Академический Документы
Профессиональный Документы
Культура Документы
50
'..1
I
'0
-" -.
..
a Signal, S; Noise, N; and Total, T b. Signal Est., Sm; Upper Bnd., Se+; Lower Snd., Se-
FIGURE 2. EFFECT OF MEASUREMENT UNCERTAINTY
hypothetical case of T known exactly and (T-N) known with =:3 dB uncertainty
erthanthespecifiedvalue-consideringthesymmetri-
cal, two-sided sha peof the sample mean dis tribution. In
many Situations, concern may be limited to the mean
valuebeingDOt morethan ornotless than some magni-
tude,i.e.,one-sided tests establishingonlyupperor
lowerbounds, respectively. Forexample, ifanupper-
bound limit is desired, then theappropriate confidence
intervalis}J. :5: X+ d
l
.
Manyproxcesses reSUlt insoundleveldistributions
whicharenon-Gaussian. However,the CentralLimit
Theoremin statistics allows thesamplingdistributionof
themean ofany random variable to be assumed normal
providedthesamplesize is sufficiently large. Theas-
sumptionofnormalityis reasonable in manycases for n
> 4 and quiteaccuratein mostcases forn > 10. (Ben-
datandPiersol)
PROPOSED CORRECTION SCHEME
Many acoustical measurementsituations involve ran-
clam processes, such as in environmental orcommunity
noiseevaluations. Thestatisticaltechniquesdescribed
abovearereadilyapplicableto these requirements. In
buildingorindustrial noise measurements, noise
sources areoften highly periodic-forexample, the
blade-passage frequency ofa fan, thehum ofan electri-
cal motor, orthe firing fate ofadiesel engine. For
theseevalua{jons,deterministictechniquesaremost
suitable. However, ifthe sourcesare operated under
fluctuatingloadingconditionsorarecombinedwith
numerousothernoisesourcesinan uncorrelatedman-
ner, theassumptionof a random process may become
applicable. Even suchapparentlysteadysinusoidal
noise sourcesas a motor-pumpset orfan may exhibit
low-magnitude-but significant-noiseemission fluctu-
ationssuchas dueto electrical-powerand pump-load
variationsfor thepumpsetorinflowturbulencetothe
fan. Furthermore, as noise propagation distances in-
crease,steadysource noiseemissionswill become modi-
fied by propagation conditions. In environmental
measurements, large amplitude fluctuations may be in-
trOduced by varying atmospheric conditions and turbu-
lence even overmoderatepropagation distances (e.g.,
fan blade-passagenoisehas been observed by theauthor
tofluctuateovera9-dB rangeata500-ftdistance).
Implementation ofa background noise correction re-
qUires repetitivemeasurementsof.hesourcesignalwith
background noise and ofthebackground noise alone to
increase confidence in the measured magnitUdes ofthe
soundpressurelevels Which defineT andN and to de-
termine thevariabilityassociated with their measure-
ment. Fromthevariabilities(andnumbersofmeas-
urements),confidenceinterval boundscan bedefined._
Thesource-signal magnitude at any instant, S., is the
"decibeldifference" in theT. and N. amplitudes, per
Equation{l}. Overtime,il c!.an ber ~ p r s n t by the
level ofthe ariLhmeric mean of"the antilogarithm differ-
ences (i.e.,theleveloftheexpectedSOurcesoundpres-
sures-squared). Confidenceintervalscanbederived
from the measurement means and standard
deviations-assumingboth the sourceofinterest and
the background noiseare uncorre13ted, random proc-
esses. Thedifficultythatarises is{hat T andN cannot
bath be measured simultaneously. However, jfthe
source and background arestationaryover the time pe-
riod spanned by the measurements, T and N can be
measured separately and the source magnitude, $, com-
puted.
ThemeasurementerrorsforT and Naresuch that
theyresullin upperand lowerbound estimatesofthe
signalwhich-whenexpressed in decibels-areasym-
3
TABLE 1. GAUSSIAN RANDOMNUMBERSTESTRESULTS
n 11EAN -ERROR- -c. I. for a - -ERR C.L for a OK
SIN Mean SD 95.0 99.0 99.5 95.0 99.0 99.5 TRlS
(dB) (dB) (dB) (dB) (dB) (dB) (%) (%) (%) (%)
6 10.1 0.5 2.0 2.0 2.9 3.3 75 91 94 100
30 10.0 0.0 0.9 0.9 1.2 1.4 81 91 91 100
100 10.1 0.1 0.5 0.5 0.7 0.8 91 94 94 100
6 0.0 -0.4 2.9 3.3 4.6 5.1 80 87 97 94
30 -0.1 0.1 1.1 1.5 2.1 2.3 91 97 97 100
100 0.1 -0.1 0.6 0.9 1.2 1.3 94 97 100 100
6 -9.8 5.1 5.9 6.5 8.2 8.8 36 55 59 69
30 -9.7 -0.6 6.7 7.2 8.4 8.7 67 67 71 75
100 -10.0 0.5 5.9 4.7 5.6 5.9 64 89 89 88
metric, i.e., the upper-bound eSlimate is more tightly
constrained than the lowerbound. This is illustrated in
Figures 2which shows the two-sided, upper- and lower-
boundestimates(Se+ and S .'respectively) ofasignal
forvariousmagnitudesofsignal-tO-noise ratio. Thus,
signalestimates may beobtained for (T- N)differences
lessthan4dB withrelatively confinedupper bounds.
