We now cxamine severa} Applications of the LMg cancellation, system modeling, and line enhancement First, we begin with the noise Cancellation Problem t the LMS adaptive FIR filter, algorithm, such as noise via application examples. 0 illustrate Operations of 10.3.1 Noise Cancellation The concept of noise Cancellation was introduced in the Previous section, Figure 10.7 shows the main idea, ‘The DSP system Consists of two ADC ch: ADC captures the Noisy Speech, d(n) speech (17) and noise n(n) due toa noisy e1 ‘annels. The first microphone with ~ s(n) + (2), which contains the clean e best estimate of noise y(n) n(n), which will output of the error Signal e(n) = (2) + n(n) — YM) & 3(n) is expected to be a best. estimate of the clean ), the cleaned digital become familiar with the setup and operations of the adaptive filter and LMS algorithm. The simulation for real adaptive noise cancellation follows, . Error signal ») a aoy=steh+r40) ela) = fe) ~ y(n) = 3(0) ADC, DAC. Lx x(n) ‘Adaptive filter Noise LMS algorithm Fy ve fil SURE 19.5 Simplest noise canceler using a one-tap adaptive filter.punTERS ANP APPLICATIONS E ava 10 ADAPTIV 3. Example 10. : - is cellation application using a: . Given the DSP system for the noise Ca aa gan adap filter with two coefficients shown in Fig! . a. Set up the LMS algorithm for the adaptive filter. b. Perform adaptive filtering to obtain outputs e(n) = 0, 1, 2 given . following inputs and outputs: x(0) = 1, x(1) = 1x@) = -1, 4) = 2, d(1) = 1, d(2) = -2 and initial weights: w(0) = w(1) = 0, convergence factor is set to be 4s = 0.1. Solution: a. The adaptive LMS algorithm is set up as: Initialization: w(0) = 0, w(1) = 0 Digital filtering: y(n) = w(0)x(m) + w(1)x(n — 1) Computing the output error = output: e(n) = d(n) — y(n). Updating each weight for the next coming sample: w@) = wi) + 2me(n)x(n — A), for i= 0, 1 or W(0) = W(0) + 2yre(n)x(nn) W() = (1) + 2ye(n)x(n — 1), Signal and noise An) = s(n) +n(n) a. Noise Output Adaptive fi fr) 0 | He ‘ FIGURE 10.8 y(n) Noise cancellation in Example 10, ple 10.3,11.3 Applications: Nolse Cancellation, System A ‘odeling, and Line &; wand Line Enhancement 475 We can see the adaptive filtering Operations as follows: 8 as follows: Kors 0 Digital filtering: POY = WOM) + wD = 1) 50. 1 +0\050 Computing the output: ©) = dO) = (0) = 2-0 Updating coefficients: WO =O) + 2N OLS cM) S022 0142. 1=04 wD s wD +2 NOLS eOa(— S042 01240500 1 al filtering: VCD) = WON) + w(D)x(0) = 0.4% 140 1 = 0.4 For Computing the output: e(l) = d(1) — (1) = 1-04 = 0.6 Updating coefficients: 20.1 x ex) = 0442 0.1N 06S L= w(0) = w(0) wh) = wd) +2x 01 x ex) = 042 N0.1N 0.6 T= 0.12 Forn =2 Digital filteri (2) = w(Ox(2) + w(Da() = 0.52 x (= D+ 0.12 Ps 04 Computing the output: e2) =dQ)—3@) =-2-(- 0 = “EE Updating coefficients: 7 z - —D=0.8. 52(0) = 6(0) £2 x O.1 x e(2)2)=HO-S2+2NOTNC 18 seen (= (1) £2 x 0.1 xe = O12 +2801 NM are listed a8 24 Fre s les *. the adaptive filter outputs for the first three samP (0) = 2. e(1) = 0.6. 2) = “16are PT T PPLICATIO pApTIVE FILTERS AND A 6 10 A ine the MSE function assuming the following Statistica] data: Next we exami \ of = ea] =4, BM] = BPO — VD] = 1, Else) —1y <9 Eld(a)x()| = 1, and Bld@)x(n — D] = -1 a = + w(1)x(n — 1). We follow R -tap adaptive filter y(n) = w(0)x(n) et follow Baus, eto to “i 0.5)t0 achieve the minimum MSE function in two dimensions as J=4+ WO) + wl) — 20(0) + 20(1). Figure 10.9 shows the MSE function versus the weights, where the optimal weights and the minimum MSE are w*(0) = 1, w (1) = 1, and Jig = 2, It the adaptive filter continues to process the data, it will Converge to the optimal weights, which locate the minimum MSE. The plot also indicates that the