I very nearly had a crash today on a quadplane running 4.3.8. During QHover takeoff, I lost my GPS module (and compass with it). As soon as the bird fell back to AHRS DCM, it started dropping like a rock. We managed to slow it down just before it hit.
QTUN.CRt
locks up at 1.5m/s when the fallback to DCM happened, so the bird thought it was climbing as it dropped like a rock, which explains why it was descending more and more rapidly (next time, I’m gonna make sure I have QStabilize on a switch). What I don’t understand is why the baro and IMU were not enough to estimate climbrate. This lockup has to be a bug, right?
Onboard log here: AQ000_2023-10-12 11-48-00_hover.bin - Google Drive
The problem is reproducible. I armed the aircraft in QHover on the ground (VTOL power disconnected), and yanked the GPS module off. 7 seconds later, the QTUN.CRt locks up. AQ000_2023-10-16 13-34-00_ground.bin - Google Drive
This incident reminded me of another GPS loss incident we had a year ago on an ALTI Transition. At the time, I didn’t look at it as closely as I should have.
The same thing occurred, though this time the canonical climb rate locked when the aircraft was in a descent, so the aircraft tended to enter an uncontrolled climb rather than descent.
Log here: AT001_2022-12-12 14-33-39_full.bin - Google Drive
Got around to digging into this more again today. This is easily reproducible in sim.
So, the DCM can only estimate velocities if the GPS is present and has a good fix. This makes some sense I suppose, at least for x/y speed. But this behavior means that QHover can be a death sentence in a GPS failure (if you lose GPS at max climb rate, you can’t stop a descent, and if you lose it at max descent rate, you can have a flyaway). I think we need one of two things to change:
- DCM should be able to measure climb rate some other way, like low-passed derivative of altitude.
- The EKF should not necessarily fall back to DCM when losing only GPS.
I think I could probably implement 1, but the right answer is probably something along the lines of 2.