Biomedical Signal processing Chapter 7 Filter Design Techniques Zhongguo Liu Biomedical Engineering School of Control Science and Engineering, Shandong University 2015/4/13 1 Zhongguo Liu_Biomedical Engineering_Shandong Univ. Chapter 7 Filter Design Techniques 7.0 Introduction 7.1 Design of Discrete-Time IIR Filters From Continuous-Time Filters 7.2 Design of FIR Filters by Windowing 7.3 Examples of FIR Filters Design by the Kaiser Window Method 7.4 Optimum Approximations of FIR Filters 7.5 Examples of FIR Equiripple Approximation 7.6 Comments on IIR and FIR DiscreteTime Filters 2 Filter Design Techniques 7.0 Introduction 3 7.0 Introduction Frequency-selective filters pass only certain frequencies Any discrete-time system that modifies certain frequencies is called a filter. We concetrate on design of causal Frequency-selective filters 4 Stages of Filter Design The specification of the desired properties of the system. The approximation of the specifications using a causal discrete-time system. The realization of the system. Our focus is on second step Specifications are typically given in the frequency domain. 5 Frequency-Selective Filters Ideal lowpass filter H lp e jw 1, w wc 0, wc w sin wc n hlp n , n n H e jw 1 2 6 wc 0 wc 2 Frequency-Selective Filters Ideal highpass filter 0, w wc H hp e 1, wc w sin wc n hhp n n , n n jw H e jw 1 2 7 wc 0 wc 2 Frequency-Selective Filters Ideal bandpass filter H bp e jw 1, wc1 w wc2 0, others H e jw 1 8 wc2 wc1 0 wc1 wc2 Frequency-Selective Filters Ideal bandstop filter H bs e jw 0, wc1 w wc2 1, others H e jw 1 9 wc2 wc1 0 wc1 wc2 Linear time-invariant discrete-time system If input is bandlimited and sampling frequency is high enough to avoid aliasing, then overall system behave as a continuous-time system: H e jT , H eff j 0, T T continuous-time specifications are converted to discrete time specifications by: w jw H e H eff j , w w T T 10 Example 7.1 Determining Specifications for a Discrete-Time Filter Specifications of the continuous-time filter: 1. passband 1 0.01 H eff j 1 0.01 for 0 2 2000 2. stopband H eff j 0.001 for 2 3000 4 H e jT , H eff j 0, 11 T 10 s T T 1 2 f max 2 2T T 2 5000 Example 7.1 Determining Specifications for a Discrete-Time Filter Specifications of the continuous-time filter: 1. passband 1 0.01 H eff j 1 0.01 for 0 2 2000 2. stopband H eff j 0.001 for 2 3000 1 0.01 2 0.001 p 2 (2000) s 2 (3000) 12 4 T 10 s 1 2 f max 2 2T T 2 5000 Example 7.1 Determining Specifications for a Discrete-Time Filter T 104 s T Specifications of the discrete-time filter in 1 0.01 2 0.001 p 0.4 p 2 (2000) s 2 (3000) 13 s 0.6 Filter Design Constraints Designing IIR filters is to find the approximation by a rational function of z. The poles of the system function must lie inside the unit circle(stability, causality). Designing FIR filters is to find the polynomial approximation. FIR filters are often required to be linearphase. 14 Filter Design Techniques 7.1 Design of Discrete-Time IIR Filters From Continuous-Time Filters 15 7.1 Design of Discrete-Time IIR Filters From Continuous-Time Filters The traditional approach to the design of discrete-time IIR filters involves the transformation of a continuous-time filter into a discrete filter meeting prescribed specification. 16 Three Reasons 1. The art of continuous-time IIR filter design is highly advanced, and since useful results can be achieved, it is advantageous to use the design procedures already developed for continuous-time filters. 17 Three Reasons 2. Many useful continuous-time IIR design method have relatively simple closed form design formulas. Therefore, discrete-time IIR filter design methods based on such standard continuous-time design formulas are rather simple to carry out. 18 Three Reasons 3. The standard approximation methods that work well for continuous-time IIR filters do not lead to simple closed-form design formulas when these methods are applied directly to the discrete-time IIR case. 19 Steps of DT filter design by transforming a prototype continuous-time filter The specifications for the continuoustime filter are obtained by a transformation of the specifications for the desired discrete-time filter. Find the system function of the continuous-time filter. Transform the continuous-time filter to derive the system function of the discrete-time filter. 