Wavelets
Wave-like oscillation localized in space (time)
−(x − b)2 /(2a2)
−(x − 2na3
b)e
First derivative of
Example: Gaussian Function
Smaller scale Larger scale (stretched)
[1] Two basic (squished)
properties:
• Scale (dilation) - how “stretched”
or “squished” wavelet is (related
to frequency)
Shifted left Shifted right and
• Location - where wavelet is squished
positioned in time
b
[1] Addison, P.S. (2005). Wavelet transforms and the ecg: A review. Physiological Measurement, 26 (5), p.
R155 1
Wavelet Transform
Decomposition of a signal using wavelets of varying scale and location
Basic idea: compute how much of
wavelet is in the signal for a
particular scale and location
i.e. evaluate convolution of signal and wavelet
at varying scales
Why wavelets?
• Traditional Fourier transform (decomposition using sine and cosine) gives global average over
entire signal, thus may obscure local information
• Wavelet transforms can extract local spectral and temporal information simultaneously
• Variety of wavelets to choose from
Addison, P.S. (2005). Wavelet transforms and the ecg: A review. Physiological Measurement, 26 (5), p. R155 2
Wavelet Transform
Two methods
Continuous Wavelet Transform (CWT) Discrete Wavelet Transform (DWT)
“Mother” (basis) wavelets are defined everywhere “Mother” (basis) wavelets are only defined on discrete grid
/ = /(t) / = /m,n(t)
Transform must be discretized for implementation Transform is already discrete!
o o
1 o
T(a, b) =
x(t) /*
(t − b)
dt
o x(t) /m,n(t) dt
a ! −oo a Tm,n =
! −oo
Coefficients (Tm,n) do not have redundant information since
T(a, b) may have redundant information since discretized wavelets can be defined to be orthonormal
same feature may be captured by multiple scales
Addison, P.S. (2005). Wavelet transforms and the ecg: A review. Physiological Measurement, 26 (5), p.
R155 3
clear; close all; clc
Example
Code to use Maximal Overlap Discrete Wavelet Transform
(MODWT) to Analyze ECG Data
Code authored by: Shawhin Talebi
The University of Texas at Dallas
Mutli-scale Integrated Remote Sensing and Simulation (MINTS)
Load Data
load('ECG.mat');
Perform Wavelet Transform
Use the maximal overlap discrete wavelet transform (MODWT) to enhance the R peaks in the ECG waveform.
The MODWT is an undecimated wavelet transform, which handles arbitrary sample sizes.
Perform MODWT for 6 different scales. The rows of wt give the wavelet (detail) coefficents for each scale. Here
we are using Symlets wavelet with 4 vanishing moments
Other wavelet families can be used. use waveletfamilies('f') to view all
% define number of levels to use
numLevels = 6;
% perform maximal overlap discrete wavelet transform (MODWT)
wt = modwt(ECG, 'sym4', numLevels);
Compare Coefficents for all MODWT Scales
% get number of Wavelet Scales
[numScales, ~] = size(wt);
% create figure
fig = figure(1);
fig.Units = 'normalized';
fig.Position = [0 0 1 1];
% plot original signal
subplot(numScales+1,1,1)
plot(ECG, 'r-')
title('Orginal ECG', 'FontSize', 14)
axis tight
% plot wavelet scale coefficents
for i=2:numScales+1
subplot(numScales+1,1,i)
plot(wt(i-1,:))
title(strcat('Wavelet Scale 2/\', string(i-1)), 'FontSize', 14)
axis tight
end 1
xlabel('Time', 'FontSize', 14)
% print figure to file
print('waveletTransform_signal_decomposition', '-dpng'); https://github.com/ShawhinT/YouTube 4
Reconstruct Signal via Inverse MODWT
Example
% create figure
fig figure(2);
fig.Units
'normalized';
fig.Position
[0 0 1 1];
% initialize an array to store the coefficents corresponding to scale
2A4 recwt zeros(size(wt));
% store coefficents corresponding to scale 4 in
array recwt(4,:) wt(4,:);
% reconstruct signal using only scale 4 with inverse
transform recECG imodwt(recwt,'sym4');
% compare original signal and reconstructed
% plot original signal
subplot(2,1,1)
plot(ECG, 'r-')
title('Orginal ECG',
2
'FontSize', 16)
xlabel('Time',
'FontSize', 14)
axis tight
% plot reconstructed signal
subplot(2,1,2)
plot(recECG)
title('Reconstructed ECG', 'FontSize', 16)
xlabel('Time', 'FontSize', 14)
axis tight
% print figure to file
print('waveletTransform_reconstructed_signal', '-dpng')
https://github.com/ShawhinT/YouTube 5
Find R Peaks Using Reconstructed Signal
% define array of time indicies
Example
t 1:length(ECG);
% square reconstructed signal
recECG abs(recECG)."2;
% find peaks in squared
reconstructed signal
[qrspeaks,locs]
findpeaks(recECG, t,
'MinPeakHeight', 0.35,
'MinPeakDistance',100);
% create figure
fig figure(3);
fig.Units 'normalized';
fig.Position [0 0 1 1];
% plot signal with detected r peaks using reconstructed signal from DWT
plot(t,ECG)
hold on
plot(locs,ECG(locs),'ro')
xlabel('Time', 'FontSize', 14)
title('R Peaks Detected Using Reconstructed Signal from Wavelet Transform',
'FontSize', 16)
axis tight
print('waveletTransform_rpeak_annotation', '-dpng')
https://github.com/ShawhinT/YouTube 6
