The Motor That Doesn't Hum
The Motor That Doesn't Hum
Your motor hums because the drive is switching in blocks, and the torque dips at every 60° boundary. Field-oriented control holds the current exactly 90° ahead of the rotor flux, so the ripple falls from about 14% to under 1%, and the harmonics stop becoming heat. Here is the arithmetic, and the one place trapezoidal still wins.
The motor hums. Not the whine of a fan, and not the grind of a bad bearing. A hum that rises and falls with the speed, that you can feel in the frame as much as hear, and that gets worse at low speed, exactly where the machine has to be precise. The drive is trapezoidal. The torque dips at every 60° boundary. The current waveform does not match the back-EMF. The harmonics become heat instead of torque.
This is the article about the fix, and the fix has a name: field-oriented control, FOC for short, also called vector control. It is not a better motor. It is not a better inverter. It is a better idea about what the current is doing, and the idea is simple enough to say in one sentence, once you know the words. Keep the stator current exactly 90° ahead of the rotor flux, always, smoothly. Everything else in this article is the arithmetic of that sentence.
Think of it like pushing a swing. A swing goes highest when you push at the right moment in the arc, when the swing is coming toward you, not when it is at the top and not when it is at the bottom. Push at the wrong moment and some of your effort goes into the ropes instead of the swing.
A motor is the same. The rotor has a magnetic field, the stator has a magnetic field, and torque comes from the angle between them. At 90°, every amp of current becomes torque.
At 0°, every amp becomes a pull straight outward on the magnet, wasted power, no torque at all. At 60° or 120°, some of the current becomes torque and the rest becomes heat. The three drive methods in this article are three ways of answering one question: how close to 90° can you hold the angle, and what does it cost you?
A brushless DC motor, in plain words, is a motor with no brushes. The old brushed motor had carbon brushes and a commutator that switched the current mechanically, and the switching kept the current at 90° to the field by construction. A brushless motor has no brushes, so the controller has to do the switching electronically.
That switching is called commutation, and it is the whole job. The controller energises the stator windings in the correct sequence to keep the rotor turning. If commutation is done in step with the rotor, the torque is smooth.
If it is done early, late, or coarsely, the result is torque ripple, acoustic noise, and loss of efficiency. That is TenxerLabs' definition, and it is the cleanest one in the literature.
Two words before the methods. Back-EMF is the voltage the spinning rotor induces in the windings, the motor acting as a generator, pushing back against the voltage the controller applies. Its shape matters: most modern BLDC motors have a sinusoidal back-EMF waveform, a smooth sine wave, and the current waveform should match it.
Torque ripple is the periodic rise and fall of the torque the motor actually produces, the dip at every commutation, felt as vibration and heard as noise. The ripple is measured as a percentage of the average torque, and it is the number this article is about.
The three methods
There are three ways to drive a brushless motor, and they sit on a ladder of smoothness, cost, and complexity. Every controller on the market is one of the three, or a blend of them.
Trapezoidal commutation, also called six-step. The oldest and simplest method. At any moment, current flows through two of the three phases and the third phase floats, it carries nothing. There are six valid combinations of two phases, so the controller steps through six states per electrical revolution, one state every 60 electrical degrees. The current waveform in each phase is roughly rectangular: on at full current, then off. The big advantage is simplicity. It needs only coarse position, six states, and three Hall sensors give this directly. The computation cost is almost nothing: an 8-bit microcontroller and basic Hall-sensor reading are enough. The disadvantage is smoothness. The current cannot jump instantly, and the torque dips at every 60° boundary, so there is noticeable torque ripple and audible noise. On a motor with sinusoidal back-EMF, which is most modern BLDC motors, block commutation gives about 14% torque ripple. Those sudden current changes also make audible electrical noise. And there is a second cost, quieter than the noise: the harmonics in the trapezoidal current waveform do not have corresponding equals in the back-EMF waveform, so they do not contribute to airgap power. Instead, they simply increase the heat in the motor windings. That is EDN's explanation, and it is the reason the efficiency gap exists.
Sinusoidal commutation. The middle ground. Instead of switching abruptly between blocks, drive all three phases at the same time with sinusoidal currents spaced 120° apart. The operation becomes much smoother, and the torque ripple and commutation noise of six-step mostly disappear. But this smoothness comes with two new costs. First, a sinusoidal drive needs continuous rotor position, not just six coarse states, so the sensing requirement goes up, an encoder or a resolver, not three Hall sensors. Second, it controls the phase currents in the stationary frame, where these currents are actual AC quantities. As the speed increases, the current control loop has to track a higher and higher frequency sinusoid. A phase lag develops between the commanded and actual current, and both efficiency and performance drop at the top of the speed range. Also, sinusoidal control does not separate the current that produces torque from the current that only builds flux, so some efficiency is left on the table. That is TenxerLabs' account, and it names the exact reason sinusoidal is a stepping stone rather than a destination: it fixes the waveform but not the frame.
Field-oriented control (FOC), also called vector control. The current state of the art for smooth and efficient brushless drives. Using the Clarke and Park transforms, FOC projects the three phase currents onto a coordinate system that rotates along with the rotor. In this frame the currents become two DC-like quantities: a direct-axis component (Id) aligned with the rotor magnetic field, and a quadrature-axis component (Iq) which actually produces the torque. The controller regulates Iq to command torque, and drives Id towards zero. That is TenxerLabs' definition. When Id is zero, all current is used for torque production, maximising efficiency. The torque equation for a surface-magnet motor is τ = (3/2)·p·λm·Iq, where p is the number of pole pairs and λm is the permanent magnet flux linkage. Read it twice: torque is proportional to Iq, and Iq alone. The current along the flux makes no torque, so for a surface-magnet motor the controller simply regulates it to zero. Control the motor reduces to control two DC currents: hold Id at zero, command Iq for the torque you want. The rest of FOC is bookkeeping that makes those two currents be DC.
The angle is the whole game
All three methods are different ways of keeping the angle between the stator field and the rotor field close to 90°, with very different levels of precision and cost. That is Volcano Motor's framing, and it is the right one, because it turns a firmware question into a geometry question.
In six-step commutation, the angle between the stator and rotor magnetic fields fluctuates between 60° and 120°. That is MathWorks' description, from its motor-control series, and it is the reason for the ripple. The maximum torque occurs when the fields align at 90°.
