Sensors, Measurement systems Signal processing and Inverse problems Exercises

Similar documents
Problem Weight Total 100

Inverse Filtering by Signal Reconstruction from Phase. Megan M. Fuller

Fourier Integral Representations Basic Formulas and facts

Digital Image and Fourier Transform

Problem Set #1 Problem Set Due: Friday, April 12

ECE 45 Homework 2. t x(τ)dτ. Problem 2.2 Find the Bode plot (magnitude and phase) and label all critical points of the transfer function

Course Web site:

System Identification

Research Article. ISSN (Print) *Corresponding author Shireen Fathima

Sampling Issues in Image and Video

MUSICAL NOTE AND INSTRUMENT CLASSIFICATION WITH LIKELIHOOD-FREQUENCY-TIME ANALYSIS AND SUPPORT VECTOR MACHINES

The Effect of Plate Deformable Mirror Actuator Grid Misalignment on the Compensation of Kolmogorov Turbulence

The Bias-Variance Tradeoff

Introduction to Signal Processing D R. T A R E K T U T U N J I P H I L A D E L P H I A U N I V E R S I T Y

Latin Square Design. Design of Experiments - Montgomery Section 4-2

ECE438 - Laboratory 4: Sampling and Reconstruction of Continuous-Time Signals

EVALUATION OF SIGNAL PROCESSING METHODS FOR SPEECH ENHANCEMENT MAHIKA DUBEY THESIS

GG450 4/12/2010. Today s material comes from p in the text book. Please read and understand all of this material!

ECE438 - Laboratory 1: Discrete and Continuous-Time Signals

Math Assignment 10

Query By Humming: Finding Songs in a Polyphonic Database

Audio-Based Video Editing with Two-Channel Microphone

A Parametric Autoregressive Model for the Extraction of Electric Network Frequency Fluctuations in Audio Forensic Authentication

Orthogonal rotation in PCAMIX

ISSN ICIRET-2014

MUSI-6201 Computational Music Analysis

Fourier Transforms 1D

Design Approach of Colour Image Denoising Using Adaptive Wavelet

A Parametric Autoregressive Model for the Extraction of Electric Network Frequency Fluctuations in Audio Forensic Authentication

Hidden Markov Model based dance recognition

Single Channel Speech Enhancement Using Spectral Subtraction Based on Minimum Statistics

NENS 230 Assignment #2 Data Import, Manipulation, and Basic Plotting

HIGH-DIMENSIONAL CHANGEPOINT ESTIMATION

Digital Signal. Continuous. Continuous. amplitude. amplitude. Discrete-time Signal. Analog Signal. Discrete. Continuous. time. time.

Chapter 6. Normal Distributions

Figure 1: Feature Vector Sequence Generator block diagram.

EE 261 The Fourier Transform and its Applications Fall 2007 Problem Set Two Due Wednesday, October 10

A NEW LOOK AT FREQUENCY RESOLUTION IN POWER SPECTRAL DENSITY ESTIMATION. Sudeshna Pal, Soosan Beheshti

Scalability of delays in input queued switches

From Fourier Series to Analysis of Non-stationary Signals - X

Design and Analysis of New Methods on Passive Image Forensics. Advisor: Fernando Pérez-González. Signal Theory and Communications Department

with - < n < +. There are two types of approximations associated with the sampling process. finite precision of the ADC finite sampling frequency.

DIGITAL COMMUNICATION

UNIVERSITY OF BOLTON SCHOOL OF ENGINEERING BENG (HONS) ELECTRICAL AND ELECTRONIC ENGINEERING SEMESTER 2 EXAMINATION 2016/2017

Using Multiple DMs for Increased Spatial Frequency Response

Research Article Instrument Identification in Polyphonic Music: Feature Weighting to Minimize Influence of Sound Overlaps

Week 14 Music Understanding and Classification

Application Of Missing Feature Theory To The Recognition Of Musical Instruments In Polyphonic Audio

Release Year Prediction for Songs

Experiment 2: Sampling and Quantization

Supervised Learning in Genre Classification

Department of Electrical & Electronic Engineering Imperial College of Science, Technology and Medicine. Project: Real-Time Speech Enhancement

Adaptive decoding of convolutional codes

Improved High-Definition Video by Encoding at an Intermediate Resolution

2. AN INTROSPECTION OF THE MORPHING PROCESS

Classification of Timbre Similarity

Design and analysis of non-uniform rate digital controllers

ADAPTIVE DIFFERENTIAL MICROPHONE ARRAYS USED AS A FRONT-END FOR AN AUTOMATIC SPEECH RECOGNITION SYSTEM

Deep Aesthetic Quality Assessment with Semantic Information

Restoration of Hyperspectral Push-Broom Scanner Data

Optimum Frame Synchronization for Preamble-less Packet Transmission of Turbo Codes