Correction Procedure.A correction procedure is
proposedconsistingofthefollowingsteps:
1. Perform ndiscretemeasurementS ofsourcesig-
nal with background noise, T., and ofback
ground noisealone, N., (for atotal of2n meas-
urements). )
2. ArbitrarilypairT. and N. measurementS, (T.,
J J 1
N.).
I
3. COmpute thesound-pressures-squared (or each
measuremenl,
p 2 =lO(XiIlO). {7}
4. (11culatethe differences,
8. =p1j2- P
N
. . {8}
I I .
5. Kepe.at Steps2-4 for the nmeasurementpairs
anddeterminethemean (8) andstandard devia-
tion(5,s) overthemeasurements.
6. Estimatethesignalmagnitudeas
S=: 10log(8). {9}
7. Calculatethesignalestimateupperboundas
Sl-a: =10log(8+k1-o-s
S
) {l0}
where k} =t} (df)/n
Y2
.",ith t (dC) thevalue
1
ofStudent's-td'istribution for--<>"d[ degrees of
freedom and l-a probability(i.e.,confidence).
LImitations.Fundamentalto thisapproaCh is theas-
sumptionthatnotonlyarethesignaland noiseuocone
lated,but that themeasurements ofthebackground
noisewith and without thesourcesignal are uncorrelat-
ed. In most situationsofinterest, this premiseis rea-
sonablesincetheau!Ocorre)ation function for random
noisegoes to zero rapidlywith increasingintra-
measurement time delay. (Bendat and Piersol) Even
narrow frequency bands ofrandom noise may be ana-
lyzed provided the sampling interv'al between the meas
urementsissufficientlylong.
A requiredconditionis thestabilityofthebackground
noisemagnitude duringthe measurementS. Stationarity
ofthe backgroundsound pressurelevels can be tested
by performing background measurements before and
after thewith-source measurementS. Ifnecessary, be-
fore andafterbackground sound levels can be averaged,
or the computation aborted ifexcessive drift is ob-
served.
EVALUATION OF PROCEDURE
TheprOced-ure was tested bysimulatingits implemen--
tationin anumberofuials. Each tri.f1l consisted ofthe
repetitivesamplingofnoise and signal-with-noise six or
more times (toevaluate the effect ofsamplesize) and
(hecomputation oftheexpectedsignalvalueand upper-
bound estimate(forvariousconfidenceintervals). For
each trial,thesignal-estimationerrorwas determined by
subtractingtheactual signal magnitude from that ex-
pected using the correction procedure;and the actual
magnitudewas compared to the upper-bound estimates.
This process was repeated for 32 trials using random
numbers and eight trials using experimentalmeasure
ments.
Theresultsofthe trials wereexamined with respect (0
fourconsiderations:
Accuracy is defined in termsofthe meanerror
experiencedin eslimatingakno'W"l1 signaL
Consistency ofIhe estimates is quantified by the
standard deviationoftheerrors.
4
TABLE2. RAYLEIGH-LIKE RANDOMNUMBERSTESTRESULTS
n MEAN ---ERROR--- ---C.I. for a --- -ERR C.I. for a OK
SiN Mean SO 95.0 99.0 99.5 95.0 99.0 99.5 TRLS
(dB) (dB) (dB) (dB) (dB) (dB) (%) (%) (%) (%)
6 9.8 0.3 0.9 1.1 1.7 2.0 78 97 97 100
30 9.9 -0.2 0.5 0.5 0.7 0.8 97 100 100 100
100 10.0 -0.0 0.2 0.3 0.4 0.4 84 97 97 100
6 0.2 -0_2 1.3 1.7 2.5 2.9 84 94 97 100
30 -0.0 0.2 0.5 0.8 1.1 1.2 84 91 94 100
100 -0.0 -0.0 0.4 0.5 0.6 0.7 91 97 97 100
6 -9.7 1.6 5.6 6.2 7.8 8.4 54 62 77 81
30 -9.8 -0.0 3.1 4.4 5.5 5.8 77 91 95 69
100 -10.0 -0.8 2.9 3.4 4.2 4.5 84 87 94 _00
Assurance that theestimateupper-boundcon-
tains theactualmagnitudedeterminesthe
lidityoftheprocedure. It is quantified by the
fractionofthe trials forwhich theconfidence
intervalisgreaterthanorequaltotheerror,
Success is considered to existwhen the mean of
the measurements is greater
than orequal to the mean ofthe noise-alone
measurements. (Assignal-to-noiseratiode-
creases,theprobability increasesofa0) < 0 dB.