At 60° or 120°, the torque is sin(60°) = 0.866 of the maximum, a dip of 13.4%, which is where the measured figure of about 14% torque ripple comes from. The dip is not a defect in the motor. It is arithmetic.
The controller holds the stator field fixed for 60° while the rotor keeps turning, so the angle drifts from 60° through 90° to 120°, and the torque follows the sine of the angle: up to full, down to 86.6%, up to full again, six times per electrical revolution. Six dips per revolution, every revolution, for as long as the motor runs. That is the hum.
In sinusoidal commutation, the stator field rotates smoothly, so the angle stays near 90°, but the current loop is chasing a sinusoid, and at high speed it lags. The phase lag is the angle error, and the angle error is the torque error. The method works well at low and moderate speed, and degrades as the electrical frequency rises.
In FOC, the angle is held at 90° by construction, because the controller works in the rotor's own frame. The merry-go-round image, from the EDN article by Dave Wilson, is the clearest way to see why. Imagine photographing a child on a merry-go-round from the ground: the child moves, and the camera has to track her, and as the merry-go-round spins faster the photograph blurs.
Now jump onto the merry-go-round with her. From there, she is standing still, and the photograph is sharp at any speed. The stationary-frame current controller is the photographer on the ground, chasing a sinusoid that spins faster and faster.
The Park transform is jumping onto the merry-go-round: in the rotating frame, the sinusoidal phase currents become two steady DC values, and an ordinary PI controller, the same kind Article #3 tuned, with its proportional, integral, and derivative knobs, can regulate them at any speed, because from the rotor's point of view nothing is moving.
That is why FOC needs the rotor angle, and needs it continuously. The Park transform rotates the frame by the rotor's electrical angle, θe, and if that angle is wrong, the whole frame is wrong, Id leaks into Iq, the torque command is wrong, and the motor hums again, for a different reason. The sensor is the price of the method.
Trapezoidal commutation is content with three Hall sensors, which give six states, 60° resolution, which is exactly the coarseness that causes the ripple. FOC needs a resolver or an encoder, which gives continuous angle. That is EDN's cost argument: if you need a shaft sensor, trapezoidal is cheaper, because it can be achieved with very low shaft angle resolution made possible by simple Hall-effect sensors, while FOC requires a more sophisticated shaft sensor.
The high cost of this sensor is one of the main reasons more and more designs are migrating to sensorless control, estimating the angle from the back-EMF instead of measuring it, which works once the motor is spinning and is the subject of the sensorless section below.
The comparison, measured
Here is the comparison on the criteria that decide the choice, with the numbers from the sources that measured them. The figures differ between sources, because they were measured on different motors, at different loads, in different applications, and the article says so, rather than pretending to a precision the literature does not have.
Torque ripple. Jeff Liu's lab notes put it at trapezoidal ~14%, sinusoidal ~5%, FOC under 1%. Volcano Motor agrees on the trapezoidal figure: about 14% on a motor with sinusoidal back-EMF. Fukushima's drone ESC comparison agrees again: about 12% for trapezoidal, a few percent or less for sinusoidal, 1% or less for FOC. eBike Engineering's table gives wider ranges for e-bike conditions: trapezoidal 15–25%, sinusoidal 5–10%, FOC under 5%. The three independent sources cluster around 12–14% for trapezoidal and under 1–5% for FOC, and the article uses the Jeff Liu figures as its working numbers, with the eBike ranges as the cross-check. The shape of the answer does not depend on which source you pick: trapezoidal ripples an order of magnitude more than FOC.
Efficiency. EBike Engineering, citing DigiKey's technical comparison: trapezoidal drive produces the highest harmonic content of the three methods, and measured efficiency is typically 3–8% below FOC at partial load, though the gap narrows at full load and high speed. TI application data for the DRV10987 reports that sinusoidal commutation can increase efficiency by up to 10% compared to trapezoidal drive in a water-pump demonstration, an application-specific upper bound, not a guaranteed gain. eBike's table, as representative estimates for e-bike operating conditions: efficiency at full load, trapezoidal 80–85%, sinusoidal 82–88%, FOC 85–92%; at partial load, trapezoidal poor, sinusoidal moderate, FOC best. The demo below uses the table's midpoints, and the correction section records that they are representative estimates, not a datasheet.
Noise. Trapezoidal is loud, sinusoidal is medium, FOC is quiet, Jeff Liu's table. eBike agrees: trapezoidal has the highest noise of the three, the motor whine is audible at low speeds, and the rectangular current waveforms excite mechanical resonances in the stator. Sinusoidal is quiet; FOC is quietest.
Low-speed performance. Trapezoidal is poor, sinusoidal is fair, FOC is excellent, Jeff Liu's table again. The reason is the angle: at low speed the rotor spends a long time in each 60° sector, so the torque dips are long and the motor coggs, the jerky pull you feel turning a brushless shaft by hand, from Article #20. FOC holds torque smoothly down to near-zero speed, because the angle is held at 90° at every speed, including standstill.
Computation and sensing. Trapezoidal: low computation, Hall sensors at 60° resolution, an 8-bit microcontroller is enough. Sinusoidal: medium computation, an encoder, a 32-bit microcontroller. FOC: high computation, an encoder, a 32-bit microcontroller with a fast ADC, and two current sensors, because FOC must measure the phase currents continuously. The current loop runs at the PWM rate, typically 16–64 kHz, and the speed loop at 1–5 kHz. That is the InnovChip FOC guide's architecture, and it is why the shunts and current-sense amplifiers from Article #23 matter: FOC needs accurate, continuous phase-current feedback, and a 1 mΩ shunt at 10 A gives 10 mV, which the amplifier multiplies up to something the ADC can read.
Cost. EBike's table puts typical controller cost at $25–50 for trapezoidal, $40–80 for sinusoidal, $40–390 for FOC, the range is wide because FOC controllers span from a hobby board to an industrial drive. EDN's cost argument is subtler: in a sensorless system, most of the cost is in the power devices and the bus capacitor, which is determined by the motor horsepower, so the total system cost of sensorless trapezoidal and sensorless FOC is almost identical. The sensor is where the money goes, and sensorless removes it.
