Prologue to Kalman Filters

Introduction to kalman filters l.jpg
1 / 24
0
0
1091 days ago, 363 views
PowerPoint PPT Presentation

Presentation Transcript

Slide 1

Prologue to Kalman Filters CEE 6430: Probabilistic Methods in Hydroscienecs Fall 2008 Acknowledgments: Numerous sources on WWW, book, papers

Slide 2

Overview What could Kalman Filters be utilized for as a part of Hydrosciences? What is a Kalman Filter? Calculated Overview The Theory of Kalman Filter (just the conditions you have to utilize) Simple Example (with bunches of yakkity yak talk through freebees)

Slide 3

A "Hydro" Example Suppose you have a hydrologic display that predicts stream water level each hour (utilizing the typical data sources). You realize that your model is not impeccable and you don't believe it 100%. So you need to send somebody to check the stream level face to face. Nonetheless, the stream level must be checked once per day around twelve and not each hour. Moreover, the individual who measures the waterway level can not be trusted 100% either. So how would you join both yields of stream level (from model and from estimation) with the goal that you get an "intertwined" and better gauge? – Kalman separating

Slide 4

Graphically

Slide 5

What is a Filter incidentally? Class – characterize numerically what a channel is (make a similarity to a genuine channel) Other utilizations of Kalman Filtering (or Filtering when all is said in done): Your Car GPS (foresee and upgrade area) Surface to Air Missile (hitting the objective) Ship or Rocket route ( Appollo 11 utilized some kind of sifting to ensure it didn't miss the Moon!)

Slide 6

The Problem in General (we should get somewhat more specialized) Black Box System Error Sources System state can't be measured specifically Need to gauge "ideally" from estimations Sometimes the framework state and the estimation might be two distinct (dislike waterway level case) External Controls System State (coveted however not known) Optimal Estimate of System State Observed Measurements Measuring Devices Estimator Measurement Error Sources

Slide 7

What is a Kalman Filter? Recursive information preparing calculation Generates ideal gauge of coveted amounts given the arrangement of estimations Optimal? For straight framework and white Gaussian blunders, Kalman channel is "ideal" gauge in view of every single past estimation For non-direct framework optimality is "qualified" Recursive? Doesn't have to store every single past estimation and reprocess all information every time step

Slide 8

Conceptual Overview Simple case to rouse the workings of the Kalman Filter The crucial conditions you have to know (Kalman Filtering for Dummies!) Examples: Prediction and Correction

Slide 9

Conceptual Overview Lost on the 1-dimensional line (envision that you are speculating your position by taking a gander at the stars utilizing sextant) Position – y(t) Assume Gaussian dispersed estimations y

Slide 10

Conceptual Overview State space – position Measurement - position Sextant is not immaculate Sextant Measurement at t 1 : Mean = z 1 and Variance =  z1 Optimal gauge of position is: ŷ(t 1 ) = z 1 Variance of blunder in gauge:  2 x (t 1 ) =  2 z1 Boat in same position at time t 2 - Predicted position is z 1

Slide 11

Conceptual Overview expectation ŷ - (t 2 ) State (by taking a gander at the stars at t2) Measurement usign GPS z(t 2 ) So we have the forecast ŷ - (t 2 ) GPS Measurement at t 2 : Mean = z 2 and Variance =  z2 Need to revise the forecast by Sextant because of estimation to get ŷ (t 2 ) Closer to more trusted estimation – would it be a good idea for us to do direct insertion?

Slide 12

Conceptual Overview expectation ŷ - (t 2 ) Kalman channel helps you combine estimation and forecast on the premise of the amount you believe every (I would believe the GPS more than the sextant) adjusted ideal gauge ŷ(t 2 ) estimation z(t 2 ) Corrected mean is the new ideal gauge of position (essentially you've "upgraded" the anticipated position by Sextant utilizing GPS New fluctuation is littler than both of the past two changes

Slide 13

Conceptual Overview (The Kalman Equations) Lessons in this way: Make expectation in view of past information - ŷ - ,  - Take estimation – z k ,  z Optimal gauge (ŷ) = Prediction + (Kalman Gain) * (Measurement - Prediction ) Variance of gauge = Variance of forecast * (1 – Kalman Gain )

