I’m playing with an app to try and implement some form of sleep detection. With the caveat that I largely have no idea what I’m doing, I’m mostly trying to match up with my Garmin…
From a completely stationary watch I’m seeing 4 -5 mg RMS per axis in the 0.25-3 Hz band (X 4.0, Y 4.3, Z 4.9). I put the watch on a pillow on the floor in an unoccupied house for 6 hours and separately put it on a hard table during the day and both agree within 2%.
How I got there: I measured the mean rectified value after filtering to that band, 3.2-3.9 mg, then converted to RMS using the Gaussian form factor of about 1.25. If the noise is uniform rather than Gaussian the factor is 1.155, so the number shifts under 10% either way. I’m thinking I’ll redo this by collected raw samples to remove the assumption entirely, just in case I’m really bad at math, but it isn’t anywhere near the gap.
The BMI270 puts accelerometer noise around 160 µg/√Hz, scaled to 2.75 Hz that works out to 0.27 mg RMS. AFAICT the watch is reporting ~15x that.
Are other people with a watch seeing something similar? I can throw together an app to make profiling watches in the wild more accessible. Then for UNA, what are the relevant configurations (ODR, bandwidth/OSR, power mode etc). I might be completely misunderstanding the calculations/metrics so I’m open to that possibility as well. Could be a low power mode set up for step counting?