International Journal of Scientific & Engineering Research, Volume 5, Issue 4, April ISSN

Tutorial on Technical and Performance Benefits of AD719x Family

DEPFET Active Pixel Sensors for the ILC

S Queueing Theory. M/G/1-LIFO-PR queue. Samuli Aalto TKK Helsinki University of Technology

E E Introduction to Wavelets & Filter Banks Spring Semester 2009

S Queueing Theory. M/G/1-PS queue. Samuli Aalto TKK Helsinki University of Technology. Lect_MG1-PS.ppt S Queueing Theory Fall 2009

DATA COMPRESSION USING THE FFT

10:15-11 am Digital signal processing

An Accurate Timbre Model for Musical Instruments and its Application to Classification

MUSIC TRANSCRIPTION USING INSTRUMENT MODEL

A Statistical Framework to Enlarge the Potential of Digital TV Broadcasting

WE treat the problem of reconstructing a random signal

A New Method for Processing Impact Excited Continuous-Scan Laser. Doppler Vibrometer Measurements

Problem Weight Score Total 100

Linear mixed models and when implied assumptions not appropriate

Bunch-by-bunch feedback and LLRF at ELSA

Spectrum Sensing by Cognitive Radios at Very Low SNR

LCD and Plasma display technologies are promising solutions for large-format

MELODY EXTRACTION BASED ON HARMONIC CODED STRUCTURE

HomeLog: A Smart System for Unobtrusive Family Routine Monitoring

Random seismic noise reduction using fuzzy based statistical filter

Rhythm together with melody is one of the basic elements in music. According to Longuet-Higgins

Exercises. ASReml Tutorial: B4 Bivariate Analysis p. 55

Signals and Systems. Spring Room 324, Geology Palace, ,

Introduction to Digital Signal Processing (Discrete-time Signal Processing) Prof. Ja-Ling Wu Dept. CSIE & GINM National Taiwan University

5.7 Gabor transforms and spectrograms

A New Method for Calculating Music Similarity

Subject-specific observed profiles of change from baseline vs week trt=10000u

06-Inverse Theory. February 28, 2018

4.4 The FFT and MATLAB

LOCOCODE versus PCA and ICA. Jurgen Schmidhuber. IDSIA, Corso Elvezia 36. CH-6900-Lugano, Switzerland. Abstract

A NOVEL CEPSTRAL REPRESENTATION FOR TIMBRE MODELING OF SOUND SOURCES IN POLYPHONIC MIXTURES

Sampling. Sampling. CS 450: Introduction to Digital Signal and Image Processing. Bryan Morse BYU Computer Science

Technical report on validation of error models for n.

Upgrading E-learning of basic measurement algorithms based on DSP and MATLAB Web Server. Milos Sedlacek 1, Ondrej Tomiska 2

Image Resolution and Contrast Enhancement of Satellite Geographical Images with Removal of Noise using Wavelet Transforms

Predictability of Music Descriptor Time Series and its Application to Cover Song Detection

Dr. David A. Clifton Group Leader Computational Health Informatics (CHI) Lab Lecturer in Engineering Science, Balliol College

Research Article Design and Analysis of a High Secure Video Encryption Algorithm with Integrated Compression and Denoising Block

Transcription:

Files: http://djafari.free.fr/cours/master_mne/cours/cours_mne_2014_01.pdf A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 1/17. Sensors, Measurement systems Signal processing and Inverse problems Exercises Ali Mohammad-Djafari Laboratoire des Signaux et Systèmes, UMR8506 CNRS-SUPELEC-UNIV PARIS SUD 11 SUPELEC, 91192 Gif-sur-Yvette, France http://lss.supelec.free.fr Email: djafari@lss.supelec.fr http://djafari.free.fr

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 2/17 Exercise 1: Fourier and Laplace transforms Consider the following signals: 1. f(t) = a sin(ωt) 2. f(t) = a cos(ωt) 3. f(t) = K k=1 [a k sin(ω k t)+b k sin(ω k t)] 4. f(t) = K k=1 a k exp[ j(ω k t)] 5. f(t) = a exp [ t 2] 6. f(t) = K k=1 a k exp [ 1 2 (t m ] k) 2 /v k 7. f(t) = a sin(ωt)/(ωt) 8. f(t) = 1, if t < a, 0 elsewhere For each of these signals, first compute their Fourier Transform F(ω), then write a Matlab program to plot these signals and their corresponding F(ω). Consider the following signals: 1. f(t) = a exp[ t/τ], t > 0 2. f(t) = 0, t 0, 1 For each of these signals, compute their Laplace Transform g(s).

