Filter Analysis and Design

Size: px

Start display at page:

Download "Filter Analysis and Design"

Antony Robertson
5 years ago
Views:

1 Filter Analysis and Design

3 Butterworth Filters The poles of a lowpass Butterworth filter lie on a semicircle of radius ω c in the open left half-plane. The number of poles is n and the angular spacing between the poles is always π / n. For n odd there is one pole on the negative real axis and all the others occur in complex conjugate pairs. For even n all the poles occur in complex conjugate pairs. The form of the transfer function is H a ( s) = 1 ( 1 s / p 1 )( 1 s / p 2 ) 1 s / p n n ( ) = 1 k=1 = 1 s / p k n k=1 p k p k s * M. J. Roberts - All Rights Reserved 3

4 Frequency Transformation A normalized lowpass filter can be transformed into an un-normalized lowpass filter or into a highpass, bandpass or bandstop filter by frequency transformation. Lowpass to lowpass s s / ω c Lowpass to highpass s ω c / s Lowpass to bandpass s s2 + ω L ω H s ω H ω L ( ) ( ) Lowpass to bandstop s s ω ω H L s 2 + ω L ω H ω c is the cutoff frequency in rad/s of a lowpass or highpass filter and ω L and ω H are the lower and upper cutoff frequencies in rad/s of a bandpass or bandstop filter. * M. J. Roberts - All Rights Reserved 4

5 Example Frequency Transformation Design a first-order bandstop Butterworth filter with a cutoff frequencies of 1 khz and 2 khz. H norm ( s) = 1 s +1 H BS ( s) = 1 ( ) s 2 + ω L ω H +1 s ω H ω L ω L = 2π 1000 = 2000π, ω H = 4000π = s 2 + ω L ω H ( ) + ω L ω H s 2 + s ω H ω L s H BS ( s) 2 + 8π = s πs + 8π 2 10 s s s Zeros at s = ± j8886 which correspond to ω = ±8886 rad/s and f = 1414 Hz Poles at s = 3142 ± j8886. * M. J. Roberts - All Rights Reserved 5

7 Chebyshev, Elliptic and Bessel Filters The Butterworth filter is "maximally flat" in its passband. This means that its frequency response in the passband is monotonic and the slope approaches zero at the maximum response. The Chebyshev filter has ripple in either its passband or stopband depending on which type of Chebyshev filter it is. Ripple is a variation of the frequency response between two limits. Allowing this ripple in the passband or stopband lets the transition from pass to stop band be faster than for a Butterworth filter of the same order. The Elliptic (or Cauer) filter has ripple in both the passband and stopband and also has an even faster transition between passband and stopband than the Chebyshev filters. * M. J. Roberts - All Rights Reserved 7

10 Impulse Invariant Design Let h δ ( t) = h a ( t)δ Ts ( t). Then H δ k= ( ) ( jω ) = f s H a j ω kω s ( ) L where h a ( t) H a ( s). Also let h d [ n] = h a ( nt s ). Then, using ω /f s = Ω ω = f s Ω H d ( e jω ) = f s H a jf s Ω 2πk k= ( ) ( ) * M. J. Roberts - All Rights Reserved 10

11 Impulse Invariant Design In the expression H d ( e jω ) = f s H a jf s Ω 2πk k = ( ) ( ) it is clear that the digital filter's frequency response consists of multiple scaled aliases of the analog filter's frequency response and, to the extent that the aliases overlap, the two frequency responses must differ. * M. J. Roberts - All Rights Reserved 11

12 Impulse Invariant Design As an example of impulse invariant design let Then h a H a ( s) = Aω c 2 s 2 + 2ω c s + ω c 2. ( t) = 2Aω c e ω ct / 2 sin( ω c t / 2 )u( t). Sample f s times per second to form h d Then and H d [ n] = 2Aω c e ω cnt s / 2 sin( ω c nt s / 2 )u[ n]. ( ) ze ωcts / 2 sin ω c T s / 2 ( z) = 2Aω c z 2 2e ω ct s / 2 cos( ω c T s / 2 )z + e 2ω ct s / 2 ( ) e H d ( e jω ) jω e ωcts / 2 sin ω c T s / 2 = 2Aω c e j 2Ω 2e ω ct s / 2 cos( ω c T s / 2 )e jω + e 2ω ct s / 2 * M. J. Roberts - All Rights Reserved 12

