4287 |
30 Jan 23 |
peter |
1 |
#ifndef theplu_kalman_filter |
4287 |
30 Jan 23 |
peter |
2 |
#define theplu_kalman_filter |
4287 |
30 Jan 23 |
peter |
3 |
|
4287 |
30 Jan 23 |
peter |
// $Id$ |
4287 |
30 Jan 23 |
peter |
5 |
|
4287 |
30 Jan 23 |
peter |
6 |
/* |
4287 |
30 Jan 23 |
peter |
Copyright (C) 2023 Peter Johansson |
4287 |
30 Jan 23 |
peter |
8 |
|
4287 |
30 Jan 23 |
peter |
This file is part of the yat library, https://dev.thep.lu.se/yat |
4287 |
30 Jan 23 |
peter |
10 |
|
4287 |
30 Jan 23 |
peter |
The yat library is free software; you can redistribute it and/or |
4287 |
30 Jan 23 |
peter |
modify it under the terms of the GNU General Public License as |
4287 |
30 Jan 23 |
peter |
published by the Free Software Foundation; either version 3 of the |
4287 |
30 Jan 23 |
peter |
License, or (at your option) any later version. |
4287 |
30 Jan 23 |
peter |
15 |
|
4287 |
30 Jan 23 |
peter |
The yat library is distributed in the hope that it will be useful, |
4287 |
30 Jan 23 |
peter |
but WITHOUT ANY WARRANTY; without even the implied warranty of |
4287 |
30 Jan 23 |
peter |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
4287 |
30 Jan 23 |
peter |
General Public License for more details. |
4287 |
30 Jan 23 |
peter |
20 |
|
4287 |
30 Jan 23 |
peter |
You should have received a copy of the GNU General Public License |
4287 |
30 Jan 23 |
peter |
along with yat. If not, see <https://www.gnu.org/licenses/>. |
4287 |
30 Jan 23 |
peter |
23 |
*/ |
4287 |
30 Jan 23 |
peter |
24 |
|
4287 |
30 Jan 23 |
peter |
25 |
#include <yat/utility/DiagonalMatrix.h> |
4287 |
30 Jan 23 |
peter |
26 |
#include <yat/utility/Matrix.h> |
4287 |
30 Jan 23 |
peter |
27 |
#include <yat/utility/Vector.h> |
4287 |
30 Jan 23 |
peter |
28 |
|
4287 |
30 Jan 23 |
peter |
29 |
namespace theplu { |
4287 |
30 Jan 23 |
peter |
30 |
namespace yat { |
4287 |
30 Jan 23 |
peter |
31 |
namespace statistics { |
4287 |
30 Jan 23 |
peter |
32 |
|
4287 |
30 Jan 23 |
peter |
33 |
/** |
4287 |
30 Jan 23 |
peter |
Data are modeled in discrete time as \f$ x_t = F_t x_{t-1} + W_t |
4287 |
30 Jan 23 |
peter |
\f$, where \a x is the state variable, \a F describe the state |
4287 |
30 Jan 23 |
peter |
transition, and \a W is the process noise. |
4287 |
30 Jan 23 |
peter |
37 |
|
4287 |
30 Jan 23 |
peter |
The state variable is not observed per se, but the observables, |
4287 |
30 Jan 23 |
peter |
\a z, are observed via \f$ z_t = H_tx_t + V_t \f$, where \a H is |
4287 |
30 Jan 23 |
peter |
a matrix describing the relation between state variable and |
4287 |
30 Jan 23 |
peter |
observables and \a V is the observation noise. |
4287 |
30 Jan 23 |
