamdani Fuzzy Inference

Mamdani inference uses fuzzy consequents. A typical rule is:

IF e is Ei AND Δe is Di THEN u is Ui.

For rule i, let wi be the antecedent firing strength. Under min implication, the rule output is clipped:

The outputs of all activated rules are aggregated using maximum:

The result is a fuzzy output set. It must be defuzzified to obtain a numerical actuator command.

4.1 Defuzzification methods

The most widely used method is the centroid, or centre of area:

The integrals extend across the output universe, and the denominator must be nonzero. The centroid often gives smooth outputs but can be influenced by distant low-level regions.

For discrete output samples uj,

The weighted-average approximation is valid when the output universe is discretized sufficiently finely.

Other methods include:

• Bisector: divides the aggregated area into two equal areas.
• Mean of maxima: averages all u values at which μ_U reaches its maximum.
• Smallest or largest of maxima: selects one extreme maximum location.

Centroid is usually smooth; mean of maxima can be simpler but may change abruptly when the set of maxima changes.

4.2 Worked Mamdani calculation

Assume two rules produce singleton-like output locations for illustration:

• Rule 1: firing strength w₁ = 0.4, representative output u₁ = 2 V.
• Rule 2: firing strength w₂ = 0.7, representative output u₂ = 5 V.

Using a weighted-average approximation,

This calculation is not the general Mamdani centroid for arbitrary overlapping output shapes; it is appropriate only when the output sets are represented by equal-shape singleton or narrow symmetric consequents. For arbitrary clipped triangular or trapezoidal sets, the aggregated membership function must be integrated or numerically sampled.

5. Takagi–Sugeno Inference

Takagi–Sugeno (TS) rules use numerical functions rather than fuzzy output sets. A zero-order rule is:

IF e is Ei AND Δe is Di THEN u = ci.

A first-order rule may be:

THEN u = ai e + bi Δe + ci.

The normalized weighted output is

where wi is the antecedent firing strength, fi is the consequent function, and N is the number of rules. The denominator must be positive. If no rule fires, the implementation should use a defined fallback, such as the previous valid command, a safe default, or a supervisory fault state.

For zero-order TS inference, fi = ci. If w₁ = 0.4, c₁ = 2 V, w₂ = 0.7, and c₂ = 5 V, the result is 4.36 V, identical to the preceding weighted-average calculation.

TS controllers are computationally efficient and naturally suited to optimization, adaptive control, and gain scheduling. First-order consequents can approximate nonlinear functions with local linear models. However, their linguistic interpretability is weaker because a consequent such as u = 1.4e − 0.3Δe + 2 is less immediately understandable than “u is Medium.”

6. Mamdani and Takagi–Sugeno Compared

Neither approach is universally superior. A safety-critical system may value transparent linguistic rules, while a high-speed embedded controller may favour TS inference. In both cases, scaling, saturation, sampling, validation, and failure handling are essential.

7. Controller Design Procedure

A disciplined design process is more reliable than choosing rules informally.

Step 1: Define the control objective

Specify the controlled variable, reference range, acceptable steady-state error, settling-time target, overshoot limit, actuator limits, and sampling rate. Determine whether the fuzzy controller generates an absolute command u or an increment Δu. Incremental control is often written:

where k is the sample index and sat[·] limits the command to the actuator range. Saturation is necessary because a physical actuator cannot deliver unlimited voltage, torque, flow, or force.

Step 2: Select inputs

Error alone may produce a proportional-like controller. Error and error change provide information about both position and trend. Additional inputs—velocity, load estimate, temperature, or operating mode—may improve performance but increase the number of rules. If each input has L linguistic terms and all combinations are included, a full rule base contains up to Lⁿ rules for n inputs. This exponential growth is a major motivation for hierarchical, sparse, or TS structures.

Step 3: Normalize variables

Map physical variables into a convenient domain, often [−1, 1]. For a variable x with symmetric scale X_max,

For asymmetric ranges, an affine transformation can be used. Scaling factors must be recorded and their units preserved. Poor scaling may cause most operating points to fall in one membership region, reducing effective resolution.

Step 4: Choose membership functions

Start with a small number of interpretable terms. Use measurement data, expert thresholds, operating limits, and simulation to position the functions. Do not tune parameters only for one trajectory; test different references, loads, disturbances, and initial conditions.