function is quadratic and that there exists only one minimum of the MSE surface. Next, a simulation example is given to illustrate this idea and its results, The noise cancellation system is assumed to have the following specifications: @ Sample rate = 8,000 Hz @ Original speech data: wen.dat1 Mote rod Apples Uae eet cortupled by Gaussian nis Witla power of 1 delayed by 4 fay the tonse teterence None referenee containing, Gaussian noise will a power of | a power 0 Adan PUR filter used £0 remove the noi a Namber of PER filter taps a Voaverzence Fietor for the LMS alpoithin chosen to be 0.016 Aly the gyeeeh wavelorms and: spectral plots for the original, corupte eewace noise and for the cleaned specch are plotted in Figsre: 10.1 Viob, Lrom the figures, it is observed that the enhanced speech wavelorin and gqgetrtint aire very elose to the original ones. The LMS algorithm converpe: after tions. ‘The method is a very effective approach for noice detailed in Program 10.1, approximately 100 itera canceling, MATLAB implementation i Progeam 10.1, MATLAB program Lor adaptive noise cancellation, ear all Given by the in: 5 Sampling rate Create the index array Lructor (won) hi s convert eat (wend) i 000000.5 1,1 8)7 5 Generate the signal plus noise Tnitialize the step size Initialize the adap} Initialize the adaptive file! output array ithm indices to time instants 4 Generate the random no the corruption 19 5 Generate coke il, length (t))i 3 Initialize the ve filtering using the LMS algor ength (t)-1 (2) + 2¢mute (m) tx (in = 37 (Continued)ERS AND APPLICATION, ava 10 ADAPTIVE FILT i i ‘trum for the o: e-sided amplitude spec’ eiginay abs (rt (wen) ) /Lengeh (wen) }WEN (1) =WEN (1) /2; Song) = 2tabs (fF ingle-sided amplitude spectrum for the corry th (d) :D(1) =D (1) /27 *£s/length (wen) pectrum for the noise-canceled 54, (1) /27 calculate the sing WE 2 Calculate the S. p= 2'abs (££t (d))/Leag f=[0:1: length (wen) /2] 4 Calculate the single-sided s f= 2'abs (£ft (e)) /length (e} 7 (1) =E 4, Plot signals and spectra ne : subplot (4, 1,1), plot (ven) ;gzidsylabel ( ‘Orig. speech’) ; subplot (4, 1,2) ,plot (da) ;grid7ylabe1 (Corrupt. speech’) subplot (4,1, 3) ,plot (x) ;grid;ylabel ( ‘Ref. noise’); subplot (4,1, 4) ,plot (e) ;grid; ylabel (‘Clean speech’) ; xlabel ( ‘Number of samples’); figure subplot (3, 1,1) ,plot (£,WEN(1: length (f)))sgrid ylabel ( ‘Orig. spectrum’) subplot (3,1,2),plot (f,D(1: length (£))) ;gridzylabel ( ‘Corrupt. spectrun') subplot (3, 1,3), plot (£,(1:length(f)))sgrid ylabel (‘Clean spectrum’) ;xlabel (‘Frequency (Hz)’); — Shgnay gnay © 200 400 600 8001000 1200 4400 1600 1800 2000 200 t 400 600 800 1000 1200 74007600 1800 2000 Ret. noise Corrupt. speech Orig. speech > Clean speech 600 800 “1000 1200 1400 1600 1600 200° FIGURE 10.104 Number of samples Waveform: . end clean so originals Pech, speech, corrupted speech, FeMaton ¢ ary | | \ i | \ \ | \ \ | i i yoo ' { | | | { \ Ha ih bo} tt EG TAN ota ceeded Se wna fun ThOD voQO HAA “OOOO A000 1 win \ ' | \ ! \ ! bef Gib Pog bE ~~ Muah al a} yon faAD 10 F000 600-800 -—-4HOO 4000 ‘ AN Whaat ‘ rest iabaotirwrtettenstnedouni:cis wes . 0 U0 1000, (noo 000 P5600 93000, A000, u Froquoney (U2) HOURE 10.100 Spect 1 for original spooch, corrupted spooch, and clean speoch Olher interference cancellations include that of 60 Tz interference in clec- Wocatdiogtaphy (HCG) (Chapter 8) and echo cancellation in long-distance telephone eitenits, which will be described in a later section, 10.3.2 System Modeling \othor application of the adaptive filter is system modeling, ‘The adaptive Her Sav heep tracking, the behavior of an unknown system by using the unknown Wslem's _ . be Deters Sstein’s input andl