bias stability

[pizza]

Hi tom, have you tried allan variance for estimation of non tuned noise/measurement covariances?

[Tom]

This is something completely new for me.
Wikipedia: http://en.wikipedia.org/wiki/Allan_variance

A google search on "allan variance gyro" returns a lot of links, most interesting ones: http://www.xbow.com/support/Support_pdf ... rement.pdf
and
http://docs.google.com/viewer?a=v&q=cac ... YPfEUiNJRg

As I understand it is a way to model the noise and bias instability of a gyro, and not a way to compensate for it?

[pizza]

i have read a few papers on how people use allan variance analysis on inertial sensors to identify the noise sources namely bias instability, ARW, RRW, quantization noise, flicker noise and ramp..then these parameters are plugged into calibration model to correct the sensor data, reducing/eliminating effects from drift and improving accuracy. ARW and RRW coefficients are used to compute the Q and R covariance matrices for Kalman filter in their implementation.

well im not sure how well it works, still trying out.
i realised the less number of bins used, the less significant those allan variances are(error increases), so i use as many bins as i can for each averaging period.

ran the av tests on sparkfun 5dof(idg500,adxl335) ..the idg500 gives really unstable data(i don't know why) while the ADXL335 gives very smooth, consistent AV curve.

tom, what has happened to your quad project from the old website? just curious why you've decided to make a jet instead

[pizza]

Hi all,

if it interests you, attached are my allan variance results ran at 100Hz for 24 hours each(total 86400*100samples), on IDG500 of sparkfun's 5DOF(gyrox, gyroy), STmicron's LY530alh gyroscope(gyroz), and Freescale's MMA3761L accelerometer(accx, accy, accz).

apparently the results are rather similar except for gyro X and gyro Y. Can anyone teach me how to identify the problem on the hardware?

thanks & best regards,
pizza

[Tom]

Hi Pizza!

It seems logical that the accelerometer is the most stable one.
Taking into account the algorithmic scale, the IDG-500 seems to have a big variance indeed. I never measured the variance over a day, but it can drift quite a lot. It is weird that the STmicro one seems so much better, while I always heared it was worse.

I could do some shorter measurements on the sensor if that would be usefull for you. What exactly is tau/s ?

PS: quadrocopter is ready and is undergoing tests this week

[pizza]

Hi Tom,

the taus that I use are
float taus[12]= {0.1, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024}; //note: unit = seconds
as these are equal intervals on the allanvariance log-log graph.
Tau is the averaging time used on the bias.

A 24 hr sample will have 86400*100 samples on a 100Hz sampling rate.
If tau = 1s, the samples are divided into 86400bins(i.e no. bins = total no. samples/samples in each tau). The samples in each bin are averaged. Then the allan variance is calculated. this is repeated for every tau. for tau=2s, there will be 43200bin etc. the less number of bins will cause higher error rate i.e ε(%)=100/√(2(bins-1)). A 2.4 hr sample will have less number of bins etc.