Active Noice Reduction headset - DiVA portal571737/FULLTEXT01.pdfheadset cup. The dynamics of the primary (cockpit, headset cups) and sec-ondary (power ampli er, speaker, acoustics

Active Noice Reduction headset

HENRIK FRANSSON

Masters’ Degree ProjectStockholm, Sweden June 2009

XR-EE-RT 2009:013

Abstract

Conventional (passive) headsets used in propeller aircrafts are reasonablygood at attenuating mid to high frequency noise, but fail to achieve good at-tenuation in the low frequency region (below approximately 300 Hz). ActiveNoise Reduction (ANR) improves the low frequency attenuation by intro-ducing an anti-noise signal creating destructive interference thus decreasingthe residual noise level. The aim of this thesis is to develop and implementa digital narrowband active noise reduction headset that works properly inaircrafts and not only in a laboratory environment. The implementation isbased on a narrowband �ltered-X least mean squares (FXLMS) algorithmwhere the tonal components in the noise spectrum are synthesized for use asreference signals to the algorithm. The controller is implemented in a paral-lel fashion where each tonal component is handled separately. The system isbuilt into a headset and laboratory tests show that the algorithm can handle�ve simultaneous tonal components with an adaptation time of less thanone second. Aircraft tests show peak attenuation of 17 dB in both single-and twin-engine aircrafts thus ful�lling the requirements. Simulations andtrue performance show some minor discrepancies which are explained anddiscussed.

Contents

1 Introduction 4

1.1 Aircraft noise . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2 Advantages with active noise reduction . . . . . . . . . . . . . 51.3 How ANR works . . . . . . . . . . . . . . . . . . . . . . . . . 51.4 Feedforward vs feedback . . . . . . . . . . . . . . . . . . . . . 61.5 Broadband vs narrowband . . . . . . . . . . . . . . . . . . . . 71.6 Previous work . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.7 Novelty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Problem de�nition 10

2.1 Problem de�nition . . . . . . . . . . . . . . . . . . . . . . . . 102.2 Non-disclosure of proprietary information . . . . . . . . . . . 10

3 Theory 11

3.1 Adaptive �ltering . . . . . . . . . . . . . . . . . . . . . . . . . 113.1.1 Mean square error . . . . . . . . . . . . . . . . . . . . 123.1.2 Steepest descent . . . . . . . . . . . . . . . . . . . . . 12

3.2 Least mean squares . . . . . . . . . . . . . . . . . . . . . . . . 133.2.1 LMS stability and convergence rate . . . . . . . . . . . 133.2.2 Leaky LMS . . . . . . . . . . . . . . . . . . . . . . . . 13

3.3 Filtered-X least mean squares . . . . . . . . . . . . . . . . . . 143.3.1 Narrowband leaky FXLMS . . . . . . . . . . . . . . . 153.3.2 Step size constraints with secondary path . . . . . . . 153.3.3 Secondary path estimation precision . . . . . . . . . . 163.3.4 FXLMS controller seen as notch �lter . . . . . . . . . 16

3.4 Identifying secondary path . . . . . . . . . . . . . . . . . . . . 17

4 Measuring performance 19

4.1 Lab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194.2 Aircraft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

5 Simulation 22

5.1 Matlab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

1

6 Hardware and software implementation 25

6.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256.2 DSP software . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

6.2.1 Secondary path estimation . . . . . . . . . . . . . . . . 276.2.2 Step size and leak factor . . . . . . . . . . . . . . . . . 27

6.3 Windows software . . . . . . . . . . . . . . . . . . . . . . . . 28

7 Results 30

7.1 Lab tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307.2 Aircraft tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

7.2.1 Single engine aircraft . . . . . . . . . . . . . . . . . . . 307.2.2 Twin engine aircraft . . . . . . . . . . . . . . . . . . . 33

8 Conclusions 35

9 Discussion 36

2

Glossary of Terms

A-weigthing Filter that approximates the human hearing frequency response [1]ADC Analog to Digital ConverterANR Active Noise ReductionCODEC COder/DECoder, a combination of an

audio ADC and audio DAC into a single chip.DAC Digital to Analog ConverterDSP Digital Signal ProcessorFXLMS Filtered X Least Mean Squares algorithm [2]LMS Least Mean Squares algorithm [2]MEMS Microelectromechanical systemsNRR Noise Reduction Rating [3]PCB Printed Circuit BoardPrimary source The unwanted noise sourcePSD Power Spectral DensitySecondary source The anti-noise sourceSNR Single Number Rating [4]SPL Sound Pressure Level, referred to 20µPa [5]

3

Chapter 1

Introduction

Conventional (passive) headsets used in propeller aircrafts are reasonablygood at attenuating mid to high frequency noise, but fail to achieve goodattenuation in the low frequency region (below approximately 300 Hz). Topassively improve the low frequency attenuation, the inner volume of theheadset cups and/or the weight of the cups must be increased. That wouldlead to obvious practical problems and uncomfortable �t. A better approachis to use Active Noise Reduction (ANR), introducing an �anti-noise� signalleading to destructive interference and thus lower noise level.

1.1 Aircraft noise

Propeller driven aircrafts have quite characteristic noise spectrums, showingsimilarities between di�erent aircrafts. The main parts are broadband ran-dom noise (e.g. wind noise) and narrowband tonal noise (e.g. rotating partsas the engine, propeller, gearbox and alternator). This can clearly be seen in�gure 1.1 where also the low frequency tilt of the spectrum is seen, makingthe noise very boring and tiresome.

