Non-parametric identification

Similar documents
Chapter 6: Nonparametric Time- and Frequency-Domain Methods. Problems presented by Uwe

6.435, System Identification

EE531 (Semester II, 2010) 6. Spectral analysis. power spectral density. periodogram analysis. window functions 6-1

Identification, Model Validation and Control. Lennart Ljung, Linköping

Time Series Analysis. Solutions to problems in Chapter 7 IMM

13.42 READING 6: SPECTRUM OF A RANDOM PROCESS 1. STATIONARY AND ERGODIC RANDOM PROCESSES

System Identification & Parameter Estimation

ECE 636: Systems identification

EL1820 Modeling of Dynamical Systems

Stochastic Process II Dr.-Ing. Sudchai Boonto

EL1820 Modeling of Dynamical Systems

Massachusetts Institute of Technology Department of Electrical Engineering and Computer Science : Discrete-Time Signal Processing

LINEAR RESPONSE THEORY

Review of Frequency Domain Fourier Series: Continuous periodic frequency components

X random; interested in impact of X on Y. Time series analogue of regression.

D.S.G. POLLOCK: BRIEF NOTES

System Identification and Parameter Estimation

Signals and Spectra (1A) Young Won Lim 11/26/12

2A1H Time-Frequency Analysis II

Systems Analysis and Control

Fundamentals of the Discrete Fourier Transform

Practical Spectral Estimation

3. ESTIMATION OF SIGNALS USING A LEAST SQUARES TECHNIQUE

System Modeling and Identification CHBE 702 Korea University Prof. Dae Ryook Yang

ENSC327 Communications Systems 2: Fourier Representations. School of Engineering Science Simon Fraser University

Spectral Analysis of Random Processes

System Identification

Linear Approximations of Nonlinear FIR Systems for Separable Input Processes

Linear and Nonlinear Oscillators (Lecture 2)

Time series models in the Frequency domain. The power spectrum, Spectral analysis

IDENTIFICATION FOR CONTROL

f-domain expression for the limit model Combine: 5.12 Approximate Modelling What can be said about H(q, θ) G(q, θ ) H(q, θ ) with

ENSC327 Communications Systems 2: Fourier Representations. Jie Liang School of Engineering Science Simon Fraser University

Continuous-Time Frequency Response (II) Lecture 28: EECS 20 N April 2, Laurent El Ghaoui

2A1H Time-Frequency Analysis II Bugs/queries to HT 2011 For hints and answers visit dwm/courses/2tf

ELEC2400 Signals & Systems

e iωt dt and explained why δ(ω) = 0 for ω 0 but δ(0) =. A crucial property of the delta function, however, is that

EEM 409. Random Signals. Problem Set-2: (Power Spectral Density, LTI Systems with Random Inputs) Problem 1: Problem 2:

e iωt dt and explained why δ(ω) = 0 for ω 0 but δ(0) =. A crucial property of the delta function, however, is that

THE FOURIER TRANSFORM (Fourier series for a function whose period is very, very long) Reading: Main 11.3

Lecture 11: Spectral Analysis

EE538 Final Exam Fall :20 pm -5:20 pm PHYS 223 Dec. 17, Cover Sheet

The Discrete Fourier Transform (DFT) Properties of the DFT DFT-Specic Properties Power spectrum estimate. Alex Sheremet.

SPECTRUM. Deterministic Signals with Finite Energy (l 2 ) Deterministic Signals with Infinite Energy N 1. n=0. N N X N(f) 2

A summary of Modeling and Simulation

Introduction to system identification

Solutions to Problems in Chapter 4

EAS 305 Random Processes Viewgraph 1 of 10. Random Processes

Centre for Mathematical Sciences HT 2017 Mathematical Statistics

Fourier Methods in Digital Signal Processing Final Exam ME 579, Spring 2015 NAME

ROBUST FREQUENCY DOMAIN ARMA MODELLING. Jonas Gillberg Fredrik Gustafsson Rik Pintelon

EE 3054: Signals, Systems, and Transforms Summer It is observed of some continuous-time LTI system that the input signal.

Homework 3 (Stochastic Processes)

Computing inverse Laplace Transforms.

LTI Systems (Continuous & Discrete) - Basics

Approximation Approach for Timing Jitter Characterization in Circuit Simulators

Statistics of Stochastic Processes

Frequency-Domain C/S of LTI Systems

Problem Sheet 1 Examples of Random Processes

IV. Covariance Analysis

Imaging with Ambient Noise II

8. Filter Bank Methods Filter bank methods assume that the true spectrum φ(ω) is constant (or nearly so) over the band [ω βπ, ω + βπ], for some β 1.

