Usually there are two specs related to the covariance matrix in a Kalman filter. Let's only take gyro for example because accel is similar. The two specs are ARW and Bias Instability. ARW reperesents the noise level of the gyro output: ARW = std*sqrt(dt), where std is the standard deviation of gyro noise and dt is the sampling interval. Depending on your system equation in the EKF, you might also need to estimate the gyro bias in real time, and then you need bias instability to model the varying gyro bias. Bias instability is hard to model and is a much difficult topic. Generally we can model it as a random walk or a first-order Gauss-Markov process. We provide an online document for your reference: https://openimu.readthedocs.io/en/latest/algorithms.html

If you want to esimate gyro bias by averaging its output, you need to put it static for some time. Static time can be determined by running an Allan Variance analysis on its output. Or you can just use 100sec.