13 Impulse Invariant Design The digital filter's frequency response is ( ) e H d ( e jω ) jω e ωcts / 2 sin ω c T s / 2 = 2Aω c and, from a previous slide, H d Therefore e j 2Ω 2e ω ct s / 2 cos( ω c T s / 2 )e jω + e 2ω ct s / 2 2 Aω H d ( e jω ) = f c s k = ( ) ( e jω ) = f s H a jf s Ω 2πk k = ( ( )). ( ) + ω c 2 jf s Ω 2πk 2 + j 2ω c f s Ω 2πk ( ) e jω e ωcts / 2 sin ω c T s / 2 = 2Aω c e j 2Ω 2e ω ct s / 2 cos( ω c T s / 2 )e jω + e 2ω ct s / 2 * M. J. Roberts - All Rights Reserved 13

14 Impulse Invariant Design Let A = 10 and ω c =100 and f s =200. Then H d and H d ( e jω ) = 2000 k = ( e jω ) = At Ω = 0, and H d ( 1) = 2000 H d ( 1) = ( ) j2 Ω 2πk * M. J. Roberts - All Rights Reserved 14 1 ( ) j2 2 Ω 2πk ( ) e jω e 1/2 2 sin 1 / 2 2 e j 2Ω 2e 1/2 2 cos( 1 / 2 2 )e jω + e 1/ 2 1 = k = 1 16π 2 k 2 j4 2πk =

18 Impulse Invariant Design Let H a Then h a and H d ( s) = Aω c H a ( j2π f ) = s + ω c ( t) = Aω c e ωct u t Aω c H a j2π f + ω c ( ). Sample at rate f s to form h d n z e ( z) = Aω c z e ω ct s H d ( e jω ) jω = Aω c e jω e ω ct s H d ( e jω ) = f s Aω c k = ( jω ) = e jω = Aω jf s ( Ω 2πk) c + ω c e jω e ω ct s Let A = 10, ω c = 50 and f s = 100 and check equality at Ω = 0. Aω c = jf s ( Ω 2πk) + ω c f s k = Aω c e jω e jω e ω ct s = = j200πk + 50 k = 1 = /2 1 e Aω c jω + ω c. [ ] = Aω c e ω cnt s u[ n] * M. J. Roberts - All Rights Reserved 18

19 Why are the two results and f s k = Impulse Invariant Design Aω c = jf s ( Ω 2πk) + ω c Aω c = j200πk + 50 k = e jω e jω e ω ct s = = /2 1 e different by almost 25%? The difference comes from sampling h a [ n] = Aω c e ω cnt s u[ n]. h a t h d ( t) to form ( ) has a discontinuity at t = 0, a sample time. If we let the first sample be Aω c / 2 instead of Aω c the two expressions for H d ( e jω ) are now exactly equal. Aω c / 2 is the average of the two limits at the point of discontinuity. This is consistent with Fourier transform theory which says that the Fourier representation at a discontinuity is the average of the two limits. * M. J. Roberts - All Rights Reserved 19

20 Impulse Invariant Design ( s) = s 2 Let H a s s s s and let f s = h a h d H d ( t) = e t cos(1199.4t 1.48) e 99.74t cos(977.27t ) [ n] = e n cos(1.1994n 1.48) e n cos( n ) ( z) = 48.4z z z z z z z u( t) u[ n] * M. J. Roberts - All Rights Reserved 20

24 Step Invariant Design In step invariant design the analog unit step response is sampled to form the digital unit sequence response. h 1a Z h 1d ( ) H ( t) = L 1 a s s ( [ n] ) = z ( s) = z 1 H d h 1d ( z) H d ( z) = z 1 z [ n ] = h 1,a ( nt s ) s 2 ( [ n] ) Z h 1d Let H a s s s s and let f s = * M. J. Roberts - All Rights Reserved 24

