The Clarke and Park Transforms: Freezing a Spinning Motor's Math

Three phases are two too many — the Clarke transform

A balanced three-phase machine is driven by currents i_a, i_b, i_c that always satisfy i_a + i_b + i_c = 0 (there is no neutral return in a star-connected motor). That constraint means the three currents carry only two independent pieces of information. Edith Clarke's 1938 transform makes this explicit by projecting the three phase axes — physically 120° apart in the stator — onto a fixed two-axis orthogonal frame she called α (aligned with phase a) and β (90° ahead).

In the common amplitude-invariant form (gain = 2/3), the forward Clarke transform is:

• iα = i_a
• iβ = (i_a + 2·i_b)/√3 (equivalently iβ = (i_b − i_c)/√3)

The 2/3 gain preserves the peak amplitude: a phase current with peak Î = 10 A produces a space vector of magnitude 10 A in the αβ plane. That's the convention almost every motor-control firmware library uses (TI's InstaSPIN, ST's MC-SDK, the C2000 IQmath routines) because it maps cleanly to the physical current a hardware sensor reads.

The alternative power-invariant form uses gain √(2/3) ≈ 0.8165. It scales αβ so that instantaneous power computes directly as P = v_α·i_α + v_β·i_β with no 3/2 factor, which is convenient in analysis and grid applications but inflates the vector magnitude by √(3/2) ≈ 1.225 relative to the phase peak. Pick one convention and stay in it — mixing a power-invariant Clarke with an amplitude-invariant torque equation is a classic firmware bug that produces a systematic 22.5% torque-constant error.

Park's trick: spin the frame until the sinusoids stop

The Clarke output is still AC. At 3,000 rpm on a 4-pole-pair (8-pole) machine the electrical frequency is f_e = p·N/60 = 4·3000/60 = 200 Hz, so i_α and i_β are 200 Hz sinusoids sitting 90° apart. Robert Park's 1929 insight (developed for synchronous-machine analysis and later described by the AIEE as the most influential power-engineering paper of its era) was to rotate the reference frame at exactly the electrical angular velocity ω_e, locking it to the rotor flux. In that spinning frame the vector that was whirling around becomes stationary — two DC numbers.

The Park transform is a plain 2×2 rotation by the measured/estimated rotor electrical angle θ_e:

• i_d = i_α·cos θ_e + i_β·sin θ_e
• i_q = −i_α·sin θ_e + i_β·cos θ_e

By convention the d-axis (direct) is aligned with the rotor's permanent-magnet flux and the q-axis (quadrature) leads it by 90° electrical. This alignment is the entire point: current on the q-axis crosses the rotor field at 90° and produces torque (like a well-timed push on a merry-go-round), while current on the d-axis is parallel to the field and produces no torque — it only strengthens or weakens the flux. The full chain abc → αβ → dq is sometimes fused into one 3-input, 2-output matrix and called the dq0 transform, where the extra 0-sequence component captures any imbalance (zero in a healthy star-connected motor).

Why the two DC axes matter: decoupled torque control

Once you're in the dq frame the electromagnetic torque of a surface-mount PMSM reduces to a strikingly clean expression:

• T_e = (3/2)·p·λ_pm·i_q

where p is the number of pole pairs and λ_pm is the rotor magnet flux linkage (Wb). Torque is proportional to i_q alone. That is the holy grail: a single DC number sets the torque, exactly like the armature current in an old brushed DC motor — which is why FOC is often called making an AC motor 'behave like a DC motor.' For an interior-PMSM with saliency (L_d ≠ L_q) a reluctance term appears: T_e = (3/2)·p·[λ_pm·i_q + (L_d − L_q)·i_d·i_q], which is exploited by MTPA (maximum-torque-per-amp) strategies that inject negative i_d.

The control architecture that falls out is elegant:

• Measure two phase currents (the third from i_a+i_b+i_c=0), run Clarke and Park to get i_d, i_q.
• Two PI controllers regulate i_d → 0 (or a field-weakening value) and i_q → torque command. Because the references are DC, a modest 1–2 kHz current-loop bandwidth with zero steady-state error is trivial — no chasing moving sinusoids.
• Add cross-coupling decoupling feedforward: v_d includes −ω_e·L_q·i_q and v_q includes +ω_e·(L_d·i_d + λ_pm), cancelling the back-EMF and speed-dependent coupling.
• Run the inverse Park (rotate v_d, v_q back by θ_e) then space vector modulation to generate the inverter gate signals.

