Single-Sensor Active Noise Cancellation Alan V. Oppenheim, Fetlow, rEBE, Ehud Weinstein, Fellow, IEEE, Kambiz C. Zangi, Student Member, IEEE, Meit Feder, Senior Member, IEEE, and Dan Gauger, Member, IEEE Absract— Active nose cancellation is an approach to nose reduction in which a secondary noise source that destructively interferes with the unwanted noise I introduced. In general a tive nase cancellation systems rely on multiple sensors to measure the unwanted noise feld and the effet of the cancellation. Tis paper develops an approach that izes a single sensor. The hole field is modeled asa stochastic process, anda time-adaptive algorithm is used to adaptively estimate the parameters of the process. Based on these parameter estimates, a canceling signal IS generated. In general, the transfer function characterises fom the canceling source tothe error sensor need to be accounted for. these can be accurately measured in advance and are invertible ‘except forthe propagation delay between the source and sensor, then the essential problem becomes one of predicting future values of the noise field. The algorithm developed is evaluated ‘with both artielally generated noise and with recordings of Airraft noise NWANTED acoustic noise is @ by-product of many industrial processes and systems. With ative noise can- cellaion (ANC), @ secondary noise source is introduced 10 _Renerate an acoustic field that interferes destructively withthe ‘unwanted noise and thereby attenuates it (2), (3) (1) ‘Conventional ANC systems typically utilize several sensors at least one to measure the noise field and a separate sensor to measure the canceled or attenuated noise. A conventional two-sensor ANC system consists of an input sensor, adaptive Str, canceling source, and error sensor, as depicted in Fg. ‘The input sensor is used to measure the unwanted noise ata location away from the error sensor and provides the input to the adaptive filter. In applications such as noise cancellation Jn a duct for which the noise propagation is essentially unidirectional, the input sensor is positioned upstream of the location at which the noise is to be canceled so that its output in effect anticipates or predicts the noise field at the Jocation of the error sensor. The error sensor measutes the Manse recive January 1192; vise Jay 1,193, Teac ator cordate reve fhe pape st proving or psa Wea De Chia Lee Ths work ws spond patty the U.S Ar Fnce-Ofles of Scie Rexeach ender Gran Nare= APOSRS IAS {2d in pat by te Oe of Naval Rear set Gran Naber NOOO adhe ‘A.V. Oppenheim sn K.C. Zang ae wih he Resch Laboratory oc Foe nd the Deparment of Elena Engng ad Compe ‘Scene, Mastachsets Ita of Tecieloy, Cambie, MA, C2130 E Weintein i wi te Deparmen of lea Engng Ses, acl of ngicrng. Te Avi Une, Tea, fel He ao eh ‘te Deparment of Applied Onan Physics aod Engineering, Most le ‘Oteatoptepie laiten, Wood's He MA’ ‘Fer tn he Deparent of Elsa Enginching Systems Faety of Eoginering Tease Dery el ave 1, Gaager with Boxe Caprio, Faminghen. MA, 0170 IEEE tog Nombe 9215040 Fig 1. Comentional astve nie canes «yen, residual acoustic field that is used 10 adjust the adaptive filter coefficients. Two-sensor systems are particularly effective when the input sensor ean anticipate the noise field a the eror ‘sensor so that processing delay in the filter and any propagation elay between the canceling source and the error sensor are easily compensated for. In addition to the isues associated ‘with deploying multiple sensors, a common difficulty with ‘multiple sensor ANC systems is that thee is typically some feedback from the canceling speaker to the input microphone. Several approaches have been proposed for dealing with this problem either by utilizing a configuration of canceling sources ‘that minimizes the feedback or by taking the effect of feedback into account in the