If we start GPS aiding before the gyro bias variances have reduced, glitches on the GPS can cause attitude disturbances that degrade flight accuracy during early flight.
Co-variances were being re-zeroed after being set. This meant that the initial declination learning was sensitive to measurement errors which could result in poor initial yaw accuracy.
Fixes bugs that prevented planes being able to reset yaw to GPS to recovery from takeoff with a bad magnetoemter.
1) If the velocity innovation check had not failed by the time the in-air transition occurred, then the yaw reset would not be performed
2) The velocity states were not being reset
3) The non fly-forward vehicle (copter) reset could occur first and effectively lock out the fly-forward vehicle (plane) yaw check.
Remember the mag bias and earth field states learned during flight when the vehicle lands.
This improves performance for vehicles that do multiple flight on one power cycle
Provide consistent overshoot of 5% across a wider range of time constants and prevent selection of larger time constants causing 'ringing' in the position and velocity outputs.
Fixes a problem observed in a flight log where rapid temperature change caused the accel bias to change faster than the EKF could keep up.
This allows the bias to be learned faster but with acceptable level of noise in the estimate
IMU data was being corrected before being used by the co-variance prediction, whereas the delta angles and velocities in the derivation were supposed to be uncorrected.
This patch creates separate variable for the corrected data
Magnetometer bias states will subject to larger errors early in flight before flight motion makes the offsets observable and the state variances reduce.
Adds a check on state variances.
Replaces the parameter check with a check of the actual filter fusion method being used.
Allow different process noise to be set for body (sensor bias) and earth field states.
This allows a stable magnetometer bias estimate to be available at end of flight whilst still allowing for external magnetic anomalies during landing.
Adjust default values to give stable mag bias learning and fast learning of external anomalies.
Automatically use the highest gain consistent with a 5% overshoot to minimise RMS tracking errors.
Provide an alternative correction method for the position and velocity states that allows the user to specify the time-constant. This can be used to fine tune the output observer for for platform specific sensor errors and control loop sensitivity estimation noise.
The toilet bowling check during early flight has been removed. This check caused problems where bad compass calibration was the cause of the toilet bowling and resetting to the compass was a bad option. The handling of simultaneous failed mag and velocity innovations is already handled outside the EKF by the failsafe.
A check for yaw errors due to a ground based magnetic anomaly has been introduced.
The logic for in-flight yaw and magnetic field resets has been cleaned up and variable names improved.
Splits in-flight yaw alignment completed status into separate yaw and magnetic field flags.
Reduce the number of places where decisions to perform a yaw and field reset are made.
Don't perform a reset unless there is is data in the buffer
Don't use 3-axis fusion if the field states still need to be reset.
When starting 3-axis fusion request a reset if not previously performed.
Ensure magnetometer and GPs heading resets are alwasy perfomred with data at teh correct time horizon.
The commencement of airspeed fusion could cause pitch errors due to small variances and large innovations. This issue is addressed by the following changes:
1) The airspeed measurement is used to set wind states to a value that reduces initial innovations.
2) The wind state variances are set to values that better reflect the wind speed uncertainty
The function used to reset magnetic field states and yaw angle should not be used when there is no magnetometer. If it is incorrectly called without a magnetometer it should not change the attitude or field states.
When changing the vehicle yaw angle, the correlation between the attitude errors and errors in other states is invalid so the corresponding co-variance terms need to be zeroed.
This needs to be done in more than one place.
M_NSE is a measurement noise
P_NSE is a observation noise
I_GATE is an innovation gate
This also ensures the new parameter values required to use the EKF2 will be enforced.
The new function can deal with a variable number of function parameters.
Additionally, I renamed the functions to norm(), because this is the
standard name used in several other projects.
Implements the following techniques to enable planes to operate without magnetometers.
1) When on ground with mag use inhibited, a synthetic heading equal to current heading is fused to prevent uncontrolled covariance growth.
2) When transitioning to in-flight, the delta between inertial and GPS velocity vector is used to align the yaw.
