[j-w9]

Adds an "Auto fit" button to the first notch that sets it up from the log: ESC telemetry first, then throttle, with peak detection to verify the tracking and the noise floor setting the attenuation.

Method: the pre-filter spectrum of each FFT window in the analysis time is resampled against the tracked frequency and averaged. Noise that follows the motors adds up into peaks at multiples of the tracked frequency, and noise that doesn't is spread out.

• ESC is rejected if the largest peak is well away from a whole multiple of the ESC frequency, with a message to check the motor poles.
• Harmonics whose peak is at least 6 dB above the local noise floor are enabled. ATT is how far the largest one stands out, limited to 10–50 dB.
• Rules of thumb: ESC FREQ is 80% of the lowest motor frequency seen. For throttle, REF is the mean throttle and FREQ is fitted from the peaks, and FM_RAT is lowered to keep tracking below REF. BW is FREQ/2. OPTS is a single notch per harmonic, multi-source for ESC. Throttle uses a single average throttle notch, as that is what the fit models.

These are all in one block at the top of AutoFit.js and are the main thing I'd like your view on.

Tested on six batch-logged copter logs, with estimated noise above 30 Hz:

• On the four flights where it picked ESC, the fit matched or beat the existing hand-set notch.
• On a ground run, it reports the ESC frequency mismatch and that throttle is zero.
• Synthetic spectra with known peaks recover the harmonics and attenuation, and a 14/12 pole error is caught.

If the notch needs more filters than F4 boards allow (24), the result says so.

Known limitation: a fixed resonance that sits near a motor harmonic in hover can be picked up as a harmonic.

[AP-Review]

Deprecated — see below for the updated review.

Previous review (2026-11-09)

Automated review note — AI-generated (Claude+Codex), validated against the live diff. Please sanity-check before acting.
Verdict: COMMENT

Reviewed at head c46ecf4014.
Full report: https://firmware.ardupilot.org/Tools/APReview/DevCallReviews/PRReviews/ardupilot/webtools/363/1.html#prWebTools-363

The auto-fit works on synthetic spectra and lints cleanly; there are no blockers, but a few things should be fixed before it leaves draft.

Should fix

• FilterReview/AutoFit.js:359: nothing bounds the total notch count. Multi-source × up to 8 harmonics × motors (× IMUs if EnableOnAllIMUs is kept) can exceed HAL_HNF_MAX_FILTERS, e.g. an octa with 4 harmonics gives 32 notches, so the firmware boots with "Too many notches". Count the notches and warn, or cap the harmonics.
• FilterReview/AutoFit.js:257: the per-motor throttle fit uses the mean root thrust across motors, but the multi-source notch it enables tracks each motor separately, so the fitted FREQ is off whenever motors run at different thrust. Resample each motor's contribution against its own thrust.

Lower priority

• FilterReview/AutoFit.js:335: calculate() and up to three fits run synchronously from the button and can take several seconds with no overlay. Wrap the fit in loading_call like the other heavy controls (index.html:472).

Minor (2)

• FilterReview/AutoFit.js:375: the auto_fit_info label is never cleared by reset() or a failed fit, so it carries over to the next log.
• FilterReview/AutoFit.js:321: rounding REF to 0.01 turns a near-idle mean thrust (< 0.005) into REF=0, which disables tracking while still reporting success. Keep more precision or reject it.

[AP-Review]

Automated review note — AI-generated (Claude+Codex), validated against the live diff. Please sanity-check before acting.
Verdict: COMMENT

Reviewed at head 110f578cdb.
Full report: https://firmware.ardupilot.org/Tools/APReview/DevCallReviews/PRReviews/ardupilot/webtools/363/2.html#prWebTools-363

Most of the previous round is fixed at 110f578, but one new should-fix problem was found in the throttle fit, along with a couple of smaller ones. Since c46ecf4, the per-motor throttle bias, the missing loading overlay, the stale status label and REF rounding to zero are all fixed, and the notch-count warning now catches the octa case.

Should fix

• FilterReview/AutoFit.js:159: the throttle fit's fundamental can come out 5-10% off, and its 2nd/3rd harmonics then get dropped. The candidate only has to be a maximum within ±10 bins, but its excess uses the highest point within ±10% of its position, so a nearby noise maximum borrows the real peak's height. On a synthetic 100 Hz spectrum this gave FREQ 105 and HMNCS 3 instead of 7. Only accept k if it is the maximum of its own peak window, or snap the position to that maximum.

Lower priority

• Notch budget (still open): FilterReview/AutoFit.js:341 counts only logged IMUs, while the firmware multiplies by every used IMU. The check at line 400 also ignores an enabled INS_HNTC2, which the firmware adds to the same limit. A quad ESC fit with HMNCS=7, EnableOnAllIMUs and only IMU0 logged gives no warning, but needs 36 filters with 3 IMUs. Use num_gyro from INS_GYRx_ID and add the second notch's count.

Minor (2)

• FilterReview/AutoFit.js:353: after a valid log, loading a log that is rejected (no PARM, or no gyro data) leaves Gyro_batch in place and Auto fit enabled. Auto fit then fits the old log into the reset parameters and reports success. Disable #auto_fit in reset() and enable it after a successful load.
• FilterReview/AutoFit.js:327: REF is rounded to 0.01 but FREQ is not rescaled. Near idle this detunes the notch, e.g. a mean of 0.006 writes REF=0.01 and the notch tracks 77 Hz for a 100 Hz peak. Keep more precision, or fit against the rounded REF.