Non-parametric estimate of the system function of a time-varying system

Lecture 9 Infinite Impulse Response Filters

Stochastic Processes. Chapter Definitions

Fourier transform representation of CT aperiodic signals Section 4.1

PASSIVE SENSOR IMAGING USING CROSS CORRELATIONS OF NOISY SIGNALS IN A SCATTERING MEDIUM

Intro to Stochastic Systems (Spring 16) Lecture 6

Digital Image Processing Lectures 13 & 14

Computational Methods for Astrophysics: Fourier Transforms

GATE EE Topic wise Questions SIGNALS & SYSTEMS

Atmospheric Flight Dynamics Example Exam 1 Solutions

Frequency estimation by DFT interpolation: A comparison of methods

Aspects of Continuous- and Discrete-Time Signals and Systems

The Spectral Density Estimation of Stationary Time Series with Missing Data

Fourier Analysis of Stationary and Non-Stationary Time Series

Assessing Structural VAR s

Response to Periodic and Non-periodic Loadings. Giacomo Boffi. March 25, 2014

ESTIMATION OF DSGE MODELS WHEN THE DATA ARE PERSISTENT

LINEAR SYSTEMS. J. Elder PSYC 6256 Principles of Neural Coding

Lecture 1: Introduction to System Modeling and Control. Introduction Basic Definitions Different Model Types System Identification

Slepian Functions Why, What, How?

Ghost Imaging. Josselin Garnier (Université Paris Diderot)

A New Approach for Computation of Timing Jitter in Phase Locked Loops

Spatial Statistics with Image Analysis. Lecture L08. Computer exercise 3. Lecture 8. Johan Lindström. November 25, 2016

Linear Filters. L[e iωt ] = 2π ĥ(ω)eiωt. Proof: Let L[e iωt ] = ẽ ω (t). Because L is time-invariant, we have that. L[e iω(t a) ] = ẽ ω (t a).

5 Analog carrier modulation with noise

Fourier Analysis and Power Spectral Density

Introduction to Biomedical Engineering

B2.III Revision notes: quantum physics

covariance function, 174 probability structure of; Yule-Walker equations, 174 Moving average process, fluctuations, 5-6, 175 probability structure of

Unstable Oscillations!

Review of Fourier Transform

STAD57 Time Series Analysis. Lecture 23

14 - Gaussian Stochastic Processes

5. THE CLASSES OF FOURIER TRANSFORMS

Frequency Domain and Filtering

Review of Linear Time-Invariant Network Analysis

arxiv: v4 [gr-qc] 15 Jul 2010

13. Power Spectrum. For a deterministic signal x(t), the spectrum is well defined: If represents its Fourier transform, i.e., if.

Transcription:

Non-parametric Non-parametric Transient Step-response using Spectral Transient Correlation Frequency function estimate Spectral System Identification, SSY230 Non-parametric 1

Non-parametric Transient Step-response using Spectral Consider the system described by or, equivalently, y(t) = y(t) =G 0 (q)u(t)+v(t) g 0 (k)u(t k)+v(t) k=1 Question: Can we determine G 0 (q) or {g 0 (k)} without parameterizing in θ? System Identification, SSY230 Non-parametric 2

Transient Non-parametric Transient Step-response using Spectral Applying the input gives the output u(t) = { k, t =0 0, t 0 y(t) =k g 0 (t)+v(t) which motivates the impulse response estimate ĝ(t) = y(t) k System Identification, SSY230 Non-parametric 3

Step-response Non-parametric Transient Step-response using Spectral Step input: u(t) = gives the output y(t) =k so that ĝ(t) = { k, t 0 0, t < 0 t g 0 (k)+v(t) k=1 y(t) y(t 1) k Problems with transient techniques: excitation disturbances nonlinearities System Identification, SSY230 Non-parametric 4

Some times one is not allowed to do anything else than small step changes. : For a tank with ideal mixing: V dc dt = Q(c i c) V :tankvolume,q: flow,c, c i : tank and flow concentration. Apply an impulsive concentration input: Stewart-Hamilton s equation: c i (t) =δ(t) c(t) = Q V e Q V t = h(t) 0 t h(t)dt = V Q Application: Blood volume measurement System Identification, SSY230 Non-parametric 5

using Non-parametric Transient Step-response using Spectral Assume that u is quasi-stationary and u and v are uncorrelated: Φ yu (ω) =G 0 (e iω )Φ u (ω) or R yu (τ) = g 0 (k)r u (τ k) k=1 Truncate the sum and solve the resulting system of equations for ĝ(k): M ˆR yu(τ) N = ĝ(k) ˆR u N (τ k) k=1 where estimates of the covariance function are formed as ˆR N u (τ) = 1 N N u(t)u(t τ) t=1 System Identification, SSY230 Non-parametric 6

Non-parametric Transient Step-response using Spectral Note: when u( ) is white noise the estimate becomes ĝ(k) = ˆR N yu(k) σ 2 u System Identification, SSY230 Non-parametric 7

Non-parametric Transient Step-response using Spectral So far we have estimated g 0 (k) what about estimating G 0 (e iω )? α cos ωt Measure amplitude and phase using : I c (N) = 1 N α y(t)cosωt N 2 G 0(e iω ) cos φ I s (N) = 1 N α y(t)sinωt N 2 G 0(e iω ) sin φ G α G cos(ωt + φ)+v(t) System Identification, SSY230 Non-parametric 8