Our measurements show that the A-weighted [1] noise level is 85�95dBSPL(A) in most aircrafts of this type, thus requiring noise attenuatingheadsets to avoid hearing loss [6] and improve radio and intercom speechintelligibility. During typical cruise in a small, 4-cylinder airplane, enginespeed is around 2300 rpm, corresponding to about 38 Hz. The tonal peaks areharmonics of this fundamental tone, for example a two-bladed propeller givesa peak at twice fundamental frequency along with additional harmonics.Except for take-o� and landing, the propeller (and engine) rpm is normallyheld constant, with only slow variations.

4

101

102

103

104

30

40

50

60

70

80

90

100

Sou

nd p

ress

ure

leve

l [dB

SP

L]

Frequency [Hz]

Figure 1.1: PSD of free �eld cabin noise recorded in a Piper PA-32 aircraft.

1.2 Advantages with active noise reduction

There are several advantages with active noise reduction, both for noisecomfort and for hearing protection. Comfort is enhanced by attenuating lowfrequency noise which is very tiresome and thus a�ecting safety. ANR helpsby reducing the residual noise level and making the audio spectrum more�at. Another advantage is that the communication speech intelligibility isenhanced, since the voice audio does not get drowned in noise.

1.3 How ANR works

Active noise reduction systems are based on superposition, where the un-wanted primary noise (from engine, propeller, wind noise etc) is combinedwith an �anti-noise� secondary signal to achieve destructive interference [2]within a spatially limited �quiet zone�, leading to a lower overall noise levelinside the headset cups, see �gure 1.2.

As seen in �gure 1.3, the ANR controller feeds an output signal y(n)through the ANR speaker to generate the anti-noise secondary signal thatgets combined with the unwanted noise d(n) to form the residual noise e(n)at the acoustic summing junction around a microphone situated in eachheadset cup. The dynamics of the primary (cockpit, headset cups) and sec-ondary (power ampli�er, speaker, acoustics inside cups) paths are modeledby the linear systems P (z) and S(z) respectively. The feedforward block

5

contains algorithms, tachometers and sensors to analyze the ambient noiseand provide synthesized reference signals to the ANR controller.

In general, the size of the quiet zone is limited by the noise frequencyto about 1/10 of the wavelength [7]. The headset cup inside dimensions areless than 100 mm in all directions, which means the highest noise frequencythat still gives full attenuation inside the whole cup is 1 m, correspondingto a frequency of around 340 Hz. As seen in �gure 1.1, all major tonalcomponents are below 200 Hz, thus ful�lling the quiet zone requirements.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1.5

−1

−0.5

0

0.5

1

1.5

Time

Unwanted noiseAnti−noiseResidual noise

Figure 1.2: Primary (unwanted) and secondary (generated) noise get combinedthrough superposition giving a lower total level.

1.4 Feedforward vs feedback

There are two basic ways of designing the ANR controller and thus generatingthe anti-noise signal: feedforward and feedback. The main external di�erenceis whether any signal path is present from the primary noise source to theANR controller. See �gure 1.4 where P (z) is the transfer function from theprimary noise source x(n) to the error microphone with output signal e(n).

The feedforward case can be seen as a system identi�cation problemwhere P (z) is the unknown system that W (z) should converge to [2]. Thefeedback case can be seen as a general prediction problem, where the con-troller converge to a solution that estimates the future values of d(n) [2].

Feedforward ANR generally has a causality constraint, meaning that thepart of the primary noise going through W (z) must reach the error micro-

6

ANRe(n)

d(n)

y(n)

P(z)

Ambient (primary) noise

x(n)

Feed

S(z)

forward

Headsetwearer

Figure 1.3: Basic ANR system inside headset cups (only one side shown). x(n) isthe primary (ambient) noise which gets �ltered through P (z) to form the unwantednoise d(n) inside the cup, y(n) is the controller output, e(n) is the combined pri-mary and secondary noise picked up by the ANR microphone (acoustic summingjunction). S(z) is the secondary path transfer function, including ANR speaker,power ampli�er and acoustic enclosure.

phone no later than the noise going though the P (z) path [2], otherwiseit is impossible to design a realizable controller giving optimal broadbandperformance.

Feedback ANR has no particular causality constraint, but larger controlloop delay gives degraded performance (decreased attenuation and lowerbandwidth) and eventually leads to stability problems.

In this thesis, x(n) is available to the DSP through external circuitry, sothe feedforward approach may be used.

1.5 Broadband vs narrowband

Depending on application, the controller can be implemented to attenuateeither broadband or narrowband (tonal) noise. Broadband controllers workon the whole frequency range of interest while narrowband controllers only

7

P ( z )

e ( n )y ( n )

d ( n )P ( z )

e ( n )y ( n )

d ( n )

N o i s es o u r c e

N o i s es o u r c e

A d a p t i v ea l g o r i t h m

W ( z )

A d a p t i v ea l g o r i t h m

W ( z )

x ( n ) x ( n )

Figure 1.4: Feedback ANR (left) and feedforward ANR (right). x(n) is the pri-mary noise (e.g. in aircraft cabin), P (z) is the transfer function from the primarynoise to the inside of the headset cups, where e(n) is the residual noise in the cups.W (z) is the ANR controller creating the anti-noise signal y(n). Secondary pathdynamics are not included.

attenuate noise at given frequencies with a narrow bandwidth but on theother hand may give better noise suppression and robustness. As seen in�gure 1.1, the present noise spectrum has strong narrowband peaks whichmakes the narrowband ANR controller suitable, along with an already im-plemented broadband analog ANR controller (not covered in this thesis).This thesis will therefore focus on a feedforward narrowband ANR system.