3) The yaw gyro bias state variance is reset following an in-flight heading reset to enable the yaw gyro bias to be learned faster.
Use an Euler yaw heading that switches between a 321 and 312 rotation
sequence to avoid areas of singularity. Using Euler yaw decouples the
observation from the roll and pitch states and prevents magnetic
disturbances from affecting roll and pitch via the magnetometer fusion
process.
The use of yaw angle fusion during startup and ground operation causes problems with tail-sitter vehicle types.
Instead of observing an Euler yaw angle, we now observe the yaw angle obtained by projecting the measured magnetic field onto the the horizontal plain.
This avoids the singularities associated with the observation of Euler yaw angle.
The innovation calculation should have been updated when the heading fusion maths was updated.
We now use a direct heading or yaw angle measurement in the derivation, not the difference between observed and published declination.
This removes a legacy design concept that is no longer required in this filter implementation. Planes will not be armed without EKF aiding and the proposed copter throw mode also requires EKF aiding to be operating.
The other problem with interrupting fusion during the launch is it doesn't reduce the corrections, it just delays them as wen the launch completes, the EKF inertial position estimate is still moving still moved and the corrections are therefore just delayed by the short launch interval.
Thank you to OXINARF for picking up the inconsistency with the previous logic
Change to user adjustable fusion of constant position (as per legacy EKF) instead of constant velocity.
Enable user to specify use of 3-axis magnetometer fusion when operating without aiding.
Don't allow gyro scale factor learning without external aiding data as it can be unreliable
Sensor bias corrections were being applied to the incoming IMU data using the wrong delta time.
This was what was driving the different tuning between plane and copter for gyro bias process noise so the same gyro bias process noise default tuning value can now be used for all platform types.
Sensor bias corrections were being applied a a second time to the output observer inertial data.
Eliminate the use of horizontal position states during non-aiding operation to make it easier to tune.
Explicitly set the horizontal position associated Kalman gains to zero and the coresponding covariance entries to zero after avery fusion operation.
Make the horizontal velocity observation noise used during non-aiding operation adjustable.
Use a fixed value of velocity noise during initial alignment so that the flight peformance can be tuned without affecting the initial alignment.
The non-GPS mode was not being activated for small height gains - eg indoor flight.
The incorrect innovation consistency check was being applied to the synthetic velocity observations.
The problem with using min() and max() is that they conflict with some
C++ headers. Name the macros in uppercase instead. We may go case by
case later converting them to be typesafe.
Changes generated with:
git ls-files '*.cpp' '*.h' -z | xargs -0 sed -i 's/\([^_[:alnum:]]\)max(/\1MAX(/g'
git ls-files '*.cpp' '*.h' -z | xargs -0 sed -i 's/\([^_[:alnum:]]\)min(/\1MIN(/g'
The airspeed observation buffer was only being checked when new data arrived instead of every frame which introduced some timing jitter. The buffer is now checked every filer update step.
The duplication and inconsistent naming of booleans used to indicate availability f data has been fixed.
These changes were pair coded an tested by Siddharth Purohit and Paul Riseborough
Fix indexing errors
Move buffer code into a separate file
Split observer and IMU/output buffers and remove duplicate sample time
Optimise observation buffer search
Reduce maximum allowed fusion age to 100 msec
This revised threshold value is still double the maximum that has been observed in flight logs so far with healthy sensors
The previous value was too slow to switch for sudden IMU gyro faults
We can afford an ocasional false trigger becasue the front end will only select another instance if it is healthy and has lower errors
The ad-hoc scaling of error growth has been replaced with a consistent method that uses the main nav filters published vertical velocity uncertainty and the terrain gradient assumption.
GPS height has been added as a measurement option along with range finder and baro
Selection of the height measurement source has been moved into a separate function
Each height source is assigned its own measurement noise
If GPS or baro alt is not able to be used, it reverts to baro
When baro is not being used, an offset is continually calculated which enables a switch to baro without a height step.
The copter method was being used for plane and the plane method was not being run due to the change in flight status not being detected.
The plane reset method did not trigger if the EKF had already dragged the velocity states along with the GPS or could align to an incorrect heading.