20 Constraints of Transformation to preserve the essential properties of the frequency response, the imaginary axis of the s-plane is mapped onto the unit circle of the jw z-plane. s j z e Im Im s plane Re 21 z plane Re Constraints of Transformation In order to preserve the property of stability, If the continuous system has poles only in the let half of the s-plane, then the discrete-time filter must have poles only inside the unit circle. Im s plane Re 22 Im z plane Re 7.1.1 Filter Design by Impulse Invariance The impulse response of discrete-time system is defined by sampling the impulse response of a continuous-time system. hn Td hc nTd Relationship of frequencies He if Hc j 0, Td jw w 2 H c j j k Td k Td w jw then H e H c j , w Td w Td for w 23 relation between frequencies Td , , w 2 H e H c j j Td k Td if Hc j 0, Td then H e jw H j w , c T d j Relationship of frequencies S plane 3 / Td / Td - / Td 24 jw Z plane k w Aliasing in the Impulse Invariance w 2 jw H j T He if Hc j 0, Td w then H e H c j , Td w jw 25 k c d j k Td Review st periodic sampling T:sample period; fs=1/T:sample rate t nT n Ωs=2π/T:sample rate xs t xc t s t xc t t nT n x[n] xc (t ) |t nT xc (nT ) 26 x nT t nT n c Review Relation between Laplace Transform and Z-transform Time domain: x(t ) Complex frequency domain: s j 2f 27 Laplace transform X ( s) x(t )e dt st j 0 s pl ane X ( s) x(t )e dt st Since So s j 0 s j j s- pl ane 0 frequency domain : X ( j) x(t )e jt dt Fourier Transform Fourier Transform is the Laplace transform when s have the value only in imaginary axis, s=jΩ 28 For discrete-time signal, x(n) x(t ) (t nT ) x(nT ) (t nT ) n n the Laplace transform L [ x(n)] st x(n)e dt x(nT ) (t nT )e dt z-transform of discretetime signal 29 n st x(nT )e snT X (e sT ) n x ( n) z X ( z ) n n 令:z e sT L [ x(n)] X ( z) x(nT )e snT X (e sT ) n x(n) z n let:z n Laplace transform continuous time signal z-transform ze e sT so: 30 e discrete-time signal ( j )T T r e T T e e jT sT relation re relation between s and z j ze e sT ( j )T T e e jT re j T 2 f f s j z re |r 1 e j X (e ) x ( n )e n j j n DTFT : Discrete Time Fourier Transform j S plane 3 / Ts / Ts - / Ts 31 Z plane Ts 2 f f s 2 f 0 2 f / 2 s : 0 s 2 f s s 2s 3 Ts Ts 3 Ts 32 f fs 0 0 2 2 4 j s pl ane fs 2 2 Ts : f 0 z plane Im[ z ] r 0 Re[ z ] discrete-time filter design by impulse invariance If input is bandlimited and fs>2fmax , : H e jT , H eff j 0, jw w Td for w w then H e H c j , Td jw 33 w 2 H e H c j j k Td k Td if Hc j 0, Td hn Td hc nTd T T w relation between frequencies Td , , w 2 H e H c j j Td k Td if Hc j 0, Td then H e jw H j w , c T d j Relationship of frequencies S plane 3 / Td / Td - / Td 34 jw Z plane k w Review st t nT n periodic sampling T:sample period; fs=1/T:sample rate Ωs=2π/T:sample rate xs t xc t s t xc t t nT n x[n] xc (t ) |t nT xc (nT ) x nT t nT n c 2 S j k s T k 1 1 2 X s j X c j * S j X c j k s d 2 2 T k 1 X c j k s T k 35 2 S j T proof of Review k k s T:sample period; fs=1/T:sample rate;Ωs=2π/T:sample rate s(t)为冲击串序列,周期为T,可展开傅立叶级数 st t nT n e jk st jk s t a e k n F 2 ( ks ) 2 S j T 36 1 jk st e T n k k s periodic sampling xs t xc t s t xc t t nT X s ( j ) n x nT t nT e n c x[n] xc (t ) |t nT xc (nT ) j X s ( j) X (e ) T X (e X (e jT 37 ) 1 ) X c j k s T k 1 X (e ) X c T k j jT 2 k j T T jt n dt X (e j ) 2 s T xc nT t nT k n xc nT e jTn xc nT e jn 1 X s j X c j k s T k if X (e jT ) 0, T 1 then X (e ) X c j T T j discrete-time filter design by impulse invariance x[n] xc (t ) |t nT xc (nT ) h[n] hc (nT ) 1 X (e ) X c T k 2 k j T T 1 X (e ) X c j T T 1 j H (e ) H c T k 2 k j T T 1 H (e ) H c j T T j j j w 2 H e H c j j hn Td hc nTd Td k Td if Hc j 0, Td jw w then H e H c j , Td jw 38 w k Steps of DT filter design by transforming a prototype continuous-time filter Obtain the specifications for continuoustime filter by transforming the specifications for the desired discrete-time filter. Find the system function of the continuoustime filter. Transform the continuous-time filter to derive the system function of the discretetime filter. 