25 Step Invariant Design h 1d h 1a ( t) = e t cos(1199.4t ) u t e 99.74t cos(977.27t ) ( )n cos(1.1994n ) ( ) n cos( n ) [ n] = H d ( z) = z z z z z z z [ ] = x[ n 1] x[ n 2] x[ n 2] x[ n 4]+1.655y[ n 1] y[ n 2] y[ n 3] y[ n 4] y n ( ) u[ n] * M. J. Roberts - All Rights Reserved 25

29 Finite-Difference Design Finite-difference design is based on approximating continuous-time derivatives by finite differences. The transfer function H a implies the differential equation d dt derivative can be approximated by d dt [ ] y d [ n] y d n ay d n T s ( y a ( t) ) + ay a t ( ) = x a t ( y a ( t) ) y n + 1 d T s ( s) = 1 s + a ( ). The [ ] y d [ n] [ ] = x d [ n] and the differential equation in. Then continuous time has become a difference equation in discrete time. Then we can form the discrete-time transfer function H d ( z) = Y d z z X d ( ) ( ) = T s ( ). z 1 at s * M. J. Roberts - All Rights Reserved 29

30 Finite-Difference Design A derivative can be approximated by a forward difference d dt ( y a ( t) ) y d n + 1 T s a backward difference d dt ( y( t) ) y n d [ ] y d [ n] [ ] y d [ n 1] T s or a central difference d ( y a ( t) ) y d n + 1 [ ] y d [ n 1] Z z 1 Y d ( z) T s Z 1 z 1 T s Y d z ( ) = z 1 ( ) zt s Y d z z z 1 Y d ( z) = z2 1 Y d ( z). 2T s 2zT s Z dt 2T s Since every multiplication by s represents a differentiation in time, finite difference design can be done by simply replacing s with a z-transform expression for a difference that approximates a derivative. * M. J. Roberts - All Rights Reserved 30

31 Finite-Difference Design Using a forward difference, H d ( z) = H a ( s) = 1 s + a Using a backward difference, H d ( z) = H a ( s) = Using a central difference, H d ( z) = H a ( s) = 1 s + a 1 s + a = s z 1 T s = s z 1 zt s = s z2 1 2zT s 1 z 1 + a T s 1 z 1 + a zt s 1 = T s z ( 1 at s ) = 1 z a 2zT s T s z 1 + at s 1 z 1 + at s = 2T s z z 2 + 2aT s z 1 * M. J. Roberts - All Rights Reserved 31

32 Finite-Difference Design Using the finite-difference technique it is possible to approximate a stable continuous-time filter by an unstable discrete-time filter. In H a ( s) = 1 s + a 1 ( z) = T s the pole is at s = a. In the digital filter H d z ( 1 at s ) the pole is at z = ( 1 at s ). If H a ( s) is stable, a > 0 and 1 at s is on the real axis of the z plane. If at s 2 the digital filter is unstable. One way to stabilize the design is to use a smaller value for T s, which implies a higher sampling rate. This design was based on using a forward difference. * M. J. Roberts - All Rights Reserved 32

33 Finite-Difference Design If we use a backward difference instead of a forward difference we get the digital filter H d ( z) = H a ( s) = = 1 s + a T s z 1+ at s s z 1 zt s ( ). z 1 / 1 + at s The pole is now at z = 1 / ( 1+ at s ) and is guaranteed to be inside the unit circle in the z plane, making this filter stable for any a > 0 and any T s. Using a backward difference in any digital filter design guarantees stability, but that guarantee comes at a price. The pole locations are restricted to a small region inside the unit circle. * M. J. Roberts - All Rights Reserved 33

34 ( s) = Finite-Difference Design s 2 Let H a s s s s and design a digital filter using the finite-difference technique with f s = H d ( z) = 0.169z 2 z 1 ( ) 2 z z z z * M. J. Roberts - All Rights Reserved 34

