Fuzzy Modeling, Identification & Control

A fuzzy inference system is a structured nonlinear map assembled from overlapping local rules. In control, the inputs are often error, error rate, measured state, or an operating-condition variable; the output may be an actuator command, a setpoint correction, or parameters for another controller. The apparent readability of the rules is valuable, but the closed-loop system still depends on scaling, sampling, saturation, delay, and plant dynamics.

The inference pipeline #

flowchart TD
  X[Measured crisp inputs] --> F[Fuzzification]
  F --> R[Rule firing strengths]
  K[Rule base and membership definitions] --> R
  R --> A[Consequent evaluation and aggregation]
  A --> D[Defuzzification or weighted output]
  D --> U[Bounded control action]
  U --> P[Plant]
  P --> X

For each input, fuzzification computes membership in the relevant terms. Rule antecedents combine those memberships into firing strengths $w_i$. Consequents are evaluated, rule outputs are aggregated, and a numerical command is produced. Input and output normalization live outside the abstract fuzzy rule system but often dominate actual behavior.

Mamdani and Takagi-Sugeno answer different questions #

In a zero-order Takagi-Sugeno model, each rule returns a constant $c_i$ and the output is

$$y(x)=\frac{\sum_{i=1}^{M} w_i(x)c_i}{\sum_{i=1}^{M}w_i(x)}.$$

With first-order consequents $f_i(x)=a_i^\top x+b_i$, replace $c_i$ by $f_i(x)$. The normalized firing strengths form a partition-like blend of local models. The denominator must remain nonzero throughout the declared input domain.

Mamdani defuzzification is part of the controller #

After rule consequents are clipped or scaled and aggregated into $\mu_U(u)$, centroid defuzzification computes

$$u^*=\frac{\int_U u\,\mu_U(u)\,du}{\int_U \mu_U(u)\,du}.$$

Changing to mean-of-maxima or another method changes the control surface. Discretization resolution, empty-output behavior, and output saturation also matter. They must be specified rather than left to a toolbox default.

Worked control surface #

Consider a tank-level controller with normalized level error $e\in[-1,1]$ and error rate $\dot e\in[-1,1]$. Use three terms on each input: negative, zero, positive. A small rule base might encode:

The table is not yet a controller. The designer must define the sign convention, physical scaling, membership overlap, the meaning of “any,” output limits, update period, and what happens when the sensor is invalid. A sign error can create positive feedback while every individual rule still looks plausible.

Identification turns rules into an empirical model #

Fuzzy identification has two coupled problems:

1. Structure identification: choose inputs, delays, rule regions, antecedent terms, and consequent form.
2. Parameter identification: fit membership locations and widths plus consequent parameters.

For a dynamic plant, candidate regressors may include

$$\varphi_k=[y_{k-1},\ldots,y_{k-n_y},u_{k-1},\ldots,u_{k-n_u}].$$

A Takagi-Sugeno model then blends local predictions from regions of this regressor space. The experiment must excite every region in which the model will be used. Randomly splitting adjacent samples is weak evidence because neighboring points share the same trajectory; validation should hold out complete runs, operating regions, disturbances, and rollout horizons.

Rule explosion and sparse coverage #

With five terms on four inputs, a full grid contains $5^4=625$ possible antecedents. Many are physically impossible or unsupported by data. Pruning them improves tractability but can leave holes where no rule fires. Practical designs therefore measure:

• minimum total firing strength over the operating envelope;
• rule occupancy in estimation and validation data;
• sensitivity to membership perturbations;
• discontinuity or excessive slope near partition boundaries;
• extrapolation behavior outside the identified domain.

Closed-loop validation is mandatory #

A low one-step prediction error does not imply good control. The model may accumulate rollout error, hide delay, or be accurate only under the controller that generated the data. A fuzzy controller may also create limit cycles through quantization, sample delay, or steep rule transitions.

Validate at least:

• reference tracking and disturbance rejection across operating regions;
• sensor noise, bias, dropout, and stale values;
• actuator saturation, rate limits, deadband, and hysteresis;
• timing jitter and worst-case inference execution time;
• mode entry and exit transients;
• the exact behavior when no rule or conflicting rules dominate.

When the representation earns its cost #

Fuzzy control is a strong candidate when operators possess useful qualitative rules, the plant is nonlinear across recognizable regions, and an inspectable smooth mapping is preferable to hard switching. It is a weak choice when a simple PID or scheduled linear controller already meets requirements, when training data do not cover the envelope, or when certification demands properties that the chosen fuzzy construction cannot demonstrate.