1.6 Previous work

Peter Rybing made a master thesis at KTH in 2003 about an Active NoiseControl headset for use in home environment using feedforward ANR imple-mented on a �oating point DSP in an open earphone type headset [8]. Theaim was to attenuate broadband noise from a noise source (e.g. a vacuumcleaner) while not a�ecting wanted signals (e.g. speech). Performance wasquite low due to the low coherence between the reference noise signal andthe noise near the headset.

Sven Johansson made a doctoral dissertation in 2000 regarding activecontrol of propeller-induced noise in aircrafts, including multiple actuatorsystems [9]. The last paper included in his dissertation is about a hybridANR headset, resembling the design presented in this thesis but the digitalANR is only implemented in Matlab and no real world performance wasanalyzed with digital ANR.

Texas Instruments has several application reports regarding active noisereduction in general, for example the application report [10]. These areoften made in a broader sense and mostly considering the usual educationalANR setup with an acoustic duct instead of a speci�c application such as aheadset.

8

Kuo and Morgan's book �Active Noise Control Systems� [2] is consideredas a reference book for ANR systems. Most of the theoretical parts in thisthesis are covered by this book.

1.7 Novelty

The novel part of this thesis is to develop a robust headset that works in thereal world in real aircrafts and not only in a lab environment, as opposedto many previous papers. I have used more than 30 di�erent airplanes andhelicopters for �eld testing and validation to make sure the headset worksas expected in all situations. Unfortunately, all design and implementationdetails can not be shown in this report to avoid disclosing any proprietaryinformation.

The narrowband ANR is implemented in a parallel fashion where eachfrequency of interest has its own sets of taps and step size, thus givingoptimal performance where each part can adapt at an optimal rate. A specialalgorithm, together with various sensors, keeps track on the current noise�eld in the aircraft cabin and determines which frequency component peaksto attenuate. Only peaks that are strong enough to be audible get selected.

9

Chapter 2

Problem de�nition

2.1 Problem de�nition

The project aim is to design and implement digital narrow band Active NoiseReduction (ANR) functionality into a personal hearing protection headsetusing a Texas Instruments �xed point TMS320C5509A [11] DSP and anAIC23a [12] CODEC. The intended use is in small aircrafts where the cabinsound pressure level is relatively constant and the periodic audio componentsfrom the engine, propeller etc are stationary or slowly varying. The headsetshould be able to attenuate up to �ve simultaneous tonal components be-tween 75�400 Hz, narrowband attenuation should be at least 15 dB on thestrongest tonal component and adaptation time should be less than threeseconds.

2.2 Non-disclosure of proprietary information

Since the headset will be launched as a commercial product, this thesis textwill only cover the basic functionality of the system. There is a great stepbetween a headset that works in a lab environment compared to a robustheadset that works in the real world. Great e�ort has been put into bothhardware and software implementation to avoid artifacts like clicks, pops andunwanted noises and also make sure the headset works as expected duringall phases of the �ight (engine start, taxing, take-o�, ascent, cruise, descent,landing etc). All these details has deliberately been left out to not discloseany proprietary information.

10

Chapter 3

Theory

As seen in �gure 1.3, the ANR system is a closed loop control system, withthe microphone as error sensor, the �lter (either analog or digital) as thecontroller and the speaker as the actuator. This means that the system canbe designed with standard control theory.

3.1 Adaptive �ltering

In order to maintain a stable destructive interference, the digital �lter needsto be constantly updated to accommodate for a non-perfect system modelor that the system changes over time, e.g. changes in headset position onthe head or di�erent users wearing the headset. There are several ways toperform this update, but the most common way is to use some variant ofthe steepest descent algorithm [2].

In general, an adaptive �lter consists of two parts: a digital �lter to per-form the �ltering process and an adaptive algorithm to adjust the coe�cientsof the �lter in order to achieve optimal performance. The �lter is usually of�nite impulse response (FIR) or in�nite impulse response (IIR) type. Thisthesis will only focus on FIR �lters and variants thereof. One advantage withFIR �lters is that they are always stable, due to only zeros and no poles (butthe complete system may however become unstable since it contains bothpoles and zeros).

Referring to �gure 3.1, for an FIR �lter of length L, the reference inputvector x(n) and coe�cient vector w(n) at time n are constructed as

x(n) = [x(n) x(n− 1) . . . x(n− L+ 1)]T (3.1)

w(n) = [w0(n) w1(n) . . . wL−1(n)]T (3.2)

The output signal y(n) at time n will then be:

y(n) = wT (n)x(n) (3.3)

11

Digitalfilter W(z)

Adaptivealgorithm

d(n)

x(n) e(n)

+

-y(n)

Figure 3.1: Block diagram of a general adaptive �lter where d(n) is the desiredsignal, x(n) is the reference signal, y(n) is the output from the �lter W (z) and e(n)is the error signal.

The residual error e(n) will then be de�ned as:

e(n) = d(n)− y(n) = d(n)− wT (n)x(n) (3.4)

3.1.1 Mean square error

To be able to update the �lter coe�cients, the current performance of the�lter must be measured, otherwise it is not possible to adapt it towardsa "better" set of coe�cients. One such measure is the mean square error(MSE), de�ned as:

ς(n) = E{e2(n)

}(3.5)

where E {·} denotes expected value. In order to achieve maximum atten-uation, which means lowest residual noise, the MSE should be as small aspossible and the optimization should therefore strive towards as small ς(n)as possible.