37 Matched z and Direct Substitution The matched-z and direct substitution digital filter design techniques use the correspondence z = e st s to map the poles and zeros of the analog filter to corresponding locations in the z plane. If the analog filter has a pole at s = a, the corresonding digital filter pole would be at z = e at s. The matched-z method replaces every factor of the form s a with the factor 1 e at s z 1 or z eat s the direct substitution method replaces it with the factor z e at s. Zeros are mapped in the same way. z and s = a z = e at s * M. J. Roberts - All Rights Reserved 37

38 Matched z and Direct Substitution Let H a ( s) = s 2 s s s s Then, using the matched-z technique y n H d ( z) = z 2 ( z 1) 2 z z z z [ ] x[ n 2] [ ] = x[ n] x n y[ n 1] y[ n 2] y[ n 3] y[ n 4] * M. J. Roberts - All Rights Reserved 38

42 The Bilinear z Method One approach to digital filter design would be to use the mapping z = e st s s ( 1 / T s )ln z z-domain expression H d ( ) to convert the s-domain expression to the ( z) = H a ( s) s ( 1/Ts ) ln( z). Unfortunately this creates a transcendental function of z with infinitely many poles. We can simplify the result by making an approximation to the exponential function. The exponential function is defined by e x = 1+ x + x2 2! + x 3 3! + = x k k! k =0 * M. J. Roberts - All Rights Reserved 42

43 The Bilinear z Transform We can approximate the exponential function by truncating its series definition to two terms e x 1 + x. Then z = 1+ st s or s z 1. This is exactly the same as the finite difference technique T s using forward differences and has the same drawback, the possibility of converting a stable analog filter into an unstable digital filter. A simple but very significant variation on this idea is to express e st s in the form est s /2 e st s /2 and then approximate both exponentials with the first two terms of the series. Then z = 1 + st s / 2 1 st s / 2 or s 2 z 1. This is T s z + 1 the basis of the bilinear digital filter design technique. * M. J. Roberts - All Rights Reserved 43

44 The Bilinear z Transform In the mapping s 2 T s z 1 z + 1 let s = jω. Then z = 2 + st s = 2 + jωt s = 1 tan 1 ( ωt s / 2) = e j 2 tan 1 ( ωt s /2) 2 st s 2 jωt s The ω axis of the s plane maps into the unit circle of the z plane. If s = σ + jω, σ < 0, the mapping is from the left half of the s plane to the interior of the unit circle in the z plane. If s = σ + jω, σ > 0, the mapping is from the right half of the s plane to the exterior of the unit circle in the z plane. * M. J. Roberts - All Rights Reserved 44

45 The Bilinear z Transform For real discrete-time frequencies Ω the correspondence between the s and z planes is s = 2 e jω 1 T s e jω + 1 = j 2 tan Ω T s 2 s = σ + jω, σ = 0 and ω= ( 2 / T s )tan Ω / 2. Then, since ( ). ( ) or Ω = 2 tan 1 ωt s / 2 * M. J. Roberts - All Rights Reserved 45

50 FIR Filter Design An analog filter can be approximated by an FIR digital filter. The impulse response of the analog filter has infinite duration in time but beyond some time it has dropped to a very low value and the remainder beyond that time can be truncated with little change in the filter characteristics. We can create a digital FIR filter by sampling over that time. * M. J. Roberts - All Rights Reserved 50

51 FIR Filter Design We can also approximate ideal filters, which are non-causal, by truncating that part of the impulse response occurring before time n = 0 and later when the impulse response has fallen to a very low value. * M. J. Roberts - All Rights Reserved 51

52 FIR Filter Design The truncated impulse response has the form h N N 1 m=0 [ n] = a m δ n m [ ] where N is the number of samples retained. This type of FIR digital filter can be realized in Direct Form II. Its transfer function is H d N 1 ( z) = a m z m. m=0 * M. J. Roberts - All Rights Reserved 52

53 FIR Filter Design [ ], 0 n < N Let h N [ n] = h n d 0, otherwise impulse response of the digital filter where h d [ n] is the = h d [ n]w[ n] be the sampled ideal analog filter impulse response and w n a window function. Then the frequency response is H N ( e jω ) = H ( d e jω ) W( e jω ). The periodic convolution with W( e jω ) changes the frequency response from ideal and introduces ripple in the frequency response. [ ] is * M. J. Roberts - All Rights Reserved 53

