Lecture 10 - Model Identification

Similar documents
References are appeared in the last slide. Last update: (1393/08/19)

Linear Gaussian State Space Models

Augmented Reality II - Kalman Filters - Gudrun Klinker May 25, 2004

CHAPTER 10 VALIDATION OF TEST WITH ARTIFICAL NEURAL NETWORK

Modal identification of structures from roving input data by means of maximum likelihood estimation of the state space model

An recursive analytical technique to estimate time dependent physical parameters in the presence of noise processes

Vectorautoregressive Model and Cointegration Analysis. Time Series Analysis Dr. Sevtap Kestel 1

Ensamble methods: Bagging and Boosting

Financial Econometrics Jeffrey R. Russell Midterm Winter 2009 SOLUTIONS

Ensamble methods: Boosting

Georey E. Hinton. University oftoronto. Technical Report CRG-TR February 22, Abstract

Distribution of Least Squares

Dimitri Solomatine. D.P. Solomatine. Data-driven modelling (part 2). 2

GMM - Generalized Method of Moments

Diebold, Chapter 7. Francis X. Diebold, Elements of Forecasting, 4th Edition (Mason, Ohio: Cengage Learning, 2006). Chapter 7. Characterizing Cycles

Lecture 1 Overview. course mechanics. outline & topics. what is a linear dynamical system? why study linear systems? some examples

The fundamental mass balance equation is ( 1 ) where: I = inputs P = production O = outputs L = losses A = accumulation

Chapter 3: Signal Transmission and Filtering. A. Bruce Carlson Paul B. Crilly 2010 The McGraw-Hill Companies

Notes on Kalman Filtering

Lectures 29 and 30 BIQUADRATICS AND STATE SPACE OP AMP REALIZATIONS. I. Introduction

Deep Learning: Theory, Techniques & Applications - Recurrent Neural Networks -

Smoothing. Backward smoother: At any give T, replace the observation yt by a combination of observations at & before T

Detecting nonlinear processes in experimental data: Applications in Psychology and Medicine

Non-parametric techniques. Instance Based Learning. NN Decision Boundaries. Nearest Neighbor Algorithm. Distance metric important

12: AUTOREGRESSIVE AND MOVING AVERAGE PROCESSES IN DISCRETE TIME. Σ j =

Stable block Toeplitz matrix for the processing of multichannel seismic data

Institute for Mathematical Methods in Economics. University of Technology Vienna. Singapore, May Manfred Deistler

PENALIZED LEAST SQUARES AND PENALIZED LIKELIHOOD

3.1.3 INTRODUCTION TO DYNAMIC OPTIMIZATION: DISCRETE TIME PROBLEMS. A. The Hamiltonian and First-Order Conditions in a Finite Time Horizon

Kriging Models Predicting Atrazine Concentrations in Surface Water Draining Agricultural Watersheds

An introduction to the theory of SDDP algorithm

Wisconsin Unemployment Rate Forecast Revisited

Probabilistic Robotics

Non-parametric techniques. Instance Based Learning. NN Decision Boundaries. Nearest Neighbor Algorithm. Distance metric important

L07. KALMAN FILTERING FOR NON-LINEAR SYSTEMS. NA568 Mobile Robotics: Methods & Algorithms

Block Diagram of a DCS in 411

SUPPLEMENTARY INFORMATION

ACE 564 Spring Lecture 7. Extensions of The Multiple Regression Model: Dummy Independent Variables. by Professor Scott H.