How FOC actually works: four steps, two transforms
FOC looks like graduate-level material, and it has the reputation for it. It is not. The entire algorithm is about forty lines of code built on one idea, and Segev Technologies, which has shipped it on everything from a 16 mm board to a 100 V, 30 A industrial drive, says so plainly. The idea is the torque equation above. The rest is bookkeeping.
The bookkeeping is two coordinate transforms, and each one does one job.
The Clarke transform collapses the three 120°-spaced phase currents into an equivalent two-axis stationary frame. Three coils evenly spaced 120° apart become two equivalent coils, one at 0° and one at 90°. The transform is amplitude-invariant, which means the two-phase waveforms have the same amplitude as the original three-phase waveforms. And because the three phase currents in a balanced machine sum to zero, ia + ib + ic = 0, only two of them need to be measured. The third is arithmetic. That is why a FOC drive has two current sensors, not three, although many systems use three anyway, to detect winding-to-frame shorts, or because three shunts are required to measure two currents when the voltage modulation reaches 100%.
The Park transform rotates that stationary frame by the rotor's electrical angle, θe, so the axes spin with the rotor. The d-axis current is aligned directly with the rotor flux, while the q-axis current is quadrature to it, at 90°. In the dq frame, the sinusoidal phase currents become two DC values, which ordinary PI controllers can regulate. That is the whole trick. Two PI loops, the same proportional-integral controllers Article #3 tuned, with the same rise time, overshoot, and settling time to trade against each other, one holding Id at zero, one commanding Iq for the torque you want. The outputs are two voltages, Vd and Vq, which go through the inverse Park transform back into the stationary frame, and then to the modulator.
The modulator is space-vector PWM, SVPWM. It turns the voltage command into the six switch states of the inverter from Article #23, the three half-bridges, each with a high-side and a low-side MOSFET, and the rule that the two devices in one half-bridge must never conduct together. SVPWM is the standard choice because it uses the DC bus better than sinusoidal PWM: about 15.47% more voltage from the same bus, which is the ratio 2/√3. In concrete terms, on a 48 V bus, sinusoidal PWM gives a phase-voltage peak of Vdc/2 = 24 V, while SVPWM gives Vdc/√3 = 27.8 V. The demo below runs the full bus-utilisation arithmetic, including the one place trapezoidal still wins.
The loop closes like this, and the order matters: measure two phase currents → Clarke transform to the stationary αβ frame → Park transform to the rotating dq frame using the measured rotor angle → two PI controllers, Id held at zero and Iq commanded for the torque wanted → inverse Park → SVPWM → the inverter → the motor → and the rotor angle and the phase currents are measured again, thousands of times a second.
Example Labs is a made-up company we use for worked examples.
The current loop runs at the PWM rate, 16–64 kHz. The speed loop, which sits outside it and commands the torque, runs at 1–5 kHz. At the Example Labs arm's shoulder speed of 1,500 RPM, the electrical frequency is 175 Hz with seven pole pairs, so a 20 kHz current loop is 114 times faster than the thing it is tracking, which is why the loop has no trouble holding the angle.
At 10,000 RPM the electrical frequency is 1,167 Hz, and the same loop is still 17 times faster. The sinusoidal controller, chasing an AC quantity in the stationary frame, does not have that margin, and its phase lag grows with speed. That is the bandwidth argument, and it is the reason FOC holds its performance across the speed range while sinusoidal does not.
Field weakening: the second use of Id
Id is zero for a surface-magnet motor, because the magnets already provide the flux and any current along the flux makes no torque. But Id has a second use, and it is the one that extends the motor's speed range: field weakening.
A motor has a top speed set by its back-EMF. As the rotor spins faster, the back-EMF rises, and it rises until it equals the voltage the inverter can supply, at which point no more current can be pushed into the windings, and the motor cannot go faster. The voltage the inverter can supply is set by the bus and the modulation: on the 48 V bench pack, SVPWM gives a line-to-line fundamental peak of 48.1 V.
Field weakening spends some of the current budget on a negative Id, a current that opposes the magnet flux, weakening the field, lowering the back-EMF, and letting the motor spin faster on the same bus. It costs torque per amp, because some of the current is now doing flux work instead of torque work, and it costs heat, because the current is still flowing through the windings. But it buys speed, and for a motor that must run past its base speed, a spindle, a traction drive, a fan that must spin up fast, it is the only way there.
The dq voltage equations show the mechanism directly: vq = Rs·iq + ωe·(Ld·id + ψf), where ψf is the magnet flux linkage. A negative id subtracts from ψf, so the speed term ωe·(Ld·id + ψf) falls, and the voltage the inverter must supply falls with it. Trapezoidal commutation has no equivalent: it cannot weaken the field, because it has no d-axis to command.
Sensorless: the angle without the sensor
FOC needs the rotor angle, continuously. An encoder or a resolver gives it directly, and costs money, wiring, and a failure point. Sensorless FOC estimates it instead, from the back-EMF, and the catch is in the physics: back-EMF is proportional to speed, so at standstill there is nothing to observe.
The controller has to start blind, in four stages. First, a windmill check: is the rotor already turning, and in which direction, a fan that is already spinning backwards is the classic case. Second, initial position detection, or align: find the rotor position by probing it, or pull the rotor to a known angle.
Third, an open-loop ramp: spin the motor up without feedback until the back-EMF is big enough to track. Fourth, closed-loop drive: hand over to the observer and run full FOC. That is the TenxerLabs sequence, from its live-hardware evaluation of the Allegro A89333 sensorless FOC controller.
The observers differ in how low they can go. A direct back-EMF observer needs roughly 10% of rated speed before it can lock on. A sliding-mode observer gets to about 5%.
A Luenberger state observer, which runs a model of the motor and corrects it with the measurements, gets to about 5–8% but needs accurate motor parameters, resistance, inductance, flux, and high tuning effort. Below those speeds the observer output becomes noisy, the handoff condition is never met, and the motor just hums and stops. That is the InnovChip guide's failure mode, and it is the reason sensorless FOC is a poor choice for a joint that must hold position at zero speed, the arm's shoulder, for example, which is why the demo below keeps its encoder.
Writing the observer, the startup sequencer, and the state machine, then tuning them for one specific motor, is why sensorless FOC firmware is often quoted as months of work. The silicon alternative, a sensorless FOC controller with the whole stack in the chip, parameters and startup profile stored in EEPROM, exists precisely because the firmware is the hard part.
