Improve magnetometer calibration

[stenbergd]

Improve magnetometer calibration using Gauss-Newton method

Read more: https://chionophilous.wordpress.com/2012/09/01/implementing-the-gauss-newton-algorithm-for-sphere-fitting-1-of-3/

[adam-at-epsilon]

Full scale is set to 1.3 Gauss as earth's magnetic field vector is .5 Gauss according to Wikipedia.

[stenbergd]

Very good, should give us more than enough wiggle-room for any magnetic interference.

There are some datasheet and application notes from ST on calibrating the magnetometer.

I previously implemented an algorithm that would compensate for hard-iron and soft-iron distortion of magnetometer readings. You basically need to turn the quadcopter/fcb board 360 degrees around all of its 3 axes to calibrate it, but will likely only need to do this once.

Anyway, the link in the OP suggests an improvement on what I did and also provides a fun mathematical optimization method!

[stenbergd]

Hard-iron offset it pure scaling of x, y, z measurement axes.

Soft-iron offset has distorted/skewed the magnetometer measurement sphere.

We will start with the hard-iron offset. If we do not have any strong magnetic sources on-board, this will most likely be sufficient to gather good measurements.

[stenbergd]

Also, we may not need to use float values (Gauss values). Raw values enough if we want to calc attitude with magnetometer?

Perhaps we will use values normed into the unit sphere?

[stenbergd]

• Use dedicated sampling task for this?
• Store sampled values dynamically?

[adam-at-epsilon]

General advice for doing numerical solutions (from link in 1st post)

Choosing parameters that are about the same order of magnitude leads to stable, reliable performance. because we avoid over/underflow of floating-point values.

Our floats are 32 bit, under&overflow is unlikely.

[adam-at-epsilon]

consider offline calibration by reading the values to a file, then doing the Gauss-Newton Least Sphere Fit in Scilab, GNU Octave or similar.

Pro: Could be useful if we are running out of program space

Con: FCB must be attached to a PC during calibration, this may not be practical.

[adam-at-epsilon]

The magnetometer is now compensated by the Gauss-Newton (offline, implemented in SciLab) method with respect to offset & scaling. (i.e. hard-iron offset)

[adam-at-epsilon]

@DanielStenis

are any scripts using the human-readable variant of the start-sensor-sampling on the Termite console? If not, I would like to remove the line breaks for easier text processing for the offline variants of the Gauss-Newton calibrations.

[stenbergd]

Not at the moment, no. But you could implement a new output encoding type named commaseparated_encoding or something. If we want to keep the human-readable as it is for now

[stenbergd]

Some more info about hard/soft-iron compensation can be found here:

http://www.sensorsmag.com/sensors/motion-velocity-displacement/compensating-tilt-hard-iron-and-soft-iron-effects-6475

[adam-at-epsilon]

So I implemented an extra "c" parameter for sensor - sampling. All it does is to print the values for XYZ on one single line.

[stenbergd]

Should we care about magnetometer declination (correction for difference between magnetic and geographic north pole)?