What Ties Return Volatilities to Price Valuations and Fundamentals? On-Line Appendix

Problem Formulation in Communication Systems

Answers to Exercises in Chapter 7 - Correlation Functions

Frequency independent automatic input variable selection for neural networks for forecasting

Learning a Class from Examples. Training set X. Class C 1. Class C of a family car. Output: Input representation: x 1 : price, x 2 : engine power

Západočeská Univerzita v Plzni, Czech Republic and Groupe ESIEE Paris, France

The general Solow model

Zürich. ETH Master Course: L Autonomous Mobile Robots Localization II

KEY. Math 334 Midterm I Fall 2008 sections 001 and 003 Instructor: Scott Glasgow

Comparing Means: t-tests for One Sample & Two Related Samples

Lecture 10 Estimating Nonlinear Regression Models

A Specification Test for Linear Dynamic Stochastic General Equilibrium Models

Forecasting optimally

Testing for a Single Factor Model in the Multivariate State Space Framework

Chapter 5. Heterocedastic Models. Introduction to time series (2008) 1

Time series Decomposition method

Resource Allocation in Visible Light Communication Networks NOMA vs. OFDMA Transmission Techniques

Using the Kalman filter Extended Kalman filter

STAD57 Time Series Analysis. Lecture 14

Random Walk with Anti-Correlated Steps

Learning a Class from Examples. Training set X. Class C 1. Class C of a family car. Output: Input representation: x 1 : price, x 2 : engine power

Distribution of Estimates

UNIVERSITÀ DI PISA DIPARTIMENTO DI INGEGNERIA MECCANICA, NUCLEARE E DELLA PRODUZIONE VIA DIOTISALVI 2, PISA

Data Fusion using Kalman Filter. Ioannis Rekleitis

Continuous Time. Time-Domain System Analysis. Impulse Response. Impulse Response. Impulse Response. Impulse Response. ( t) + b 0.

Explaining Total Factor Productivity. Ulrich Kohli University of Geneva December 2015

Nature Neuroscience: doi: /nn Supplementary Figure 1. Spike-count autocorrelations in time.

STATE-SPACE MODELLING. A mass balance across the tank gives:

Multi-scale 2D acoustic full waveform inversion with high frequency impulsive source

( ) ( ) if t = t. It must satisfy the identity. So, bulkiness of the unit impulse (hyper)function is equal to 1. The defining characteristic is

Online Convex Optimization Example And Follow-The-Leader

NMR Spectroscopy: Principles and Applications. Nagarajan Murali 1D - Methods Lecture 5

Robert Kollmann. 6 September 2017

14 Autoregressive Moving Average Models

Introduction to Mobile Robotics

State-Space Models. Initialization, Estimation and Smoothing of the Kalman Filter

FAULT DETECTION FOR NONLINEAR SYSTEMS WITH MULTIPLE PERIODIC INPUTS. Z. Y. Yang and C. W. Chan

Vector autoregression VAR. Case 1

Adaptive Noise Estimation Based on Non-negative Matrix Factorization

Unit Root Time Series. Univariate random walk

Model Reduction for Dynamical Systems Lecture 6

Application of a Stochastic-Fuzzy Approach to Modeling Optimal Discrete Time Dynamical Systems by Using Large Scale Data Processing

Empirical Process Theory

Components Model. Remember that we said that it was useful to think about the components representation

Problem Set 5. Graduate Macro II, Spring 2017 The University of Notre Dame Professor Sims

Course Notes for EE227C (Spring 2018): Convex Optimization and Approximation

Chapter 2. First Order Scalar Equations

The electromagnetic interference in case of onboard navy ships computers - a new approach

Lecture 33: November 29

EE 435. Lecture 31. Absolute and Relative Accuracy DAC Design. The String DAC

OBJECTIVES OF TIME SERIES ANALYSIS

Dynamic Effects of Feedback Control!

INTRODUCTION TO MACHINE LEARNING 3RD EDITION

Tracking. Many slides adapted from Kristen Grauman, Deva Ramanan

Article from. Predictive Analytics and Futurism. July 2016 Issue 13

DIRECTIONAL CHANGE IN A PRIORI ANTI-WINDUP COMPENSATORS VS. PREDICTION HORIZON

The Rosenblatt s LMS algorithm for Perceptron (1958) is built around a linear neuron (a neuron with a linear

Applications in Industry (Extended) Kalman Filter. Week Date Lecture Title

Understanding the asymptotic behaviour of empirical Bayes methods

Mechanical Fatigue and Load-Induced Aging of Loudspeaker Suspension. Wolfgang Klippel,