39 Transformation from discrete to continuous In the impulse invariance design procedure, the transformation is He jw w H c j , Td w Assuming the aliasing involved in the transformation is neglected, the relationship of transformation is w Td 40 Steps of DT filter design by transforming a prototype continuous-time filter Obtain the specifications for continuoustime filter by transforming the specifications for the desired discrete-time filter. Find the system function of the continuoustime filter. Transform the continuous-time filter to derive the system function of the discretetime filter. 41 Continuous-time IIR filters Butterworth filters Chebyshev Type I filters Chebyshev Type II filters Elliptic filters 42 Steps of DT filter design by transforming a prototype continuous-time filter Obtain the specifications for continuoustime filter by transforming the specifications for the desired discrete-time filter. Find the system function of the continuoustime filter. Transform the continuous-time filter to derive the system function of the discretetime filter. 43 Transformation from continuous to discrete N Ak H s k 1 s sk N Ak e sk t , t 0 hc t k 1 0, t0 N hn Td hc t Td Ak e k 1 N sk nTd un Td Ak e k 1 N Td Ak H z sk Td 1 1 e z k 1 pole : s sk z e sk Td two requirements for transformation 44 un sk Td n Example 7.2 Impulse Invariance with a Butterworth Filter Specifications for the discrete-time filter: 0.89125 H e jw 1, 0 w 0.2 0.17783, He jw 0.3 w let Td 1 w Td Assume the effect of aliasing is negligible 0.89125 H c j 1, 0 0.2 H c j 0.17783, 45 0.3 Example 7.2 Impulse Invariance with a Butterworth Filter 0.89125 H c j 1, 0 0.2 H c j 0.2 0.89125 H c j 0.17783, H c j 0.3 0.17783 H c j 0.3 1 2 1 j j c 2N 1 j j c 2N 1 H c j 2N 0.2 1 1 0.89125 c 2N 0.2 46 0.3 0.3 1 1 0.17783 c 2 2 2 Example 7.2 Impulse Invariance with a Butterworth Filter H c j H c j 0.2 0.89125 1 2 1 j j c 2N 2N 2 0.2 1 1 1.25893 c 0.89125 2N 0.2 0.25893 c 2N 2N 0.3 1 2 1 31.62204 c 0.17783 0.3 30.62204 c N 5.8858, c 0.70470 47 H c j 0.3 0.17783 2N 3 118.26378 2 N 6, c 0.7032 Example 7.2 Impulse Invariance with a Butterworth Filter 1 2 1 H c j H c s H c s H c s 1 j j 2 c sk j c 1 1 2N 1 s j c j ce N 6, c 0.7032 H c s Plole pairs: 0.182 j 0.679, 0.497 j 0.497, 0.679 j 0.182 48 2N 2 N 2k N 1 , 2N k 0,1, ,2 N 1 Example 7.2 Impulse Invariance with a Butterworth Filter H c s H c s H c s 2 H c s Plole pairs: 1 1 s j c 2N c2 N s c 2N 2N 0.182 j 0.679, 0.497 j 0.497, 0.679 j 0.182 cN 0.70326 0.12093 Hc s s 0.182 j0.679 s 0.182 j0.679 s 0.497 j0.497 1 s 0.497 j0.497 s 0.679 j0.182 s 0.679 j0.182 H c s 49 0.12093 s 2 0.3640s 0.4945 s 2 0.9945s 0.4945 s 2 1.3585s 0.4945 Example 7.2 Impulse Invariance with a Butterworth Filter 0.12093 Hc s s 0.182 j0.679 s 0.182 j0.679 s 0.497 j0.497 1 s 0.497 j0.497 s 0.679 j0.182 s 0.679 j0.182 N Ak k 1 s sk N H z k 1 50 N Td Ak 1 e skTd z 1 k 1 Ak sk 1 e z 1 Td 1 0.2871 0.4466 z 1 2.1428 1.1455 z 1 1 2 1 1.2971z 0.6949 z 1 1.0691z 1 0.3699 z 2 1.8557 0.6303z 1 1 0.9972 z 1 0.2570 z 2 Basic for Impulse Invariance To chose an impulse response for the discrete-time filter that is similar in some sense to the impulse response of the continuous-time filter. If the continuous-time filter is bandlimited, then the discrete-time filter frequency response will closely approximate the continuous-time frequency response. The relationship between continuous-time and discrete-time frequency is linear; consequently, except for aliasing, the shape of the frequency response is preserved. 