peter |
42 |
|
4287 |
30 Jan 23 |
peter |
\since New in yat 0.21 |
4287 |
30 Jan 23 |
peter |
44 |
*/ |
4287 |
30 Jan 23 |
peter |
45 |
class KalmanFilter |
4287 |
30 Jan 23 |
peter |
46 |
{ |
4287 |
30 Jan 23 |
peter |
47 |
public: |
4287 |
30 Jan 23 |
peter |
48 |
/** |
4287 |
30 Jan 23 |
peter |
\param x initial value of state variable |
4287 |
30 Jan 23 |
peter |
\param P covariance of state variable |
4287 |
30 Jan 23 |
peter |
51 |
*/ |
4287 |
30 Jan 23 |
peter |
52 |
KalmanFilter(const utility::Vector& x, const utility::Matrix& P); |
4287 |
30 Jan 23 |
peter |
53 |
|
4287 |
30 Jan 23 |
peter |
54 |
/** |
4287 |
30 Jan 23 |
peter |
\brief Predict new state |
4287 |
30 Jan 23 |
peter |
56 |
|
4287 |
30 Jan 23 |
peter |
The predicted state is calculated as |
4287 |
30 Jan 23 |
peter |
58 |
|
4287 |
30 Jan 23 |
peter |
\f$ \hat{x} = F_t x_{t-1} \f$ |
4287 |
30 Jan 23 |
peter |
60 |
|
4287 |
30 Jan 23 |
peter |
and the covariance of the state variable |
4287 |
30 Jan 23 |
peter |
62 |
|
4287 |
30 Jan 23 |
peter |
\f$ \hat{P} = F_t P_{t-1} F_t^T + Q_t \f$ |
4287 |
30 Jan 23 |
peter |
64 |
|
4287 |
30 Jan 23 |
peter |
\param F state transition model |
4287 |
30 Jan 23 |
peter |
\param Q process noise |
4287 |
30 Jan 23 |
peter |
67 |
|
4287 |
30 Jan 23 |
peter |
\return predicted state, xhat. |
4287 |
30 Jan 23 |
peter |
69 |
|
4287 |
30 Jan 23 |
peter |
\see xhat() and Phat() |
4287 |
30 Jan 23 |
peter |
71 |
*/ |
4287 |
30 Jan 23 |
peter |
72 |
const utility::Vector& |
4287 |
30 Jan 23 |
peter |
73 |
predict(const utility::MatrixBase& F, const utility::MatrixBase& Q); |
4287 |
30 Jan 23 |
peter |
74 |
|
4287 |
30 Jan 23 |
peter |
75 |
/** |
4287 |
30 Jan 23 |
peter |
\brief update the state from an observation, z. |
4287 |
30 Jan 23 |
peter |
77 |
|
4287 |
30 Jan 23 |
peter |
It is assumed that predict() has been called so predicted values, |
4287 |
30 Jan 23 |
peter |
\f$ \hat{x} \f$ and \f$ \hat{P} \f$, have been updated. |
4287 |
30 Jan 23 |
peter |
80 |
|
4287 |
30 Jan 23 |
peter |
The state variable is updated as |
4287 |
30 Jan 23 |
peter |
82 |
|
4287 |
30 Jan 23 |
peter |
\f$ x_t = \hat{x} + K_t \hat{y}_t \f$ |
4287 |
30 Jan 23 |
peter |
84 |
|
4287 |
30 Jan 23 |
peter |
and covariance as |
4287 |
30 Jan 23 |
peter |
86 |
|
4287 |
30 Jan 23 |
peter |
\f$ P_t = \left( I - K_t H_t\right) \hat{P}_t \f$ |
4287 |
30 Jan 23 |
peter |
88 |
|
4287 |
30 Jan 23 |
peter |
where |
4287 |
30 Jan 23 |
peter |
90 |
|
4287 |
30 Jan 23 |
peter |
\f$ K_t = \hat{P}_t H_t S_t^{-1} \f$ |
4287 |
30 Jan 23 |
peter |
92 |
|
4287 |
30 Jan 23 |
peter |
\f$ S_t = H_t \hat{P}_t H_t^T + R_t \f$ |
4287 |
