Are the parameters required to change over time?

No. A filter may estimate fixed unknown parameters; allowing drift is a modeling choice.

Is a dual EKF the same as a Kalman smoother?

No. Filtering uses observations available so far, whereas smoothing can use later observations to revise earlier states.

Does a neural network have to be involved?

No. Neural models are one application; physical nonlinear models can also be used.

AI Glossary / Probabilistic machine learning / Advanced

Dual extended Kalman filter

Definition

A dual extended Kalman filter (dual EKF or DEKF) uses two coupled estimation processes to track both a system's changing hidden state and uncertain model parameters from noisy measurements. One filter estimates the state using current parameter estimates; the other updates parameters using information from the state-estimation process.

What it means

A system's state describes what is happening now, such as an internal temperature. A parameter describes how the system behaves, such as a heat-transfer coefficient. If the coefficient is uncertain, temperature predictions may be wrong; if temperature estimates are poor, learning the coefficient is difficult.

Dual estimation addresses this interaction. “Extended” indicates that nonlinear relationships are locally approximated using derivatives, as in an ordinary EKF. “Dual” refers to the coupled state and parameter estimation tasks, not simply two sensors. The estimates remain approximate and depend on the model, available measurements and treatment of uncertainty between the two filters.

How it works

The state filter predicts the next hidden state from the dynamics, known inputs and a parameter estimate. It compares the predicted observation with the actual measurement and uses an uncertainty-dependent gain to correct the state.

The parameter filter treats unknown coefficients as quantities to estimate, often using constant or slowly changing parameter dynamics. Its observation relationship depends on the estimated state. Derivatives with respect to states and parameters provide local sensitivities for the updates.

The filters exchange estimates, but their errors are not automatically independent. Update order, propagated sensitivities and treatment of coupling vary across implementations. A joint EKF instead places state and parameters in one augmented vector with a joint covariance. Neither architecture is universally superior; numerical stability and identifiability matter as much as the choice of label.

Three practical examples

1. A thermal-monitoring system estimates an internal temperature while updating a heat-transfer parameter that is not known precisely. Independent reference measurements are needed to assess accuracy.

2. A vehicle-model experiment estimates motion states together with uncertain dynamics parameters using sensor streams. The measurements must distinguish the effects of the parameters being learned.

3. A nonlinear time-series model estimates a cleaner underlying signal while updating model weights. Noise assumptions and the relationship between state and weight errors affect the result.

Think of it this way

Think of a navigator estimating both the boat's current position and how strongly a persistent current pushes it. The position estimate depends on the current estimate, and subsequent observations help revise both. The analogy stops short of the mathematics: a dual EKF uses explicit dynamics, measurement equations, derivatives and covariance approximations, not two independent guesses that automatically become correct when combined.

Worked example

Consider a simplified scalar system x at the next step equals a times the previous x, and the sensor measures x plus noise. Let the previous state estimate be 2 and the coefficient estimate be 0.8. The predicted state and observation are therefore 1.6. A new observation is 1.9, giving innovation 0.3.

Suppose the state prediction variance is 0.25 and measurement-noise variance is 0.25. The scalar state gain is 0.25/(0.25 + 0.25) = 0.5, so the corrected state is 1.6 + 0.5 × 0.3 = 1.75.

For a separate illustrative parameter update, hold the previous state estimate fixed at 2, use coefficient variance 0.04 and effective observation-noise variance 0.25. The sensitivity to a is 2. The parameter gain is 0.04 × 2/(2² × 0.04 + 0.25) = 0.08/0.41, approximately 0.1951. Its update is 0.8 + 0.1951 × 0.3, approximately 0.8585.

These two local calculations demonstrate the different roles. They deliberately condition on supplied variances and omit cross-filter uncertainty; they are not a complete, statistically consistent DEKF implementation or a reproduction of a published experiment.

Illustrative case study

Fictional case: Kiln Meadow monitors a small heated enclosure. Engineers can measure surface temperature but want an estimate of internal temperature and an uncertain cooling coefficient. They build a physical model and test a dual EKF using laboratory runs with additional reference sensors.

Some steady operating periods reveal little about the coefficient, despite producing a smooth temperature trace. The team uses varied controlled inputs to investigate identifiability and checks whether the filter's uncertainty matches observed errors. It retains independent operating limits. This example claims neither a real installation nor a validated safety system.

Common misconceptions and mistakes

Two filters do not provide two independent confirmations. They are coupled through the same model and often the same measurements. Ignoring their interaction can produce excessive confidence.

A smooth estimate is not necessarily accurate. Incorrect noise settings, a poor initial guess or strong nonlinearity can cause divergence or biased estimates. An apparently drifting parameter may actually be compensating for a missing physical effect or sensor bias.

The extended Kalman filter is an approximation, and a dual EKF does not generally produce the exact Bayesian posterior. Nor can it identify parameters that the measurement process does not distinguish.

Business use

This approach can be useful when online state monitoring depends on parameters that are uncertain or slowly varying. Before implementation, define which quantities are states, which are parameters, and why measurements contain enough information about each.

Validate against independent references and realistic disturbances. Monitor innovations, parameter bounds and uncertainty consistency, not just the plotted state. Compare with fixed-parameter filtering and a joint estimator where appropriate. Document covariance coupling and update order so another engineer can reproduce the actual method rather than infer it from the phrase “dual EKF.”

Check your understanding

If the previous state were zero in the worked dynamics, would that observation directly reveal a through the sensitivity 2 used above?

Answer: No. The local sensitivity would be zero, so this simplified measurement would carry no direct information about a at that step.

Related terms

Extended Kalman filter
Jacobian
state estimation
parameter estimation
joint estimation
state-space model
innovation
identifiability

Learn next

Work through a scalar Kalman measurement update, then compare separate and augmented state-parameter covariance representations. Pay particular attention to what uncertainty each construction retains or approximates.

Sources

Primary sources: Wan and Nelson, Dual Kalman Filtering Methods for Nonlinear Prediction, Smoothing, and Estimation, Section 2.2.2: https://papers.nips.cc/paper/1202-dual-kalman-filtering-methods-for-nonlinear-prediction-smoothing-and-estimation.pdf ; battery thermal application abstract: https://arxiv.org/abs/1501.06160 . Independent review checked Section 2.2.2 coupling and Jacobians; the numerical calculation is explicitly simplified and original.

Previous: Dropping conditions Next: Dual graph