51 7.1.2 Bilinear Transformation Bilinear transformation can avoid the problem of aliasing. Bilinear transformation maps onto w 1 2 1 z Bilinear transformation: s Td 1 z 1 2 H z Hc Td 52 1 z Hc s 2 1 s 1 z Td 1 1 z 1 1 1 z 7.1.2 Bilinear Transformation 2 s Td Td 2 s(1 z ) 1 z 1 1 Td 2 s]z 1 Td 2 s 1 Td 2s z 1 Td 2s 1 z 1 1 z 1 s j 1 1 Td 2 j Td 2 z 1 Td 2 j Td 2 0 z 1 for any 0 z 1 for any 53 1 7.1.2 Bilinear Transformation 1 Td 2 j Td 2 z 1 Td 2 j Td 2 s j 1 j Td 2 z 1 j Td 2 Im s plane Re 54 0 z 1 for any 0 z 1 for any j axis s j jw 1 j Td 2 e z 1 1 j Td 2 Im z plane Re 7.1.2 Bilinear Transformation 2 j Td jw 1 e jw 1 e e jw/2 (e jw/2 e jw/2 ) 2 jw/2 jw/2 jw/2 e Td ( e e ) 2 tanw 2 Td 2 Td w 2 tan1 Td 2 55 1 2 1 z s 1 Td 1 z 2 2 j sin w 2 j tan w 2 2 cos w 2 Td relation between frequency response of Hc(s), H(z) H (e j ) H c ( j) 2 tan Td 2 prewarp : 56 p 2 tan p Td 2 2 tan s s Td 2 Comments on the Bilinear Transformation It avoids the problem of aliasing encountered with the use of impulse invariance. It is nonlinear compression of frequency axis. 2 tanw 2 Td j S plane 3 / Td w 2 tan Td 2 1 / Td - / Td 57 Z plane Comments on the Bilinear Transformation The design of discrete-time filters using bilinear transformation is useful only when this compression can be tolerated or compensated for, as the case of filters that approximate ideal piecewise-constant magnitude-response characteristics. H e jw 1 2 58 wc 0 wc 2 Bilinear Transformation of 2 s Td 1 z 1 1 1 z 2 tan w 2 Td 59 Td 2 tanw 2 Td e e s j Comparisons of Impulse Invariance and Bilinear Transformation The use of bilinear transformation is restricted to the design of approximations to filters with piecewise-constant frequency magnitude characteristics, such as highpass, lowpass and bandpass filters. Impulse invariance can also design lowpass filters. However, it cannot be used to design highpass filters because they are not bandlimited. 60 Comparisons of Impulse Invariance and Bilinear Transformation Bilinear transformation cannot design filter whose magnitude response isn’t piecewise constant, such as differentiator. However, Impulse invariance can design an bandlimited differentiator. 61 7.1.3 Example of Bilinear Transformation Butterworth Filter, Chebyshev Approximation, Elliptic Approximation 0.99 H e jw 1.01, 0.001, He 62 jw w 0.4 0.6 w Example 7.3 Bilinear Transformation of a Butterworth Filter 0.89125 H e jw 1, 0 w 0.2 2 tanw 2 Td jw H e 0.17783, 0.3 w 2 0.2 0.89125 H c j 1, 0 tan Td 2 2 0.3 H c j 0.7783, tan Td 2 For convenience, we choose Td 1 H c j 2 tan0.1 0.89125, H c j 2 tan0.15 0.17783, 63 0.01 0.016 Example 7.3 Bilinear Transformation of a Butterworth Filter H c j 2 tan0.1 0.89125, H c j 2 tan0.15 0.17783, H c j 2 1 1 j jc 2 tan 0.1 1 c 2N 2N 3 tan 0.15 1 c 64 1 0.89125 2N 0.01 0.016 2 1 0.17783 N 5.305 2 N 6, c 0.766 Locations of Poles H c j 1 2 1 j j c sk j c 1 1 2N 2N j ce N 6, c 0.766 H c s Plole pairs: 0.1998 j 0.7401, 0.5418 j 0.5418, 0.7401 j 0.1998 65 H c s H c s H c s 2 2 N 2k N 1 , 1 1 s j c 2N k 0,1, ,2 N 1 Example 7.3 Bilinear Transformation of a Butterworth Filter 2N 2 1 c H c s H c s H c s 2N 2N 2N s c 1 s j c H c s Plole pairs: 0.1998 j 0.7401, 0.5418 j 0.5418, 0.7401 j 0.1998 0.20238 H c s 2 s 0.3996s 0.5871 s 2 1.0836s 0.5871 s 2 1.4802s 0.5871 1 6 0.00073781 z H z 1 1.2686z 1 0.7051z 2 1 1.0106z 1 0.3583z 2 1 1 2 1 z 1 2 s 1 0.9904z 0.2155z 1 66 Td 1 z Ex. 7.3 frequency response of discrete-time filter 67 Example 7.4 Butterworth Approximation (Hw) 0.99 H e jw 1.01, 0.001, He 68 jw w 0.4 0.6 w order N 14 Example 7.4 frequency response 69 Chebyshev filters C Chebyshev filter (type I) 1 2 | H c ( j) | 2 2 1 VN ( / c ) 1 1 1 VN ( x) cos(N cos x) Chebyshev polynomial Chebyshev filter (type II) 1 | H c ( j) | 2 2 1 [ VN ( / c )]1 c 1 2 70 c Example 7.5 Chebyshev Type I , II Approximation 1.01, 0.99 H e jw H e jw 0.001, Type I 71 w 0.4 order N 8 0.6 w Type II Example 7.5 frequency response of Chebyshev Type I 72 Type II elliptic filters E Elliptic filter 1 | H c ( j) | 1 2U N2 () 2 Jacobian elliptic function 1 1 1 2 p s 73 Example 7.6 Elliptic Approximation 0.99 H e jw 1.01, 0.001, He 74 jw w 0.4 0.6 w order N 6 Example 7.6 frequency response of Elliptic 75 *Comparison of Butterworth, Chebyshev, elliptic filters: Example -Given specification 0.99 | H (e j ) | 1.01 | H (e j ) | 0.001 | | 0.4 0.6 | | (s ) 1 0.01, 2 0.001 p 0.4 , s 0.6 -Order B Butterworth Filter : N=14. ( max flat) C Chebyshev Filter : N=8. ( Cheby 1, Cheby 2) E Elliptic Filter : N=6 ( equiripple) 76 -Pole-zero plot (analog) B C1 C2 E C2 E -Pole-zero plot (digital) B (14) 77 C1 (8) -Group delay -Magnitude C1 20 B B E C1 E C2 5 C2 0.4 78 0.6 0.4 0.6 7.2 Design of FIR Filters by Windowing FIR filters are designed based on directly approximating the desired frequency response of the discretetime system. Most techniques for approximating the magnitude response of an FIR system assume a linear phase constraint. 