DynamicalModelCompensation#

class DynamicalModelCompensation(continuous_time_covariance: CovarianceMatrix, state_definition=None, sequence_definition=None, beta: ndarray | None = None, max_interval=None, body=None)#

Bases: ProcessNoise

Dynamic Model Compensation (DMC) process noise model.

An alternative to SNC that models process noise through direct augmentation of the state vector with stochastic parameters representing unmodeled accelerations or forces, allowing their estimation alongside the spacecraft state.

Augmented state and dynamics

DMC augments the state vector with a stochastic acceleration vector \(\mathbf{w} \in \mathbb{R}^3\):

\[\mathbf{x}^T = [\,\mathbf{r}^T \quad \mathbf{v}^T \quad \mathbf{w}^T\,]\]

The stochastic acceleration obeys a first-order Gauss-Markov (exponentially correlated) process:

\[\dot{\mathbf{w}} = -\mathbf{B}\,\mathbf{w} + \mathbf{u}(t)\]

where \(\mathbf{B} = \mathrm{diag}(\beta_x, \beta_y, \beta_z)\) contains the inverse correlation times (decay rates), and \(\mathbf{u}(t)\) is a zero-mean input held constant over each step, with covariance \(\mathbf{Q}\) (3, 3), in km²/s⁶.

Discretization (gamma matrix)

The response of the state to a constant input \(\mathbf{u}\) over a step \(\Delta t\) is \(\mathbf{\Gamma}_k = \int_{t_k}^{t_{k+1}} \mathbf{\Phi}(t_{k+1}, \tau)\,\mathbf{B}_u\,d\tau\) (9, 3), with \(\mathbf{B}_u = [\mathbf{0}\;\mathbf{0}\;\mathbf{I}_3]^T\):

\[\mathbf{\Gamma}_r = \frac{\Delta t^2}{2}\,\mathbf{B}^{-1} - \Delta t\,\mathbf{B}^{-2} + \mathbf{B}^{-3}\Bigl(\mathbf{I} - e^{-\mathbf{B}\Delta t}\Bigr)\]
\[\mathbf{\Gamma}_v = \Delta t\,\mathbf{B}^{-1} - \mathbf{B}^{-2}\Bigl(\mathbf{I} - e^{-\mathbf{B}\Delta t}\Bigr)\]
\[\mathbf{\Gamma}_w = \mathbf{B}^{-1}\Bigl(\mathbf{I} - e^{-\mathbf{B}\Delta t}\Bigr)\]

All matrices are diagonal because \(\mathbf{B}\) is diagonal, so every block is computed element-wise. The full discrete-time process noise covariance is:

\[\mathbf{Q}_k = \mathbf{\Gamma}_k\,\mathbf{Q}\,\mathbf{\Gamma}_k^T\]

This treats \(\mathbf{u}\) as piecewise constant, as SNC does with the acceleration. Integrating white noise over the step instead gives a full rank \(\mathbf{Q}_k\) that grows more slowly with \(\Delta t\); the form above has rank three, which the SRIF needs to factor it through \(\mathbf{\Gamma}_k\).

Decay rates

The filters discretize the noise with the decay rates the dynamics propagate the state with: the beta_fogm state of body when there is one, else the fogm_beta of the ForceModelTranslation of that body. They are evaluated on each interval and again on every iteration, so an estimated decay rate is followed. beta is used only when no reference decay rates are given.

Covariance propagation

The propagated covariance is augmented identically to SNC:

\[\bar{\mathbf{P}}_{k+1} = \mathbf{\Phi}_{k+1,k}\,\mathbf{P}_k\,\mathbf{\Phi}_{k+1,k}^T + \mathbf{Q}_k\]

where \(\mathbf{\Phi}_{k+1,k}\) is now the STM of the augmented \([\mathbf{r}, \mathbf{v}, \mathbf{w}]\) system, including the exponential decay \(e^{-\mathbf{B}\Delta t}\) in the \(\mathbf{w}\) partition.

Comparison with SNC

Unlike SNC, which treats the unmodeled acceleration as a purely random impulse reset at each epoch, DMC retains memory of the stochastic acceleration across epochs via the exponential decay \(e^{-\mathbf{B}\Delta t}\). As \(\beta \to \infty\) (very short correlation time) DMC approaches SNC; as \(\beta \to 0\) (long correlation time) the stochastic acceleration becomes nearly constant and is effectively solved for as a deterministic parameter.

Parameters:
continuous_time_covarianceCovarianceMatrix

Covariance \(\mathbf{Q}\) of the input driving the stochastic acceleration, held constant over each interval, in km²/s⁶.

state_definitionoptional

State vector definition. Defaults to None.

sequence_definitionoptional

Sequence definition for multi-leg trajectories. Defaults to None.

betanp.ndarray, optional

Decay rates (inverse correlation times, 1/s) of the Gauss-Markov process, a scalar (isotropic) or a 3-element vector for x, y, z components. Used only when _compute_discrete_time_covariances() is given no reference decay rates, which the filters always give. Defaults to None.

max_intervalArrayWUnits or float, optional

Longest interval the noise is accumulated over, in seconds when a float; a longer one, such as a data gap, receives the noise of one this long (see _noise_intervals()). Defaults to None, no limit.

bodyBody or int, optional

Body whose position, velocity and a_fogm the noise drives, or its SPICE ID. Defaults to None, the owner of the first position in the state.

Raises:
DynamicalModelCompensation.BetaSizeError

If beta is neither a scalar nor a 3-element vector.

DynamicalModelCompensation.NonPositiveBetaError

If a component of beta is not strictly positive.

See also

scarabaeus.StateNoiseCompensation

Simpler piecewise-constant SNC variant.

scarabaeus.ProcessNoise

Abstract base class.

scarabaeus.ProcessNoiseSettings

Configuration object consumed by the filter.

Notes

As \(\beta \to \infty\) (very short correlation time) DMC approaches SNC; as \(\beta \to 0\) (long correlation time) the stochastic acceleration becomes nearly constant and is effectively estimated as a deterministic parameter.

References

[1]

Tapley, B. D., Schutz, B. E., & Born, G. H. (2004). Statistical Orbit Determination. Elsevier Academic Press. ISBN 978-0-12-683630-1.

exception BetaSizeError#

Bases: ValueError

Raised when beta is neither a scalar nor a 3-element vector.

exception MissingBetaError#

Bases: ValueError

Raised when the noise is discretized with neither reference decay rates nor a beta given to the constructor.

exception NonPositiveBetaError#

Bases: ValueError

Raised when a component of beta is not strictly positive.