Comparing a fuzzy logic controller against a PI controller is one of the most common final year projects in drives, and most versions of it prove nothing, because the PI they compare against was never tuned properly. This project does it honestly: the same BLDC plant, the same speed command, the same load disturbance, a PI controller tuned from the plant model rather than by eye, and a fuzzy controller whose complete rule base is published in the report. Both are judged on the same numbers — rise time, overshoot, settling time, steady-state error and integral of time-weighted absolute error — and the conclusion follows the numbers wherever they go.
Not just the parts. Teams budget from a component list, then run out in week six on wire, connectors and the driver board they destroyed. These are the lines that catch them.
| Line | Cost |
|---|---|
Components and materials From the bill of materials | Rs 0 |
Report, printing and binding Two or three bound copies, plots and the plagiarism check | Rs 400 to Rs 1,200 |
| Total | Rs 400 to Rs 1,200 |
| Each, split 3 ways | Rs 133 to Rs 400 |
Indicative bands last checked in August 2026 — a guide, not a quotation. Prices move and vary by seller, and anything you can borrow from the lab comes straight off the total. The 3 week figure assumes 7 person-weeks of work shared between 3; students past 4 add coordination rather than output, so the estimate stops improving there.
The same chain in words
The plant is the same for both controllers — same inertia, same friction, same torque constant, same inverter model — so the only variable is the controller itself.
The PI controller is tuned from the plant's own parameters rather than by trial and error, and the tuning method is stated so the result can be reproduced.
The fuzzy controller takes two inputs: the speed error and its rate of change. Each is divided into membership functions, and a rule table maps every combination onto a change in the control output.
Fuzzification converts the crisp inputs into membership grades, the rules are evaluated, and defuzzification converts the resulting fuzzy set back into a single number.
Both controllers face the same test sequence: a speed step from rest, a load torque applied mid-run, and a speed reversal. The metric block computes the same figures for each.
This is the chapter that separates an engineering project from an assembly job. Every value below carries the formula that produced it and the assumptions it rests on — check them against your own ratings before you use them, because your motor is not this motor.
error and rate of change of error, each in seven sets
Seven sets each — negative big through zero to positive big — gives a 7 by 7 rule table. Fewer sets is coarser and easier to tune; more sets rarely helps and quickly becomes impossible to justify rule by rule.
adjacent triangular sets should overlap at about half height
Too little overlap leaves input regions where no rule fires strongly and the output jumps. Too much and the controller becomes sluggish because everything fires at once.
centroid of the aggregated output set
Centroid is smooth and is what almost every published design uses. Mean of maximum is faster to compute but produces a discontinuous output, which on a motor drive shows up as torque steps.
gains derived from the plant's mechanical time constant
This is the line that makes the comparison honest. A fuzzy controller beating an arbitrarily chosen PI proves only that the PI was bad. State the tuning method, the target damping and the resulting gains.
ITAE = integral of t x |e(t)| dt
ITAE weights late error more heavily than early error, so it rewards fast settling without rewarding an aggressive initial response that overshoots. Quoting it alongside rise time and overshoot prevents cherry-picking whichever measure flatters your preferred controller.
the same current limit applied to both controllers
If one controller is allowed more current than the other, it will look faster, and the comparison is meaningless. Saturate both identically and say so.
| Item | Qty | Unit cost |
|---|---|---|
MATLAB with Simulink and Fuzzy Logic Toolbox College or student licence Use the institutional licence — check the Fuzzy Logic Toolbox is included before starting, as it is a separate product. | 1 | — |
Computer Any machine that runs the simulation in reasonable time Your own laptop or a lab machine — nothing to buy. | 1 | — |
Specifications are written as ratings rather than brands, so any equivalent part works. Where a line says to borrow from the lab, do — it is usually the largest number on the page, and the total drops to Rs 400 at the low end without them.
None of these blocks you — they are learnable inside a semester. Each one you do not have costs a couple of weeks, so count them into your plan.
Anything marked required is a blocker, not a difficulty. Arrange the access before you commit to the project, not after.
Use these when your guide says the scope is too small, or when you want something in the results chapter that nobody else in your batch will have.
