Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing...
Transcript of Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing...
![Page 1: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/1.jpg)
Learning basis vectors and weights for EEG signals
vibhor kumar
T-61-181 Biomedical Signal Processing
Sections 4.5.3 – 4.6.2
![Page 2: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/2.jpg)
Finding optimal basis function to denoise
1. Karhunen-Loeve Expansion
2. Interprtaion as Linear , Time-Variant Filtering
3. Damped Sinusoids : Finding Optimal function
Adaptive Analysis Using Basis Functions
1. Instantaneous LMS algorithm
2. Block LMS algorithm
![Page 3: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/3.jpg)
Karhunen-Loeve Expansion – optimal basis function
To design the basis function such that the signal part is efficientlyrepresented witha small number of functions
Derivation:
Decompostion of signal x in to two sums representing signal and noise
Aim is to find set of such that resembles original signal as closely as possible.
Can be done by minimising the noise power estimate:
![Page 4: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/4.jpg)
PROOF
considering that the actual signal is composed of the noise power estimate becomes.
= = as noise is zero-mean with correlationmatrix
as signal noise are assumed to be uncorrelated.
![Page 5: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/5.jpg)
Since noise power is not dependent on . The minimisation of is equivalent to making resemble S
![Page 6: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/6.jpg)
Optimal basis function and correlation matrix of data
We start frim the equation of noise power estimate again:
=
since = = =
Minimisation by the use of Lagrange multiplier to ensure that orthonoramiltyof is mentained. we get the function to be minimised as :
lagrange multipliers
![Page 7: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/7.jpg)
MSE as Sum of Eigen values
The minimization of the above function with lagrange multipliers caneb done by taking gradient with respect to and setting it to zero, which gives:
or
so mean square error can be represented in terms of lagrange coefficients oreigen values
So is minmised when N- K smallest eigenvalues are chosen
or
When signal is represented in terms of basis vectorcorresponding to largest K eiegenvales
![Page 8: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/8.jpg)
Interpretations
It can be viewed as findingdirection of maximum variance in n dimensional space
Interpretation as Linear, Time Variant Filtering
when signal is estimated as means of different basis function i.e.
where
![Page 9: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/9.jpg)
Interpretation in frequency domain
orthonormal basis
functions
Instantaneousfrequency responses
![Page 10: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/10.jpg)
Modeling with Damped Sinusoids
Sinusoidal basis function with exponential damping can be used to make a variety of basis functions, such as data x can be represented as :
frequencyamplitude
phase
Damping factor
Since x(n) is generally real valued signal, so data can be represented by damped cosines because sum of conjugate complex pairs is cosine.
![Page 11: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/11.jpg)
Pros and cons of damped sinusoid as basis function
.Damping add an extra degree of freedom , so fewer basis function represent the data
No orthonormality
Non linearity of equations makes it tough to find the solutions for best basis function
![Page 12: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/12.jpg)
Prony Method
The model
Can be represented as where
This model can be viewed as homogenous soultion to a linear difference equation
with fixed parameters. i.e.
x(n) + a1x(n-1) +. . . + akx(n-k) = 0
In prony’s original method it assumed that number of available samples is equal
to unknown parameters, so the difference equation is valid for n = k, . . . 2k-1
![Page 13: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/13.jpg)
The difference equation can be represented as k x k matrix equation as:
On solving the above equation we get values of a1 . . .ak which can be used to find the roots of the polynomial
The roots can be used to determine the parameters h1, . . . hk from the eq.
![Page 14: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/14.jpg)
Prony method for real problems: least square Prony method
When we want to use less parameters k than available number of observations N we have relax the requirement of difference equation as given below
x(n) + a1x(n-1) +. . . + akx(n-k) = e(n) error
And to minimise the error e(n) w.r.t the parameters a1, . . .ak
The parameters h1 . . .hk is determined by minimizing the function:
And the least square solution comes out to be
![Page 15: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/15.jpg)
Adaptive analysis using Basis Functions
Instanteneous LMS Algorithm
The observed signal is made from concatenation of successive Evokedpotentials
The siganl estimate is calcukated from truncated series expansion butthis time with time varying weight vector w(n). For this purpose minimsation of MSE is done at every sample. For this the differenital of the function
is taken to yeild : -
![Page 16: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/16.jpg)
using the gradient to update weights and using mean µ instead of expectationthe update formula is :
which when expressed in terms of initial weight vector w(0) can bewritten as
where
![Page 17: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/17.jpg)
A representation of adaptive analysis using LMS algorithm
![Page 18: Learning basis vectors and weights for EEG signalsvibhor kumar T-61-181 Biomedical Signal Processing Sections 4.5.3 – 4.6.2 Finding optimal basis function to denoise 1. Karhunen-Loeve](https://reader034.fdocuments.in/reader034/viewer/2022052613/5f265c7060c1665f317cf6c2/html5/thumbnails/18.jpg)
The Block LMS algorithm
Its a extension of singla trial LMS method using truncated series expansion.
the error function is quiet simmilar
but wi-1 is used instead of wi
so the updateing of weights is done by
wi = (1-µ)wi-1 + µF Ts xi or wi = wi-1 + µF T
s ( xi – F swi-1 )