30 Jan 23 |
peter |
94 |
|
4287 |
30 Jan 23 |
peter |
\f$ \hat{y}_t = z_t - H_t \hat{x}_t \f$ |
4287 |
30 Jan 23 |
peter |
96 |
|
4287 |
30 Jan 23 |
peter |
\param z observed values |
4287 |
30 Jan 23 |
peter |
\param H observation model |
4287 |
30 Jan 23 |
peter |
\param R observation noise |
4287 |
30 Jan 23 |
peter |
100 |
|
4287 |
30 Jan 23 |
peter |
\return estimated state variable, x. |
4287 |
30 Jan 23 |
peter |
102 |
*/ |
4287 |
30 Jan 23 |
peter |
103 |
const utility::Vector& |
4287 |
30 Jan 23 |
peter |
104 |
update(const utility::VectorBase& z, const utility::MatrixBase& H, |
4287 |
30 Jan 23 |
peter |
105 |
const utility::MatrixBase& R); |
4287 |
30 Jan 23 |
peter |
106 |
|
4287 |
30 Jan 23 |
peter |
107 |
/** |
4287 |
30 Jan 23 |
peter |
Equivalent with calling predict() followed by update(). |
4287 |
30 Jan 23 |
peter |
109 |
*/ |
4287 |
30 Jan 23 |
peter |
110 |
void operator()(const utility::VectorBase& z, |
4287 |
30 Jan 23 |
peter |
111 |
const utility::MatrixBase& F, |
4287 |
30 Jan 23 |
peter |
112 |
const utility::MatrixBase& Q, |
4287 |
30 Jan 23 |
peter |
113 |
const utility::MatrixBase& H, |
4287 |
30 Jan 23 |
peter |
114 |
const utility::MatrixBase& R); |
4287 |
30 Jan 23 |
peter |
115 |
/** |
4287 |
30 Jan 23 |
peter |
\brief logarithm of marginal likelihood |
4287 |
30 Jan 23 |
peter |
117 |
|
4287 |
30 Jan 23 |
peter |
The marginal log likelihood is calculated as |
4287 |
30 Jan 23 |
peter |
119 |
|
4287 |
30 Jan 23 |
peter |
\f$ -0.5 \left(\hat{y}^TS^{-1}\hat{y} + \ln |S| + d \ln 2\pi \right) \f$ |
4287 |
30 Jan 23 |
peter |
where \a d is the dimension of the model. |
4287 |
30 Jan 23 |
peter |
122 |
*/ |
4287 |
30 Jan 23 |
peter |
123 |
double logL(void) const; |
4287 |
30 Jan 23 |
peter |
124 |
|
4287 |
30 Jan 23 |
peter |
125 |
/** |
4287 |
30 Jan 23 |
peter |
\brief Kalman gain |
4287 |
30 Jan 23 |
peter |
127 |
|
4287 |
30 Jan 23 |
peter |
K is calculated in update() as |
4287 |
30 Jan 23 |
peter |
129 |
|
4287 |
30 Jan 23 |
peter |
\f$ K = \hat{P} H^T S^{-1} \f$ |
4287 |
30 Jan 23 |
peter |
131 |
|
4287 |
30 Jan 23 |
peter |
\return K, Kalman gain |
4287 |
30 Jan 23 |
peter |
133 |
*/ |
4287 |
30 Jan 23 |
peter |
134 |
const utility::Matrix& K(void) const; |
4287 |
30 Jan 23 |
peter |
135 |
|
4287 |
30 Jan 23 |
peter |
136 |
/** |
4287 |
30 Jan 23 |
peter |
\brief Predicted estimate |
4287 |
30 Jan 23 |
peter |
138 |
|
4287 |
30 Jan 23 |
peter |
Estimate is calculated in predict() as |
4287 |
30 Jan 23 |
peter |
\f$ x_t = \hat{x} + K_t \hat{y}_t \f$ |
4287 |
30 Jan 23 |
peter |
141 |
*/ |
4287 |
30 Jan 23 |
peter |
142 |
const utility::Vector& xhat(void) const; |