The method has been reworked so that it resets to the GPS course, but only if there are inconsistent angles and large innovations.
To stop a failed magnetometer causing a loss of yaw reference later in flight, if all available sensors have been tried in flight and timed out, then no further magnetoemter data will be used
Down-sample the IMU and output observer state data to 100Hz for storage in the buffer.
This reduces storage requirements for Copter by 75% or 6KB
It does not affect memory required by plane which already uses short buffers due to its 50Hz execution rate.
This means that the EKF filter operations operate at a maximum rate of 100Hz.
The output observer continues to operate at 400Hz and coning and sculling corrections are applied during the down-sampling so there is no loss of accuracy.
Large magnetometer innovations on the ground could be caused by factors that will disappear when flying, eg:
a) Bad initial gyro bias
b) External magnetic field disturbances (adjacent metal structures, placement of hatches with magnets, etc)
To avoid unnecessary switches, we inhibit switching until off-ground and when sufficient time has lapsed from power on to learn gyro bias offsets.
If the magnetometer fails innovation consistency checks for too long (currently 10 sec), then the next available sensor approved for yaw measurement will be used.
The original design intent was to require all axes to pass because severe errors are rarely constrained to a single axis.
This was not achieved with the previous implementation.
These changes move the innovation consistency checks for all three axes to the top before any axes are fused.
Unnecessary performance timers have been removed.
This was problematic to implement with magnetometer switching. It is likely that slow magnetometer learning can still be performed externally (eg plane) but this will need to be monitored to see if it causes issues.
The setting of the EKF origin and the entry into GPS aiding mode have been separated to make the logic clear.
The order of operations has been changed to ensure that when a reset to GPS is performed, a valid GPS measurement is available in the buffer
Declaration of GPS availability is not made unless the GPS data has been entered into the buffer
Only applied to interfaces required for data logging.
If an invalid instance is requested, the data for the primary instance is returned. This allows the primary data to be returned by calling with a -1 instance value.
Apply filtering to baro innovation check and and don't apply innovation checks once aiding has commenced because GPS and baro disturbances on the ground and during launch could generate a false positive
Prevents frame over-runs due to simultaneous fusion of measurements on each instance.
The offset is only applied if less than 5msec available between frames
Vibration in the 400Hz delta angles could cause the angular rate condition check for in-flight magnetic field alignment to fail.
The symptons were failure to start magnetic field learning as expected when EK2_MAG_CAL=3 was set.
Use the more robust, but less accurate compass heading fusion up to 5m altitude
Wait for the magnetometer data fusion time offset to be correct before using data to reset states
Don't reset magnetic field states if the vehicle is rotating rapidly as timing offsets will produce large errors
When doing the yaw angle reset, apply the reset increment to all quaternions stored in the output buffer to avoid transients produced by yaw rotations and the 0.25 second fusion time horizon offset.
Only do the one yaw and mag reset at 5m, not two at 1.5 and 5.0m
Always re-do the yaw and mag reset when leaving the ground.
Now variables don't have to be declared with PROGMEM anymore, so remove
them. This was automated with:
git grep -l -z PROGMEM | xargs -0 sed -i 's/ PROGMEM / /g'
git grep -l -z PROGMEM | xargs -0 sed -i 's/PROGMEM//g'
The 2 commands were done so we don't leave behind spurious spaces.
AVR-specific places were not changed.
The PSTR is already define as a NOP for all supported platforms. It's
only needed for AVR so here we remove all the uses throughout the
codebase.
This was automated with a simple python script so it also converts
places which spans to multiple lines, removing the matching parentheses.
AVR-specific places were not changed.
This increase in assumed altimeter noise and reduction in accel noise has been flight tested by L.Hall with noticeable reduction in the immediate response to Baro errors during moving flight.
This increases the time constant of response to baro errors such that the pilot can more easily compensate.
The previous gain from rate to magnetometer error was excessive. The revised value is equivalent to a magnetic field length of 0.5 with a timing uncertainty of 0.01 sec
Testing on different platforms has shown that the new EKF has smaller innovations enabling innovation consistency checks that reject GPS and baro errors to be tightened.
