kalmanfilteringinR(R语言实现卡尔曼滤波）_平方根卡尔曼滤波资源-CSDN文库

4星 · 超过85%的资源需积分: 49 88 浏览量 2018-07-13 12:10:40 上传评论 4 收藏 842KB PDF 举报

在讨论R语言实现卡尔曼滤波时，我们需要首先了解卡尔曼滤波算法本身及其在R语言中的具体应用。卡尔曼滤波是一种高效的递归滤波器，用于估计线性动态系统的状态。它在各种领域如信号处理、航天、经济时间序列分析等领域有着广泛的应用。 **卡尔曼滤波算法的原理** 卡尔曼滤波算法的核心思想是在给定当前观测和上一时刻状态估计的情况下，对系统的状态变量进行最优估计。该算法利用了系统的动态信息和观测信息，通过预测和更新两个步骤交替进行，以实现对系统状态的最优估计。其基本步骤通常包括： 1. 预测（Predict）：基于模型的动态方程和先前的状态估计，对当前状态进行预测，同时计算预测的误差协方差。 2. 更新（Update）：将观测数据与预测值结合，利用卡尔曼增益对状态估计进行修正，并更新误差协方差。 **卡尔曼滤波在R语言中的实现** R语言是一种用于统计计算和图形表示的编程语言和软件环境。R语言通过一系列的工具包（Packages）来支持状态空间估计和卡尔曼滤波的实现。在R语言中实现卡尔曼滤波，通常使用以下几种工具包： 1. **stats包**：这是R的基础包，其中包含了简单的卡尔曼滤波实现。 2. **KFAS（Kalman Filter and Smoothers for Exponential Family State Space Models）**：用于状态空间模型的卡尔曼滤波和平滑。 3. **dse包**：提供了对动态系统估计（Dynamic Systems Estimation）的支持。 4. **dlm包**：主要用于动态线性模型（Dynamic Linear Models）的滤波、平滑和预测。以上提及的这些包和更多其他包，如**MARSS包**，它们中一些还提供了对非高斯模型的滤波和估计、模拟平滑和扰动平滑以及帮助动态线性模型的贝叶斯分析的功能。 **不同类型的卡尔曼滤波** 在R语言中实现的卡尔曼滤波算法可以有多种变体，其中包括： - **标准卡尔曼滤波**：适用于线性高斯系统的标准形式。 - **平方根卡尔曼滤波**：通过使用Cholesky分解等技巧来避免数值计算中的不稳定性。 - **序贯卡尔曼滤波**：在每个时间点依次进行观测更新和状态预测，以减少计算负担。 - **扩展卡尔曼滤波**：适用于非线性系统的卡尔曼滤波算法，通过局部线性化处理非线性系统。 - **区间卡尔曼滤波**：用于处理模型不确定性和观测噪声的滤波方法。 **R语言实现卡尔曼滤波的关键知识点** - 状态空间模型：包含状态方程和观测方程两个部分，分别描述了系统的内部动态和如何从状态生成观测数据。 - 状态估计和误差协方差：算法需要估算系统的当前状态以及其误差协方差，以此来表示估计的不确定性。 - 卡尔曼增益：在更新步骤中用于加权预测值和观测值，以得到更准确的状态估计。 - 预测和更新的迭代过程：是卡尔曼滤波的核心，每一次迭代都会基于新观测更新状态估计。 - 平滑和模拟平滑：除了滤波过程外，一些算法还提供平滑功能，可以利用未来观测值的信息来改善过去状态的估计。在应用中，为了使得卡尔曼滤波的实现更加高效和准确，通常需要对具体的系统模型进行参数化和调试。R语言中的工具包为我们提供了进行状态空间模型分析的强大工具集，使得在R环境下实现卡尔曼滤波变得方便和灵活。同时，对于初学者和非专业人士来说，理解这些工具包的文档和使用示例也非常重要，以确保正确和有效地利用这些工具来解决实际问题。

资源详情

资源评论

JSS

Journal of Statistical Software

March 2011, Volume 39, Issue 2. http://www.jstatsoft.org/

Kalman Filtering in R

Fernando Tusell

University of the Basque Country

Abstract

Support in R for state space estimation via Kalman ﬁltering was limited to one package,

until fairly recently. In the last ﬁve years, the situation has changed with no less than

four additional packages oﬀering general implementations of the Kalman ﬁlter, including in

some cases smoothing, simulation smoothing and other functionality. This paper reviews

some of the oﬀerings in R to help the prospective user to make an informed choice.

Keywords: state space models, Kalman ﬁlter, time series, R.

1. Introduction

The Kalman ﬁlter is an important algorithm, for which relatively little support existed in R

(R Development Core Team 2010) up until fairly recently. Perhaps one of the reasons is the

(deceptive) simplicity of the algorithm, which makes it easy for any prospective user to throw

in his/her own quick implementation.