The honest counterpoint: where trapezoidal wins
FOC is not better at everything, and an article that says otherwise is selling something. EDN names the two places trapezoidal wins, and both are real.
First, the inverter. Trapezoidal commutated control at full speed can be achieved without PWM, which significantly reduces switching losses in the inverter. With FOC, the motor usually runs more efficiently; but with trapezoidal commutation at full speed, the inverter runs more efficiently. Which is best depends on the particular motor and inverter. At full load and high speed, the efficiency gap narrows, eBike Engineering says the same, because the motor's harmonics matter less when the motor is working hard, and the inverter's switching losses matter more.
Second, the top speed. The fundamental harmonic of the voltage waveform from trapezoidal commutation is larger than the highest voltage sinewave possible with FOC, which allows the motor to achieve higher speeds. On the 48 V bus, the six-step line-to-line fundamental peak is (2√3/π)·Vdc = 53.0 V, against 48.1 V for linear SVPWM, about 10% more voltage, and therefore about 10% more speed, before the inverter runs out of voltage.
FOC can close the gap with over-modulation: advanced FOC systems allow over-modulating the voltage waveform until it gradually morphs into a trapezoidal form, increasing the modulation index by 33% compared to sinusoidal modulation alone. And field weakening, above, extends it further. But the plain statement stands: on the same bus, with the same motor, trapezoidal spins faster than linear FOC.
The demo quantifies it.
DEMO
The demo uses the Example Labs bench arm from Article #20: two links, 0.425 m each, link 1 is 4 kg, link 2 is 3 kg, payload 5 kg at 0.85 m, horizontal worst case. The shoulder needs 126.8 N·m at the joint, static 68.8 N·m plus dynamic 15.7 N·m, times a 1.5 safety factor. Through a 100:1 two-stage planetary gearbox at 93% efficiency, the motor needs 126.8 ÷ (100 × 0.93) = 1.36 N·m, spinning at 15 RPM × 100 = 1,500 RPM.
The bench pack is the 48 V pack from Article #6: 13S1P NMC, 48.1 V nominal, 54.6 V full, 39.0 V empty, 168 Wh, and the arm's shift pack is 48 V, 100 Ah, 4,800 Wh. The drive stage is Article #23's: a 48 V three-phase inverter, six MOSFETs, gate driver, shunts for current feedback. The motor is a BLDC with a sinusoidal back-EMF, seven pole pairs, the pole-pair count is the article's own labelled assumption, chosen because seven pairs is a common outrunner figure and it makes the arithmetic concrete.
These are the same figures the earlier articles used, and the arithmetic below was run in Python before writing.
Step 1, the angle, and where 14% comes from. In six-step commutation the angle between the stator and rotor fields swings from 60° to 120°, and torque follows the sine of the angle. sin(60°) = 0.866, sin(90°) = 1.000. The torque dips to 86.6% of its peak at every boundary, a dip of 13.4%, which is the measured figure of about 14% torque ripple, derived here from geometry rather than quoted from a table. Six dips per electrical revolution. At 1,500 RPM the motor turns 25 revolutions a second; with seven pole pairs the electrical frequency is 175 Hz; six commutations per electrical revolution gives a ripple frequency of 1,050 Hz, a kilohertz whine, audible, and felt in the frame. At 60 RPM, the ripple is 42 Hz, a low growl, one dip every 24 milliseconds, which is why the hum is worst at low speed, exactly where the arm has to be precise.
Step 2, the ripple at the joint. The shoulder motor produces 1.36 N·m. On trapezoidal drive at ~14% ripple, the torque swings by ±0.19 N·m at the motor. Through the 100:1 gearbox at 93% efficiency, that is ±17.7 N·m at the joint, on a load of 126.8 N·m, the joint torque swings by 14%, six times per electrical revolution, 1,050 times a second. On sinusoidal drive at ~5%, the swing is ±0.07 N·m at the motor, ±6.3 N·m at the joint. On FOC at under 1%, it is under ±0.014 N·m at the motor, under ±1.3 N·m at the joint. The arm's positioning budget from Article #20 is set by the gearbox backlash, 1 arcminute moves the tool by about 0.29 mm at one metre, and a joint torque swinging by 17.7 N·m is a positioning error the controller has to fight, 1,050 times a second, while the FOC swing of 1.3 N·m is inside the noise.
Step 3, the heat, and the battery. The motor's mechanical power is 1.36 N·m × 157.08 rad/s = 213.6 W. (The joint power is 126.8 N·m × 1.57 rad/s = 199.1 W; divided by the gearbox efficiency of 0.93, that is 214.1 W, the same number, allowing for rounding.) Using eBike Engineering's full-load efficiency table, midpoints: FOC at 88.5% draws 241.4 W and makes 27.8 W of heat; trapezoidal at 82.5% draws 258.9 W and makes 45.3 W of heat. The difference is 17.6 W of extra heat in the motor, every second the shoulder is working, heat that comes from the harmonics, which do not contribute to airgap power and instead increase the heat in the windings. Over an eight-hour shift at that load, the extra energy is 140.4 Wh, 2.9% of the 4,800 Wh shift pack, spent on nothing. At the ends of the table's ranges the gap is wider: trapezoidal at 80% against FOC at 92% is 267.0 W against 232.2 W, a difference of 34.8 W. And the counterpoint, stated honestly: at full speed the trapezoidal inverter runs without PWM, so its switching losses fall, and the gap narrows, which is why the article does not claim the efficiency advantage at every operating point.
Step 4, the speed ceiling, on the 48.1 V bus. Sinusoidal PWM gives a phase-voltage peak of Vdc/2 = 24.05 V. SVPWM gives Vdc/√3 = 27.77 V, 15.47% more, the ratio 2/√3. The six-step line-to-line fundamental peak is (2√3/π)·Vdc = 53.04 V, against 48.1 V for linear SVPWM and 41.66 V for sinusoidal PWM. So on the same bus, trapezoidal reaches about 10.3% higher speed than linear SVPWM FOC, and about 27.3% higher than sinusoidal PWM, before the inverter runs out of voltage. FOC closes the gap two ways: over-modulation, which increases the modulation index by 33% compared to sinusoidal modulation alone, morphing the waveform gradually toward trapezoidal; and field weakening, a negative Id that opposes the magnet flux and lowers the back-EMF. The arm's shoulder does not need either, it runs at 1,500 RPM, far below the voltage ceiling, but a spindle or a traction drive would, and the choice of drive method decides whether the speed is available at all.
