Source code for stonesoup.predictor.asd

import copy
from itertools import islice
from operator import itemgetter

import numpy as np

from ._utils import predict_lru_cache
from .kalman import KalmanPredictor
from ..types.prediction import ASDGaussianStatePrediction
from ..types.state import GaussianState


[docs] class ASDKalmanPredictor(KalmanPredictor): """Accumulated State Densities Kalman Predictor A linear predictor for accumulated state densities, for processing out of sequence measurements. This requires the state is represented in :class:`~.ASDGaussianState` multi-state. References ---------- 1. W. Koch and F. Govaers, On Accumulated State Densities with Applications to Out-of-Sequence Measurement Processing in IEEE Transactions on Aerospace and Electronic Systems, vol. 47, no. 4, pp. 2766-2778, OCTOBER 2011, doi: 10.1109/TAES.2011.6034663. 2. F. Govaers and W. Koch, Generalized Solution to Smoothing and Out-of-Sequence Processing in IEEE Transactions on Aerospace and Electronic Systems, vol. 50, no. 3, pp. 1739-1748, JULY 2014, doi: 10.1109/TAES.2014.130009. """ def _predict_over_interval(self, prior, timestamp): """Private function to get the prediction interval (or None) Parameters ---------- prior : :class:`~.ASDState` The prior state timestamp : :class:`datetime.datetime`, optional The (current) timestamp Returns ------- : :class:`datetime.timedelta` time interval to predict over if the timestamp is in the past then it returns the interval to the next timestamp in the state : :class:`datetime.datetime` time from which the interval is calculated """ predict_over_interval = timestamp - prior.timestamp timestamp_from_which_is_predicted = prior.timestamp if predict_over_interval.total_seconds() < 0: predict_over_interval, timestamp_from_which_is_predicted = min( ((timestamp - t, t) for t in prior.timestamps if (timestamp - t).total_seconds() >= 0), key=itemgetter(0)) return predict_over_interval, timestamp_from_which_is_predicted
[docs] @predict_lru_cache() def predict(self, prior, timestamp, **kwargs): r"""The predict function Parameters ---------- prior : :class:`~.ASDGaussianState` :math:`\mathbf{x}_{k-1}` timestamp : :class:`datetime.datetime`, :math:`k` **kwargs : These are passed, via :meth:`~.KalmanFilter.transition_function` to :meth:`~.LinearGaussianTransitionModel.matrix` Returns ------- : :class:`~.ASDState` :math:`\mathbf{x}_{k|k-1}`, the predicted state and the predicted state covariance :math:`P_{k|k-1}` """ correlation_matrices = copy.copy(prior.correlation_matrices) # Get the prediction interval predict_over_interval, timestamp_from_which_is_predicted = \ self._predict_over_interval(prior, timestamp) # Build the first correlation matrix just after starting the # predictor the first time. if not correlation_matrices: correlation_matrices.append({'P': prior.covar}) # Consider, if the given timestep is an Out-Of-Sequence measurement t_index = prior.timestamps.index(timestamp_from_which_is_predicted) if t_index == 0: # case that it is a normal prediction # Prediction of the mean and covariance x_pred_m = self._transition_function( prior, time_interval=predict_over_interval, **kwargs) p_pred_m = self._predicted_covariance( prior, predict_over_interval=predict_over_interval, **kwargs) transition_matrix = self._transition_matrix( prior=prior, time_interval=predict_over_interval, **kwargs) # Generation of the combined retrodiction matrices combined_retrodiction_matrices = self._generate_C_matrices( correlation_matrices, prior.ndim) # normal case x_pred = np.concatenate([x_pred_m, prior.multi_state_vector]) W_P_column = np.vstack([c @ prior.covar for c in combined_retrodiction_matrices]) correlated_column = W_P_column @ transition_matrix.T correlated_row = correlated_column.T p_top = np.hstack((p_pred_m, correlated_row)) p_bottom = np.hstack((correlated_column, prior.multi_covar)) p_pred = np.vstack((p_top, p_bottom)) # add new correlation matrix with the present time step correlation_matrices[t_index] = time_corr_matrices = \ correlation_matrices[t_index].copy() time_corr_matrices['P_pred'] = p_pred_m time_corr_matrices['F'] = transition_matrix time_corr_matrices['PFP'] = \ time_corr_matrices['P'] \ @ time_corr_matrices['F'].T \ @ np.linalg.inv(time_corr_matrices['P_pred']) correlation_matrices.insert(t_index, {'P': p_pred_m}) else: # Below based on equations from 69 to 75 in reference 2. # case of out of sequence prediction case timestamp_m_plus_1 = prior.timestamps[t_index - 1] time_interval_m_to_m_plus_1 = timestamp_m_plus_1 - timestamp ndim = prior.ndim # Normal Kalman-like prediction for the state m|m-1. # prediction to the timestamp m|m-1 prior_at_t = prior[t_index] x_pred_m = self._transition_function( prior_at_t, time_interval=predict_over_interval, **kwargs) p_pred_m = self._predicted_covariance( prior_at_t, predict_over_interval=predict_over_interval, **kwargs) state_pred_m = GaussianState(x_pred_m, p_pred_m, timestamp) # prediction to the timestamp m + 1|m-1 x_pred_m_plus_1 = self._transition_function( state_pred_m, time_interval=time_interval_m_to_m_plus_1, **kwargs) p_pred_m_plus_1 = self._predicted_covariance( state_pred_m, predict_over_interval=time_interval_m_to_m_plus_1, **kwargs) # transitions for timestamp m transition_matrix_m = self._transition_matrix( prior=prior_at_t, time_interval=predict_over_interval, **kwargs) # transitions for timestamp m+1 transition_matrix_m_plus_1 = self._transition_matrix( prior=state_pred_m, time_interval=time_interval_m_to_m_plus_1, **kwargs) t_minus2t = slice((t_index-1) * ndim, t_index * ndim) x_m_plus_1_given_k = prior.multi_state_vector[t_minus2t] x_diff = x_m_plus_1_given_k - x_pred_m_plus_1 p_m_plus_1_given_k = prior.multi_covar[t_minus2t, t_minus2t] p_diff = p_m_plus_1_given_k - p_pred_m_plus_1 W = p_pred_m @ transition_matrix_m_plus_1.T @ np.linalg.inv(p_pred_m_plus_1) x_pred_m_given_k = x_pred_m + W@x_diff p_pred_m_given_k = p_pred_m + W@p_diff@W.T # build full state x_pred = np.concatenate([prior.multi_state_vector[:t_index * ndim], x_pred_m_given_k, prior.multi_state_vector[t_index * ndim:]]) P_right_lower = prior.multi_covar[t_index * ndim:, t_index * ndim:] # Calculate what would be the orignal predicted covariance p_state_pred = GaussianState( x_pred_m, correlation_matrices[t_index]['P'], prior.timestamp) p_p_pred = self._predicted_covariance( p_state_pred, predict_over_interval=predict_over_interval) # add new correlation matrix with the present time step correlation_matrices[t_index] = pred_from_corr_matrices = \ correlation_matrices[t_index].copy() pred_from_corr_matrices['P_pred'] = p_p_pred pred_from_corr_matrices['F'] = transition_matrix_m pred_from_corr_matrices['PFP'] = ( pred_from_corr_matrices['P'] @ transition_matrix_m.T @ np.linalg.inv(p_p_pred)) correlation_matrices.insert(t_index, {}) correlation_matrices[t_index]['F'] = transition_matrix_m_plus_1 correlation_matrices[t_index]['P_pred'] = p_pred_m_plus_1 correlation_matrices[t_index]['P'] = p_pred_m correlation_matrices[t_index]['PFP'] = \ p_pred_m @ transition_matrix_m_plus_1.T @ np.linalg.inv(p_pred_m_plus_1) # generate prediction matrix p_pred = np.zeros((ndim * (prior.nstep+1), ndim * (prior.nstep+1))) p_pred[(t_index+1) * ndim:, (t_index+1) * ndim:] = P_right_lower # get all timestamps which has to be recalculated beginning # with the newest one timestamps_to_recalculate = prior.timestamps[:t_index] timestamps_to_recalculate.append(timestamp) covars = \ [prior.multi_covar[i * ndim:(i+1) * ndim, i * ndim:(i+1) * ndim] for i in range(t_index)] covars.append(p_pred_m_given_k) for i, ts in enumerate(timestamps_to_recalculate): corrs = correlation_matrices[i:] combined_retrodiction_matrices = self._generate_C_matrices(corrs, ndim) combined_retrodiction_matrices = combined_retrodiction_matrices[1:] W_column = np.vstack([c @ covars[i] for c in combined_retrodiction_matrices]) W_row = W_column.T i2i_plus = slice(i * ndim, (i + 1) * ndim) i_plus2end = slice((i+1) * ndim, None) # set covar p_pred[i2i_plus, i2i_plus] = covars[i] # set column p_pred[i_plus2end, i2i_plus] = W_column # set row p_pred[i2i_plus, i_plus2end] = W_row timestamps = sorted(prior.timestamps + [timestamp], reverse=True) # the act_timestamp parameter is used for the updater to # know for which timestamp the prediction is calculated predicted_state = ASDGaussianStatePrediction( multi_state_vector=x_pred, multi_covar=p_pred, correlation_matrices=correlation_matrices, timestamps=timestamps, max_nstep=prior.max_nstep, act_timestamp=timestamp) self.prune_state(predicted_state) return predicted_state
def _generate_C_matrices(self, correlation_matrices, ndim): combined_retrodiction_matrices = [np.eye(ndim)] for item in islice(correlation_matrices, 1, None): combined_retrodiction_matrices.append( combined_retrodiction_matrices[-1] @ item['PFP']) return combined_retrodiction_matrices
[docs] def prune_state(self, predicted_state): r"""Simple ASD pruning function. Deletes timesteps from the multi state if it is longer then max_nstep Parameters ---------- predicted_state : :class:`~.ASDState` :math:`\mathbf{x}_{k|k-1}` Returns ------- : :class:`~.ASDState` :math:`\mathbf{x}_{k|k-1}`, the pruned state and the pruned state covariance :math:`P_{k|k-1}` """ if predicted_state.nstep > predicted_state.max_nstep != 0: index = predicted_state.max_nstep * predicted_state.ndim predicted_state.multi_state_vector = \ predicted_state.multi_state_vector[:index] predicted_state.multi_covar = \ predicted_state.multi_covar[:index, :index] predicted_state.timestamps = \ predicted_state.timestamps[:predicted_state.max_nstep] predicted_state.correlation_matrices = \ predicted_state.correlation_matrices[:predicted_state.max_nstep] return predicted_state else: return predicted_state