Exercise 2: Input-Output modeling, Transfert function Consider the following system: with RC = 1. f(t) R C g(t) Write the expression of the transfer function H(ω) = G(ω) F(ω) Write the expression of the impulse response h(t) Write the expression of the relation linking the output g(t) to the input f(t) and the impulse response h(t) Write the expression of the relation linking the Fourier transforms G(ω), F(ω) and H(ω) Write the expression of the relation linking the Laplace transforms G(s), F(s) and H(s) A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 3/17

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 4/17 Exercise 2 (continued) Give the expression of the output when the input is f(t) = δ(t) Give the expression of the { output when the input is a step 0 t < 0, function f(t) = u(t) = 1 t 0 Give the expression of the output when the input is f(t) = asin(ω 0 t) Give the expression of the output when the input is f(t) = k f k sin(ω k t) Give the expression of the output when the input ist f(t) = j f jδ(t t j )

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 5/17 Exercise 3: Averaging to increase accuracy Let note x N = 1 N N n=1 x(n), v N = 1 N N n=1 (x(n) x N) 2 x N 1 = 1 N 1 N 1 n=1 x(n), v N 1 = 1 N 1 N 1 n=1 (x(n) x N) 2 Show that Updating mean and variance: x N = N 1 N x N 1 + 1 N x(n) = x N 1 + 1 N (x(n) x N 1) v N = N 1 N v N 1 + N 1 N (x(n) x 2 N ) 2 Updating inverse of the variance: v 1 N = N N 1 v 1 N 1 + with ρ N = (x(n) x N ) 2 v 1 N 1 Vectorial data x n N (N 1)(N+ρ N ) (x(n) x N) 2 v 2 N 1 x N = N 1 N x N 1 + 1 N x(n) = x N 1 + 1 N (x(n) x N 1) V N = N 1 N V N 1 + N 1 (x(n) x N 2 N )(x(n) x N ) V 1 N = N N 1 V 1 N 1 + N (N 1)(N+ρ N ) V 1 N 1 (x(n) x N)(x(n) x N ) V 1 N 1 with ρ N = (x(n) x N ) V 1 N 1 (x(n) x N)

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 6/17 Exercise 4: Forward modeling Consider the following system: with H(ω) = 1 1+jω. Find h(t). f(t) H(ω) g(t) For a given input f(t) give the general expression of the output g(t). f(t) give the general expression of the output g(t). Give the expression of the { output when the input is a step 0 t < 0, function f(t) = u(t) = 1 t 0 Give the expression of the output when the input is f(t) = asin(ω 0 t)

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 7/17 Exercise 5: Discretization and forward computation Consider the following general system: f(t) h(t) g(t) For a given input f(t) give the general expression of the output g(t). Give the expression of the output when the input is f(t) = N n=0 f nδ(t jn ) with = 1. K Suppose that h(t) = h k δ(t k ) with = 1, Compute k=0 the output g(t) for t = 0,,M with = 1 and M > N. Show that if g(t) is sampled at the same sampling period δ = 1, we have g(t) = M m=0 g mδ(t m ). Then show that g m = K h k f nk k=0

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 8/17 Exercise 5 (continued) Show that the relation between f = [f 0,,f N ], h = [h 0,,h K ] and g = [g 0,,g M ] can be written as g = Hf or as g = Fh. give the expressions and the structures of the matrices H and F. What do you remark on the structure of these two matrices? Write a Matlab programs which compute g when f and h are given. Let name this program g=direct(h,f,method) where method will indicate different methods to use to do the computation. Test it with creating different inputs and different impulse responses and compute the outputs.

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 9/17 Exercise 6: Least Squares and Regularisation In a measurement system, we have established the following relation: g = Hf +ǫ where g is a vector containig the measured data {g m, m = 1,M}, ǫ is a vector representing the errors {ǫ m, m = 1,M}, f is a vector representing the unknowns {f n, n = 1,N}, and H is a matrix with the elements {a mn } depending on the geometry of the measurement system and assumed to be known. Suppose first M = N and that the matrix H be invertible. Why the solution f 0 = H 1 g is not, in general, a satisfactory solution? What relation exists between δ f 0 f 0 and δg g?

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 10/17 Exercise 6 (continued) Let come back to the general case M N. Show then that the Least Squares (LS) solution, i.e. f1 which minimises J 1 (f) = g Hf 2 is also a solution of equation H Hf = H g and if H H is invertible, then we have f 1 = [H H] 1 H g What is the relation between δ f 1 f 1 and δg g? What is the relation between the covarience of f 1 and covarience of g?