3.1.2 Steepest descent

To reach the optimum set of coe�cients which minimizes ς(n), the methodof steepest descent is used, mainly since it is widely known and can beimplemented on a sample-by-sample basis. The objective with this methodis to adjust the �lter coe�cients w(n) in small steps in the direction of thenegative gradient of ς(n) to make the MSE as small as possible. The gradientis de�ned as the partial derivate of ς(n):

∇ς(n) = ∇E{e2(n)

}= ∇E {e(n)e(n)} = 2E {∇(e(n))e(n)} (3.6)

The de�nition of e(n) in equation 3.4 then gives:

∇e(n) = ∇{d(n)− wT (n)x(n)

}= −x(n) (3.7)

∇ς(n) = −2E {x(n)e(n)} (3.8)

12

This lets us update the �lter weights in the negative gradient direction:

w(n+ 1) = w(n)− µ

2∇ς(n) = w(n) + µE {x(n)e(n)} (3.9)

where µ2 is the step size for each coe�cient vector update.

3.2 Least mean squares

To use equation 3.9, an estimate of E {x(n)e(n)} must be found. A simpleapproach is to approximate the expected value as:

E {x(n)e(n)} ≈ x(n)e(n) (3.10)

which gives us the least mean squares (LMS) algorithm, widely used foradaptive �lters due to its simplicity and low computational e�ort. Equation3.10 then gives the following LMS update equation:

w(n+ 1) = w(n) + µx(n)e(n) (3.11)

3.2.1 LMS stability and convergence rate

In order to use equation 3.11 the step size µ must be determined. It can beshown [2] that the step size with a no-delay secondary path is bounded by:

0 < µ <2

LPx(3.12)

where L is the �lter length and Px is the instantaneous power of x(n). Asmall value of µ will lead to slow convergence and a large value will eitherlead to excessive noise (�lter weights will �uctuate around the true weightscausing audible artifacts) or that the algorithm diverges. A compromisemust therefore be made if using the standard LMS algorithm.

3.2.2 Leaky LMS

Since the electro-acoustic open loop system, with microphone and speaker,inherently is AC-coupled (contains at least one null at S-plane origin), the�lter weights w(n) may try to adapt to a solution containing a DC com-ponent, which can not be reproduced by the speaker. The DC componentleads to practical problems such as reduced �lter tap calculation headroomand numerical over�ow. To make sure the �lter taps does not get a too largeDC component, the LMS update algorithm equation 3.11 is modi�ed as:

w(n+ 1) = αw(n) + µx(n)e(n) (3.13)

where α is a constant close to, but less than, unity. This give two "forces"acting on the taps; one is the tap update µx(n)e(n) making the taps grow

13

until optimum solution is found and the other is the α constant which makesthe taps strive towards zero. α must be chosen with care, since all valuesless than 1 (which is the standard LMS algorithm) give a higher MSE andthus higher residual noise. Another practical advantage with the leaky LMSalgorithm is that the robustness is improved and excessively large tap sizesgets decreased [2], which reduces peak power delivered to the speakers, thusincreasing battery life.

3.3 Filtered-X least mean squares

A feedforward adaptive �lter considered as a system identi�cation problemas in �gure 1.4 does not really re�ect the reality when it comes to activenoise reduction, since the summing junction is acoustic (see �gure 1.3), andtherefore need to take other parts, such as loudspeakers, microphones andthe acoustic environment, into account. Since electret microphones usuallyhave very �at frequency response compared to speakers and the primary pathP (z) is not perfectly known, it is convenient to include also the microphonedynamics into the secondary path model S(z) and the primary path P (z),leading to the same result (but with simpli�ed calculations) as if the micro-phone dynamics had been placed after the summing junction in the model.The secondary path �lters the adaptive �lter output y(n) to y′(n) as seen in�gure 3.2. The addition of S(z) makes the error signal e(n) no longer look

e(n)y’(n)

d(n)

Noisesource

LMS

W(z)

P(z)

S(z)y(n)x(n)

Figure 3.2: Adaptive �lter including secondary path S(z).

like in equation 3.4, but rather like:

e(n) = d(n)− y′(n) = d(n)− s(n)∗ y(n) = d(n)− s(n)∗ [wT (n)x(n)] (3.14)

which makes the LMS update equation 3.11 change to:

w(n+ 1) = w(n) + µs(n) ∗ x(n)e(n) (3.15)

However, S(z) is normally unknown and must therefore be estimated as S(z),leading to the �ltered-X least mean squares (FXLMS) update equation andblock diagram in �gure 3.3:

w(n+ 1) = w(n) + µs(n) ∗ x(n)e(n) (3.16)

14

e(n)y’(n)

d(n)

Noisesource

LMS

W(z)

P(z)

S(z)y(n)x(n)

S(z)

x’(n)

Figure 3.3: Adaptive FXLMS �lter including secondary path estimate S(z).

3.3.1 Narrowband leaky FXLMS

Since the digital ANR in this thesis is aimed towards tonal (narrowband)components, the standard broadband LMS/FXLMS implementation can bemodi�ed into a narrowband algorithm by only using the periodic parts ofthe reference signal x(n). By using external circuitry monitoring the noisesource, a synthesized reference signal can be created with only narrowbandcomponents and without broadband noise. This solution has several advan-tages such as the LMS stability constraints in equation 3.12 getting simpli�edsince the reference signal now has a constant amplitude and therefore con-stant signal power, thus letting the step size µ be constant. Furthermore,the supervising control system may select which narrowband components toattenuate, based on e.g. frequency and perceived loudness.