output, as depicted in Vigure 10.11. | Unknown ae — yn) Output laptivo, “ FUR Hiltor Unput vy Ne URE 1 11 Adaptive filtor for system modeling.Ss AND APREREMESE EONS FILTER prive 480 10 ADA is goi be as close as the unkno : n) is going to i 910 sy As shown a fete system and the adaptive filter use yj ms output. Since j i ‘ 8a he transfer function of the adaptive filter will approximate that fig input, the tr unknown system. Example 10.4. / ' ; Given the system modeling described and using i single-weight adaptive file y(n) = wx(n) to perform the system modeling task, a. Set up the LMS algorithm to implement the adaptive filter, assuming thay the initial w =0 and p = 0.5. b. Perform adaptive filtering to obtain y(0), (1), »(2), and y(3) given (0) = 1, d(1) = 2, d(2) = -2, d3) =2, x(0) = 0.5, x(1) = 1, x(2) = -1, x(3) = 1. Solution: a. Adaptive filtering equations are set up as w=Oand 24=2x05=1 y(n) = wx(n) e(n) = dn) — y(n) w= w+ e(n)x(n) b. Adaptive filtering: n=0, y(0) = wx0) = 0x 0.5 =0 ) =d0)-y0)=1-9 =4 W= w+ eO)x0) =041%0.5=0.5 "= 1) = wx) =0.5%1 20.5 &) =) ~ yt) = 2-05 = 15 W= w+ e(l)x(1) = 0.541, _ "=2.90)=Q)=2« (yo =2.0 €) = 2) ~ yay — ~2-(-2)=0 Y= w+ e(2)x(2) = 2 "a — 7 - "=3:9G)=0x@) 225125 °° am °8) = 43) ~ 3) 9 “9 _ wawy e(3)x(3 7 . y=240x15 7 ed Particular cas ; . ip? Se, the system ig actually a digital amplifiercations: Noise Cancettatioy 10.3 Appllcatio Cancetation, System Modeling, and t fe Enh ine Enhancement 481 : ume that the unknown system j. Next, We ass! 'ystem is a fourth-orde: : eiose 3 dB lower and upper cutoff frequencies are 1,400 Haute fie ing at 8,000 Hz, We use an input 4 eae consisting of tones of ing J es of 500, 1,500, 1500 Hz, The unknown system’s frequency responses are shown in Figure TO °° ye input saveform x(7) with three tones is shown asthe fist ples n Figure 19.3. We can predict that the output of the unknown system will contain a 1.300 Hz tone only, since the other two tones are rejected by the unknown system. Now, let us look at adaptive filter results. We use an adaptive FIR filter witb the number of taps being 21, and a convergence factor set to be 0.01. In time domain, the output waveforms of the unknown system d(n) and adaptive fiter output y(7) are almost identical after 70 samples when the LMS algorithm converges. The error signal e(7) is also plotted to show that the adaptive filter keeps tracking the unknown system’s output with no difference after the first 50 samples. Figure 10.14 depicts the frequency domain comparisons. The first plot displays the frequency components of the input signal, which clearly shows 500, 1,500, and 2,500 Hz. The second plot shows the unknown system’s output spectrum, which contains only a 1,500 Hz tone, while the third plot displays the spectrum of the adaptive filter output. As we can see, in frequency domain, the g 3 & g & g 5 ioe : 3600 4000 : 600 1000 1500 2000 2500 3000 Frequency (Hz) 200 F 100 2 fo 8 100 }- i i : : 0 $0900 15002000 2600 $00 Frequency (H2) nses- The unknown system’s frequency 'esP°!PEND AERLICATIONs ILT ERS pTIVe F aun 19 BOK The waveforms for the unknown system’s output, adaptive fite output, and error output. ° 3500 000 B14 ] : 2, BOB omer | 3 ~ | a2 3500 00? z oS Bose gE | BE Pb 8 ° 5001000 “t500 2000 2500-3000 «35 Frequency (Hz) FIGURE 10.14 Spectrum fo input « sett filter outpun, "> PPUt Signal, unknown system output,14.4 Applications: Noite Conesllation, sy sem Modeti ling, and Line Enh Shoncement 423 Jfiltor tracks the characteristics Of the ANON Vs. . ven in Program 1.9, We S™EPO%N s3stem. The MATLAB yyypatt 10.2. MATLAB program for adaptiy system ication. PUVE system identifi eI “ ne ES) 5 s([0 £s/2-801)); 00*t) Produce the unknown