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 11/17 Exercise 6 (continued) Consider now the case M < N. Evidently, g = Hf has infinite number of solutions. The minimum norm solution is: { f = arg min f 2 } Hf=g Show that this solution is obtained via: [ ][ ] I H t f = H 0 λ [ 0 g ] which gives: f 2 = H t (HH t ) 1 g if HH t is invertible. Show that with this solution we have: ĝ = H f 2 = g. What is the relation between the covarience of f 2 and covarience of g?

where D is a matrix approximating the operator of derivation. Show that this solution is given by: f 2 = argmin f {J 2 (f)} = [ H H +λd D ] 1 H g A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 12/17 Exercise 6 (continued) Let come back to the general case M N and define f = argmin f {J(f)} with J(f) = g Hf 2 +λ f 2 Show that for any λ > 0, this solution exists and is unique and is obtained by: f = [H H +λi] 1 H g What relation exists between ĝ = H f and g? What is the relation between the covarience of f and covarience of g? Another regularized solution f 2 to this problem is to minimize a criterion such as: J 2 (f) = g Hf 2 +λ Df 2,

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 13/17 Exercise 6 (continued) Suppose that H and D be circulent matrices and symmetric. Then, show that the regularised solution f 2 can be written using the DFT by: where F(ω) = 1 H(ω) H(ω) 2 H(ω) 2 +λ D(ω) 2G(ω) H(ω) is the DFT of the first ligne of the matrix H, D(ω) is the DFT of the first ligne of the matrix D F(ω) is the DFT of the solution vector f2, et G(ω)is the DFT of the data measurement vector g. Comment the expressions of f 2 in the question 3. and F(ω) in the question 4. when λ = 0 and when λ.

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 14/17 Exercise 7: Sensor output noise filtering: Bayesian approach We have a sensor output signal g(t) which is very noisy. Let note the non-noisy signalf(t), then we have: g(t) = f(t)+ǫ(t). Let, first assume the noise to be modelled by a centered and Gaussian probability law with known variance σ 2 ǫ. We have observed this signal at times: t = 1,2,,n and let note by g = [g 1,,g n ], f = [f 1,,f n ] and ǫ = [ǫ 1,,ǫ n ], the vectors containing, respectively, the samples of g(t), f(t) and ǫ(t). Part 1: First we model the signal by a separable Gaussian model: p(f j ) = N(f 0,σ 2 f ), j. Give the expressions of p(ǫ j ) et p(g j f j ) and then p(f j g j ) using the Bayes rule. Show that p(f j g j ) = N( µ j, v j ) and give the expressions of µ j et v j. Give the expression of the Maximum A posteriori estimate f j of f j.

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 15/17 Part 2: Now, let model the input signal by a first order AR model: f j = af j 1 +ξ j where we assume ξ j N(0,σ 2 f ). write the expression of p(f j f j 1 ) and then p(f) and show that it can be [ written as: ] p(f) exp 1 Df 2 where D is a matrix that you give σf 2 the expression. Write the expressions of p(g f) and then using the a priori p(f) give the expression of the a posteriori p(f g). Show that the a posteriori law p(f g) is given by: p(f g) exp [ 1 2 J(f)] where you give the expression of J(f)and the expression of the Maximum A posteriori (MAP estimate: fmap = argmax f {p(f g)}.

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 16/17 Exercise 8: Identification and Deconvolution Consider the problem of deconvolution where the measured signal g(t) is related to the input signal f(t) and the impulse response h(t) by g(t) = h(t) f(t)+ǫ(t) and where we are looking to estimate h(t) from the knowledge of the input f(t) and output g(t) and to estimate f(t) from the knowledge of the impulse response h(t) and output g(t). Given f(t) and g(t), describe different methods for estimating h(t). Write a Matlab program which can compute h given f and g. Let name it: h=identification(g,f,method). Test it by creating different inputs f and outputs g. Think also about the noise. Once test your programs without noise, then add some noise on the output g and test them again.

A. Mohammad-Djafari, Sensors, Measurement systems, Signal processing and Inverse problems, Master MNE 2014, 17/17 Exercise 8: (continued) Given g(t) and h(t), describe different methods for estimating f(t). Write a Matlab program which can compute f given g and h. Let name it: f=inversion(g,h,method). Test it by creating different inputs f and outputs g. Think also about the noise. Once test your programs without noise, then add some noise on the output g and test them again. Bring back your experiences and comments.