79 Window Method An ideal desired frequency response h ne 1 h n H e e 2 H d e jw n d wc jw d H e jw 0 H lp e d 1 jwn wc jwn dw jw 1, w wc 0, wc w sin wc n hlp n n Many idealized systems are defined by piecewise-constant frequency response with discontinuities at the boundaries. As a result, these systems have impulse responses that are noncausal and infinitely long. 80 Window Method The most straightforward approach to obtaining a causal FIR approximation is to truncate the ideal impulse response. hd n, 0 n M hn otherwise 0, hn hd nwn 1, wn 0, 1 jw H e 2 81 0nM otherwise H d e jw W e j w d Windowing in Frequency Domain Windowed frequency response He j 1 j j H e W e d d 2 The windowed version is smeared version of desired response 82 Window Method If wn 1 n wne W e jw jwn 2 n 1 H e 2 jw Hd e j W e j w 83 1 4 5 10 15 k 2 2 0 w 2k W e jw 4 2 15 10 5 6 wc d H d e jw H e jw 0 wc Choice of Window wn is as short as possible in duration. This minimizes computation in the implementation of the filter. 1, wn 0, 0nM otherwise W e jw approximates an impulse. W e jw w n e n jw M 1 1 e jw 1 e 84 jwn e jwM 2 M e M 1 W e jw jwn n 0 sin w M 1 2 sin w 2 2 M 1 2 M 1 Window Method If wn is chosen so that W e jw is concentrated in a narrow band of frequencies around w 0 then H e jw would look like H d e jw , except where H d e jw changes very abruptly. He jw M 1 85 2 M 1 1 2 W e 2 M 1 jw H d e jw W e j w d H d e jw 1 wc H d e jw 0 wc Rectangular Window W e jw for the rectangular window has a generalized linear phase. M M 1 W e jw e jwM 2 sin w M 1 2 sin w 2 M M 1 As M increases, the width of the “main lobe” decreases. wm 4 M 1 While the width of each lobe decreases with M, the peak amplitudes of the main lobe and the side lobes grow such that the area under each lobe is a constant. M 1 86 2 M 1 2 M 1 Rectangular Window H d e jw W e j w d will oscillate at the discontinuity. The oscillations occur more rapidly, but do not decrease in magnitude as M increases. The Gibbs phenomenon can be moderated through the use of a less abrupt truncation of the Fourier series. 87 Rectangular Window By tapering the window smoothly to zero at each end, the height of the side lobes can be diminished. The expense is a wider main lobe and thus a wider transition at the discontinuity. 88 7.2 Design of FIR Filters by Windowing Method Review To design an ilowpass FIR Filters H e 1 jw H lp e jw 1, 0, sin wc n hlp n n w wc wc w sin wc n M 2 n M 2 h n hd n wn w n 1, 0 n M He jw 0, otherwise 1 2 M 1 89 H d e jw W e j w d 2 M 1 W e jw 2 M 1 wc 0 0 M0 2 wc M 0 M 2 M 0 M 2 M 7.2.1 Properties of Commonly Used Windows Rectangular 1, 0 n M wn 0, otherwise Bartlett (triangular) 2n M , 0 n M 2 wn 2 2n M , M 2 n M 0, otherwise 90 7.2.1 Properties of Commonly Used Windows Hanning 0.5 0.5 cos2 n M , 0 n M wn otherwise 0, Hamming 0.54 0.46cos2 n M , 0 n M wn otherwise 0, 91 7.2.1 Properties of Commonly Used Windows Blackman 0.42 0.5 cos2 n M wn 0.08cos4 n M , 0nM otherwise 0, 92 7.2.1 Properties of Commonly Used Windows 93 Frequency Spectrum of Windows (a) Rectangular, (b) Bartlett, (c) Hanning, (d) Hamming, (e) Blackman , (M=50) 94 (a)-(e) attenuation of sidelobe increases, width of mainlobe increases. 