Step 5, the decision. The shoulder joint holds position, moves slowly, and must be smooth: FOC, with its encoder, its two current sensors, and its 20 kHz current loop. The ripple falls from ±17.7 N·m to under ±1.3 N·m at the joint, the heat falls by 17.6 W, and the motor is quiet. A fan or a pump at full speed, cost-sensitive, with no positioning requirement: trapezoidal, on an 8-bit microcontroller with three Hall sensors, and the inverter runs without PWM at full speed. The rule is the same as the article's opening: if the load feels the ripple, buy FOC. If the load is a fan, buy trapezoidal. And if the load is a joint that must hold still at zero speed, do not buy sensorless, the observer has nothing to lock onto at standstill, and the motor will hum and stop.
The decision test
Five questions. Answer them honestly, and the drive method chooses itself.
- Does the load feel the ripple? If a human holds the tool, if a camera sits on the arm, if a pipette must not shake, the load feels 14% torque ripple, and the answer is FOC. If the load is a fan, a pump, or a wheel, it does not, and trapezoidal is on the table.
- What is the back-EMF shape? If the motor has a sinusoidal back-EMF, and most modern BLDC motors do, trapezoidal commutation will ripple, because the current waveform does not match. If the motor has a trapezoidal back-EMF, trapezoidal commutation is the matched drive, and the ripple is small. Read the datasheet before choosing the drive.
- How slow must it go, smoothly? If the motor must hold torque at near-zero speed, a joint holding position, a gimbal, a slow slew, FOC is the only method of the three that holds the angle at 90° at standstill. Trapezoidal coggs at low speed. Sensorless FOC cannot start from standstill at all.
- What does the top speed need? If the motor must spin faster than the bus allows with linear modulation, trapezoidal gives about 10% more speed than SVPWM FOC on the same bus, and over-modulation or field weakening is the FOC answer. Work out the voltage ceiling before choosing.
- What does the sensor cost, and what does it save? Three Hall sensors and an 8-bit microcontroller against an encoder, two current sensors, and a 32-bit microcontroller with a fast ADC. But in a sensorless system the total cost is almost identical, because most of the cost is in the power devices and the bus capacitor. Price the system, not the controller.
Objections
"FOC is overkill. Trapezoidal works fine." It works fine for loads that do not feel the ripple. A fan does not care about 14% torque ripple. A robot arm holding a tool does. The question is not whether trapezoidal works, it does, and it has driven motors for decades. The question is whether the load feels the difference between ±17.7 N·m and ±1.3 N·m at the joint. If it does, trapezoidal does not work fine. It works loudly.
"FOC needs an expensive encoder and a fast processor." It needs continuous angle and two current measurements, and that costs money, but the cost argument has a limit. In a sensorless system, most of the cost is in the power devices and the bus capacitor, determined by the motor horsepower, so the total system cost of sensorless trapezoidal and sensorless FOC is almost identical. And the processor is not the constraint it was: the whole algorithm is about forty lines of code, and it has shipped on a 16 mm board. The cost is in the tuning, not the silicon.
"Sinusoidal is smooth enough. Why go further?" Sinusoidal is smooth at low and moderate speed, and it is a genuine improvement over trapezoidal, the ripple falls from ~14% to ~5%, and the noise falls with it. But it controls AC quantities in the stationary frame, so the current loop must track a sinusoid whose frequency rises with speed, and a phase lag develops. At 1,500 RPM the electrical frequency is 175 Hz; at 10,000 RPM it is 1,167 Hz, and a 1 kHz speed loop is slower than the thing it is tracking. FOC turns the problem into two DC currents, and the loop holds at any speed. Sinusoidal is the stepping stone. FOC is the destination.
"FOC is always more efficient." Not always. With FOC, the motor usually runs more efficiently; but with trapezoidal commutation at full speed, the inverter runs more efficiently, because it can run without PWM. At full load and high speed the gap narrows, and which is best depends on the particular motor and inverter. The article's efficiency numbers are at partial load and at the table's midpoints, and the demo says so. Do not buy FOC for efficiency alone at full speed. Buy it for smoothness, for low-speed torque, and for control.
"Sensorless FOC removes the sensor cost, so it is the best of both." Sensorless removes the sensor and adds the startup problem. At standstill there is no back-EMF, so the observer has nothing to lock onto, and the controller must start blind, windmill check, align, open-loop ramp, handoff, and below about 5–10% of rated speed the observer cannot lock on at all. For a fan that spins up from rest under no load, sensorless is the right answer. For a joint that must hold position at zero speed, it is the wrong one, and the motor will hum and stop.
"Our motor has a trapezoidal back-EMF, so FOC is pointless." If the back-EMF is genuinely trapezoidal, trapezoidal commutation is the matched drive, and the ripple is small, EDN says so explicitly, and the article agrees. But check the datasheet before believing it. Most modern BLDC motors have a sinusoidal back-EMF, because a trapezoidal back-EMF needs the airgap flux held constant, which is harder to achieve in motor design. The motor you have is probably sinusoidal, and the drive you have is probably trapezoidal, and that mismatch is the hum.
