Kalman Filter
A recursive state estimator that combines model prediction with noisy measurements while explicitly tracking estimation uncertainty.
Predict and Update
At each step, a Kalman filter first predicts the next system state from a model and then corrects that prediction when a measurement arrives.
The classical filter provides a closed-form solution for linear models with Gaussian noise assumptions.
Covariance Matters
The filter tracks not only the estimated state but also uncertainty through covariance. A noisy sensor should influence the update differently from a highly trusted measurement.
System Boundary
Incorrect process or measurement-noise assumptions can produce a visually smooth estimate without making it accurate. Bias and model error need separate consideration.
The innovation is the difference between an observed measurement and its prediction. Its magnitude and consistency with expected covariance can reveal sensor bias or model mismatch, but a large innovation alone does not identify which component failed.