While direct transcription of the equations of the Kalman ﬁlter as they appear in many

engineering or time series books may be suﬃcient for some applications, an all-around im-

plementation requires more complex coding. In Section 2, we provide a short overview of

available algorithms. It is against this background that we discuss in Section 3 the particular

choices made by four packages oﬀering fairly general Kalman ﬁltering in R, with some mention

of functions in other packages which cater to particular needs.

Kalman ﬁltering is a large topic. In the sequel we focus on linear Gaussian models and

their estimation, which is what the packages we review oﬀer in common (and the foundation

on which most anything else rests). Other functionalities present in some of the packages

examined include ﬁltering and estimation of non-Gaussian models, simulation and disturbance

smoothing and functions to help with the Bayesian analysis of dynamic linear models, etc.

none of which are assessed.

2 Kalman Filtering in R

2. Kalman ﬁlter algorithms

We shall consider a fairly general state-space model speciﬁcation, suﬃcient for the purpose

of the discussion to follow in Section 3, even if not the most comprehensive. The notation

follows Harvey (1989). Let

= c

+ T

t−1

+ R

(1)

= d

+ Z

+ 

(2)

where η

∼ N(0, Q

) and 

∼ N(0, H

). The state equation (1) describes the dynamics of

the state vector α

, driven by deterministic (c

) and stochastic (η

) inputs. The observation

(or measurement) equation links the observed response y

with the unobserved state vector,

with noise 

and (possibly) deterministic inputs d

. Matrices T

, R

, Z

, Q

and H

may

depend on a vector of parameters, θ, and be time varying or constant over time. The noises η

and 

are assumed serially and also mutually uncorrelated, i.e., E[η



] = 0 for all t, s. The

last assumption and the gaussianity of η

and 

can be dispensed with; see e.g., Anderson

and Moore (1979).

2.1. The Kalman ﬁlter.

The Kalman ﬁlter equations, with slight notational variations, are standard in any textbook:

see, e.g., Anderson and Moore (1979), Simon (2006), Durbin and Koopman (2001), Grewal

and Andrews (2001), West and Harrison (1997) or Shumway and Stoﬀer (2006), to name only

a few. We reproduce those equations here, however, as repeated reference is made to them in

the sequel. Deﬁne

t−1

= E[α

t−1

, . . . , y

t−1

] (3)

t−1

= E[(α

t−1

− a

t−1

)(α

t−1

− a

t−1

)

] ; (4)

estimates of the state vector and its covariance matrix at time t with information available

at time t − 1, a

t|t−1

and P

t|t−1

respectively, are given by the time update equations

t|t−1

= T

t−1

+ c

(5)

t|t−1

= T

t−1

+ R

. (6)

Let F

= Z

t|t−1

+ H

. If a new observation is available at time t, then a

t|t−1

and

t|t−1

can be updated with the measurement update equations

= a

t|t−1

+ P

t|t−1

−1

− Z

t|t−1

− d

) (7)

= P

t|t−1

− P

t|t−1

−1

t|t−1

. (8)

Equations (5)–(6) and (7)–(8) taken together make up the Kalman ﬁlter. Substituting (5)

in (7) and (6) in (8), a single set of equations linking a

t−1

and P

t−1

to a

and P

can be

obtained. In order to start the iteration we need initial values of a

−1

and P

−1

(or a

0|−1

and

0|−1

The ﬁlter equations (6) and (8) propagate the covariance matrix of the state, and are said

to deﬁne a covariance ﬁlter (CF). Equivalent equations can be written which propagate the

matrix P

−1

, giving an information ﬁlter (IF). Information ﬁlters require in general more

Journal of Statistical Software 3

computational eﬀort. One possible advantage is that they provide a natural way to specify

complete uncertainty about the initial value of a component of the state: we can set the

corresponding diagonal term in the information matrix to zero. With a covariance ﬁlter, we

have to set the corresponding variance in P

0|−1

to a “large” number, or else use exact diﬀuse

initialization, an option described below.

Direct transcription of the equations making up the Kalman ﬁlter into computer code is

straightforward. It was soon noticed, though, that the resulting programs suﬀered from

numerical instability; see for instance Bucy and Joseph (1968). In particular, buildup of

ﬂoating point errors in equation (8) may eventually yield non symmetric or non positive

deﬁnite P

matrices. An alternative to equation (8) (more expensive from the computational

point of view) is:

= (I − K

t|t−1

(I − K

)

+ K

(9)

with K

= P

t|t−1

−1

; but even this (“Joseph stabilized form”, see Bucy and Joseph

(1968), p. 175) may prove insuﬃcient to prevent roundoﬀ error degeneracy in the ﬁlter. A

detailed reference describing the pitfalls associated with the numerical implementation of the

Kalman ﬁlter is Bierman (1977); see also Grewal and Andrews (2001), Chap. 6 and Anderson

and Moore (1979), Chap. 6. Square root algorithms, that we discuss next, go a long way in

improving the numerical stability of the ﬁlter.