The position and velocity thresholds for plane have been left the same because planes are less sensitive to GPS glitches as they fly higher and with more separation to surrounding objects. They are also more prone to bad inertial data due to the installation practices.
The altitude noise has been increased on plane to allow for the larger baro disturbances that result from the higher speeds and lack of a proper static pressure source. The innovation consistency gate has been adjusted to provide the same baro error limit of ~20m before baro is rejected.
Extended GPS loss can result in the earth field states becoming rotated and making it difficult for the EKF to recover its heading when GPS is regained.
During prolonged GPS outages, the position covariance can become large enough to cause the reset function to continually activate. This is fixed by ensuring that position covariances are always reset when the position is reset.
The innovation variance was being used incorrectly instead of the state variance to trigger the glitch reset.
Because we have changed the yaw angle and have taken a point sample on the magnetic field, covariances associated with the magnetic field states will be invalid and subsequent innovations could cause an unwanted disturbance in roll and pitch.
The reset of the Euler angles to a new yaw orientation was being done using roll and pitch from the output observer states, not the EKF state vector which meant that when roll and pitch were changing, the reset to a new yaw angle would also cause a roll and pitch disturbance.
This sets the fusion of the synthetic position and velocity to occur at the same time as the barometer
This makes filter tuning more consistent between GPS and non-GPS useage
High measurement data rates can fill buffers with data that is always new and never fused because it is over-written before it falls behind the measurement time horizon.
The legacy EKF switches GPs aiding on on arming, whereas the new EKF switches it on based on GPS data quality.
This means the decision to arm and therefore the predicted solution flags must now reflect the actual status of the navigation solution as it will no longer change when motor arming occurs.
If high vibration levels cause offsets between the two, it switches to the accelerometer with lower vibration levels. The default behaviour is to use the average of both accelerometers.
Don't fuse other measurements on the same frame that magnetometer measurements arrive if running at a high frame rate as there will be insufficient time to complete other operations.
This parameter is a compromise between numerical accuracy of the covariance prediction and sensor timing jitter
Further testing has shown that doing covariance prediction and sensor fusion every 10msec has no observable effect on fusion health and reduces timing hitter noise on magnetometer observations during high rate maneovures
The values chosen ensure that up to consistent 250 msec of sensor delay compensation is available for different platform types
The revised values also ensure that fusion occurs at different time to when the 10Hz magnetometer measurements are read
Adds fusion of the declination when there are no earth relative measurements so that the declination angle and therefore the copters yaw angle have an absolute reference.
This enables the length (but not the declination) of the earth field North/East states to change along with the magnetometer offsets.
Provide an option to always do learning
Make field learning decision logic clearer
Change defaults so that plane learns when airborne
Change defaults so that Rover does not learn (large external magnetic interference)
Magnetic interference whilst on the ground can adversely affect filter states. This patch ensures that the simpler and more robust magnetic heading observation method is used until the vehicle has cleared the ground.
If an external gyro calibration has been performed, we should assume that it has been done under static conditions
Otherwise it is pointless and we should allow the EKF to find its own gyro bias offsets.
The failure to initialise the magnetometer bias states to zero can result in a large jump in yaw gyro bias and heading when a heading reset is performed.
MPU6000 data sheet indicates that variation on gyro ZRO across temperature range from -40 to +85 is +-20 deg/sec.
The limits on the gyro bias states have been increased to allow for this.
To enable the EKF to accommodate such large gyro bias values in yaw without the yaw error wrapping, leading to continual heading drift, an unwrap function has been applied to the compass heading error.
priseborough
Jonathan Challinger
Andrew Tridgell
Tom Pittenger
Ricardo de Almeida Gonzaga
dgrat
Lucas De Marchi
Peter Barker
Don Gagne
Siddharth Purohit and Paul Riseborough
Siddharth Bharat Purohit
Caio Marcelo de Oliveira Filho
Paul Riseborough
Randy Mackay