The structure below is what a project report is marked against, with the mistake that costs marks in each chapter. There is no template to download and there will not be one: the writing is your work, and a site that did it for you would be helping you cheat rather than helping you pass.
A mini project report runs to about 40 to 50 pages, and the literature survey is usually brief. Guides expect one clear objective, met and measured. A final year report is normally 60 to 90 pages with a substantial literature survey, and many departments expect a paper submission alongside it.
The full report guide — front matter, conventions and page countsOne paragraph covering what the problem is, what you built, and what the result was. Written last, even though it appears first.
Loses marks: Describing the problem and the method but never stating a result. If there is no number in your abstract, it is not finished.
Why the problem matters, the context it sits in, and a clear statement of the objectives your project set out to meet.
Loses marks: Objectives written so vaguely that nobody can tell at the end whether they were met. Write objectives you can tick off in the conclusion.
What has already been done, what each approach achieved, and the specific gap your work addresses.
Loses marks: A list of paper summaries with no comparison and no gap. The survey exists to justify your design choice, so end it by saying what you chose and why.
The system as a whole: the block diagram, what each block does, and the signal or power path between them.
Loses marks: A block diagram copied from a datasheet or another report. Draw yours to match what you actually built, including the parts that are inconvenient.
Every component value and rating, with the formula, the assumptions and the arithmetic that produced it. This is the chapter that distinguishes an engineering project from an assembly job.
Loses marks: Component values with no justification. If a resistor is 10 kOhm, the report must say why it is not 1 kOhm.
The circuit as built, the PCB or wiring, the flowchart, and the parts of the code that carry the actual logic.
Loses marks: Twenty pages of pasted source listing. Include the flowchart and the few functions that matter, and put the full listing in an appendix.
What you measured, in tables and graphs, compared against what you predicted in the design chapter — including where they disagreed.
Loses marks: Photographs of the working model presented as results. A photograph is evidence that it exists, not evidence that it works. Results are measurements.
Whether each objective was met, what the measured performance was, and what the honest next step would be.
Loses marks: Future scope written as a wish list of unrelated features. It should follow from a limitation you actually hit.
Every source in a consistent style, plus datasheets, full code listings and the complete bill of materials.
Loses marks: Citing a blog or a video where a datasheet or a standard exists. Cite the primary source for anything you relied on.
The demonstration is half the assessment. These are the questions this project invites, and the answers are written the way you would give them out loud.
It is converting a crisp measured value into degrees of membership of the linguistic sets defined for that input. A speed error of a certain size might be, say, 0.7 a member of the set called small positive and 0.3 a member of medium positive. Those grades are what the rules operate on, rather than the raw number.
When the plant is nonlinear, when its parameters change during operation, or when a good model is not available. A PI controller is tuned for one operating point and one set of parameters; a well-designed fuzzy rule base can encode different behaviour in different regions of the error space. On a linear, well-modelled plant at a fixed operating point, a properly tuned PI is very hard to beat, and an honest report says so.
Centroid takes the centre of area of the whole aggregated output set, so it uses information from every rule that fired and produces a smooth, continuous output as the inputs vary. Mean of maximum only considers the rules with the highest activation, so the output jumps as the dominant rule changes, which on a drive appears as a step in torque.
It is the integral of time multiplied by the absolute error. Weighting by time means an error that persists is penalised much more heavily than the same error early in the response, so it rewards a controller that settles quickly and does not reward one that merely starts fast and then oscillates. It reduces a whole response to one comparable number.
The plant model, the reference profile, the load disturbance, the sampling rate and the current limit are identical for both controllers. The PI gains were derived from the plant model to a stated damping target rather than adjusted by eye, and the fuzzy rule base is published in full. Both are then judged on the same set of metrics rather than on whichever one favours a chosen conclusion.
Because between them they describe where the response is and where it is heading, which is what a proportional and derivative action use. The rule base can then encode behaviour such as easing off the drive when the error is still large but shrinking fast, which is exactly the situation where a proportional controller would still be pushing hard and causing overshoot.
Normally taken as a B.E. mini project (5th/6th sem) or B.E. final year project (7th/8th sem) project. If yours is a different course, the extensions above scope it up and the simpler half of the design scopes it down — the arithmetic does not change. Back to all projects.