sys nce factor 5 (1, length (t)) ; Initialize the alize the error vector aptive filtering using the LMS algorithm the input late the single-sided amplitude spectrum for abs (fee 7X (1) =X (1) /22 | eee spectra x a Bs (ft (dy) /length (d) 7D(1 =D(1) /27 wr the adaptive filter output @te the single-sided amplitude spectrum fo! eee (££t(y) ) length (y) axxo ae the Erequency index to its frequency § Length (x) /2] *£s/Length (X) 7 (Continued)VE FILTERS AND APPLICKTIONng apaPT! apa 10 10.3.3 Line Enhancement Using Linear Predic We study adaptive filtering via another applicatior Ifa signal frequency content is very narrow compa S$ with time. then the signal can effici: al by the white Ga ed line consists of the delay element to delay the corrupted A samples to produce an input to the adaptive filter. The adap actually a linear predictor of the desired narrow band signal. A two-tp 22%: FIR filter can predict one sinusoid (proof is beyond the scope of this value of A is usually determined by experiments or experience achieve the best enhanced signal. O(n) = A cos(2ztn/ f,) ~ n(n) FIGURE 10.15 Une enhancement using an adaptive filter.10.3 Applications: Noise Cancellation, System Modeling, and line Enhancement 485 Noisy signal ADF output (enhanced signal) 0 100 200 300 400 500 600 700 800 Number of samples FIGURE 10.16 Noisy signal and enhanced signal. Our simulation example has the following) specifications: 0 Sampling rate = 8,000 Hz ® Corrupted signal = 500 Hz tone with unit amplitude added with white Gaussian noise 4 Adaptive filter = FIR type, 21 taps © Convergence factor = 0.001 3 Delay value A =7 " LMS algorithm applied | ete Fisure 10.16 shows time domain results. The first plot 7 ie noe eee ee ka 12" plot clearly demonstrates the enhanced signal. Figur Signal is shown in feauency domaty saint of view. The spectrum of {Re PONY TT! er the he top plot, he e can see that white noise is ee cae ihe CHtite ban aia ay oe plot is the enhanced signal al paeasy Method ie adavitve: fis copay effective when the CONETE gram for this ‘ctanging with time, Program 10.3 lists the MA Simulations .piLTeRs AND APPLICATIONS E gas 10 ADAPTIY 3, MATLAB program for adaptive line enhancement, Program 10.3. ; tee 4 sampling rate and sampling period 1000; T= é 1 second time instants % Generate the Gaussian random noise (2"pi7500°t) + nF % Generate the 500-Hz tone plus noise < feiver({00000002), 1, ai pelay filter = 0.0017 4 Initialize the step size for the LMS algorithms re zeros (1, 227 4 Initialize the adaptive filter coefficients y= zeros (1, length (t)) 7 % Initialize the adaptive filter outpy =yi 4 Initialize the error vector erform adaptive filtering using the LMS algorithm for m= 22:1:length(t)-1 sum = 0; for i=1:1:21 sum = sum+w(i)*x(m—4i)7 o:t:0.17 neranda(1,length(t))# end y(m) = sum; e(m) =d(m) ~y (m) 7 for i=1:1:21 w(i) =w(i) + 2*mu*e (m) *x (m-i) 7 end end % Calculate the single-sided amplitude spectrum for the corrupted signal D= 2*abs (fft (d)) /length(d) ;D(1) =D(1) /2; % Calculate the single-sided amplitude spectrum for the enhanced signal = 2*abs (f£t (y))/length(y) #¥(1) =¥ (1) /2; % Map the frequency index to its frequency in Hz :1:length (x) /2] *8000/length (x) ¢ % Plot the signals and spectra subplot (2, 1,1) , plot (d) ;grid;axis ( {0 length (x) -2.52.5]) ;ylabel (noisy signal’)? subplot (2,1, 2) plot (y) ;grid;axis((0 length (y) —2.52.5]); ylabel (‘ADF output (enhanced signal) ') ;xlabel (‘Number of samples’) figure subplot (2,1,1),plot (£,D(1:len : i ylabel ( ‘Noisy signal ea i eo subplot (2, 1,2) »plot (£,¥ (1: length (£))) ;grid;axis({0 £s/201-5])? ylabel (‘ADF output spectrum’ ) ; xlabel ( ‘Frequency (Hz) ")7 This section conti ont et without