Finite Horizon Control Design for Optimal Model Discrimination

Time Series Forecasting using CCA and Kohonen Maps - Application to Electricity Consumption

Transcription:

Lecure - odel Idenificaion Wha is ssem idenificaion? Direc impulse response idenificaion Linear regression Regularizaion Parameric model ID nonlinear LS Conrol Engineering -

Wha is Ssem Idenificaion? Experimen Plan Daa Ssem Idenificaion odel Rarel used in real-life conrol Whie-box idenificaion esimae parameers of a phsical model from daa Example: aircraf fligh model Gra-box idenificaion given generic model srucure esimae parameers from daa Example: neural nework model of an engine Black-box idenificaion deermine model srucure and esimae parameers from daa Example: securi pricing models for sock marke Conrol Engineering -

Indusrial Use of Ssem ID Process conrol - mos developed ID approaches all plans and processes are differen need o do idenificaion canno spend oo much ime on each indusrial idenificaion ools Aerospace whie-box idenificaion speciall designed programs of ess Auomoive whie-box significan effor on model developmen and calibraion Disk drives used o do horough idenificaion shorer ccle ime Embedded ssems simplified models shor ccle ime Conrol Engineering -3

Impulse response idenificaion Simples approach: appl conrol impulse and collec he IPULSE RESPOSE daa 3 4 5 6 7 IE Difficul o appl a shor impulse big enough such ha he response is much larger han he noise.8.6.4. OISY IPULSE RESPOSE.5 -.5 3 4 5 6 7 IE FIR modeling can be used for building simplified conrol design models from complex sims Conrol Engineering -4

Sep response idenificaion Sep bump conrol inpu and collec he daa used in process conrol.5 SEP RESPOSE OF PAPER WEIGH Acuaor bumped.5 4 6 8 IE SEC Impulse esimae: impulse sep-sep- Sill nois.3.. IPULSE RESPOSE OF PAPER WEIGH 3 4 5 6 IE SEC Conrol Engineering -5

oise reducion oise can be reduced b saisical averaging: Collec daa for muliple sep inpus and perform more averaging o esimae he sep/pulse response Use a parameric model of he ssem and esimae a few model parameers describing he response: dead ime rise ime gain Do boh in a sequence done in real process conrol ID packages Pre-filer daa Conrol Engineering -6

Linear Regression - univariae Simple fiing problem: Given model sep response And process sep response Find he gain facor + e.5.5 SEP RESPOSE OF PAPER WEIGH 4 6 8 IE SEC Φ + e Φ e e e Soluion assuming uncorrelaed noise: Φ Φ Φ Conrol Engineering -7

Conrol Engineering -8 Linear Regression Linear regression is one of he main Ssem ID ools e j j j + Daa Regression weighs Regressor Error of he fi + e Φ Φ e e e O

Linear regression - leas squares akes sense onl when marix Φ is all > more daa available han he number of unknown parameers. Saisical averaging Leas square soluion: e min ˆ L L Φ Φ min Φ Φ Φ Φ Φ Can be compued using alab pinv or lef marix division \ Conrol Engineering -9

Conrol Engineering - Linear regression - leas squares Correlaion inerpreaion of he leas squares soluion Φ Φ Φ ˆ c R O R c ˆ Informaion marix Correlaion vecor Φ Φ R c Φ

Example: Firs-order ARA model a + gu + e Linear regression represenaion a u g ˆ Φ Φ Φ his pe of approach is considered in mos of he echnical lieraure on idenificaion alab Idenificaion oolbox Limied indusrial use Fundamenal issue: + + e Lennar Ljung Ssem Idenificaion: heor for he User nd Ed 999 Small error in a migh mean large change in he ssem response Conrol Engineering -

Regularizaion Linear regression where Φ Φ is ill-condiioned Insead of e min solve a regularized problem e + r min Φ + e where r is a small regularizaion parameer A..ikhonov 963 see hp://solon.cma.univie.ac.a/~neum/ms/reguorial.pdf Regularized soluion ˆ ri Φ Φ + Φ Cu off he singular values of Φ ha are smaller han r Conrol Engineering -

