Explainer
Mounting a Motor Encoder Behind a Belt Reduction
Short answer
No, not for commutation: field-oriented control switches phase current against the rotor's electrical angle, which only a sensor on the motor shaft reports directly. Behind a 10:1 belt, a 4096-count encoder gives about 410 counts per motor revolution — enough counts per pole to turn a 6-pole motor, but belt lash and multiplied noise put a load-dependent error inside the current loop. Use the belt-side sensor for velocity only; commutate from the motor shaft and add an output encoder if you need load position.
On this page
- Does the commutation encoder have to sit on the motor shaft?
- How many counts per motor revolution survive a 10:1 belt?
- Is 68 counts per pole enough to commutate smoothly?
- Why does belt lash end up inside the current loop?
- When is an encoder behind the belt good enough?
- How do I commutate and still read true load position?
- What confirms the placement once the machine runs?
- Questions asked
No — not for commutation. Field-oriented control times its phase currents against the rotor's electrical angle, and that angle is fixed to the windings only at the motor shaft. An absolute encoder on the driven pulley reports output angle; the drive has to infer motor angle through the belt ratio, and everything the belt adds — lash, stretch, tooth-mesh error — lands in the current loop as angle error that moves with load. Run the arithmetic before buying hardware: a 4096-count magnetic encoder behind a 10:1 reduction yields about 410 counts per motor revolution, roughly 68 per pole on a 6-pole motor, which clears the usual 10-counts-per-pole rule of thumb. Resolution is not what kills this arrangement; the belt is. Commutate from the motor shaft, and let a behind-the-reduction encoder carry velocity or load position only.
Does the commutation encoder have to sit on the motor shaft? #
Yes, whenever the drive commutates by rotor angle: field-oriented control needs the electrical angle between rotor field and stator windings, and that relationship is fixed only at the motor shaft, upstream of every gear and belt.
A field-oriented drive does two things every control cycle: read where the rotor is, then aim the stator field at it. Torque comes from holding the stator current vector ninety electrical degrees off the rotor field, so any angle error costs torque immediately. Put the sensor behind a reduction and the drive no longer reads rotor angle at all — it reads load angle and divides by the ratio, so the angle it commutates with is rotor angle plus every error in the path between. A fixed offset can be calibrated out; a load-dependent one cannot, because it changes sign each time the torque reverses. The trade-off here is the same one that settles timing belt versus direct mount for a rotary encoder: a belt buys freedom of placement and pays for it in angle fidelity.
How many counts per motor revolution survive a 10:1 belt? #
A 4096-count encoder behind a 10:1 reduction delivers 4096 ÷ 10 = 409.6, about 410 counts per motor revolution — the number to judge against the motor's pole count, not against the encoder's nameplate.
Referred resolution is simply encoder counts divided by ratio. The values below are example inputs for a calculated example, not measured results: take the exact counts-per-turn from the sensor datasheet and the ratio from the pulley tooth counts before judging your own axis.
| Quantity | Example value | Where it comes from |
|---|---|---|
| Encoder resolution | 4096 counts/rev | Example input; confirm on the encoder datasheet |
| Belt reduction | 10:1 | Example input; pulley tooth counts |
| Counts per motor revolution | 4096 ÷ 10 = 409.6 | Calculated |
| Motor poles | 6, which is 3 pole pairs | Example input; motor datasheet |
| Counts per pole | 409.6 ÷ 6 ≈ 68 | Calculated |
| Counts per pole pair | 409.6 ÷ 3 ≈ 137 | Calculated |
Read the pole-count line carefully: datasheets quote both poles and pole pairs, and the result moves by a factor of two between the two conventions.
Is 68 counts per pole enough to commutate smoothly? #
On resolution alone it passes: 68 counts per pole is roughly seven times the working rule of thumb of about 10 counts per pole, the level below which commutation turns coarse — six-step Hall commutation, three Hall signals with six states per pole pair, runs near 3 counts per pole and still spins motors.
Two checks refine that pass. A long-standing encoder-selection rule of thumb keeps speed times counts above 100,000 — RPM × counts per turn > 100,000 — and with 409.6 effective counts at the motor that needs about 244 RPM at the motor shaft, 24 RPM at the output, before velocity feedback goes coarse. Quantization referred to the motor is 360° × 10 ÷ 4096 = 0.88° mechanical per count, which is 2.6° electrical on the 6-pole machine, so ±2 counts of sensor jitter is about ±5° electrical of commutation error. Workable at speed, thin at low speed. The discipline is the same one behind choosing encoder resolution to match a stepper's step response: state the demand first, then divide the counts against it.
Why does belt lash end up inside the current loop? #
Because the belt is a spring in the feedback path: when torque reverses, lash and belt stretch change the angle between the encoder pulley and the rotor, so the inferred electrical angle swings with load instead of sitting at a fixed, calibratable offset.