Step 5: Build and audit rules

Begin with physical reasoning. In a position loop, positive error normally implies a positive correcting command. Use the trend input to add damping: if positive error is decreasing rapidly, reduce the command or apply braking. Check coverage, conflicting consequents, redundant rules, and transitions between adjacent regions.

Step 6: Select inference and defuzzification

Mamdani with min–max inference and centroid defuzzification is a transparent baseline. TS weighted averaging is attractive for real-time implementation and parameter optimization. The choice should be justified by computational resources, explainability requirements, and the available knowledge.

Step 7: Validate progressively

Test first with a static rule-surface inspection, then with a plant simulation, hardware-in-the-loop, and finally controlled physical experiments. Examine tracking, overshoot, control effort, noise sensitivity, saturation, and behaviour under sensor or actuator faults.

8. Applied Case: Fuzzy Position Control of a DC Motor

The following is an illustrative engineering case, not a report of a particular commercial system.

Background and problem

A small DC motor drives a lightweight positioning mechanism. Its load varies because interchangeable fixtures are attached to the shaft. Coulomb friction and backlash make a single linear model inaccurate near zero speed. The controller must track position commands while limiting overshoot and avoiding excessive voltage changes.

Let e be position error in degrees and Δe be its sampled change. The motor driver accepts u from −12 V to +12 V. A controller based only on proportional gain performs acceptably for one fixture but becomes oscillatory for another.

Analysis

The error sign determines the direction of correction. The error trend indicates whether the mechanism is approaching the target. Therefore, a two-input controller is selected. The normalized variables are

and, assuming a design maximum of 10° per sample,

Both variables are limited to [−1, 1] before fuzzification. Five terms are used: NL, NS, ZE, PS, and PL. The output terms are NB, NS, ZE, PS, and PB, later mapped to the voltage range.

Approach and solution

A representative rule is:

IF e is PL AND Δe is ZE THEN u is PB.

This commands strong positive voltage when the shaft is far below the target and not yet approaching it. A second rule is:

IF e is PS AND Δe is PS THEN u is NS.

This applies braking when the shaft is slightly below target but moving further in the positive direction. The apparently counterintuitive negative command is a damping action, not a correction of the instantaneous error alone.

Mamdani min–max inference with centroid defuzzification is selected because maintenance personnel need to inspect the rules. The output is saturated at ±12 V. A rate limiter is also applied so that the voltage change per sample does not exceed the driver specification.

Outcome evaluation

The design is evaluated using step commands, repeated commands, load changes, and sensor noise. Important metrics include:

• maximum absolute error;
• settling time within a specified tolerance band;
• percentage overshoot;
• integral absolute error, defined for sampled data as

where K is the final sample index;

• peak and RMS actuator effort; and
• number and duration of saturation events.

The fuzzy controller is expected to reduce overshoot under the heavier load because the trend input activates braking rules. If testing reveals chatter near zero, membership overlap, filtering, dead zones, or a less aggressive output rule may be adjusted. Any filter introduces delay, so its effect on phase margin and transient response must be evaluated rather than assumed harmless.

Learning points

The case demonstrates that fuzzy rules can encode qualitative damping knowledge, but they do not eliminate the need for plant testing. Rule changes must be assessed against actuator constraints, sampling effects, noise, and all relevant load conditions. A visually plausible rule table is not evidence of stability or safety.

9. Interpretability, Stability, and Consistency

Interpretability has several dimensions. A controller is easier to interpret when its variables have physical meanings, membership labels are ordered and well separated, rules are few enough to inspect, and consequents correspond to understandable actions. A large automatically generated rule base may achieve low error while becoming impossible to audit.

Fuzzy control does not automatically guarantee closed-loop stability. Stability depends on the plant, sampling, nonlinearities, saturation, delays, and the complete control surface. Local analysis may approximate a fuzzy controller by a gain in a region, while Lyapunov-based or sector-based methods can provide stronger guarantees for particular structures. At bachelor level, the essential engineering practice is to combine simulation, bounded commands, robust testing, and—where required—formal stability analysis.

Consistency can be assessed through:

1. Coverage maps: identify regions where no rule fires or firing strength is too small.
2. Conflict matrices: compare rules with overlapping antecedents and different consequents.
3. Control-surface plots: inspect whether output changes smoothly with inputs.
4. Monotonicity checks: test expected directional relationships.
5. Sensitivity tests: perturb membership parameters and examine performance changes.