Regularizaion Analsis hrough SVD singular value decomposiion R U R ; S diag{ s Φ USV U U VV I Regularized soluion s j ri V ˆ Φ Φ + Φ diag U s j r + j Cu off he singular values of Φ ha are smaller han r V ; j} j Inverse singular values /s Regularized inverse s values s +. IVERSE - - - - SIGULAR VALUE Conrol Engineering -3 s

Linear regression for FIR model Idenifing impulse response b.5 appling muliple seps PRBS exciaion signal -.5 - FIR impulse response model 3 4 5 h k u Linear regression represenaion k u u Φ u u k Regularized LS soluion: u + e u ˆ O u ri Φ Φ + Φ PRBS EXCIAIO SIGAL PRBS Pseudo-Random Binar Sequence See IDIPU in alab u h + h Conrol Engineering -4

Example: FIR model ID PRBS exciaion inpu.5 -.5 PRBS exciaion Simulaed ssem oupu: 4 samples random noise of he ampliude.5-4 6 8.5 -.5 - SYSE RESPOSE 4 6 8 IE Conrol Engineering -5

Example: FIR model ID Linear regression esimae of he FIR model Impulse response for he simulaed ssem:..5..5 -.5 3 4 5 6 7..5..5 FIR esimae Impulse Response -.5 3 4 5 6 7 ime sec H f[.5][. ]; P cdh.5; Conrol Engineering -6

onlinear parameric model ID Predicion model depending on he unknown parameer vecor u ODEL ˆ Opimizer Loss Index L onlinear regression: loss index L ˆ min Ieraive numerical opimizaion. Compuaion of L as a subrouine L ˆ u odel including he parameers sim Lennar Ljung Idenificaion for Conrol: Simple Process odels IEEE Conf. on Decision and Conrol Las Vegas V Conrol Engineering -7

Parameric SsID of sep response Firs order process wih deadime os common indusrial process model Response o a conrol sep applied a B τ g γ + g e B D / τ for for > B B D D D γ g τ D Example: Paper machine process Conrol Engineering -8

Conrol Engineering -9 Sep: Gain and Offse Esimaion wo-sep approach: linear regression + nonlinear regression For given he modeled sep response can be presened in he form his is a linear regression Parameer esimae and predicion for given ; D D g τ γ τ + ; k k k D τ τ D D Φ Φ Φ ˆ ˆ τ ˆ ˆ ˆ D D g τ γ τ + g D τ γ τ D

Sep : Rise ime & Dead ime Esimaion For an given τ D he loss index is L ˆ τ D Grid τ D and find he minimum of L L τ D Conrol Engineering -

Examples: Sep Response ID Idenificaion resuls for real indusrial process daa his algorihm works in an indusrial ool used in 5+ indusrial plans man processes each.6 Process parameers: Gain.34; del.; rise 9.8969.8.6 onlinear Regression ID.4..4. -. Linear Regression ID of he firs-order model.8.6.4. onlinear Regression ID -.4 3 4 5 6 7 8 -. 3 4 5 6 7 8 ime in sec.; D response - solid; esimaed response - dashed Conrol Engineering -

Linear Filering in SsID A rick ha helps: pre-filer daa Consider daa model h * u + e u Plan SsID ĥ F is a linear filering operaor usuall LPF { F F h * u + Fe { f F h * u e Fh * u f h * Fu Plan Can esimae h from filered and filered u Or can esimae filered h from filered and raw u Pre-filer bandwidh limis he esimaion bandwidh u Conrol Engineering - F SsID F ĥ

ulivariable Idenificaion Sep/impulse response idenificaion is a ke par of he indusrial mulivariable odel Predicive Conrol packages Appl SISO ID o various inpu/oupu pairs eed n ess: excie each inpu in urn and collec all oupus a ha Conrol Engineering -3