Problem
Calibration results currently vary with DC bus voltage. Observed while bring-up testing on feature/encoder-distortion-compensation:
- Inductance cal (biased-AC injection) produces different Ld/Lq curves at different bus voltages (e.g. 50 V vs lower bus). Old values also confounded by the graph-FOC ISR fight fixed in d44975d.
- Flux linkage cal (
cal Motor.PMSM.FluxLinkage) hits its voltage ceiling at very low RPM on lower bus (58-96 RPM), leaving too few/too clustered samples for a robust (psi_m, V_off) least-squares fit. psi_m estimates have swung ~10x between runs (0.0986 vs 0.72-0.96 Wb) under different conditions (bus, ellipse fit on/off).
- The (psi_m, V_off) joint fit is badly conditioned when samples cluster in a narrow low-speed band; median vs fit disagreed by ~30%.
Voltage-derived calibration quantities (psi_m, Ld/Lq) should be motor constants and must not depend on the bus.
Known bugs / confounds to untangle
- Ellipse fit (Layer 1) phi is not repeatable run-to-run (observed 1.25-3.7 deg); a saved fit with a wrong phi measurably degrades commutation and corrupts downstream cal (flux cal ceiling RPM improved 58 -> 96 RPM immediately after
encfit clear). Do not trust/save a fit until phi repeats within ~0.2 deg and capture is lock-gated.
PWM_FindSafeSamplePoint (f96e749) can select the all-high-side window where low-side shunts read zero once duties leave the centered region (saturation); turns any PI railing event into feedback collapse. Needs review/fix.- Inductance cal sometimes needs AC amplitude retries ("weak response"); measured L may depend on the AC perturbation amplitude on a saturating machine - define a standard amplitude per bias point for comparability.
- ENCLIN (encoder linearity sweep) measurement exists and runs, but its residual is dominated by torque-lag/cogging in single-direction mode; FWD+REV dual-pass averaging is implemented but the INL result is not yet validated/used anywhere.
Acceptance
- Repeating
cal all (or each stage) at two different bus voltages (e.g. 24 V and 50 V) yields poles, cycles/rev, offset, resistance, Ld/Lq curve, and psi_m within stated tolerances of each other.
- Flux cal documents minimum speed-range/sample-spread requirements and fails with a clear message when the bus is too low to meet them, instead of returning a garbage fit.
Problem
Calibration results currently vary with DC bus voltage. Observed while bring-up testing on
feature/encoder-distortion-compensation:cal Motor.PMSM.FluxLinkage) hits its voltage ceiling at very low RPM on lower bus (58-96 RPM), leaving too few/too clustered samples for a robust (psi_m, V_off) least-squares fit. psi_m estimates have swung ~10x between runs (0.0986 vs 0.72-0.96 Wb) under different conditions (bus, ellipse fit on/off).Voltage-derived calibration quantities (psi_m, Ld/Lq) should be motor constants and must not depend on the bus.
Known bugs / confounds to untangle
encfit clear). Do not trust/save a fit until phi repeats within ~0.2 deg and capture is lock-gated.PWM_FindSafeSamplePoint(f96e749) can select the all-high-side window where low-side shunts read zero once duties leave the centered region (saturation); turns any PI railing event into feedback collapse. Needs review/fix.Acceptance
cal all(or each stage) at two different bus voltages (e.g. 24 V and 50 V) yields poles, cycles/rev, offset, resistance, Ld/Lq curve, and psi_m within stated tolerances of each other.