Slide 14

Conceptual Overview What if the watercraft was presently moving? ŷ(t 2 ) Naïve Prediction (sextant) ŷ - (t 3 ) At time t 3 , vessel moves with speed dy/dt=u Naïve approach: Shift likelihood to one side to anticipate This would work on the off chance that we knew the speed precisely (culminate demonstrate)

Slide 15

Conceptual Overview Naïve Prediction ŷ - (t 3 ) But you may not be so certain about the correct speed ŷ(t 2 ) Prediction ŷ - (t 3 ) Better to expect defective model by including Gaussian commotion dy/dt = u + w Distribution for expectation moves and spreads out

Slide 16

Conceptual Overview Corrected ideal gauge ŷ(t 3 ) Updated Sextant position utilizing GPS Measurement z(t 3 ) GPS Prediction ŷ - (t 3 ) Sextant Now we take an estimation at t 3 Need to by and by right the forecast Same as before

Slide 17

Conceptual Overview Lessons learnt from theoretical review: Initial conditions ( ŷ k-1 and  k-1 ) Prediction ( ŷ - k ,  - k ) Use starting conditions and model (eg. steady speed) to make expectation Measurement (z k ) Take estimation Correction ( ŷ k ,  k ) Use estimation to right forecast by "mixing" forecast and lingering – dependably an instance of combining just two Gaussians Optimal gauge with littler change

Slide 18

Blending Factor If we are certain about estimations: Measurement mistake covariance (R) abatements to zero K reductions and weights leftover more intensely than forecast If we are certain about expectation Prediction blunder covariance P - k declines to zero K increments and weights forecast more vigorously than remaining

Slide 19

Correction (Measurement Update) Prediction (Time Update) (1) Compute the Kalman Gain (1) Project the state ahead K = P - k H T (HP - k H T + R) - 1 ŷ - k = Ay k-1 + Bu k (2) Update assess with estimation z k (2) Project the blunder covariance ahead ŷ k = ŷ - k + K(z k - H ŷ - k ) P - k = AP k-1 A T + Q (3) Update Error Covariance P k = (I - KH)P - k The arrangement of Kalman Filtering Equations in Detail

Slide 20

Assumptions behind Kalman Filter The model you use to foresee the "state" should be a LINEAR capacity of the estimation (so how would we utilize non-direct precipitation overflow models?) The model mistake and the estimation mistake (commotion) must be Gaussian with zero mean

Slide 21

What if the clamor is NOT Gaussian? Given just the mean and standard deviation of clamor, the Kalman channel is the best straight estimator. Non-straight estimators might be better. Why is Kalman Filtering so well known? · Good results by and by because of optimality and structure. · Convenient frame for online continuous preparing. · Easy to plan and actualize given a fundamental comprehension. · Measurement conditions require not be modified. Likewise prominent in hydrosciences, climate/oceanography/hydrologic displaying, information osmosis

Slide 22

Now ..to comprehend the languages (You may start the freebees ) First read the pass out by PD Joseph Next, read the pass out by Welch and Bishop titled 'An Introduction to the Kalman Filter'. (you can skip pages 4-5, 7-11). Pages 7-11 are on 'Augmented Kalman Filtering' (for non-direct frameworks). Perused the tackled case from pages 11-16.

Slide 23

Homework (theoretical) Explain in NO MORE THAN 1 PAGE the illustration that you read from pages 11-16 in the freebee by Welch and Bishop. Fundamentally, I need you to give me a straightforward theoretical diagram of why and how "sifting" was connected utilizing the past similarity on a watercraft lost in ocean. DUE – Same date as the Class extend report. Additional CREDIT 5% marks– If you audit (3-4 pages) the great paper in 1960 by Kalman (pass out) EXTRA CREDIT 5% marks – in the event that you turn in a nitty gritty synopsis of the STEVE programming (stars/cons, what it is and so on.)

Slide 24

References Kalman, R. E. 1960. "A New Approach to Linear Filtering and Prediction Problems", Transaction of the ASME- - Journal of Basic Engineering, pp. 35-45 (March 1960). Welch, G and Bishop, G. 2001. "A prologue to the Kalman Filter", http://www.cs.unc.edu/~welch/kalman/By the way Dr. Rudolf Kalman is alive and living great today

SPONSORS