Sources
- TenxerLabs, "BLDC Motor Control: From Six-Step to Sensorless FOC" (via Hackster)
What we used
"Since a BLDC motor has no brushes, the electronic controller must energize the stator windings in the correct sequence to keep the rotor turning. This is called commutation. If commutation is done in step with the rotor, the torque is smooth. If it is done early, late or coarsely, the result is torque ripple, acoustic noise and loss of efficiency." "Six-step (trapezoidal) commutation: At any point of time, current flows through two of the three phases and the third phase floats. There are six valid combinations, so the controller steps through six states per electrical revolution, one state every 60 electrical degrees." "The big advantage is simplicity. It needs only coarse position, six states, and three Hall sensors give this directly. The computation cost is almost nothing." "The disadvantage is smoothness. The current cannot jump instantly, and the torque dips at every 60 degree boundary, so there is noticeable torque ripple and audible noise." "Sinusoidal commutation: drive all three phases at the same time with sinusoidal currents spaced 120 degrees apart, instead of switching abruptly between blocks. The operation becomes much smoother, and the torque ripple and commutation noise of six-step mostly disappear. But this smoothness comes with two new costs. First, a sinusoidal drive needs continuous rotor position, not just six coarse states, so the sensing requirement goes up. Second, it controls the phase currents in the stationary (stator) frame, where these currents are actual AC quantities. As the speed increases, the current control loop has to track a higher and higher frequency sinusoid. A phase lag develops between the commanded and actual current, and both efficiency and performance drop at the top of the speed range. Also, sinusoidal control does not separate the current that produces torque from the current that only builds flux, so some efficiency is left on the table." "Field-Oriented Control (FOC), also called vector control, is the current state of the art for smooth and efficient brushless drives. Using the Clarke and Park transforms, FOC projects the three phase currents onto a coordinate system that rotates along with the rotor. In this frame the currents become two DC-like quantities: a direct-axis component (Id) aligned with the rotor magnetic field, and a quadrature-axis component (Iq) which actually produces the torque. The controller regulates Iq to command torque, and drives Id towards zero." "When Id = 0, all current is used for torque production, maximizing efficiency." "Here’s the core issue: back-EMF is proportional to speed. At standstill there is nothing to observe, so the controller has to start blind, in four stages: Windmill check: is the rotor already turning, and in which direction? IPD or align: find the rotor position by probing it (initial position detection), or pull the rotor to a known angle. Open-loop ramp: spin the motor up without feedback until the back-EMF is big enough to track. Closed-loop drive: hand over to the observer and run full FOC." "It all has to be tuned per motor. Writing the observer, the startup sequencer and the state machine, then tuning them for one specific motor, is why sensorless FOC firmware is often quoted as months of work." Checked via browser.search, 2026-11-20
- EDN, "Field-oriented control by the numbers" (Dave Wilson; opened in full, 2026-11-20)
What we used
"Most BLDC motors in fact have a sinusoidal back-EMF waveform. If you employ trapezoidal commutation on a motor like this, torque ripple will occur which is synchronous to the commutation process." "To achieve low-torque ripple with trapezoidal commutation (especially at lower speeds), the back-EMF waveform should be trapezoidal in shape. But this back-EMF shape is harder to achieve in motor design than sinusoidal back-EMF since the airgap flux would need to be held constant." "In most cases, a BLDC motor with sinusoidal back-EMF will run more efficiently with FOC than with trapezoidal commutation. This is because the higher harmonics in the trapezoidal current waveform do not have corresponding equals in the back-EMF waveform, and therefore don’t contribute to airgap power. Instead, they simply increase the heat in the motor windings." "On the other hand, trapezoidal commutated control at full speed can be achieved without PWM, which significantly reduces switching losses in the inverter. So with FOC, the motor usually runs more efficiently; but with trapezoidal commutation at full speed, the inverter runs more efficiently. Which is best depends on your particular motor and inverter." "The fundamental harmonic of the voltage waveform resulting from trapezoidal commutation is larger than the highest voltage sinewave possible with FOC, which allows the motor to achieve higher speeds. You can increase the FOC voltage using modulation enhancements such as third harmonic injection and space vector modulation, increasing the FOC voltage amplitude 15% by dynamically shifting the common-mode voltage of the output waveforms. But this is still not equal to the fundamental voltage harmonic with trapezoidal commutation. More advanced FOC systems (such as InstaSPIN motor control technology from Texas Instruments) allow you to over-modulate the voltage waveform until it gradually morphs into a trapezoidal form. This increases the modulation index by 33% compared to sinusoidal modulation alone." "If you need a shaft sensor for your design, then trapezoidal commutation is going to be less expensive. This is because it can be achieved with very low shaft angle resolution made possible by simple Hall Effect sensors." "But FOC requires a more sophisticated shaft sensor, such as a resolver or encoder. The high cost of this sensor is one of the main reasons why more and more designs are migrating to sensorless control. If you can eliminate the shaft sensor, sensorless trapezoidal commutation is still less expensive than sensorless FOC because the processor requirements are lower. But the total system cost is almost identical. This is because in a sensorless system, most of the cost is in your power devices and bus capacitor(s), which is determined by the motor horsepower." "Considering the enhanced performance and flexibility possible with FOC, it may very well turn out to be the most cost-effective solution for your application. This is why many low-cost, high-volume applications (such as appliances) are abandoning trapezoidal control altogether and embracing FOC faster than ever before." "The Clarke transform takes three coils evenly spaced 120° apart, and turns them into two equivalent coils, one at 0° and one at 90°." "The Park transform rotates that frame by the rotor's electrical angle, so the axes spin with the rotor. The d-axis current is aligned directly with the rotor flux, while the q-axis current is quadrature to the d-axis. The angle in these equations is that of the rotor flux. For a permanent magnet synchronous motor, the rotor flux does not move with respect to the rotor. So by accurately measuring the angle of the rotor with an encoder or a resolver, we can accurately determine the angle of the rotor flux." "Many FOC systems actually do utilize three current sensors anyway, however, so as to detect winding to frame shorts, or because three shunts are required in order to measure two currents when the voltage modulation reaches 100%." "These four steps are also used to do FOC on an AC motor. The main difference is that, since AC motors do not have brushes and a commutator, the process must also include the orientation part of the job." "The brushes and commutator on the DC motor mechanically perform field orientation. They keep the rotor current vector always oriented by 90° with respect to the stator flux vector, to achieve maximum torque per amp."