For a single frequency ω0, the narrowband leaky FXLMS reference and�lter weight vectors are de�ned as:

x(n) = [sin(ω0) cos(ω0)] (3.17)

w(n) = [w0(n) w1(n)] (3.18)

where ω0 is the digital frequency of interest, giving a tap update equation:

w(n+ 1) = αw(n) + µs(n) ∗ x(n)e(n) (3.19)

which can be expanded to:

w0(n+ 1) = αw0(n) + µs(n) ∗ sin(ω0)e(n) (3.20)

w1(n+ 1) = αw1(n) + µs(n) ∗ cos(ω0)e(n) (3.21)

Multiple single-frequency ANR blocks (with individual ω0 frequencies) maybe paralleled to be able to attenuate a composite signal consisting or morethan one frequency component.

3.3.2 Step size constraints with secondary path

When the secondary path S(z) is present, the step size constraints in equa-tion 3.12 will be modi�ed and can be approximated by [13]:

15

sin

cos

LMS

w

w

e(n)y(n)

S(z)

S(z)

1

0

S(z)

d(n)

x (n)0

x (n)1

f (n)0

x’ (n)1

x’ (n)0

Figure 3.4: Block diagram of a single frequency narrowband FXLMS implemen-tation.

0 < µ <1

Px′(L+ ∆)(3.22)

where L is the �lter length (2 in the case of narrowband FXLMS), Px′ isthe instantaneous power of x′(n) and ∆ is the delay of the secondary pathexpressed as number of samples.

3.3.3 Secondary path estimation precision

As shown in [2], for small step sizes, the closed-loop transfer function fromthe primary noise d(n) to the residual noise e(n) can be approximated as:

E(z)D(z)

≈ z2 − 2z cosω0 + 1z2 − [2 cosω0 − µAs cos(ω0 − φ∆)]z + 1− µAs cosφ∆

(3.23)

where As is the gain of the secondary path at ω0 and φ∆ is the phase di�er-ence between S(z) and S(z) at frequency ω0. For small step sizes, the poleradius of equation 3.23 is:

rp ≈√

1− µAs cosφ∆ (3.24)

which can only be greater than 1 (instability) if cosφ∆ is negative, thus theconstraints on φ∆ are:

−90◦ < φ∆ < 90◦ (3.25)

This means that it is enough to determine the secondary path phase estimatewithin±90 degree of the true phase to maintain stability, although it is shownin [2] that increasing phase estimation error a�ects adaptation time.

3.3.4 FXLMS controller seen as notch �lter

As shown in equation 3.23, the closed loop transfer function takes the formof a second order system, having two complex poles and zeros respectively.

16

Both zeros are located on the unit circle and the poles are slightly insidethe unit circle and, if φ∆ is zero, at the same frequency as the zeros. Thetransfer function therefore takes the form of a notch �lter, as seen in �gure3.5. Step size a�ects both bandwidth and depth of the notch, where largerstep size gives a wider and deeper notch, which is not always feasible as willbe shown later during simulations.

70 80 90 100 110 120 130−60

−50

−40

−30

−20

−10

0

10

Frequency [Hz]

Mag

nitu

de [d

B]

Step size 0.03Step size 0.12

Figure 3.5: Narrowband FXLMS transfer function from primary noise d(n) toresidual noise e(n) for two di�erent step sizes.

3.4 Identifying secondary path

In order to use the FXLMS algorithm, the secondary path S(z) must beestimated. Since we are only interested in the secondary path transfer func-tion at discrete frequencies, a frequency analysis method as in [14] is usedto identify both amplitude and phase shift at a number of frequencies andthen make a linear least squares �tting of the amplitude and phase responsesrespectively that the DSP algorithms will use during runtime. As shown inequation 3.25, the phase response does not need to be very precisely deter-mined so linear interpolation is good enough. A sinusoid output signal u(n)is generated by the DSP and fed through the secondary path to the errormicrophone resulting in signal eu(n).

u(n) = Aident cosω0 (3.26)

Eu(z) = U(z)S(z) (3.27)

17

where Aident is output amplitude of u(n) and ω0 is the current frequency tobe analyzed. A leaky LMS algorithm as in equation 3.13 is used to determineamplitude and phase of eu(n) (and thus S(z)) at each discrete frequency.

18

Chapter 4

Measuring performance

The performance of passive personal hearing protection headsets are usuallymeasured using Noise Reduction Rating (NRR) [3] in the US or Single Num-

ber Rating (SNR) [4] in Europe. Both methods use test persons to get asubjective performance �gure condensed into a single number. The attenua-tion is A-weighted [1] before calculating the NRR/SNR value, which meansthat the low frequency ANR contribution to the total attenuation is almostignored, thus making this measure inappropriate for the active part of ANRheadsets. When developing an ANR headset, an objective, non-weightedand more detailed measurement method must be used to be able to makerepetitive measurements and get detailed performance �gures.

4.1 Lab

A soundproof room equipped with hi� speakers to reproduce an arbitraryexcitation signal is used for lab tests, see �gure 4.1. An arti�cial head as in�gure 4.2 is placed in the middle of the room. The head is made of solidaluminum according to [15] containing electret mic capsules mounted �ushto the �ear� surface to be able to measure the residual noise inside the cups.

An audio analyzer is used to generate the excitation signal and analyzethe signals from the head microphones. The most common excitation signalsare e.g. pink noise, white noise, one or more sinusoids or a swept sinusoid,depending on which kind of performance to measure. For the present applica-tion (aircraft noise) single or multiple sinusoid signals, optionally embeddedin noise, resemble the real-world noise appropriately and are therefore usedto analyze the performance of the headset. Attenuation (also called insertionloss) is measured as the level di�erence between ANR on and ANR o� ineach frequency bin of the resulting FFT or PSD spectrum.

