代做ECE5550: Applied Kalman Filtering THE LINEAR KALMAN FILTER代做Java程序

- 首页 >> Algorithm 算法

ECE5550: Applied Kalman Filtering

THE LINEAR KALMAN FILTER

4.1: Introduction

The principal goal of this course is to learn how to estimate the

present hidden state (vector) value of a dynamic system, using noisy measurements that are somehow related to that state (vector).

We assume a general, possibly nonlinear, model

xk = fk1(xk1, uk1,wk1)

z k = hk (xk, uk ,vk ),

where uk is a known (deterministic/measured) input signal, wk is a    process-noise random input, and vk is a sensor-noise random input.

SEQUENTIAL PROBABILISTIC INFERENCE: Estimate the present state xk of a dynamic system using all measurements Zk = {z0 , z 1 , ··· , z k } .

This notes chapter provides a unified theoretic framework to develop a family of estimators for this task: particle filters, Kalman filters, extended Kalman filters, sigma-point (unscented) Kalman filters. . .

A smattering of estimation theory

There are various approaches to “optimal estimation” of some unknown quantity x .

One says that we would like to minimize the expected magnitude

(length) of the error vector between x and the estimatex(ˆ) .

This turns out to be the median of the a posteriori pdf f (x | Z).

A similar result, but easier to derive analytically minimizes the expected length squared of that error vector.

This is the minimum mean square error (MMSE) estimator

We solve forx(ˆ) by differentiating the cost function and setting the

result to zero

Another approach to estimation is to optimize a likelihood function

Yet a fourth is the maximum a posteriori estimate

In general,x(^)MME  /=x(^)MMSE  /=x(^)ML  /=x(^)MAP , so which is best”?

Answer: It probably depends on the application.

The text gives some metrics for comparison: bias, MSE, etc.

Here, we usex(^)MMSE  = E[x | Z] because it “makes sense” and works

well in a lot of applications and is mathematically tractable.

Some examples

In example 1, mean, median, and mode are identical. Any of these statistics would make a good estimator of x .

In example 2, mean, median, and mode are all different. Which to choose is not necessarily obvious.

In example 3, the distribution is multi-modal. None of the estimates is likely to be satisfactory!

4.2: Developing the framework

The Kalman filter applies the MMSE estimation criteria to a dynamic system. That is, our state estimate is the conditional mean

where Rxk is the set comprising the range of possible xk .

To make progress toward implementing this estimator, we must break f (xk | Zk ) into simpler pieces.

We first use Bayes’ rule to write:

We then break up Zk into smaller constituent parts within the joint probabilities as Zk1  and z k

Thirdly, we use the joint probability rule f (a , b) = f (a | b)f (b) on the numerator and denominator terms

Next, we apply Bayes’ rule once again in the terms within the [   ]

We now cancel some terms from numerator and denominator

Finally, recognize that zk is conditionally independent of Zk1  given xk

So, overall, we have shown that

KEY POINT #1: This shows that we can compute the desired density recursively with two steps per iteration:

The first step computes probability densities for predicting xk given all past observations

The second step updates the prediction via

Therefore, the general sequential inference solution breaks naturally into a prediction/update scenario.

To proceed further using this approach, the relevant probability densities may be computed as

KEY POINT #2: Closed-form. solutions to solving the multi-dimensional integrals is intractable for most real-world systems.

For applications that justify the computational expense, the integrals may be approximated using Monte Carlo methods (particle filters).

But, besides applications using particle filters, this approach appears to be a dead end.

KEY POINT #3: A simplified solution may be obtained if we are willing to

make the assumption that all probability densities are Gaussian.

This is the basis of the original Kalman filter, the extended Kalman

filter, and the sigma-point (unscented) Kalman filters to be discussed.





站长地图