- eBike Engineering, "FOC vs Sinusoidal vs Trapezoidal: Controller Drive Modes Explained" (published 9 November 2026)
What we used
"The drive mode, how a controller sequences current through the motor's three phases, is the single largest factor in ride quality, efficiency, and noise." "Trapezoidal drive is the oldest and simplest method for driving a brushless DC motor. The controller energizes exactly two of the three motor phases at any instant and switches to the next pair every 60 electrical degrees. The current waveform in each phase is roughly rectangular, on at full current, then off." "Because the commutation points are 60° apart, there is a discontinuity at every transition. The instantaneous torque drops and then surges as the new phase pair takes over. This causes torque ripple, periodic oscillations in the force the motor applies, which the rider feels as vibration and hears as cogging noise at low speeds." "According to DigiKey's technical comparison, trapezoidal drive produces the highest harmonic content of the three methods. These harmonics represent wasted energy, current flowing in patterns that generate heat rather than torque. Measured efficiency is typically 3–8% below FOC at partial load, though the gap narrows at full load and high speed." "TI application data for the DRV10987 reports that sinusoidal commutation can increase efficiency by up to 10% compared to trapezoidal drive in a water-pump demonstration, though this figure is application-specific and represents an upper bound rather than a guaranteed gain in e-bike motors." Side-by-side table: efficiency at full load, trapezoidal 80–85%, sinusoidal 82–88%, FOC 85–92%; torque ripple, trapezoidal high (15–25%), sinusoidal low (5–10%), FOC lowest (<5%); audible noise, noticeable cogging / quiet / quietest; processing, 8-bit MCU / 32-bit MCU / 32-bit MCU + fast ADC; typical controller cost, $25–50 / $40–80 / $40–390. "Efficiency ranges in the table are representative estimates for typical e-bike operating conditions and vary by motor, load, and speed. The DigiKey comparison provides qualitative rankings (Excellent/Moderate/Low) rather than specific percentages; the TI DRV10987 data reports up to 10% efficiency gain for sinusoidal over trapezoidal in a water-pump application." Checked via browser.search, 2026-11-20
- Volcano Motor, "BLDC Motor Control Strategies: From Six-Step Commutation to Field-Oriented Control"
What we used
"All three control methods are just different ways of keeping that angle close to 90°, with very different levels of precision and cost." "You get maximum torque when the stator magnetic field is exactly 90° ahead of the rotor field." "On a motor with sinusoidal back-EMF (which is most modern BLDC motors), block commutation gives you about 14% torque ripple." "Sinusoidal control does not perform the vector transformation into a rotating reference frame that true FOC does. Torque and flux remain coupled, you can't control them independently." "In FOC, Iq and Id are independently regulated. The control system keeps Id near zero (because the magnets already provide the flux) and adjusts Iq to control torque. Because torque and flux are decoupled, FOC achieves minimal torque ripple, higher efficiency (especially at high speed), superior dynamic response, and precise torque control." Switching losses: trapezoidal low (1× PWM), sinusoidal high (3× PWM). Checked via browser.search, 2026-11-20
- Fukushima LLC, "ESC3, Comparison of Drive Methods"
What we used
"Torque fluctuates momentarily when switching steps, causing about 12% ripple." "Torque ripple is significantly smaller than that of trapezoidal waves (a few percent or less)" for sine wave drive. "Current for the same torque is minimized, resulting in maximum efficiency. Torque ripple is minimal (can reach 1% or less). Smooth across the entire range from low to high speeds." "Trapezoidal wave is simple and strong at high speeds, while FOC is smooth and highly efficient across the entire range, but requires stricter hardware and control design." Checked via browser.search, 2026-11-20
- Jeff Liu Lab, "Brushless Motors and FOC" (engineering notes)
What we used
"FOC is currently the most advanced control algorithm for BLDC/PMSM motors, using coordinate transforms to convert AC quantities into DC quantities for control." "Decompose three-phase AC currents into: Id (direct-axis current): Controls magnetic flux (typically set to 0). Iq (quadrature-axis current): Directly controls torque. τ = (3/2)·p·λm·Iq, where p is the number of pole pairs and λm is the permanent magnet flux linkage. When Id = 0, all current is used for torque production, maximizing efficiency." Comparison table: torque ripple, trapezoidal ~14%, sinusoidal ~5%, FOC <1%; efficiency, medium / fairly high / highest; low-speed performance, poor / fair / excellent; noise, loud / medium / quiet; position sensor, Hall (60° resolution) / encoder / encoder; computation, low / medium / high. Checked via browser.search, 2026-11-20
- MathWorks, "Understanding Field-Oriented Control" (Motor Control, Part 4)
What we used
"One of the drawbacks with this type of control are the ripples that we observe in the speed and torque responses of the motor. We clearly see that the angle between the stator and rotor magnetic fields fluctuates between 60 and 120 degrees. This is the reason behind the speed and torque ripples. It also prevents us from getting the maximum torque, which occurs when the fields align at 90 degrees." "Field-oriented control greatly reduces the ripples of the system response and leads to smoother operation of the motor." "We simply force the direct axis component to be zero while allowing the quadrature axis component to grow. Once the direct component diminishes completely, our stator field vector is at exactly 90 degrees with the reference vector." Checked via browser.search, 2026-11-20
- Mouser / "Basics of Motor Field-Oriented Control"
What we used
"FOC is a technique used to control the torque and speed of three-phase AC motors. The purpose of FOC is to direct the motor to achieve maximum torque at a given speed by ensuring that the rotor field is always lagging by 90° concerning the stator, even under continually varying load conditions." "In FOC, the current is resolved into two perpendicular parts: the part that causes the desired vertical pull, called torque or quadrature current (Iq), and one that causes the undesired outward pull, called flux or direct current (Id). Once these constituent present components are determined, FOC adjusts these two components by adjusting the voltage phase so the direct current becomes zero and only the desired torque current remains." "For practical motor control, a pair of proportional-integral (PI) closed-loop feedback controllers are employed to regulate both the phase and voltage to the coils, ensuring the desired Iq and Id current values." Checked via browser.search, 2026-11-20