design of the adaptive filter [3] In this paper, we present a new ANC system utilizing only single sensor. As with conventional systems, a canceling source is used to generate « second acoustic field, However, Inthe system we develop, a single sensor is used to provide an estimate of both the original noise field and the canceled noise field, ‘To compensate forthe propagation delay between the can- cling speaker and the sensor, predicted Values of the noise field are used, The prediction is based on modeling the noise field as a stochastic autoregressive process whose parameters are adaptively estimated. Because the parameter estimation is adaptive, the resulting ANC system can be used to cancel Stationary a§ well as nonstationary noise, Tn a recent paper by Zangi (13), the performance of & single sensor ANC algorithm is compared with that of the ‘wo sensor algorithm proposed by Burgess (1). The results in this paper suggest thatthe noise attenuation levels obtained by the single sensor algorithm are typically 10-15 dB higher than the ones obtained by the two sensor algorithm. Note that in the two sensor algorithm, the canceling signal is derived from the output of the input Sensor, wheress in the signal sensor algorithm, the canceling single is derived from the outpat of | the error sensor Its argued in [13] thatthe output of the error sensor caries much more information about the future values 1068-46740450400 © 1994 IEEEcote tntge Fig 2. Generic single miroptone sete ose cancelation sytem, of the unwanted noise atthe error sensor than the output of the input sensor located away from the ertor sensor. This is particularly true whenever the noise field is omnidirectional So thatthe output of the input sensor can no longer anticipate the noise atthe error sensor. Using the algorithms developed in this paper, a number of experiments were performed, both with artificially generated noise and with recordings of aircraft noise. With artificially generated noise, the performance of the proposed ANC system in a mean square sense was found to be close to that of the system that would exploit the exact noise characteristics rather than adaptively estimate them. Furthermore, in the ease of nonstationary noise, the proposed system is able to adapt 10 the changing noise statistics Inthe context of aircraft noise, the algorithms were evalu- ated in a simulated environment to cancel the noise generated by a helicopter, a propeller siplane, and a jet airplane ‘Assuming thatthe canceling speaker and the microphone were three centimeters apart and that the transfer function between the canceling speaker and the microphone is a pure delay, the algorithm is able to atenuate the overall noise power by 45, 40, and 35 aB, respectively In Section I, we present our model for the cancellation ‘environment and the unwanted noise. In Section Il, we derive the ANC algorithm based on this model. Setion IV discusses the performance ofthe algorithm on recorded airraft noise. IL Monet. Speciricarion ‘A generic single-ensor ANC system is depicted in Fig. 2, Where the microphone output is m(¢), and r(e isthe input tothe canceling loudspeaker. The microphone measures the sum ofthe unwanted noise s(t) and the canceling signal et. ‘The objective isto generate r() based on the measurements ‘of raft) in such a way thatthe energy of n() is minimized. ‘The block diagram forthe sngle-sensor ANC stem that we propose and develop in this paper is depicted in Fig A. The system G(z) represents the overall wanser function from the canceling soure input r(t) 0 the sensor output ‘m(2) and incorporates the transfer fonctions of the source and the sensor together with the propagation delay between the source and senor. The microphone outpat mit) 8 the sum of the unwanted noise s(), the canceling signal (), aud the measurement noise u(), The overall