19

4.2 Aircraft

Aircraft measurements were mainly carried out by using small KnowlesMEMS microphones mounted on ear plugs inserted into the wearer's as in[16], see �gure 4.3. Together with a portable digital sound recorder, theresidual noise inside the headset cups is recorded for further post-processingafter �ight where the attenuation is calculated in the same way as for thelab case.

Figure 4.1: Schematic view of soundproof room used for lab tests.

Figure 4.2: Arti�cial head used for lab tests.

20

Figure 4.3: In-ear microphones used for aircraft tests.

21

Chapter 5

Simulation

5.1 Matlab

The narrowband FXLMS algorithm described in equations 3.20 and 3.21 wastested with a single sinusoid embedded in white noise giving a signal-to-noiseratio of 20 dB, since a pure sinusoid with no noise gives an unrealisticallylarge attenuation during simulation (in the real world, there is always noisepresent).

Unity reference signal amplitude, step size of 0.03 and 0.12, secondarypath delay of 11 samples, disturbance frequency of 100 Hz and a sample rateof 1 kHz gives attenuation results as shown in �gure 5.1. The disturbancesignal peaks at 0 dB and step size 0.03 gives an attenuation of 22 dB ofthe 100 Hz peak. When step size is increased to 0.12, the peak attenuationis increased to 41 dB, although at the cost of excessive out-of-band noisemaking the overall (broadband) level higher than with the small step size.

Residual error signal power and tap sizes as a function of time are shownin �gure 5.2, where the larger step size gives higher error magnitude in the�rst graph as a consequence of the excessive tap adjustments seen in thethird graph. It can also be seen that the adaptation time (time for each tapto reach 1− e−1 of �true� value) with step size 0.03 is approximately 200 ms.

The e�ect of step size was further simulated by using the same parametersas above, but varying the step size in smaller amounts and measuring boththe overall (broadband) and narrowband attenuations. Results are shown in�gure 5.3. The narrowband attenuation increases with increased step size upto a certain point but the broadband residual level is approximately constantat step size 0.03�0.11, which means the out-of-band noise is increased. Thealgorithm �nally diverges at a step size of approximately 0.14. This is a bithigher than the theoretical step size de�ned in equation 3.22 which yieldsan upper bound of 0.08. The reason is probably that the theoretical valueis an approximation and made for wideband FXLMS systems, with up tohundreds of taps.

22

0 20 40 60 80 100 120 140 160 180 200−45

−40

−35

−30

−25

−20

−15

−10

−5

0

5

Frequency [Hz]

Sig

nal p

ower

[dB

]Disturbance signalError signal, step size 0.03Error signal, step size 0.12

Figure 5.1: Matlab simulation PSD of disturbance and residual error signals. Theerror signals are the residual �noise� inside the headset cups.

0 500 1000 1500 2000 2500 3000 3500 4000 4500 50000

0.05

0.1

0.15

0.2

Err

or m

agni

tude

squ

ared

Step size 0.03Step size 0.12

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000−1

−0.5

0

0.5

1

Tap

mag

nitu

de

w0 (step size 0.03)w1 (step size 0.03)

0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000−1

−0.5

0

0.5

1

Samples

Tap

mag

nitu

de

w0 (step size 0.12)w1 (step size 0.12)

Figure 5.2: Matlab simulation of LMS algorithm. Error magnitude squared andtap magnitudes at two di�erent step sizes. Sample rate is 1 kHz. With step size0.03, algorithm has converged after approximately 200 ms.

23

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2−45

−40

−35

−30

−25

−20

−15

−10

−5

0

5

Step size

Res

idua

l err

or le

vel [

dB]

BroadbandNarrowband

Figure 5.3: Matlab simulation of LMS algorithm. Broadband and narrowbandresidual levels as a function of step size. Algorithm starts to severely diverge atstep size 0.14.

24

Chapter 6

Hardware and software

implementation

6.1 Hardware

The system was �rst implemented using a custom designed PCB containingall electronic parts (DSP, CODEC, �ash memory, ANR microphone andspeaker ampli�ers etc). After veri�cation, the hardware was re-designed to�t inside the headset cups. Debugging was done through the JTAG port ofthe DSP in combination with an asynchronous RS232 port connected to acustom written software running on a PC.

The main parts of the hardware are:

• DSP (TI TMS320C5509A)

• CODEC (TI AIC23a)

• Flash memory

• Serial port

• Microphone ampli�er

• Audio power ampli�er

6.2 DSP software

The DSP software was developed using Texas Instrument Code ComposerStudio, which is an Integrated Development Environment, including editor,compiler and debugger with JTAG support. All code is entirely written in Cfrom scratch, except for the device drivers for the serial port and the codec,which are from TI:s libraries.

25

Figure 6.1: PCB with DSP and CODEC.

The codec and its built-in anti-alias and reconstruction �lters are con�g-ured to run in 16 bit stereo mode at 8 kHz sample rate, see �gure 6.2. TheDSP is con�gured to read audio packets from the codec consisting of 8 stereosamples via DMA into RAM and then trigger an interrupt taking care of allaudio processing. Since the highest frequency of interest is considered to be400 Hz, the �rst processing step is to decimate the input audio. Accordingto Shannon's sampling theorem, the bandwidth of a sampled signal must beless than half the sampling frequency, so decimation by a factor 8 is suitablegiving a processing rate of 1 kHz. Due to the narrowband implementation ofthe ANR functionality, it is more suitable to implement the decimation �l-ters as band-pass �lters centered around each frequency of interest instead oflow-pass �lters (which is most common). The algorithm creates biquadratic(two poles, two zeros) �lters for each frequency component of interest and�lter the incoming audio signals before decimation. The decimation �ltercoe�cients are continuously updated to track the instantaneous frequencycomponents in the noise inside the headset cups. This approach helps theANR algorithms by �ltering out signal components that are far from thecenter frequency.

