Wavelet basis functions in biomedical signal processing - SPAN LAB
Wavelet basis functions in biomedical signal processing - SPAN LAB
Wavelet basis functions in biomedical signal processing - SPAN LAB
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
J. Rafiee et al. / Expert Systems with Applications 38 (2011) 6190–6201 6195Raw bio<strong>signal</strong>s(selected class)Segmentation of <strong>signal</strong>s <strong>in</strong>to smaller units for each subjectSegmented <strong>signal</strong>s foreach six class324 motherwavelet <strong>functions</strong>Select<strong>in</strong>g a mother wavelet and one segmented <strong>signal</strong> from one class andfrom one specific subjectS= sum of absolute value of CWC <strong>in</strong> all scales for eachsegmented <strong>signal</strong> and for each classAS= average of S <strong>in</strong>segmented samplesEC= average of AS forsubjectsEC Max <strong>in</strong>class 1EC Max <strong>in</strong>class 2EC Max <strong>in</strong>class 3EC Max <strong>in</strong>class 4EC Max <strong>in</strong>class NMother wavelet for each classFig. 8. Search<strong>in</strong>g algorithm to f<strong>in</strong>d the most similar mother wavelet based on evaluation criterion (for bio<strong>signal</strong>s). R 11 w rðxÞdx ¼ R 11 xr w r ðxÞdx ¼ 0.Daubechies’ wavelets provide appropriate results <strong>in</strong> <strong>signal</strong> process<strong>in</strong>gtechniques due to the above-mentioned properties. Awavelet function with compact support can be implemented easilyby f<strong>in</strong>ite length filters. Moreover, the compact support enables spatialdoma<strong>in</strong> localization. S<strong>in</strong>ce the wavelet <strong>basis</strong> <strong>functions</strong> havecont<strong>in</strong>uous derivatives, they decompose a cont<strong>in</strong>uous functionmore efficiently while avoid<strong>in</strong>g edge artifacts. S<strong>in</strong>ce the motherwavelets are used to characterize details <strong>in</strong> a <strong>signal</strong>, they shouldhave a zero <strong>in</strong>tegral so that the trend <strong>in</strong>formation is stored <strong>in</strong> thecoefficients obta<strong>in</strong>ed by the father wavelet. A Daubechies’ waveletrepresentation of a function is a l<strong>in</strong>ear comb<strong>in</strong>ation of the wavelet<strong>basis</strong> <strong>functions</strong>. Daubechies’ wavelet transforms are usually implementednumerically via quadratic mirror filters. Multiresolutionanalysis of the trend and fluctuation of a function is implementedby convolv<strong>in</strong>g it with a low-pass filter and a high-pass filter thatare versions of the same wavelet.As shown <strong>in</strong> bio<strong>signal</strong>s, the <strong>signal</strong> with a sharp spike is properlyanalyzed by Daubechies’ wavelets, because much less energy (ortrend) is stored <strong>in</strong> the high-pass bands. Hence, Daubechies’ waveletsare ideally suited for natural <strong>signal</strong>s. In general, Daubechies’wavelets with long-length filters give more appropriate energyconcentration than those with short-length filters. Therefore, thementioned reason supports the use of EC to f<strong>in</strong>d the most similarfunction to bio<strong>signal</strong>s. Besides, as illustrated <strong>in</strong> Fig. 9, db44 is theonly Daubechies function with near-symmetric characteristic aswell as hav<strong>in</strong>g sharp spike that would be matched with natural<strong>signal</strong>s.4. Results and discussionThe result shows that, among all wavelet families considered <strong>in</strong>Tables 1A–2B (Appendix A), Daubechies 44 (db 44) provides a betterfit to the tested bio<strong>signal</strong>s. Figs. 10–12 show the EC acrossmother wavelet candidates for surface EMG, EEG, and VPA <strong>signal</strong>s,respectively. For example, <strong>in</strong> Fig. 10, all classes of EMG (for 10 differenthand motions) have a peak whose amplitude is greater thanthose calculated by the other candidate mother wavelets. The peakrepresents the largest EC which is calculated by db44. The samepattern was achieved for other bio<strong>signal</strong>s. The difference betweendb44 and the other 323 wavelet <strong>functions</strong> towards bio<strong>signal</strong>s is