2.2. Square root algorithms

Consider a matrix S

such that P

= S

; square root covariance ﬁlter (SRCF) algorithms

propagate S

instead of P

with two main beneﬁts (cf. Anderson and Moore (1979), § 6.5):

(i) Re-constitution of P

from S

will always yield a symmetric non negative matrix, and (ii)

The numerical condition of S

will in general be much better than that of P

. If instead of

the covariance matrix we choose to factor the information matrix P

−1

we have a square root

information ﬁlter algorithm (SRIF).

It is easy to produce a replacement for the time update equation (6). Consider an orthogonal

matrix G such that:





= G



t−1

1/2

)



(10)

where M is an upper triangular matrix. Matrix G can be constructed in a variety of ways,

including repeated application of Householder or Givens transforms. Multiplication of the

transpose of (10) by itself produces, taking into account the orthogonality of G,

M = T

t−1

+ R

1/2

)

(11)

= T

t−1

+ R

; (12)

comparison with (6) shows that M is a possible choice for S

t|t−1

. Likewise, it can be shown

(Simon 2006, Section 6.3.4) that the orthogonal matrix G

∗

such that



+ Z

t|t−1

) K

∗

0 M

∗



= G

∗

1/2

)

t|t−1

(13)

produces in the left-hand side of (13) a block M

∗

that can be taken as S

, and thus performs

the measurement update.

4 Kalman Filtering in R

Although the matrices performing the block triangularizations described in equations (10)

and (13) can be obtained quite eﬃciently (Lawson and Hanson 1974; Gentle 2007), clearly

the computational eﬀort is greater than that required by the time and measurement updates

in equations (6) and (8); square root ﬁltering does have a cost.

In the equations above G and G

∗

can be chosen so that M and M

∗

are Cholesky factors

of the corresponding covariance matrices. This needs not be so, and other factorizations

are possible. In particular, the factors in the singular value decomposition of P

t−1

can be

propagated: see for instance Zhang and Li (1996) and Appendix B of Petris et al. (2009). Also,

we may note that time and measurement updates can be merged in a single triangularization

(see Anderson and Moore 1979, p. 148, and Vanbegin and Verhaegen 1989).

2.3. Sequential processing

In the particular case where H

is diagonal, the components of y

= (y

, . . . , y

) are

uncorrelated. We may pretend that we observe one y

at a time and perform p univariate

measurement updates similar to (7)–(8), followed by a time update (5)–(6) – see Durbin and

Koopman (2001, Section 6.4) or Anderson and Moore (1979) for details.

The advantage of sequential processing is that F

becomes 1 ×1 and the inversion of a p ×p

matrix in equation (7)–(8) is avoided. Clearly, the situation where we stand to gain most

from this strategy is when the dimension of the observation vector y

is large.

Sequential processing can be combined with square root covariance and information ﬁlters,

although in the latter case the computational advantages seem unclear (Anderson and Moore

1979, p. 142; see also Bierman 1977).

Aside from the case where H

is diagonal, sequential processing may also be used when H

is block diagonal – in which case we can perform a sequence of reduced dimension updates

– or when it can be reduced to diagonal form by a linear transformation. In the last case,

assuming full rank, let H

−1/2

be a square root of H

−1

. Multiplying (2) through by H

−1/2

we get

∗

= d

∗

+ Z

∗

+ 

∗

(14)

with E[

∗



∗

] = I. If matrix H

is time-invariant, the same linear transformation will

decorrelate the measurements at all times.

2.4. Smoothing and the simulation smoother

The algorithms presented produce predicted a

t|t−1

or ﬁltered a

values of the state vector

. Sometimes it is of interest to estimate a

t|N

for 0 < t ≤ N, i.e., E[α

, . . . , y

], the

value of the state vector given all past and future observations. It turns out that this can

be done running once the Kalman ﬁlter and then a recursion backwards in time (Durbin and

Koopman 2001, Section 4.3, Harvey 1989, Section 3.6).

In some cases, and notably for the Bayesian analysis of the state space model, it is of interest

to generate random samples of state and disturbance vectors, conditional on the observations

, . . . , y

. Fr

uhwirth-Schnatter (1994) and Carter and Kohn (1994) provided algorithms to

that purpose, improved by de Jong (1995). Durbin and Koopman (2001, Section 4.7) draws

on work of the last author; the algorithm they present is fairly involved. Recently, Durbin

and Koopman (2002) have provided a much simpler algorithm for simulation smoothing of

the Gaussian state space model; see also Strickland et al. (2009).