7.2.1 Properties of Commonly Used Windows Table 7.1 smallest,the sharpest transition biggest,high oscillations at discontinuity 95 7.2.2 Incorporation of Generalized Linear Phase In designing FIR filters, it is desirable to obtain causal systems with a generalized linear phase response. The above five windows are all symmetric about the point M 2 ,i.e., wM n, 0 n M wn otherwise 0, 96 7.2.2 Incorporation of Generalized Linear Phase Their Fourier transforms are of the form jw jw jwM 2 W e We e e jw We e is a real and even functionof w hn hd nwn : causal if hd M n hd n h n hd n wn M 2 h M n h n : generalized linear phase A e e He 97 jw jw e jwM 2 M 7.2.2 Incorporation of Generalized Linear Phase if hd M n hd n hn hd nwn hM n hn : generalized linear phase H e jw jA e e jw jwM 2 o M 2 98 M Frequency Domain Representation if hd M n hd n W e e wn w M n 1 H e 2 jw 1 2 H e j d j Ae e jw jw jw j M 2 jwM 2 e j w W e d e 99 We H e e Ae e jw e jwM where H d e jw He e jw e jwM 2 h n hd n wn j w j w M 2 We e d e 2 1 2 H e e jw j w We e d Example 7.7 Linear-Phase Lowpass Filter The desired frequency response is jwM 2 e , w wc jw H lp e 0, wc w H e jw 1 0 w wp 1 wc jwM 2 jwn jw hlp n e e dw H e ws w w c 2 sin wc n M 2 for n n M 2 hlp M n 100 sinwc n M 2 hn wn n M 2 M 2 magnitude frequency response H e 1 0 w w p 20log10 p H e w w jw p p jw s 20log10 s w ws wp ws s s wp H e jw 1 0.05 0 w 0.25 H e jw 0.1 0.15 w s 20dB p 20log10 0.05 26dB w ws wp 0.1 101 7.2.1 Properties of Commonly Used Windows biggest,high oscillations at discontinuity 102 smallest,the sharpest transition 7.2.3 The Kaiser Window Filter Design Method 2 I 0 1 n w n I0 0, where M 2, 12 , 0nM otherwise u 2 I0 u 1 r ! r 1 r 2 I0 u : zero order modified Bessel function of the first kind two parameters : shape parameter: Trade side-lobe amplitude for main-lobe width length : M 1, 103 M=20 =6 As increases, attenuation of sidelobe increases, width of mainlobe increases. As M increases, attenuation of sidelobe is preserved, width of mainlobe decreases. 104 Figure 7.24 (a) Window shape, M=20, (b) Frequency spectrum, M=20, (c) beta=6 Table 7.1 Transition width is a little less than mainlobe width 105 Comparison If the window is tapered more, the side lobe of the Fourier transform become smaller, but the main lobe become wider. Increasing M wile holding constant causes the main lobe to decrease in width, but does not affect the amplitude of the side lobe. M=20 =6 M=20 106 Filter Design by Kaiser Window 1 0 w w H e w w He jw jw s ws wp 107 w ws wp A 20log10 p Filter Design by Kaiser Window 2 I 0 1 n w n I0 0, w ws wp 12 , 0nM otherwise A 20log10 0.1102 A 8.7 , A 50 0.4 0.5842 A 21 0.07886 A 21, 21 A 50 0.0, A 21 M=20 108 A8 M 2.285 w 2 Example 7.8 Kaiser Window Design of a Lowpass Filter 0.99 H e jw 1.01, H e jw 0.001, w 0.4 0.6 w 2 I 0 1 n sin wc n n I0 h n 0, otherwise 12 , 0 n M where M 2 18.5 0.1102 A 8.7 , A 50 0.5842 A 210.4 0.07886 A 21, 21 A 50 0.0, A 21 109 A8 M 2.285w A 20log10 w ws wp Example 7.8 Kaiser Window Design of a Lowpass Filter 0.99 H e jw 1.01, w 0.4 0.001, He jw 0.6 w step 1 : wp 0.4 , ws 0.6 , 1 0.01, 2 0.001, min 1 , 2 0.001 110 ws wp 0.5 step 2 : cutoff frequency wc step 3 : w ws wp 0.2 A 20log10 60 0.5653 M 37 2 Example 7.8 Kaiser Window Design of a Lowpass Filter step 3: w ws wp 0.2 A8 M 37 2.285w A 20log10 60 0.5653 0.1102 A 8.7 , A 50 0.5842 A 210.4 0.07886 A 21, 21 A 50 0.0, A 21 2 12 I 0 1 n sin wc n , 0 n M n I0 h n 0, 111 otherwise where M 2 18.5 u 2 I0 u 1 r 1 r! r 2 Ex. 7.8 Kaiser Window Design of a Lowpass Filter 12 2 I 0 1 n sin wc n h n n I0 112 , 0 n M 7.3 Examples of FIR Filters Design by the Kaiser Window Method The ideal highpass filter with generalized linear phase H hp e jw 0, w wc jwM 2 , wc w e Hhp e jw e jwM 2 Hlp e jw sin n M 2 sin wc n M 2 hhp n , n n M 2 n M 2 hn hhp n wn 113 Example 7.9 Kaiser Window Design of a Highpass Filter Specifications: , w w 1 H e 1 , He jw 2 s jw 1 1 wp w where ws 0.35 , wp 0.5 , 1 21 0.021 By Kaiser window method 2.6, M 24 114 Example 7.9 Kaiser Window Design of a Highpass Filter Specifications: , w w 1 H e 1 , He jw 2 s jw 1 1 wp w where ws 0.35 , wp 0.5 , 1 21 0.021 By Kaiser window method 2.6, M 24 115 7.3.2 Discrete-Time Differentiator Hdiff e jw jwe jwM 2 , w cos n M 2 sin n M 2 hdiff n , n 2 n M 2 n M 2 hn hdiff nwn hn hM n: type III or type IV generalized linear phase 116 Example 7.10 Kaiser Window Design of a Differentiator Since kaiser’s formulas were developed for frequency responses with simple magnitude discontinuities, it is not straightforward to apply them to differentiators. Suppose M 10 2.4 117 Group Delay Phase: M w 5w 2 2 2 Group Delay:M 2 118 5 samples Group Delay Phase: M 5 w w 2 2 2 2 Group Delay:M 5 samples 2 2 Noninteger delay 119 7.4 Optimum Approximations of FIR Filters Goal: Design a ‘best’ filter for a given M In designing a causal type I linear phase FIR filter, it is convenient first to consider the design of a zero phase filter. he n he n Then insert a delay sufficient to make it causal. 