The previous practice of relying on the off-diagonals being zero caused problems with conditioning of the magnetometer fusion on one flight. By storing the variances when the learning inhibit becomes active and ensuring that the rows and columns in the covariance matrix for the inhibited states are always zero, the observed numerical conditioning error has been eliminated for replay of the problem flight log .
Doing so can casue large jumps in GPS position and innovation check errors after landing and also reduces the effectiveness of pre-flight innovation consistency checks.
Previously GPS quality checks were only run until the EKF origin was set. This meant that they could not be used by other pre-flight checks.
This change ensures that checks will always be run when the vehicle on-ground or not using GPS to enable use by external preflight checks.
When using a union of flags and integer value it is safer to initialise the value to 0 rather than memset the flags because the flags may not define all bits in the integer.
This bug caused X and Y delta velocity bias state variance to be reset to the same value as the Z axis when learning was inhibited.
Documentation has also been updated.
This change removes the following compiler error when building using the ARM cross compiler.
/Users/paul/src/Firmware/src/lib/ecl/EKF/ekf_helper.cpp:45:12: error: 'std::abs' has not been declared
using std::abs;
Make the target EKF rate an integer multiple of the IMU rate. This slightly increases the average prediction time step for the EKF from just over 10msec to 12msec, but the variation reduces significantly which makes filter tuning more deterministic.
Improve the algorithm used to adjust the collection time criteria to reduce jitter in the correction.
This was incorrectly using the IMU (1/250 sec) timestamp instead of the EKF (1/100 sec) value.
The corresponding accelerometer limit has been made a parameter and adjusted to match previous behaviour.
This is a functionally equivalent. It moves all of the code for the terrain estimator into a single function call from the main filter update, making it clear that it is independent of the main filter.
The tilt compensation being applied previously was based on a flat earth geometric model assuming perfect tilt knowledge which reduces the effect of range errors on height error as the vehicle tilts. however in the real world, variations in terrain gradient and uncertainty in vehicle tilt and sensor alignment tend to increase height error with tilt, so the adjustment of observation variance with tilt has been removed given we do not have a valid mathematical model on which to base it.
Paul Riseborough