ANR processing is performed at a rate of 1 kHz and the taps are updatedon each iteration. Each frequency component has its own set of taps andstep size which makes it possible to let each component adapt at an optimalrate.

The anti-noise generation is implemented through sine table lookup witha phase accumulator for each frequency component. By looking up 8 con-

26

secutive samples per iteration we get the upsampling action �for free� andno interpolation �lters are needed (since we haven't actually made any up-sampling at all). Secondary path delay (phase angle) is looked up from theestimated and linearized response, as previously described, and added as ano�set to the phase accumulator. This o�set makes it possible to use a very�ne resolution of the secondary path delay and also saves memory space,since no old reference values need to be stored.

AD converter(CODEC)

ANR mics

Decimationfilters

Hardware

Analysis

FXLMS

sin/coslookup

8:1

Decimation Tapgains

sin

8:1

8:1

8:1

8:1

Software Hardware

ANR speakers

Software

8:1

8:1

8:1

8:1

8:1

DA converter(CODEC)

FXLMS

FXLMS

FXLMS

FXLMS

FXLMS

FXLMS

FXLMS

FXLMS

FXLMS

sin

sin

sin

sin

sin

sin

sin

sin

sin

Figure 6.2: DSP software �ow overview.

6.2.1 Secondary path estimation

Secondary path transfer function, including decimation bandpass �lters, wasestimated as described in the Theory section. Figure 6.3 shows the phaseresponse, which is also converted into corresponding group delay. Magnitudeis not shown here, since (as shown in equation 3.25) it is only the phaseresponse that a�ects the FXLMS algorithm stability. Since all frequencies ofinterest are below the resonance frequency of the speaker, the phase responsecan be adequately approximated by a linear function which is also shown in�gure 6.3 by the group delay being almost constant at 11 samples regardlessof frequency.

6.2.2 Step size and leak factor

To avoid excessive audible artifacts and ensure algorithm stability, the stepsize and leak factors described in the Theory part can not be constant val-ues for all frequencies and circumstances. Both parameters are thereforefunctions of several variables such as the current noise spectrum shape, per-ceived sound level and the present type of aircraft (single engine, twin engine,helicopter) carefully designed during in-�ight evaluations.

27

50 100 150 200 250 300 350 400−2000

−1500

−1000

−500

0

Pha

se a

ngle

[deg

ree]

50 100 150 200 250 300 350 4000

2

4

6

8

10

12

Frequency [Hz]

Gro

up d

elay

[sam

ples

]

Figure 6.3: Secondary path phase response and group delay.

6.3 Windows software

To be able to control, con�gure and monitor the DSP, a custom Windowssoftware was developed using Borland C++ Builder. The program uses theRS232 interface to communicate with the DSP using a custom protocol givingaccess to relevant DSP data structures and setting parameters in runtime,for example the step size and leak factors.

28

Figure 6.4: Windows software screen shot when connected to a headset during�ight monitoring various algorithm signals. Top-left is noise inside cups, top-rightis anti-noise output, middle-left is one set of taps.

29

Chapter 7

Results

7.1 Lab tests

An ambient sound consisting of �ve simultaneous frequency components (80,120, 160, 200, 240 Hz) with a total level of 86 dBSPL(A) was used for testin the sound-proof room. The sound level is in the range of that of a normalaircraft. As can be seen in �gure 7.1, the narrowband attenuation is 27 dBat 80 Hz dropping to 20 dB at 200 Hz and 11 dB at 240 Hz. The seeminglygreat number of harmonics over 200 Hz are due to the ambient loudspeakersand not the ANR processing. The 40 Hz peak is a sub-harmonic producedby the ambient loudspeakers.

The headset has no problems with �ve tones and the adaptation time isnot a�ected by the number of tones, since each tone is processed in parallelby the DSP and adapt at its own optimal rate. The overall adaptation timewas measured by turning the ambient sound on and o� and measure theoverall level. The o� time is long enough to let the algorithm bleed to zero.Result is shown in �gure 7.2 showing that the algorithm has converted 400ms after the ambient sound reached �nal level. The 50 ms slope starting attime zero is due to the audio generator ramping up the generated sound toavoid damaging the ambient loudspeakers.

7.2 Aircraft tests

Aircraft tests have been carried out in over 30 di�erent aircrafts, two of themwill be presented here.

7.2.1 Single engine aircraft

The aircraft shown in �gure 7.3 is a Piper PA-32 with a six cylinder engineand a two-bladed constant speed prop and is a typical aircraft where ANRheadsets are suitable. Figure 7.4 shows the sound pressure level with and

30

101

102

103

0

10

20

30

40

50

60

70

80

Frequency [Hz]

Sou

nd p

ress

ure

leve

l ins

ide

cup

[dB

SP

L]

No ANRWith ANR

Figure 7.1: Residual noise level inside headset cup with and without digital ANR.Measured on arti�cial head in sound-proof room with ambient sound consisting of�ve simultaneous frequency components with a total level of 86 dBSPL(A).

−0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.645

50

55

60

65

70

75

80

85

90

Time [s]

Sou

nd p

ress

ure

leve

l ins

ide

cup

[dB

SP

L]

Figure 7.2: Sound pressure level as a function of time to show ANR adaptationtime when exposed to ambient sound consisting of �ve simultaneous frequency com-ponents with a total level of 86 dBSPL(A). Sound is started at time zero, rampedup during 50 ms (to avoid damaging the ambient loudspeakers) and algorithm hasconverged fully after another 400 ms.

