Understanding EKF Innovation Spikes in ArduPilot Logs

TL;DR: EKF innovations are the heart of position-control safety in ArduPilot. XKF3.IVN/IVE/IVD measure the disagreement between what the EKF expected and what the sensors reported. XKF4.SV/SP/SH are normalised test ratios — a value above 1 means that measurement was rejected. Sustained XKF4.SP above 0.5 with rising XKF4.SS bits is what the EKF check uses to trigger lane switch and ultimately EKF failsafe. Most innovation spikes trace to GPS multipath, magnetometer interference, or vibration corrupting IMU integration.

What "innovation" means in EKF terms

Innovation is the residual between the EKF's prediction of a measurement and the actual measurement. When the GPS reports velocity north, the EKF compares it to its own predicted north velocity and writes the difference into XKF3.IVN. Small innovations mean the model and sensors agree; large innovations mean either the model is wrong, the sensor is wrong, or both.

The same machinery runs for position, magnetometer readings, yaw, and airspeed. Each measurement family has its own innovation channel in XKF3 (for EKF3, the default) and the equivalent in NKF3 if you're running EKF2.

The XKF3 fields you'll actually read

XKF3  TimeUS  C  IVN  IVE  IVD  IPN  IPE  IPD  IMX  IMY  IMZ  IYAW  IVT  RErr  ErSc

• C — which EKF core this row belongs to. ArduPilot runs multiple cores in parallel; XKF4.PI tells you which is primary.
• IVN/IVE/IVD — velocity innovations in m/s, north/east/down.
• IPN/IPE/IPD — position innovations in metres, north/east/down.
• IMX/IMY/IMZ — magnetometer innovations.
• IYAW — yaw innovation, in radians.
• IVT — airspeed innovation (Plane).
• RErr — this core's accumulated relative error against the active primary core.
• ErSc — consolidated error score; higher means less healthy.

The XKF4 test ratios — the most useful number in the log

XKF4  TimeUS  C  SV  SP  SH  SM  SVT  errRP  OFN  OFE  FS  TS  SS  GPS  PI

The headline numbers are SV, SP, and SH. These are not raw innovations — they're squared innovation test ratios normalised against the EKF's expected uncertainty for that measurement. A value below 1 means the measurement passed the gate and was used; a value above 1 means it was rejected. That single rule makes XKF4.SP the cleanest health indicator in the log.

• SV — velocity test ratio.
• SP — position test ratio. The most watched.
• SH — height test ratio.
• SM — magnetometer test ratio.
• SVT — airspeed test ratio (Plane).
• FS — filter fault status bitmask.
• SS — solution status bitmask; tells you which sensors are actively trusted right now.
• PI — primary core index. A change here is a lane switch.

Confirming a spike in Mission Planner

Plot XKF3.IPN, XKF3.IPE, and XKF4.SP on one axis. A real disturbance shows up as a position innovation excursion followed by the SP test ratio crossing 1 — that's the firmware logging "I just rejected a GPS update". The EKF is allowed to clear test ratios over time as fresh measurements come in; if SP stays above 0.5 for sustained periods you have a stuck disagreement, not a transient glitch.

Confirming it in MAVExplorer

MAV> graph XKF3.IVN XKF3.IVE XKF3.IVD
MAV> graph XKF4.SV XKF4.SP XKF4.SH
MAV> graph XKF4.PI

If XKF4.PI goes from 0 to 1 (or 1 to 0) mid-flight, that's a lane switch. Two switches within a few seconds is a sign the EKF can't pick a primary it trusts.

Why innovations spike — ranked by what we see most often

1. GPS multipath or RFI. Position innovations climb because the GPS reports a position physics says you can't be at. See our GPS glitch post for the receiver-side story.
2. Magnetometer interference. Current-draw peaks bias the compass; XKF3.IMX/Y/Z grow as a function of throttle. COMPASSMOT calibration and physical separation are the fixes.
3. Severe vibration. VIBE.VibeZ above 30 m/s² corrupts IMU integration, so the EKF's predicted velocity grows incorrectly — innovations chase that error. The cause is mechanical, not EKF-tunable. See our VIBE message guide for the mitigation chain.
4. Aggressive manoeuvres. Sharp pitches can transiently spike XKF3.IVD. If the spike is short and SP doesn't cross 1, this is normal.
5. Bad airspeed (Plane). XKF3.IVT climbing during cruise is your tell that the pitot or ARSPD_RATIO calibration is off.

Tuning the gates — carefully

Each measurement family has an innovation gate parameter expressed in tenths of standard deviations:

• EK3_POS_I_GATE — GPS position. Default 500 (5.0 sigma).
• EK3_VEL_I_GATE — GPS velocity.
• EK3_HGT_I_GATE — barometer height.
• EK3_MAG_I_GATE — magnetometer.
• EK3_YAW_I_GATE — compass yaw.

Raising a gate makes the filter more tolerant of noisy sensors but lets through bad measurements that would otherwise have been rejected. Lower one only if you understand exactly what your sensor's noise floor is — tightening the gate on a vehicle with real EMI gives you constant EKF failsafes. The gates are a backstop, not a fix.

From innovation spike to lane switch to failsafe

ArduPilot escalates in three steps:

1. Innovation arrives high. EKF rejects the measurement; XKF4.SP crosses 1 for one update.
2. If multiple rejections accumulate, the EKF check (running at 10 Hz) writes ERR Subsys=16 ECode=2 (EKFCHECK_BAD_VARIANCE).
3. If the bad variance condition persists past FS_EKF_THRESH for FS_EKF_ACTION-determined duration, the autopilot triggers EKF failsafe — ERR Subsys=17 ECode=1 — and lane-switches the EKF core (XKF4.PI flips). If the secondary core is also bad, the autopilot drops to Land.

Reading that chain top-to-bottom in the log tells you whether you saw a transient (single SP spike that cleared) or a degradation (failsafe fired).

When LogHat helps — and when it doesn't

LogHat plots XKF3 and XKF4 with the rejection thresholds drawn on the axis and annotates lane switches in the 3D replay timeline. What we can't do is tell you which physical sensor degraded — that needs you to swap hardware or run a hover-stand measurement. The log narrows the search to magnetometer, GPS, or vibration with high confidence; the screwdriver work is yours.