A rule base can be locally consistent but globally unsuitable. For example, each neighbouring rule may appear reasonable while the combined surface creates a steep ridge that causes actuator chatter.

10. Implementation, Safety, and Governance

A practical implementation should specify numerical precision, sampling time, lookup-table resolution, handling of missing sensors, output saturation, watchdog behaviour, and recovery after reset. Unit conversion and sign conventions should be tested independently. A common failure is reversing the error sign, turning negative feedback into positive feedback.

Fuzzy rules are knowledge artifacts. Their source should be recorded: experiment, expert interview, model analysis, or optimization. Changes should be version controlled and validated. In regulated or safety-related applications, the controller requires traceability, hazard analysis, independent testing, and defined fail-safe behaviour. Explainability is valuable, but linguistic transparency does not replace verification.

Fuzzy control can be combined with machine learning. Data-driven methods may tune membership parameters, discover rules, or identify TS consequent coefficients. This can improve performance but introduces risks of overfitting, distribution shift, hidden bias in operating data, and loss of interpretability. A hybrid design should preserve physical constraints and retain a safe fallback controller.

11. Limitations and Contemporary Directions

Fuzzy controllers have several limitations:

• rule count grows rapidly with the number of inputs;
• membership functions and rules may be subjective;
• performance can degrade outside the design domain;
• formal stability analysis may be difficult;
• tuning can be time-consuming; and
• a fuzzy controller cannot compensate for an inadequately designed sensor, actuator, or mechanical system.

Current directions include adaptive neuro-fuzzy inference systems, interval and type-2 fuzzy sets for higher-order uncertainty, sparse rule bases, online parameter adaptation, event-triggered implementation, and TS models for nonlinear control and fault diagnosis. These methods are useful when their additional complexity is justified. The basic design principles remain unchanged: define meaningful variables, ensure coverage, preserve physical constraints, validate across operating conditions, and document the knowledge representation.

Best-Practice Checklist

Before deployment, verify that:

• every input and output has a stated physical unit and range;
• normalization and sign conventions are documented;
• membership functions cover the intended domain;
• the rule base has been checked for gaps and contradictions;
• inference and defuzzification choices are explicit;
• zero-rule activation has a safe response;
• saturation and rate limits match the actuator;
• noise, delay, sampling, and quantization have been tested;
• performance is measured under disturbances and parameter changes; and
• software, rules, tuning data, and validation results are traceable.

Chapter Summary

Fuzzy logic represents gradual linguistic concepts through membership functions rather than binary set membership. A fuzzy controller fuzzifies numerical measurements, evaluates a knowledge-based rule base, combines the activated rules, and produces a numerical command. Triangular, trapezoidal, and Gaussian membership functions offer different compromises among interpretability, smoothness, and computational cost.

Mamdani controllers use fuzzy output sets and require aggregation and defuzzification, commonly by the centroid method. Takagi–Sugeno controllers use constant or functional consequents and calculate a normalized weighted output. Mamdani methods are often highly interpretable, while TS methods are efficient and well suited to optimization and nonlinear approximation.

Effective design requires meaningful variables, careful scaling, adequate rule coverage, consistency checks, actuator constraints, and progressive validation. Fuzzy logic can encode useful human and engineering knowledge, but it does not by itself guarantee stability, safety, or performance outside the design domain.

Key Terms

Fuzzy set: A set whose membership grades range continuously from 0 to 1.

Membership function: A function assigning a membership grade to each value of a variable.

Linguistic variable: A variable described by qualitative terms such as Small or High.

Fuzzification: Conversion of a crisp numerical input into membership grades.

Rule base: Collection of fuzzy IF–THEN control rules.

Antecedent: The IF portion of a fuzzy rule.

Consequent: The THEN portion of a fuzzy rule.

Firing strength: Degree to which a rule’s antecedent is satisfied.

T-norm: An operation, such as minimum or product, representing fuzzy AND.

Aggregation: Combination of outputs from multiple activated rules.

Mamdani inference: Fuzzy inference with fuzzy-set consequents.

Takagi–Sugeno inference: Inference with constant or functional numerical consequents.

Defuzzification: Conversion of an output fuzzy set into a crisp value.