31

Figure 7.3: Piper PA-32, single engine airplane used for testing.

without ANR at the ear entrance when �ying at normal cruise speed andaltitude. Four major tonal components can be seen in the low-frequencyregion at approximately 40, 80, 120 and 160 Hz. The lowest frequency is at40 Hz, where the human hearing is not very sensitive so this component isleft out by the DSP algorithm. The 80 Hz component has largest amplitudeand gets attenuated by approximately 17 dB. The residual spectrum showsalmost no tonal components left below 400 Hz. The subjective impressionis that the engine noise is almost completely removed and the noise soundsmore as from a jet aircraft.

101

102

103

104

45

50

55

60

65

70

75

80

85

90

95

Sou

nd p

ress

ure

leve

l at e

ar e

ntra

nce

[dB

]

Frequency [Hz]

ANR offANR on

Figure 7.4: Sound pressure level at ear entrance in a Piper PA-32 airplane.

32

Figure 7.5: Cessna 310, twin engine airplane used for testing.

7.2.2 Twin engine aircraft

The Cessna 310 twin engine aircraft shown in �gure 7.5 was also used fortesting. The in-ear sound pressure level in �gure 7.6 shows a peak attenuationof 17 dB on the strongest component and 13 dB on the smaller peak locatedat approximately 150 Hz.

101

102

103

104

50

55

60

65

70

75

80

85

90

95

100

Sou

nd p

ress

ure

leve

l at e

ar e

ntra

nce

[dB

]

Frequency [Hz]

ANR offANR on

Figure 7.6: Sound pressure level at ear entrance in a Cessna 310.

33

Figure 7.7: Headsets in use during approach to Bromma airport in Stockholm.

34

Chapter 8

Conclusions

The results show that the system works well in the lab and also in �ight,ful�lling the performance requirements stated in the problem de�nition; Nar-rowband attenuation should be at least 15 dB, result shows 17 dB and �vesimultaneous frequency components are attenuated in the lab (no aircrafttests made with �ve major frequency components since no such aircraft hasbeen found yet). Adaptation time is less than one second.

35

Chapter 9

Discussion

The theoretical maximum step size given in equation 3.22 yields an upperbound of 0.08, while the simulations showed that the algorithm was stable(although generates artifacts) with step size up to 0.14. The reason for thisdiscrepancy is probably that the theoretical value is an approximation madefor wideband FXLMS systems, with up to hundreds of taps, instead of onlytwo taps as used in my algorithm.

The lab tests showed a peak attenuation of 27 dB while aircraft resultsshowed a peak attenuation of 17 dB (which nevertheless is within the spec-i�cations). This is due to vibration components in the aircraft, which getspicked up by the ANR microphones and converted into corresponding elec-trical signals, a�ecting algorithm performance. Great e�ort has been putinto mitigating these kind of issues in the �nal product. This is only one ofthe big di�erence between laboratory work and real-world implementations.

The adaptation time in the lab tests and simulations show a di�erenceof a factor two. This is due to that the simulations were performed on theFXLMS algorithm with leak factory set to unity, while the lab tests wereperformed with a leak factor of less than unity.

When it comes to consumer products, hard facts are not always theselling point. Regarding aviation headsets, many pilots put greater value on,for example, a comfortable �t or just that it looks nice instead of buyingit because it is the highest attenuating headset on the market. Also, someANR headsets may leave audible artifacts, such as hiss, high frequency noiseor pops, which may be irritating even though the attenuation itself is good.What this work has shown me is that the best way of designing or buyingan ANR headset is to try it out, along with competitor headsets, in realaircrafts.

36

Bibliography

[1] ANSI S1.42-2001 (R2006).

[2] Dennis R. Morgan Sen M. Kuo. Active Noise Control Systems: Al-

gorithms and DSP Implementations. Wiley-Interscience, 1996. ISBN0471134244.

[3] ANSI S3.19-1974.

[4] ISO 4869 2.2 (1992).

[5] Brian C.J. Moore. An introduction to the Psychology of Hearing. Else-vier, �fth edition, 2004. ISBN 012505628-1.

[6] Arbetsmiljöverket. Arbetsmiljöverkets författningssamling Buller AFS

2005:16. Publikationsservice, Solna, Sweden, 2005. ISBN 91-7930-455-9.

[7] A. David and S.J. Elliot. Numerical studies of actively generated quietzones. Applied Acoustics, 1993.

[8] Peter Rybing. Active noise control in home environment. Royal Institueof Technology, 2003.

[9] Sven Johansson. Active Control of Propeller-Induced Noise in Aircraft.Kasern tryckeriet AB, Karlskrona Sweden, 2000. ISBN 91-631-0172-6.

[10] Texas Instruments. Design of active noise control systems with theTMS320 family. focus.ti.com/lit/an/spra042/spra042.pdf, 1996.

[11] Texas Instruments. Tms320vc5509a �xed-point digital signal processordata manual. http://focus.ti.com/lit/ds/symlink/tms320vc5509a.pdf.

[12] Texas Instruments. TLV320AIC23b Data Manual.http://focus.ti.com/lit/ds/symlink/tlv320aic23b.pdf.

[13] S.J. Elliot and P.A. Nelson. Active noise control. IEEE Signal Process.

Mag, Oct 1993.

37

[14] Torkel Glad Lennart Ljung. Modellbygge och simulering. Studentlitter-atur, Lund, Sweden, 1991. ISBN 91-44-31871-5.

[15] EN24869-3 (1993).

[16] EN352-5.

38