57 FIR Filter Design The digital filter impulse and frequency responses on the previous three slides were all formed by using a rectangular window of the form, w n. Other window 0, otherwise functions could be used to reduce the ripple in the frequency response. [ ] = 1, 0 n < N * M. J. Roberts - All Rights Reserved 57

58 FIR Filter Design von Hann Bartlett Hamming Blackman Kaiser w n Frequently-Used Window Functions [ ] = cos 2πn w n w n w[ n] = N 1, 0 n < N 2n, 0 n N 1 N n N 1, N 1 n < N 2 2πn N 1, 0 n < N [ ] = cos [ ] = cos w[ n] = 2πn 4πn N cos N 1, 0 n < N 2 N 1 I 0 ω a 2 n N 1 2 N 1 I 0 ω a 2 2 * M. J. Roberts - All Rights Reserved 58

61 FIR Filter Design In studying the common types of windows in the last two slides it becomes apparent that in designing FIR filters two design goals are in conflict. The ideal filter has no ripple in the passband and the transition from the passband to the stopband is abrupt in frequency; it has no width. For a fixed window width, if we optimize the transition to be as quick as possible we must accept large ripple in the pass and stop bands. If we want to minimize passband and stopband ripple we must accept a slower transition between the passband and stopband. * M. J. Roberts - All Rights Reserved 61

62 FIR Filter Design The general form of an FIR impulse response is h d [ n] = h d [ 0]δ [ n] + h d [ 1]δ [ n 1] + + h d [ N 1]δ n N 1 and its z transform is H d ( z) = h d [ 0] + h d [ 1]z h d [ N 1]z ( N 1) Let N be an even number and let the coefficients be chosen such that h d 0 ( ) [ ] = h d [ N 1], h d [ 1] = h d [ N 2],, h d [ N / 2 1] = h d [ N / 2] * M. J. Roberts - All Rights Reserved 62

63 FIR Filter Design The frequency response is H ( d e jω ) = h d [ 0] + h d [ 0]e j( N 1)Ω + h d [ 1]e jω + h d [ 1]e j( N 2)Ω + + h d [ N / 2 1]e j( N /2 1)Ω + h d [ N / 2 1]e jnω/2 which can be expressed in the form N 1 Ω h d [ 0]cos 2 Ω + h N 3 d [ 1]cos 2 Ω + + h d [ N / 2 1]cos( Ω) The phase of the frequency response is linear. This is usually a desirable H ( d e jω ) = 2e j N 1 2 result because it is the same as an ideal filter. By a similar process it can be shown that if the coefficients are anti-symmetric meaning [ ] = h d [ N 1], h d [ 1] = h d [ N 2],, h d [ N / 2 1] = h d [ N / 2] h d 0 that the phase is also linear and this holds true for N even or odd. * M. J. Roberts - All Rights Reserved 63

64 Optimal FIR Filter Design The Parks-McClellan algorithm is a technique for directly designing a filter by specifying its frequency response by multiple linear functions of frequency in multiple frequency bands separated by transition bands. In MATLAB the syntax is B = firpm(n,f,a) where B is a vector of N + 1 real symmetric coefficients in the impulse response of the FIR filter which is the best approximation to the desired frequency response described by F and A. (See the MATLAB help file for a more detailed description.) * M. J. Roberts - All Rights Reserved 64

Optimal FIR Filter Design Example Design an optimal FIR filter to

66 Optimal FIR Filter Design Example { } and the desired The band edges are Ω = 0, 0.6, 0.7, 2.2, 2.3,π amplitude responses at those band edges are A = { 0,0,1,1,0,0 }. The vector F should be F = Ω / π = { 0, 0.191, , , ,1}. N is the length of the filter and can be found by trying different values until filter specification is met. In this case N = 70 meets the specification. * M. J. Roberts - All Rights Reserved 66

Digital Signal Processing

COMP ENG 4TL4: Digital Signal Processing Notes for Lecture #24 Tuesday, November 4, 2003 6.8 IIR Filter Design Properties of IIR Filters: IIR filters may be unstable Causal IIR filters with rational system