showing eas €© xP other adaptive filter applications Ft eee Simulations. The topics include periodic iM ast interference cancellation, and echo cancellation in 1on810. i 10.4 Other Application Examples 487 Noisy signal spectrum, ADF output spectrum ‘3500 4000 0 0 500 1000 1500 2000 2500 3000 Frequency (Hz) HGURE 10.17 Spectrum plots for the noisy signal and enhanced signal. japon circuits. Detailed information can also be explored in Haykin (1991), lfeachor and Jervis (2002), Stearns (2003), and Widrow and Stearns (1985). dic Interferences diction qiudio signal may be corrupted by periodic interference and no noise refer- the 7 available. Such examples include the playback of speech or music with ee of tape hum, turntable rumble, or vehicle engine or power line Figue 1018 We can use the modified line enhancement structure as shown in mae oe filter uses the delayed version of the expe te periodic interference. The number of del FIR fie of the adaptive filter performance. Note incr_can predict a one sinusoid, as noted earlier. Aft e filter would predict the interference 4S 10.4.1 Canceling Perio Using Linear Pre the corrupted signal x(n) to layed samples is selected by that a two-tap adaptive fer convergence, the dat ni) = Sworn 9A cos(2afn/f) i=010 ADAPTIVE FILTERS AND APPLICATIONS 488 ‘Audio and periodic interference (7) = 8(7) + A cos (2xfn/f,) E — x(n) Adaptive y(n) FIR filter FIGURE 10.18 Canceling periodic interference using the adaptive filter, Therefore, the error signal contains only the desired audio signal en) = s(n). (10.17) 10.4.2 Electrocardiography Interference Cancellation from magnetic induction, displacement currents in Jeads or in the body of the patient, and equipment interconnecti and imperfections. Figure 10.19 illustrates the application of adaptive noise canceling in ECG. The primary input is taken from the ECG Preamplifier, while a 60-Hz reference To 60-Hz wall outlet x(n) y(n) Adaptive Reference . Et ith retard Pay signa | ECS err in sone tt interference Iz sie interference ane, s(n) + n(n) ECG Preamplifier and ADC Ld FIGURE 10.19 Mlustration of canceling 60-Hz interference in ECG.10.4 Other Application Examplos 489 gaken from a wall outlet with proper attenuation. After prope ns Hg the digital interference x(n) is auequired by te digital signal (DS) iv The digital adaptive filter uses this reference input signal to produce on which approximates the 60-Hz interference 1() sensed from the og amplifier y(n) & n(@)- (10.18) pee FIR adaptive filter with NV taps and the LMS algorithm can be used for ieapplcation: ya) = ww(0)x(72) + w(L)x(a — I) to + w(N = Lx — N+1). (10.19) nce of the adaptive filter, the estimated interference is ice the ‘Then after convergence e ubtracted from the primary signal seat ale), in which the 0-H intexferen ato) = da) — yn) = (a) +902) ~ 0 s(n). with enhanced ECG recording, doctors in clinics can give more accurate diagnoses for patients. | of the ECG preamplifier to produ .ce is canceled: (10.20) 10.4.3 Echo Cancellation in Long-Distance Telephone Circuits en suffers from impedance mismatches. t interface. Balancing electric networks Long-distance telephone transmission oft fh the hybrid to the subscriber loop due This occurs primarily at the hybrid circu within the hybrid can never perfectly mate! to temperature variations, degradation of the transmission line, and so on. As @ ret, a small portion of the received signal is Jeaked for transmission. For example, in Figure 10.20a, if speaker B talks, the speech indicated as xp(n) will Ee re transmission line to reach user A, and a portion of xp(n) at site A is and transmitted back to the user B, forcing caller B to hear his or her own Local 