The rotor angle is everything — sizing and sourcing θ_e

The Park transform lives or dies on the accuracy of the rotor electrical angle θ_e. A 10° electrical angle error costs cos(10°) ≈ 1.5% of available torque and dumps the rest into a wasteful, heating d-axis current; a 90° error and the motor produces zero useful torque while drawing full current. So the angle source is a first-order design decision:

• Incremental encoder — e.g. a 2,500 line/rev unit gives 10,000 counts/rev after quadrature = 0.036° mechanical resolution; multiply by pole pairs for electrical. Needs an index or alignment routine to find absolute angle at startup.
• Resolver — the aerospace/EV workhorse, robust to 150 °C and vibration; a resolver-to-digital converter (e.g. AD2S1210) outputs 12–16 bit angle, ~0.005° resolution, with tracking bandwidth in the kHz.
• Sensorless observers — a sliding-mode or Luenberger flux observer, or a PLL riding on the back-EMF, estimates θ_e from the measured currents and the voltage model. Excellent above ~5–10% of rated speed; below that the back-EMF vanishes and high-frequency signal injection (exploiting L_d ≠ L_q saliency) takes over.

Timing is a hidden trap. At 200 Hz electrical, the vector advances 360°·200·T_s per control period. Sampling currents and applying the resulting voltage one PWM period later — say T_s = 100 µs (10 kHz loop) — the rotor has already rotated 200·360·100e-6 = 7.2° electrical. Good firmware compensates by predicting θ_e forward by 1.5 sample periods to align the transform with the instant the voltage is actually applied.

Numbers on real hardware

Consider an EV traction inverter driving an 8-pole (p = 4) interior-PMSM, 150 kW peak, 400 V DC bus, base speed 4,500 rpm. At base speed f_e = 4·4500/60 = 300 Hz. The control loop runs at 10 kHz on a 200 MHz automotive MCU (TI TMS320F28379D or Infineon AURIX); the entire Clarke, Park, two PI loops, inverse Park and SVM chain must finish in well under 100 µs. On a C2000 with the trig lookup and IQ-math, that pipeline is on the order of a few microseconds — small next to the 100 µs budget, leaving room for the observer.

• Phase current peak at 300 A_rms ≈ 424 A; the current sensors (Hall or shunt+isolated amp) must resolve to ~1 A to keep i_d, i_q ripple low.
• In field weakening above base speed, the controller commands negative i_d (often −100 to −200 A) to buck the magnet flux and keep the q-axis voltage below the DC-bus ceiling of ~V_dc/√3 ≈ 231 V per phase peak. This is only possible because the dq frame separates flux (d) from torque (q).

At the small end, the same math runs on a $1 STM32G4 spinning a 3,000 rpm drone motor at a 20–40 kHz loop, or a hobby FOC board (VESC, ODrive, SimpleFOC). The transform equations are identical from a 3.7 V drone ESC to a megawatt wind-turbine grid inverter — only the scaling, the angle source, and the modulation index change.

Failure modes, limits, and best practice

The transforms themselves are exact linear algebra — they don't 'fail.' What fails is the surrounding assumptions:

• Convention mismatch. Amplitude-invariant Clarke (2/3) with a torque equation derived for power-invariant scaling gives a fixed √(3/2) ≈ 1.225 gain error. Symptom: measured torque constant off by ~22% versus datasheet. Fix: derive the whole chain in one convention.
• Angle sign / offset error. A swapped encoder direction or a magnet-alignment offset makes θ_e count backward or lead by a fixed angle. Symptom: motor draws current but produces little torque, or spins the wrong way, or oscillates. Always run a d-axis alignment (inject i_d, let rotor snap to θ = 0) at commissioning.
• Sensor imbalance / DC offset. A gain mismatch between the two current channels injects a 2·f_e ripple into i_d, i_q; a channel offset injects a 1·f_e ripple. These beat against the PI loops and cause audible whine and torque ripple. Calibrate offsets at zero current every power-up.
• Low-speed sensorless dropout. Back-EMF observers lose the angle near standstill because the estimated flux signal vanishes into noise; below ~5% speed you must switch to HF injection or accept open-loop I/f startup.
• Angle latency at high f_e. Without the 1.5-sample angle-advance compensation, the dq frame lags the rotor, cross-couples d and q, degrades current-loop phase margin, and can destabilize the loop at high speed.

Best practice: pick the amplitude-invariant convention to match hardware sensors, verify i_a+i_b+i_c ≈ 0 as a live health check, hold i_d = 0 below base speed on surface-PMSMs, and always compensate transport delay in θ_e. Do that, and three whirling 300 Hz sinusoids collapse into two rock-steady DC numbers you can control with a textbook PI loop.

Amplitude-invariant vs. power-invariant Clarke transform, and what each reference frame looks like to the controller

Property	abc (stationary phases)	αβ (Clarke)	dq (Park)
Number of variables	3 (redundant, sum=0)	2 orthogonal	2 orthogonal
Time behavior at speed	Sinusoidal (f_e = p·N/60)	Sinusoidal, 90° apart	DC (constant in steady state)
Frame	Stationary	Stationary	Rotating at ω_e
Amplitude-invariant gain (2/3)	peak = Î	|αβ| = peak Î	d,q = phase peak
Power-invariant gain (√(2/3))	peak = Î	|αβ| = η√(3/2)	power computes directly as v_d·i_d+v_q·i_q