120 7.4 Optimum Approximations of FIR Filters he n he n h ne Ae e jw L n L jwn e , LM 2 L Ae e jw he 0 2he ncoswn : real, even, periodic function n 1 A causal system can be obtained from he n by delayingit by L M 2 samples. hn he n M 2 hM n A e e He 121 jw jw e jwM 2 7.4 Optimum Approximations of FIR Filters Designing a filter to meet these specifications is to find the (L+1) impulse response values he n, 0 n L Packs-McClellan algorithm is the dominant method for optimum design of FIR filters. In Packs-McClellan algorithm, L, wp , ws , and 1 2 is fixed, and 1 or 2 is variable. 122 7.4 Optimum Approximations of FIR Filters 1 coswn Tn cosw cos n cos cosw cosw0 T cosw cos0 cos cosw 1 cosw1 T cosw cos1cos cosw cosw 1 0 1 1 cosw2 T2 cos w 2 cos w 1 2 coswn Tn cos w 2cos wTn1 cos w Tn2 cos w cosw3 2cos wcos2w cos w 2 cos w 2 cos2 w 1 cos w 4 cos3 w 3 cos w 123 7.4 Optimum Approximations of FIR Filters h 0 2h ncoswn a cos w L Ae e L k jw e n 1 e k 0 k L where Px ak x k k 0 Define an approxim ation error function A e E w W w H d e jw jw e where W w is the weighting function 124 Px x cos w 7.4 Optimum Approximations of FIR Filters Hd e jw 1, 0 w w p 0, ws w 1 2 , 0 w wp W w K 1 1, ws w 125 Minimax criterion Within the frequency interval of the passband and stopband, we seek a frequency response Ae e jw that minimizes the maximum weighted approximation error of Ew W wH e A e jw d min max Ew he n :0 n L 126 wF jw e Other criterions H 1 min E w dw he n :0 n L 0 2 H 2 min E w dw he n :0 n L 0 H 127 min max Ew he n :0 n L wF Alternation Theorem Let Fp denote the closet subset consisting of the disjoint union of closed subsets of the real axis rx. k Px ak x is an r th-order polynomial. k 0 DP x denotes a given desired function of x that is continuous on Fp WP x is a positive function, continuous on Fp The weighted error is EP x WP xDP x Px The maximum error is defined as E max EP x 128 xFP Alternation Theorem A necessary and sufficient condition that be the unique rth-order polynomial that Px minimizes E is that EP x exhibit at least (r+2) alternations; i.e., there must exist at least (r+2) values xi in FP such that x1 x2 xr 2 EP xi EP xi1 E and such that for i 1,2,, r 1 129 Example 7.11 Alternation Theorem and Polynomials Each of these polynomials is of fifth order. The closed subsets of the real axis x referred to in the theorem are the regions 1 x 0.1 and 0.1 x 1 WP x 1 130 7.4.1 Optimal Type I Lowpass Filters For Type I lowpass filter L Pcos w ak cos w k k 0 The desired lowpass frequency response cos wp cos w 1 0 w wp 1, D p cos w 0, 1 cos w cos ws ws w Weighting function 1 , cos w p cos w 1 0 w w p W p cos w K 1, 1 cos w cos ws ws w 131 7.4.1 Optimal Type I Lowpass Filters The weighted approximation error is EP cos w WP cos wDP cos w Pcos w The closed subset 0 w wp EP x is and ws w or coswp cosw 1 and 1 w cosws 132 7.4.1 Optimal Type I Lowpass Filters The alternation theorem states that a set of coefficients ak will correspond to the filter representing the unique best approximation to the ideal lowpass filter with the ratio fixed at K and with passband and stopband edge wp and ws if and only if EP (cosw) EP (cosw) exhibits at least (L+2) alternations on , i.e., if and only if FP alternately equals plus and minus its maximum value at least (L+2) times. Such approximations are called equiripple approximations. 1 133 2 7.4.1 Optimal Type I Lowpass Filters The alternation theorem states that the optimum filter must have a minimum of (L+2) alternations, but does not exclude the possibility of more than (L+2) alternations. In fact, for a lowpass filter, the maximum possible number of alternations is (L+3). 