2-wire Repeaters Local 2-wire ‘customer loop > customer loop (0) ——J] f x(n) ICentral L{ ICentratl office office — wire trunk A A SURETO.20a si Simplified long-distance circuiA PLICATION DAPTIVE FILTERS ND AP : 490 10 A echo for speaker B. A similar echo illustration, can When the telephone call is made over a long distan : conducted pals nk as with geostationary satellites), the echo gg ae o much as 540 ms. The echo impairment can be annoying tg te le i i e distance. Se of echo in long-distance communications, an ane filter is applied at each end of the et hae aS shown j Figure 10.20b. Let us examine the adaptive filter installed at the speaker A site, The incoming signal is xg(n) from speaker B, while the outgoing Signal contains the speech from speaker A and a portion of leakage from the hybrid circuit dj(v) = xa(n) + Xo(n). If the leakage Xp(7) returns back to speaker B, it becomes an annoying echo. To prevent the echo, the adaptive filter at the Speaker A site uses the incoming signal from speaker B as an input and makes its output approximate to the leaked speaker B signal by adjusting its filter Coefficients, that is, voice. This is known as an N-1 val) = S> wxa(n — id) & Xp(n). (10.21) i=0 As shown in Figure 10.20(b), the estimated echo y4(n) & Xp(n) is subtracted from the outgoing signal, thus producing the signal that contains only speech A; that is, e4(n) © x4(n). As a result, the echo of speaker B is removed. We can illustrate similar operations for the adaptive filter used at the speaker B site. In Practice, the FIR adaptive filter with several hundred coefficients or more is Echo of speaker B %2l") Transmitting site oo et DS Fal) e4(n) = x4(n) customer loop Channel ADF Hybrid Yaln) 4 — Channel oa ae Local air SalM~ He) dn) eh + ‘customer loop : 4-wire trunk Transmitting site x(n) B Echo of speaker A F 'GURE 10.208 Adaptive echo cancelers,10.6 Probloms 491 gnonly used L0 effectively cancel the eho, nontineari y : Neariti ich pall © corresponding nontingar adaptive canceler oe ncerned in I me the performance of the echo ¢ rcellation ‘nceler can be used to int ation, sof adaptive filters and other applications Myer fort olerre c fc ed Lo the erences for f book. ‘The reader is. are beyond the scope urther development, of (his jo.5 Summary |. Adaptive fillers can be applied to gi J ; ignal-changing environments, spectral overlap between noise and signal, ai ind unknown, or time-varying, noises, 2. Wiener filler theory provides optimal weight solutions based on statistics. I involves collection of a large block of data, calculation of an auto- correlation matrix and a cross-correlation matrix, and inversion of a large size of the autocorrelation matrix. The steepest descent algorithm can find the optimal weight solution using an iterative method, so a large matrix inversion is not needed. But it still requires calculating an autocorrelation and cross-correlation matrix. The LMS is a sample-based algorithm, which does not need collection of data or computation of statistics and does not involve matrix inversion. . The convergence factor for the LMS algorithm is bounded by the recip- rocal of the product of the number of filter coefficients and input signal Power, The LMS adaptive FIR filter can be effectively applied for noise cancel- lation, system modeling, and line enhancement. ications such as cancellation of ignal enhancement, and adaptive + Further exploration includes | other app! Periodic interference, biomedical ECG si telephone echo cancellation. 10.6 Problems i filter: Moy, Given a quadratic MSE function for the Wiener 2 7 = 50—40w + 10w", * to achieve # the minimum MSE Jinin and find the optimal solution for ‘etermine Jinin-