l!0 < adB, then thesignal cannotbe estimated
andthemeasurementtrialfails.)
In evaluatingthecorrectionmethod,theprocedurere-
sultSwereconsidered useful iftheassuranceandsuccess
wereatleast90%.
Random NumbersTests
GaussianDistribution.Thecorrectionprocedure
was tested using normallydistributed, random numbers.
Therandomnumberswereused tocreatesound-Ievel-
likemagnitudesby meansofa constant piUS theprodUCt
ofa dispersionfaCtor andarandomfunction. Thedis-
persionfactoryieldeda standarddeviationofapproxi-
mately3dB, in otherWOrds a differencebetween10th-
and 90th-percentilesound levels (L\O-L
90
) ofabout
8dB. [Experiencewithenvironmenta nOlse measure-
ments shows that L -L varies between about 4-
10 90
10 dBAwith 8 dBAtypical. Nearsteadily operatingfa-
cilities (e.g-, powerorprocess plants), this rangeis like-
ly to be much smaller.] From such random nlJmbers,
valuesforT and N were generated. Theconstantswere
selectedtoyielddesirednominalsignal-to-noiseratios
(S!l'l"). TestS weresimulatedtocorrespond to 6, 30, and
100 repetitive measurements with nominal signal-tO-
noiseratios of +10dB, 0 dB, and -10 dB. Foreach
numberofmeasurements,errorsand 95.0,99.0and
99.5% confidence inlerv'al bounds were computed over
32 trials.
TheresultS,summarizedin Table1,are:
Accuracy-meanerrormagnitudes typically
0.6dB;
Consistency-standard deviation oferrors <
3.0dBfor nominalSIN 2: adB and6-7 dB for
SIN ::::: -10dB;
intervalcontainserror
> 90%oftime for nominal SIN 2: 0 dB and
confidence 1-0: 2: 99%(except SIN ::::: 0 dB and
n =6), and confidenceintervalcontainserror
about90%oftimefor SIN =:; -10dBwith 1-0: 2:
99%and n=100;and
94%successful trials for nominal
SiN 2: 0dB andabout90%success ratefor SIN
<>:: -10dB.
From these testS, the correction procedurewas deter-
mined to provide useful estimates even for very poor
signal-to-noiseratios(i.e.,-10dB). Forhighsignal-to-
noiseratios(about +10dB),even six measurementS are
adequate; for poor signal-to-noise (about 0dB), at least
30measurementsarenecessary;and,forvery poorsig-
nal-to-noiseratios, 100 measurementSand 99%confi-
dence intervals are needed to obtain marginally accept-
ableconsistencyandassurance.
Rayleigh-Like Distribution.Environmentalsound
levelsoftendo not distributenormallybutexhibita
Rayleigh distribution. Suchadistributionispositively
skewedwith 3 finite lower bound but an asymptotic
upper bound. A rna thema IicaI expression was de-
veloped to generate random numbers eXhibiting a Ray-
leigh-likedistribution. When randomnumberswith
sound-Level-like magnitudes were createdwith these-
lectionorasuitableconstant,theresultantdistributions
5
"
'"
.,
I'.
-----,
' ' I
---I
,""
,
,I
.,
10 11 'I: u ,. .,
"'''
1;:-
--
-,-
a. 15-mintimehistory b. 2-min segmentof"l.vhitenoise"excitation
FIGURE3. EXPERMENTALMEASUREMENTS-SPEAKER-EXCITATIONTIME HISTORIES
TABLE 3. SAMPLING ERROR IN ABSENCE OF
BACKGROUNDNOISE(overeighttrials)
n ----ERROR (dB) for SAMPLING TYPE---
-----Random-----
mean SO mean SD
6 1.8 3.8
10 1.4 2.7
30 0.1 1.9
56 0.2 1.1
-0.1 1.6
-0.2 1.5
-0.0 0.5
+0.2 0.4
yieldedstandarddeviationsofabout 1.6dB (ifnormally
distributed,L -L :::: 4dB). Asfor theGaussian raD-
10 90
dam numbers,valuesofTand Nweregenerated to );eJd
nominal +10dB, 0dB, and 10dB signal-to-noise ratios
for testswith6,30,and100repetitivemeasurements.