Centroid: Defuzzification method based on the centre of the aggregated area.

Normalization: Scaling a physical variable to a convenient numerical domain.

Rule coverage: Extent to which the rule base addresses the input domain.

Rule consistency: Absence of unjustified contradictory outputs for overlapping antecedents.

Control surface: Mapping from controller inputs to its output.

Saturation: Limitation of a command to physical actuator bounds.

TS consequent: The constant or mathematical function in a Takagi–Sugeno rule.

Review Questions

Short Answer Questions

1. What is the difference between a fuzzy membership grade and a probability?
2. Why are overlapping membership functions useful in a controller?
3. State the main computational difference between Mamdani and Takagi–Sugeno inference.
4. What is the purpose of output saturation?
5. Name two methods for auditing rule-base consistency.

Long Answer Questions

1. Explain the complete signal flow of a fuzzy feedback controller, from sensor measurement to actuator command.
2. Compare triangular, trapezoidal, and Gaussian membership functions, including practical advantages and disadvantages.
3. Derive and explain centroid defuzzification for a continuous aggregated output set.
4. Discuss how error and error change can be used together to provide both correction and damping.
5. Explain why a fuzzy controller does not automatically guarantee closed-loop stability.

Analytical / Application Questions

1. Design a two-input rule base for temperature control using temperature error and error change. State your linguistic terms, membership-function strategy, and at least nine rules.
2. Two TS rules have firing strengths 0.25 and 0.75, with constant consequents 4 V and 8 V. Calculate the controller output and explain the assumption required for the weighted average.
3. A motor controller exhibits oscillation near zero error. Analyse at least four possible fuzzy-design and implementation causes, and propose a testable corrective action for each.

Multiple Choice Questions with Answers

Question 1

Which statement best describes a fuzzy membership grade?

A. The probability that an event will occur

B. The degree to which a value belongs to a linguistic concept

C. The percentage error of a sensor

D. The gain of a feedback controller

Correct Answer: B. The degree to which a value belongs to a linguistic concept

Question 2

For a triangular membership function with parameters a = 0, b = 2, and c = 4, what is μ(1)?

A. 0

B. 0.25

C. 0.5

D. 1

Correct Answer: C. 0.5

Question 3

Under min-based fuzzy AND, antecedent memberships 0.6 and 0.8 produce which firing strength?

A. 0.14

B. 0.48

C. 0.6

D. 1.4

Correct Answer: C. 0.6

Question 4

Which defuzzification method uses the first moment of the aggregated membership area?

A. Centroid

B. Smallest of maxima

C. Mean of maxima

D. Random selection

Correct Answer: A. Centroid

Question 5

A zero-order Takagi–Sugeno controller has firing strengths 0.25 and 0.75 and consequents 4 V and 8 V. What is its output?

A. 5 V

B. 6 V

C. 7 V

D. 8 V

Correct Answer: C. 7 V

Question 6

Which feature is characteristic of a Mamdani controller?

A. Every consequent must be a differential equation

B. Consequents are fuzzy output sets

C. No defuzzification is ever needed

D. It cannot use linguistic rules

Correct Answer: B. Consequents are fuzzy output sets

Question 7

What is a major reason for using error change as a fuzzy-controller input?

A. It measures actuator voltage directly

B. It provides information about the direction and rate of approach

C. It removes the need for feedback

D. It guarantees stability

Correct Answer: B. It provides information about the direction and rate of approach

Question 8

If no rule fires in a Takagi–Sugeno controller, what issue must the implementation address?

A. The output denominator becomes zero

B. The sampling period becomes infinite

C. All membership grades become probabilities

D. The actuator automatically becomes linear

Correct Answer: A. The output denominator becomes zero

Question 9

Which practice most directly helps detect contradictory fuzzy rules?

A. Increasing sensor noise

B. Removing all overlap between membership functions

C. Comparing consequents of rules with overlapping antecedents

D. Reducing the actuator voltage rating

Correct Answer: C. Comparing consequents of rules with overlapping antecedents

Question 10

Why can a large number of controller inputs be problematic?

A. It always eliminates nonlinear behaviour

B. The number of full rule combinations can grow exponentially

C. It prevents normalization

D. It makes all membership grades equal to zero

Correct Answer: B. The number of full rule combinations can grow exponentially