Non-parametric Transient Step-response using Spectral implying I c (N) i I s (N) α 2 G 0(e iω ) (cos φ + i sin φ) so that Ĝ N = 2 α (I c(n) i I s (N)) System Identification, SSY230 Non-parametric 9

Interpretation of frequency response by the method: Recall that Y N (ω) = 1 N N t=1 y(t)e iωt = N(I c (N) i I s (N)) so that Ĝ N = 2 α (I c(n) i I s (N)) = 2 α N Y N(ω) = Y N(ω) U N (ω) where the last equality holds for ω = 2π N k. Hence, the estimate obtained using frequency by the method is simply the ratio between the DFT of the output and the input. This can be generalized... System Identification, SSY230 Non-parametric 10

Non-parametric Transient Step-response using Spectral Consider the open-loop system y(t) = G o (q)u(t)+v(t) Y N (ω) = G o (e iω )U N (ω)+v N (ω)+r N (ω) where R N (ω) 1/ N (or =0if u is periodic). function estimate () is: Ĝ(e iω )= Y N(ω) U N (ω), ω = k 2π N System Identification, SSY230 Non-parametric 11

Non-parametric Transient Step-response using Spectral Properties: u periodic number of ω fixed unbiased variance 1/N EĜ N (e iω ) = G 0 (e iω )+ R N (ω) U N (ω) VarĜ N (e iω Φ v (ω) ) U N (ω) 2 u stochastic process number of ω increases asymptotically unbiased variance (SNR) 1 as. uncorrelated System Identification, SSY230 Non-parametric 12

Spectral The is comprised of uncorrelated estimates at different ω but we know that the frequency response is smooth in reality! Idea: (1) Compute Ĝ(eiω ) as a weighted sum of the Ĝ around ω = ω 0 : Ĝ(e iω 0 )= k2 k=k 1 α k Ĝ(e iω k) k2 k=k 1 α k (1) where 2πk 1 /N = ω 0 Δ and 2πk 2 /N = ω 0 +Δ (2) Now choose {α k } inversely proportional to the variance of Ĝ(e iω k): Ĝ(e iω 0 )= UN (ω k ) 2 Φ v (ω k ) Ĝ(e iω k) UN (ω k ) 2 Φ v (ω k ) ω0 +Δ U N (ω) 2 ω 0 Δ Φ v (ω) ω0 +Δ U N (ω) 2 ω 0 Δ Φ v (ω) Ĝ(e iω )dω dω System Identification, SSY230 Non-parametric 13

(3) Put weight according to distance from ω 0 : Ĝ(e iω 0 )= π π W γ(ω ω 0 ) U N (ω) 2 Φ v (ω) Ĝ(e iω )dω π π π π W γ(ω ω 0 ) U W γ(ω ω 0 ) U N (ω) 2 Ĝ(e iω )dω N (ω) 2 π Φ v (ω) dω π W γ(ω ω 0 ) U N (ω) 2 dω π π = W γ(ω ω 0 )Y N (ω)u N (ω)dω π π W γ(ω ω 0 ) U N (ω) 2 = ˆΦ yu (ω 0 ) (2) dω ˆΦ u (ω 0 ) This is the estimate obtained by spectral according to Blackman-Tukey. The estimate is obtained by a natural replacement of the spectra in Φ yu (ω) =G 0 (e iω )Φ u (ω) (3) by the corresponding smoothed periodograms. System Identification, SSY230 Non-parametric 14

Non-parametric Transient Step-response using Spectral Consider the spectral estimate ˆΦ u (ω) = π π W γ (ω ω 0 ) U N (ω) 2 dω where the frequency window, W γ, is narrow for large γ. System Identification, SSY230 Non-parametric 15

Non-parametric Transient Step-response using Spectral The convolution above is transformed into multiplication in the time domain, so that where w γ (τ) = 1 2π ˆR u (τ) = 1 2π ˆΦ u (ω) = π π π π w γ (τ) ˆR u (τ)e iωτ W γ (ω)e iωτ dω U N (ω) 2 e iωτ dω = 1 N N u(t)u(t τ) Note: The time (lag) window w γ (τ) is wide when the frequency window W γ (ω) is narrow, and vice versa. System Identification, SSY230 Non-parametric 16

Non-parametric Transient Step-response using Spectral When γ increases (narrow frequency window), the bias decreases but the variance increases. This can be expressed in terms of the functions M(γ) = W (γ) =2π π π π π when γ.also, W (γ) γ. ω 2 W γ (ω)dω 0 (4) W 2 γ (ω)dω (5) System Identification, SSY230 Non-parametric 17

Non-parametric Transient Step-response using Spectral Theorem: Assume that y(t) =G 0 (q)u(t)+v(t) (6) where v( ) is a stochastic process, independent of u( ), whichis quasi-stationary. Then the following asymptotic results hold when γ,n,γ/n 0: [ ] 1 EĜN (e iω ) G 0 (e iω ) M(γ) 2 G 0 + G 0 Φ u Φ u E ĜN EĜN 2 1 N W γ Φv Φ u ReĜN, ImĜN are as. uncorrelated Estimates at different ω are as. uncorrelated System Identification, SSY230 Non-parametric 18