I am implementing a Kalman filter (for a linear ODE system for now).

My model represents a physical device that has 6 "parameters", i.e. those values of the device do not evolve over time (within a time-domain simulation environment for now).

I have 100 such "slightly differing devices", i.e. whose parameter values differ from the nominal, but the distribution of all 6 of the model parameters are known apriori (Gaussian, with known mean (nominal values) and standard deviations).

I am assuming that there is no other source of process noise (i.e. no unmodeled states or other dynamics are present)

I am implementing a Kalman Filter with a single tuning parameter across all these 100 devices. I think that these deviations in model parameters (from the nominal) can be treated as process noise. Is there a deterministic way, i.e. a recipe, to compute the "Process noise covariance matrix", typically denoted as $Q$ in Kalman Filter literature?

I am thinking that one can somehow exploit the affine properties of Guassian distributions to arrive at the error covariance matrix, although I am a bit rusty with the math.

Partially related, but different question asked by someone else: https://dsp.stackexchange.com/questions/21796/question-about-q-matrix-noise-process-covariance-in-kalman-filter

  • $\begingroup$ I think you don't have process noise, hence it's not really a Kalman filter model. It's closer to RLS formulation where Q=0. $\endgroup$ – Memming Sep 26 '17 at 16:44
  • $\begingroup$ @Memming That is not true. The deviations in model parameters (from the nominal) can indeed be treated as a process noise. I shall edit the post to clarify this. $\endgroup$ – Dr Krishnakumar Gopalakrishnan Sep 26 '17 at 22:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.