strategy i based on the observation that if G(=) is known or canbe adaptively estimated the, since r() is known exactly, an estate of the tcanceled noise at the sensor can be obtained by subtracting cout the component ofthe sensor output due tothe canceling oe 2a Lea Hy Cancellation Algorithm Fe. Ow so mine he soe cea em oom |p fr Sen source, i.e, we can extract =(0) = (0) + v(é) from the output of the microphone m(t). The input to the canceling loudspeaker is then generated based on 2(°). Although itis eventually important and of interest t fully ‘develop the algorithm by adaptively estimating G2), we focus in this paper on the more idealized and simpler problem in which we only account forthe propagation delay between the source and sensor. Since this delay is the result of the source and sensor separation, it is reasonable to assume that itis ‘known and constant. ‘A block diagram representation forthe overall system, with G(z) = 2-™, is shown in Fg. 4. In this figure, o(¢) is the unwanted noise tthe microphone and is modeled asthe output ‘of an all-pole transfer funetion driven by white noise, i... a(t) is modeled as an autoregressive (AR) process. v(@) represents the measurement noise in the microphone Recall that the objective is to choose r(t) such that the energy of tbe residual signal measured by the microphone rm(t) is minimized; furthermore, r(t) must be generated based on 2(r) : 7 = 1,...,t It is then easy to see that in the system of Fig. 4, the minimizing choice for r(t) is given by the following conditional expectation: r() a(t)} where 2(0) = 9(8) + (2). Note that 2(0) is what the microphone would have measured, if the canceling signal was tumed off. By choosing r(t) according to (1), the output ofthe microphone mt) becomes equal to the following prediction B{e(t + M)|2(),2(¢- 1), au(t) = 2(0)-B(e(H)[2(t-30),2(0-M=1),...,2(0)). @) (Our approach isto estimate the parameters of the AR model through the adaptive algoritum derived in Section Il. These parameters along with the measurements of the microphone fare then used to predict 2(t). These predicted values are in tum used 10 obtain the input to the canceling loudspeaker according to (1). ‘Although, in this paper. we focus specifically on signal prediction and noise cancellation, we note thatthe algorithms presented here can also be used for the purpose of signal enhancement since they generate the signal and parameter estimates in the presence of noise. Since the algorithms are sequental/adaptive in nature, they may be particularly useful in the context of enhancing nonstaionary signals such as speech in the presence of nonstationary noise. IL, SINGLE SENSOR ADAPTIVE ALGORITHM, ‘The model for the signal =(t) is 2) = a(t) +0(t) @ where =~ Dans -#) tue) ® where u(t) and v(t) are statistically independent zero mean white Gaussian processes with average powers of o2 and o2, respectively. Given signal observations up to time ¢, the minimum mean square error (mms. estimate of s(t +m) is given by the conditional expectation: (C+ m) = Bfo(t-+ m)l2(1),202),.--2}- Note that E{a(t + M)|s(t),...,2(1)} = a(¢ + M) since 2(U+M) = s{t+Af)+0(t+M) and v(t+ Af) is independent Of 2(1)v...,2(0). This implies that (1) can be rewritten as r(Q) = “a(t-+ M). Hence, inthe remainder of this section, ‘we will concentrate on adaptively calculating 4(¢-+ M), If we assume that the parameters a1,02,---»09,08, and 2 are precisely known, then this conditional expectation can be computed efficiently using the Kalman filtering equation. ‘Toward this end, we represent (3) and (4) in state-space form x(0) = Sx{t—1) +eru(e) © Fx(0) + 0(0) ® where x) isthe (p +1) x 1 state vector defined by x(t) = [olt)slt = 1) (tp)? ® isthe (P+ 1) x (p+ 1) transition matrix may map ay 0 1 ° o #=|° o ° o 16 and ey isthe (p +1) 1 unit vector e=l 0 oF ao) Denote by (U0) = E(x( O20), 210} ay