- Segev Technologies, "FOC Controller From Scratch: Clarke, Park and SVPWM"
What we used
"Field-Oriented Control has a reputation as graduate-level material, yet the entire algorithm is about forty lines of code built on one idea. We've shipped it on everything from a 16 mm board to a 100 V / 30 A industrial drive." "FOC instead drives the motor like the synchronous machine it is: keep the stator current vector exactly 90° ahead of the rotor flux, always, smoothly. Improved torque response, stable torque at near-zero speed, four-quadrant operation: the properties on our controller spec sheets all come from this one discipline." "Seen from the rotor, only the component of stator current that's perpendicular to the rotor's magnetic flux makes torque. For a surface-magnet machine: τ = (3/2)·p·λpm·iq. Read it twice, because this is the entire game: torque is proportional to iq. The current along the flux (id) makes no torque: for a surface-magnet motor we simply regulate it to zero." "In dq, the sinusoidal phase currents become two DC values: which ordinary PI controllers can regulate." "In salient machines (IPM), a reluctance term, (3/2)·p·(Ld − Lq)·id·iq, joins the torque equation and makes id useful too; this article assumes the surface-magnet case, where it vanishes." Checked via browser.search, 2026-11-20
- InnovChip, "BLDC Motor Control with FOC Guide"
What we used
"Field-Oriented Control (FOC), also called Vector Control, treats the three-phase motor as a decoupled DC-like system: it independently controls torque (q-axis) and flux/field (d-axis), enabling smooth, efficient, low-noise operation across the entire speed range." "Torque equation (surface PMSM): Te = (3/2) × p × [λpm·iq + (Ld − Lq)·id·iq]. For surface-mounted PM (SPMSM, Ld = Lq): Te ∝ iq, torque is directly proportional to q-axis current. This is the heart of FOC: control iq to control torque, control id = 0 to maximize torque-per-amp (MTPA for SPMSM)." "The current loop runs at the PWM/ISR rate (16–64 kHz); the speed loop typically runs at 1–5 kHz (every N ISRs)." Sensorless observers: back-EMF direct, minimum speed ~10% rated, low robustness; sliding-mode observer, ~5% rated, high robustness; Luenberger, ~5–8% rated, medium robustness, needs accurate R, L, flux. "At standstill, the rotor has no measurable BEMF, the observer has nothing to lock onto. Therefore sensorless drives must use an open-loop forced-commutation startup sequence, then hand off to closed-loop FOC once BEMF is observable." "At low speed, the BEMF is tiny (proportional to ω). The observer output becomes noisy, and the handoff condition is never met, the motor just hums and stops." Checked via browser.search, 2026-11-20
- MDPI, "Co-Simulation Comparison Analysis of SPWM and SVPWM in FOC of an Outer Rotor PMSM" (15 July 2026)
What we used
"SPWM is simple to build, but it has low DC bus utilization (about 78.5% of the DC-link voltage) and increased total harmonic distortion (THD). SVPWM is the preferred choice for high-performance drives due to its greater DC-link utilization (approximately 15.5% higher) and lower harmonic distortion." "SVPWM beats SPWM in all measured parameters, including current THD, DC-link voltage usage, and torque ripple." Checked via browser.search, 2026-11-20
- ElectrifyParts, "From PWM to SPWM and SVPWM"
What we used
"SVPWM offers about 15.47% greater DC-bus voltage utilization than conventional SPWM." "SPWM: Vref,max = Vdc/2. SVPWM: Vref,max = Vdc/√3." "In a simplified steady-state PMSM dq model: vd = Rs·id − ωe·Lq·iq; vq = Rs·iq + ωe·(Ld·id + ψf)." "SVPWM can delay voltage saturation and allow more room for high-speed current control or field weakening." Checked via browser.search, 2026-11-20
- Technblogy, "Why SVPWM Gives About 15% More Voltage Than SPWM"
What we used
"The usual shorthand is 'about 15% more.' That comparison is real, but only in the linear range, and only if you compare SVPWM, or third-harmonic injection PWM, with SPWM. It is not 15% more than six-step. Six-step is still higher." "At the linear SPWM limit the modulation wave just fills the carrier (M = 1). The pole voltage is then a sine of peak Vdc/2, and the line-to-line fundamental peak is (√3/2)·Vdc ≈ 0.866·Vdc." "At the linear SVPWM limit the reference vector sits on the inscribed circle of the switching hexagon, |Vref| = Vdc/√3. The line-to-line fundamental peak is then exactly Vdc. The ratio of those two limits is 2/√3 ≈ 1.1547, which is 15.47% more." "Six-step, square-wave poles, 120° apart, has a still larger fundamental, (2√3/π)·Vdc ≈ 1.1027·Vdc. Linear SVPWM is about 9.3% below six-step (Mi ≈ 0.9069), not 15% above it." Checked via browser.search, 2026-11-20
- MESC Documentation (Molony Electronic Speed Control), "Control loops"
What we used
"All the FOC gets done in the fastLoop." "The sensorless observer is very simple. It works (as most successful observers do) on the basis of flux integration, that is the assumption that for a spinning magnet passing a coil, the voltage is given by: V = turns × dφ/dt." "ST Micro's FOC library uses a Luenberger observer and Alex Evers' UNIMOC uses a Kalman filter." Checked via browser.search, 2026-11-20
- Preprints.org, "Advanced Control Techniques for Wide-Speed-Range Sensorless Permanent Magnet Motor Drives"
What we used
"Permanent Magnet Synchronous Motor (PMSM) drives employing sensorless control offer reduced cost, improved reliability and increased power density by eliminating mechanical position sensors. However, reliable operation from standstill to the field-weakening region remains challenging because of startup synchronization difficulties, electromechanical oscillations, and rotor position estimation errors at high speeds." "FOC has become the dominant control strategy for high-performance PMSM drives because it enables the independent regulation of torque- and flux-producing current components, allowing the machine to exhibit dynamic characteristics comparable to separately excited dc motors." Checked via browser.search, 2026-11-20
- Cluster figures reused from earlier articles in this series (read back from the files, not re-derived): Example Labs bench arm
What we used
Two links 0.425 m each, link 1 is 4 kg, link 2 is 3 kg, payload 5 kg at 0.85 m, shoulder 126.8 N·m, elbow 40.7 N·m, wrist 11.0 N·m, 100:1 two-stage planetary at 93% efficiency, motor 1.36 N·m at 1,500 RPM, shoulder mechanical power 199 W at the joint (#20, #29); 48 V bench pack 13S1P NMC, 48.1 V nominal, 54.6 V full, 39.0 V empty, 168 Wh (#6); 48 V, 100 Ah shift pack, 4,800 Wh (#20); BLDC 85–92% efficiency, 2,000–10,000 RPM typical (#20); three-phase inverter, six MOSFETs, gate driver, shunts (1 mΩ at 10 A gives 10 mV), current-sense amplifier gains 5/10/20/40 V/V, DRV8323 and DRV8305, bootstrap against charge pump, shoot-through and dead time (#23); PID, output = Kp × error + Ki × (error added up over time) + Kd × (how fast the error is changing), Ziegler–Nichols 1942 (#3); 1 arcminute of backlash moves the tool about 0.29 mm at one metre (#20, #21)