4287 |
30 Jan 23 |
peter |
143 |
|
4287 |
30 Jan 23 |
peter |
144 |
/** |
4287 |
30 Jan 23 |
peter |
\brief Predicted estimate covariance |
4287 |
30 Jan 23 |
peter |
146 |
|
4287 |
30 Jan 23 |
peter |
Covariance is calculated in predict() as |
4287 |
30 Jan 23 |
peter |
\f$ \hat{P} = F P F^T + Q \f$ |
4287 |
30 Jan 23 |
peter |
149 |
*/ |
4287 |
30 Jan 23 |
peter |
150 |
const utility::Matrix& Phat(void) const; |
4287 |
30 Jan 23 |
peter |
151 |
|
4287 |
30 Jan 23 |
peter |
152 |
/** |
4287 |
30 Jan 23 |
peter |
\brief Covariance of pre-fit residual |
4287 |
30 Jan 23 |
peter |
154 |
|
4287 |
30 Jan 23 |
peter |
Covariance is calculated in evaluate() as |
4287 |
30 Jan 23 |
peter |
156 |
|
4287 |
30 Jan 23 |
peter |
\f$ S = H \hat{P} H^T + R \f$ |
4287 |
30 Jan 23 |
peter |
158 |
*/ |
4287 |
30 Jan 23 |
peter |
159 |
const utility::Matrix& S(void) const; |
4287 |
30 Jan 23 |
peter |
160 |
|
4287 |
30 Jan 23 |
peter |
161 |
/** |
4287 |
30 Jan 23 |
peter |
\brief post-fit residual |
4287 |
30 Jan 23 |
peter |
163 |
|
4287 |
30 Jan 23 |
peter |
It is calculated in evaluate() as |
4287 |
30 Jan 23 |
peter |
165 |
|
4287 |
30 Jan 23 |
peter |
\f$ y = z - Hx \f$ |
4287 |
30 Jan 23 |
peter |
167 |
*/ |
4287 |
30 Jan 23 |
peter |
168 |
const utility::Vector& y(void) const; |
4287 |
30 Jan 23 |
peter |
169 |
|
4287 |
30 Jan 23 |
peter |
170 |
/** |
4287 |
30 Jan 23 |
peter |
\brief prefit residual |
4287 |
30 Jan 23 |
peter |
172 |
|
4287 |
30 Jan 23 |
peter |
It is calculated in evaluate() as |
4287 |
30 Jan 23 |
peter |
174 |
|
4287 |
30 Jan 23 |
peter |
\f$ \hat{y} = z - H\hat{x} \f$ |
4287 |
30 Jan 23 |
peter |
176 |
*/ |
4287 |
30 Jan 23 |
peter |
177 |
const utility::Vector& yhat(void) const; |
4287 |
30 Jan 23 |
peter |
178 |
private: |
4287 |
30 Jan 23 |
peter |
179 |
utility::Vector x_; |
4287 |
30 Jan 23 |
peter |
180 |
utility::Matrix P_; |
4287 |
30 Jan 23 |
peter |
181 |
utility::DiagonalMatrix I_; |
4287 |
30 Jan 23 |
peter |
182 |
utility::Vector xhat_; |
4287 |
30 Jan 23 |
peter |
183 |
utility::Matrix Phat_; |
4287 |
30 Jan 23 |
peter |
184 |
utility::Vector yhat_; |
4287 |
30 Jan 23 |
peter |
185 |
utility::Matrix S_; |
4287 |
30 Jan 23 |
peter |
186 |
double ln_det_S_; |
4287 |
30 Jan 23 |
peter |
187 |
utility::Matrix S_inverse_; |
4287 |
30 Jan 23 |
peter |
188 |
utility::Matrix K_; |
4287 |
30 Jan 23 |
peter |
189 |
utility::Vector y_; |
4287 |
30 Jan 23 |
peter |
190 |
}; |
4287 |
30 Jan 23 |
peter |
191 |
|
4287 |
30 Jan 23 |
peter |
192 |
}}} // end of namespaces statistics, yat, and theplu |
4287 |
30 Jan 23 |
peter |
193 |
|
4287 |
30 Jan 23 |
peter |
194 |
#endif |