the state estimate based on d to time t and by P(tlt) = BUREe) ~ x(ONER(ee) — x(a), 20) (2 the associted error covariance matrix ‘Then, using the standard Kalman filter formulation, &(¢|t) and P(t | 4) can be computed sequentially in time, in two stages, as follows: Propagation Equations: (tt 1) = @R(¢— 11-1) Pitlt-1) = @P(C~ Ile 1)8" + oleef a, (sy Updating Equations: X(t) = (ule 1) + (DLAC) — FRCL) AS) P(t) = [= k(®eF P(e = 1) 6) where k(t isthe Kalman gain given by Ke) : POlt—t)e. 7) SPCR Dei +e ‘The frst component of 5(¢|t isthe estimate of s(t) based con data up to time ¢. To obtain the signal estimate at ime (¢-+m) as required by (8), we use (4) together with £(¢[) to obtain the predicted values &(t+1[t),a(¢+2lt),...4(¢+mlt), specifically Hee rl) == Danser A) 7 =1,2,...m. 8) ‘The Kalman filtering equations require ,03, and o2. Since these parameters are not available, they must be estimated as wel ‘The signal parameters satisfy the Yule-Walker equation aywhere 0 isthe px 1 vector of zeros, a is the p x 1 vector of the AR parameters = fanya2y.-s05]", co) and R isthe (p+ 1) x (p41) signal comelation matrix = Blx(0x70), where x(t) isthe state vector defined in (8). ‘The estimate of the noise spectral level is given by a8 = Blla(®)~ a?) “El the vector of unknown parameters and by a . (2 HH = JRO 240) its estimate based on data up to time ‘Thea, in accordance with (19), we consider generating the estimates of the signal parameters in such a way that the following equation is satisfied: sofaeen] = FO] wer R(t) isthe estimate of RL which is obtined by performing the weighted averaging ey 2 Denote by a ey 0s) Roe LR) a6 where x(x O*R(e)87 ld) + PC) @ and where 3(¢[t) and P(¢|¢) are the estimate of the state and its covariance, which are computed using the Kalman filtering equations (13}-{16), where instead of 8, we use the current cstimate 4(0), Le Propagation Equations: 3(¢}t—1) = B(oR(¢~ 1) es) P(tlt—1) = HPC Ae 1)87(0) + BE(tereP 29) Updating Equations Xt) = ACHE) +K(O[e(0) ~ FREHE~Y] GO) PEt) = [7 - ket] Pee — 1) on where (2) is the maui defined in () computed at a = &(), aia K(0) isthe vector define in (17) with of replaced by #0. Let (lt) and x(t}x"(@) be paritioned as follows: ata ith Frey =f 80 stosF-1) 10 Oy oie ‘To obtain a sequential procedure for updating the signal parameter estimates, we define x= Je ay (ee gain) : PT [Qa Qa) fine fou ati, Ame Math Sir, vo 37. pp. a9, 196, luo) 8 Vrsman, “Admin fk cnaded fe sae et ad secuence compound desion robles J Malaria el Vl 10, ps4, 1980. ion med” Anal (01) ©. Waa. Pole, and J. Tey, “Active steno” US Pet 1453 Som Spe 25 oka 112] E Weasei, MF, aod A. Open, "Sign enhancement sing Single tnd melste senor masucns® MIELE Tech Rep. Si Dee 1, (13) Zang “A new twos stive ass aneain alo a Proc ICASSP "inaegals), ApS 193 op I-88 we ‘Alan. Oppentci (77, oc shorograph and bogey. plese ee pe $13 ofthe etter 193 la of thi RANSACTIONS ud Welnstin (S86. 98), fr photograph and Dogrphy plete se age 413 of he Octaber 1989 se fhe TRANSACTIONS ambi C: Zang (91) mci the SM, dere inclecrialengieerng in 190 from he Masta ‘ett of Technolgy, Cambridge, wee is curery working toate PD, deo Deparment of lee! Engng an Caper Some is min ae of scar nil processing leon for wo snd eo applets, Mee Feder (M¥6-SM93), for photograph and biography, pies page| 413 of he Ober 199 sae of ths TRANSACTIONS Dan Gauges (M8) eeved the SB, and SM greta elec egicesng rom te Mat tehoseus Ite of Techoaogy, Cambridge Feary 1981 He je the enpnesng staf at Bose Corpor sos, Framingham, MA, st 580 Ip his 13 Jeon a Bow, eta worked pearly o ot toclaton td poer procesng ts toe oe pape ‘nveach Bed He ha suo worked in yoga ‘manager’ sd makeing a sare Boe a {Cet Eine for Nowe Redaction Tol Profesional, his ime ae cncetaed i col sims. Sip sang, and et doug ‘Mes Gage iss member of Sigma Xian he Space States ni