134 7.4.1 Optimal Type I Lowpass Filters Because all of the filters satisfy the alternation theorem for L=7 and for the same value of K 1 2 , it follows that wpand/or ws must be different for each ,since the alternation theorem states that the optimum filter under the conditions of the theorem is unique. 135 Property for type I lowpass filters from the alternation theorem The maximum possible number of alternations of the error is (L+3) Alternations will always occur at wp and ws All points with zero slop inside the passband and all points with zero slop inside stopband will correspond to alternations; i.e., the filter will be equiripple, except possibly at w and w 0 136 7.4.2 Optimal Type II Lowpass Filters For Type II causal FIR filter: hn 0 n M The filter length (M+1) is even, ie, M is odd Impulse response is symmetric hM n hn The frequency response is e He jw jwM 2 M 1 2 n 0 1 e bncos w n 2 n 1 where bn 2hM 1 2 n, n 1,2, , M 1 2 jwM 2 137 M 1 2 M 2hncos w n 2 7.4.2 Optimal Type II Lowpass Filters M 1 2 n 1 M 1 2 ~ 1 bncosw n cosw 2 b ncoswn 2 n 0 e He jw jwM 2 cosw 2Pcos w L where Pw ak cos w k and L M 1 2 k 0 ~ find ak b n bn bn 2hM 1 2 n 138 7.4.2 Optimal Type II Lowpass Filters For Type II lowpass filter, Hd e jw 1 , 0 w wp DP cos w cosw 2 ws w 0, cosw 2 , 0 w wp W w WP cos w K cosw 2, s w 139 7.4.3 The Park-McClellan Algorithm From the alternation theorem, the optimum filter Ae e jw will satisfy the set of equation i 1 jw jw W w Hd e Ae e 1 140 2 L 1 x x x 1 1 1 2 L 1 x x x 2 2 2 2 L 1 x x x L2 L2 L2 where xi cos wi i 1,2,, L 2 1 W w1 a H e jw1 0 d 1 jw2 a H e 1 W w2 d jwL 2 L2 H e 1 d W wL 2 7.4.3 The Park-McClellan Algorithm Guessing a set of alternation frequencies wi for i 1,2,, L 2 L2 141 jwk b H e k d k 1 L2 bk 1 k 1 W wk k 1 and wl wp , wl 1 ws L2 where 1 bk , xi cos wi i 1 xk xi ik 7.4.3 The Park-McClellan Algorithm L 1 Ae e jw Pcos w d x x C k 1 L 1 k k 1 dk bk xk xL 2 i 1 xk xi ik 142 k d x x k 1 L 1 k k , xk x cos wk 7.4.3 The Park-McClellan Algorithm For equiripple lowpass approximation 10log10 1 2 13 M 2.324w where w ws wp Filter length: (M+1) 143 7.5 Examples of FIR Equiripple Approximation 7.5.1 Lowpass Filter 0.99 H e jw 1.01, H e jw 0.001, w 0.4 0.6 w M 26 unweightedapproxim ation error jw 1 Ae e , 0 w w p E w E A w W w 0 Ae e jw , ws w 144 Comments M=26, Type I filter The minimum number of alternations is (L+2)=(M/2+2)=15 7 alternations in passband and 8 alternations in stopband The maximum error in passband and stopband are 0.0116 and 0.0016, which exceed the specifications. 145 7.5.1 Lowpass Filter M=27, , Type II filter, zero at z=-1 w The maximum error in passband and stopband are 0.0092 and 0.00092, which exceed the specifications. The minimum number of alternations is (L+2)=(M-1)/2+2=15 7 alternations in passband and 8 alternations in stopband 146 Comparison Kaiser window method require M=38 to meet or exceed the specifications. Park-McClellan method require M=27 Window method produce approximately equal maximum error in passband and stopband. Park-McClellan method can weight the error differently. 147 7.6 Comments on IIR and FIR Discrete-Time Filters What type of system is best, IIR or FIR? Why give so many different design methods? Which method yields the best result? 148 7.6 Comments on IIR and FIR Discrete-Time Filters 149 Generalized Linear Phase Order IIR ClosedForm Formulas Yes No Low FIR No Yes High 7.2.1 Properties of Commonly Used Windows Their Fourier transforms are concentrated around w 0 They have a simple functional form that allows them to be computed easily. The Fourier transform of the Bartlett window can be expressed as a product of Fourier transforms of rectangular windows. The Fourier transforms of the other windows can be expressed as sums of frequency-shifted Fourier transforms of rectangular windows.(Problem7.34) 150 Homework Simulate the frequency response (magnitude and phase) for Rectangular, Bartlett, Hanning, Hamming, and Blackman window with M=21 and M=51 151 Chapter 5 HW 7.2, 7.4, 7.15, 152 返 回2015/4/13 上一页 Zhongguo Liu_Biomedical Engineering_Shandong Univ. 下一页