Journal of Statistical Software 5

2.5. Exact diﬀuse initial conditions

As mentioned above, the Kalman ﬁlter iteration needs starting values a

−1

and P

−1

. When

nothing is known about the initial state, a customary practice has been to set P

−1

with

“large” elements along the main diagonal, to reﬂect our uncertainty. This practice has been

criticized on the ground that

“While [it can] be useful for approximate exploratory work, it is not recommended

for general use, since it can lead to large rounding errors.”

(Durbin and Koopman 2001, p. 101). Grewal and Andrews (2001, Section 6.3.2) show how

this may come about when elements of P

−1

are set to values “too large” relative to the mea-

surement variances. An alternative is to use an information ﬁlter with the initial information

matrix (or some diagonal elements of it) set to zero. Durbin and Koopman (2001, Section 5.3)

advocate a diﬀerent approach: see also Koopman and Durbin (2003) and Koopman (1997).

The last reference discusses several alternatives and their respective merits.

2.6. Maximum likelihood estimation

An important special case of the state space model is that in which some or all of the matrices

, R

, Z

, Q

and H

in equations (1)–(2) are time invariant and depend on a vector of

parameters, θ that we seek to estimate.

Assuming α

∼ N (a

, P

) with both a

and P

known, the log likelihood is given by,

L(θ) = log p(y

, . . . , y

|θ)

= −

(N + 1)p

log(2π) −

t=0



log |F

| + e

−1



(15)

where e

= y

− Z

. (Durbin and Koopman 2001, Section 7.2; the result goes back to

Schweppe 1965). Except for |F

|, all other quantities in (15) are computed when running

the Kalman ﬁlter (cf. equations (5)–(8)); therefore, the likelihood is easily computed. (If

we resort to sequential processing, Section 2.3, the analog of equation (15) does not require

computation of determinants nor matrix inversions.)

Maximum likelihood estimation can therefore be implemented easily: it suﬃces to write a

routine computing (15) as a function of the parameters θ and use R functions such as optim

or nlminb to perform the optimization. An alternative is to use the EM algorithm which quite

naturally adapts to this likelihood maximization problem, but that approach has generally

been found slower than the use of quasi-Newton methods, see for instance Shumway and

Stoﬀer (2006), p. 345.

3. Kalman ﬁltering in R

There are several packages available from the Comprehensive R Archive Network (CRAN)

oﬀering general Kalman ﬁlter capabilities, plus a number of functions scattered in other

packages which cater to special models or problems. We describe ﬁve of those packages in

chronological order of ﬁrst appearance on CRAN.

剩余26页未读，继续阅读

评论收藏

内容反馈

weixin_40508312

2019-10-06

好实用的资源！

kalman filtering in R(R语言实现卡尔曼滤波）

评论2

最新资源

kalman filtering in R(R语言实现卡尔曼滤波）

评论2

最新资源

相关推荐

卡尔曼滤波原理及应用仿真__MATLAB.rar_kalman；matlab_卡尔曼滤波

kalman filtering

Kalman Filtering

Kalman filtering_卡尔曼滤波_

卡尔曼滤波算法（KalmanFilter）学习

mathnet-filtering：Math.NET过滤（以前称为Neodym）

卡尔曼滤波.rar_卡尔曼_卡尔曼滤波

R包动态线性模型

Kalman Filtering - Theory and Practice Using MATLAB 4th

kalman_filter.rar_卡尔曼滤波实现

kalman卡尔曼滤波器的各种应用实现

kalman-filtering.rar_Kalman filter_kalman_oja_卡尔曼_卡尔曼滤波

vb语言实现的卡尔曼滤波源程序.rar_kalman + VB_kalman filter vb_卡尔曼_卡尔曼滤波 vb

卡尔曼滤波c语言实现

卡尔曼滤波（Kalman Filter）原理介绍

卡尔曼.rar_卡尔曼_卡尔曼滤波_卡曼尔滤波

Matlab实现卡尔曼滤波(Kalman Filtering)、卡尔曼平滑器(Kalman smoother)和缺失数据插值

Matlab实现卡尔曼滤波(Kalman Filtering)时间序列分析

卡尔曼滤波.zip经典卡尔曼Kalman filtering和UKF滤波实验代码

卡尔曼滤波（Kalman filter）

卡尔曼滤波(KALMAN）

卡尔曼滤波工具箱(kalman filter box)

Kalman滤波c++实现

卡尔曼滤波简介及其实现(附C代码)

卡尔曼滤波算法流程

卡尔曼滤波原理-Kalman Filtering Dan Simen

MATLAB在卡尔曼滤波器中应用的理论与实践Kalman