Hello,
We encountered an issue with the bias of the accelerometer of the LSM6DSV device and I wonder if anyone else sees that. Before getting into details, I would like to disclose we already corresponded with ST, got full cooperation, but came out with no real solution.
We produce compact AR glasses, and are using LSM6DSO as our IMU on the glasses to track head direction. As we experience inconvenient temperature sensitivity of the gyro's x axis, and LSM6DSV seem to be a virtually a drop-in and better-by-datasheet replacement, we tried it on our boards, somewhat success-oriented.
We realized the accelerometer does not calibrate well (calibrate = 3x3 matrix and 3x1 vector that makes the norm of gravity constant regardless of device orientation). Investigating, we realized the bias of the accelerometer has an odd behavior over time.
We went back to test with ProfiMEMS and observed the same issue there. Applying to ST via a sales representative of the supplier, ST tested with their lab setup and failed to replicate, but did replicate using ProfiMEMS. ST's response insisted that ProfiMEMS heats the component via the DIL24 connector and therefore "The ProfiMEMS tool is not suitable for sensor characterization but only for a general evaluation of the product".
The measurements below, taken from a lab report ST kindly provided as a response to our application, are similar in nature to our own measurements.
One can (hopefully, given the tiny images) see that even though the temperature changes very little (not sure why it has less samples than the acceleration) some accelerometer axes show a very slow convergence. The amount of change in the accelerometer reading is far larger than what is expected according to the datasheet numbers for temperature sensitivity.
We, per the claim above, enlarged the decoupling of the component from the main board by longer wires:
As one can see below, the behavior of the z axis is hard to correlate with the temperature. Note here the x axis of all graphs is given in samples (at 240Hz) and not seconds:
We can not explain why the lab setup of ST does not exhibit this behavior. Maybe the sensors are always on and already relaxed to the asymptote value when the recording was made. This is something we cannot do in real life.
Has anyone experienced such behavior? Any ideas how to mitigate?