Torque per amp holds up only while commutation angle error stays small: an error of δ degrees costs torque roughly as cos δ and pushes current onto the direct axis, where it only heats the motor. Take 0.5° of lash at the encoder pulley as an example input — hold the motor and turn the output by hand to measure yours. Referred to the motor it is 0.5° × 10 = 5° mechanical, and × 3 pole pairs = 15° electrical, so torque per amp falls to cos 15° ≈ 0.97, about a 3% loss as a calculated example for those inputs. The loss is not the problem; the sign change is. At low speed the motor hunts across the dead band, and raising loop gains turns that nonlinearity into a limit cycle, because the controller is now correcting its own sensor. Referred noise stacks on top of lash: ±1 count at the output reads as ±10 counts at the motor, which is why a precision ring-style absolute encoder tolerates this arrangement better than a generic board-level magnetic sensor, and why neither matches the same sensor on the motor shaft.
When is an encoder behind the belt good enough? #
For velocity control at speed where nothing needs low-speed smoothness, position hold or anti-cogging — the e-bike class of duty — the example above clears every rule of thumb, which is why Hall commutation at about 3 counts per pole already serves that class of machine.
Commit to the belt-side sensor only when all of these hold: velocity is the controlled variable; the axis stays above the speed floor set by the 100,000 rule, about 244 RPM at the motor in this example; the load does not slam between directions; and nobody will ask the axis to creep or hold position. Put the encoder on the motor shaft when any low-speed or standstill torque, position hold, anti-cogging compensation or torque lineularity matters, and also before running the drive's motor calibration: those routines find the electrical offset by moving the motor, and running them through a belt smears the result by the lash. That is the decision rule, not a preference — below the speed floor or with reversing loads, it is the motor shaft.
How do I commutate and still read true load position? #
Use two encoders: the motor-shaft sensor feeds commutation and the inner velocity loop, and a second absolute encoder behind the reduction reports load position and sets the machine zero.
That is the standard dual-loop split. The outer position loop closes on the load-side encoder, the fast inner loops on the motor-side sensor, and the belt's lash sits between the two — visible to the position loop, where the controller can profile around it, and outside the current loop, where it would corrupt commutation. Calibrate the machine zero against the output encoder after homing. When you fit the motor-shaft encoder, plan the joint between its shaft and the motor's rear shaft around a flexible shaft coupling, so small misalignment never loads the encoder bearing.
What confirms the placement once the machine runs? #
Three logged signals settle it: torque-producing current at constant low speed should sit flat across a revolution, the reported angle must not step when load torque reverses, and jitter at standstill with the shaft held is sensor noise reaching the loop, not motion.
- Run the drive's encoder calibration with the motor-shaft sensor connected, and repeat it after any belt or pulley change.
- Hold a low constant speed and log torque-producing current; ripple that follows load direction points at lash or stretch in the sensing path.
- Reverse the load torque gently and watch the reported angle; a step means backlash is inside the feedback path.
- Disable the drive, hold the shaft, and watch the angle readout; jitter there shows up as current chatter or hum once enabled.
One safety note: disable the drive and follow the machine's lockout procedure before touching the belt, pulleys or encoder, and run the reversal check with the guards closed.
Questions asked
My motor datasheet gives pole pairs instead of poles. How does that change the math?
Divide by poles, not pole pairs. A 6-pole motor is 3 pole pairs, so 410 counts per motor revolution is about 68 counts per pole or about 137 per pole pair. The 10-counts guideline is quoted per pole, which equals about 20 counts per pole pair; mixing the two conventions shifts your result by a factor of two, enough to turn a marginal pass into a fail.
Does a higher-resolution encoder fix the belt-side problem?
No. Resolution fixes quantization, not lash. Counts per turn set the smallest angle step the sensor can report, but the load-dependent error from belt spring and lash is set by the mechanical train and is multiplied by the ratio no matter how fine the sensor is. A finer encoder behind the belt still commutates through the same spring, so spend the upgrade on a motor-shaft sensor instead.
Does the same argument apply behind a gearbox instead of a belt?
Yes for placement: commutation sensing still belongs on the motor shaft, and the counts-per-revolution arithmetic is identical for any ratio. The error character differs. A gear train holds a fairly fixed mesh lash, while a belt adds spring wind-up that grows with torque, so a belt's angle error moves more with load. Either way the error sits inside the current loop and is multiplied by the ratio.
Can the output encoder close velocity while the motor encoder only commutates?
It can, but keep the fast inner velocity loop on the motor-shaft sensor and use the output encoder for the slower outer position loop. Closing velocity on the far side of the belt puts the belt's resonance and lash inside the velocity loop, which shows up as rough speed regulation and audible growl at low speed.