The results,summarized in Table2,are:
Accuracy-mean errormagnitudes typically :s
0.8dB;
Consistency-standarddeviationoferrorsgen-
erallyless thanabout3dB;
Assurance-> 90%for nominal SIN 0 dB
and greater thanabout 90%for SiN <:::: 10dB
withl-a 2: 99%andn;:: 30;and
Success-100% forall butSiN =:: 10 dB and n
530.
Theresultsfrom these testsagain showed the proce.
dureto pro\-;deuseful estimates. Theaccuracy'with the
non-Gaussian variables was about the same as with the
Gaussian-distributedvariables. Theconsistencyofthe
results was greater and the confidence intervals smaller
withtheRayleigh-likedistribution-probablyas a result
oftbereducedvariabilit:yofthetest parameters. In
spiteofthe tighterconfidenceintervals, theassurance
remained high. Foreven poorsignal-la-noiseratios
(approx.adB), as few as six measurementsweresuffi
cient;while,forverypoorsignal-la-noise,100meas-
urements and 99% confidence intervals appear to be
necessary.
ExperimentaIMeasurements
Furtheranalyses ofthecorrectionprocedure were de-
signed based upon the results ofthe random numbers
tests. Thesetests were performed indoorsand consisted
oftwo loudspeakersand acentrally placed microphone.
Speaker A was excited with filtered random noise
(Standard deviation, approx. 0.7 dB; L L approx.
10 90
,
2dB),as shown in Figure3a. Theexcitatlon inuoduced-
toSpeakerB, as alsoshown in Figure3a,was oneoftwo
approximatelyIS-min. recordingsoftrafficnoisemade
at about 100 ft from theedgeofavery busy freeway
(standard deviation, approx. 2.3 dB; L -L
9Q
, approx..
10
6dB). (As can beseen from inspection ofthe white-
noise time history, that excitation was neither perfectly
stationary throughout the measurements nor truly ran
dom. Closerexamination ofa2-min. interval in Figure
3b shows theexcitationin fact to be quasi-random wilh
aperiodicityofabout 4/min.) Thegains ofthe loud-
speaker amplifiers were such that.nominal
+10-dB, QdB, and -10-dB signal-to-noiseratios were
obtained. All ofthe speaker outputs were measured 3t
I-sec intervals over(!bou{ 15 min. for a total ofapprox
imately 900 measurements in terms of i-sec equivalenr
sound level, L> (1 s), values. A, A + Ban.d Bwere
measured for e1h test condition with Aand BarbiWlri
Iy definedassignalornoise, Graphicalmethodscan be
used toshowboth excitationsto beverynearlyGausian,
6
TABLE 4. PERIODICALLY SAMPLED MEASUREMENTS TEST RESULTS
n MEAN -ERROR- ---e. 1. for a - -ERR .:: C. I. for a OK
SiN Mean SD 95.0 99.0 99.5 95.0 99.0 99.5 TRLS
(dB) (dB) (dB) (dB) (dB) (dB) (%) (%) (%) (%)
6 11. 6 1.8 3.2 1.2 1.8 2.1 38 38 38 100
10 12.0 1.3 2.6 1.0 1.5 1.7 50 63 63 100
30 12.6 0.0 1.9 0.6 0.9 1.0 75 75 75 100
56 12.7 0.2 1.1 0.5 0.8 0_8 75 75 75 100
6 1.2 1.5 3.8 1.3 1.9 2.2 38 38 63 100
10 1.6 1.4 2.7 1.1 1.5 1.7 50 63 63 100
30 2.3 0.1 2.2 0.7 0.9 1.0 63 63 63 100
56 2.4 0.3 1.3 0.5 0.7 0.8 63 63 63 100
6 -8.8 2.3
3_9 3.0 4.2 4_7
63 75 75 100
10 -8.5 1.4 2.7 2.7 3.6 3.9 63 63 88 100
30 -7.8 0.5 2.3 1.7 2.2 2.4 63 75 75 100
56 -7.7 0.7 1.2 1.2 1.6 1.8 63 75 75 100
The measured data were sampled such that effeCtively
trials of 6, 10,30, and 56 L (1 s) measurements were
obtained for each noise condition.
The sampling was done both periodically (every 2 sec)
and randomly. For each number of measurements, er-
rors and 95.0, 99.0 and 99.5% con fidence interval
bounds over 8 trials were computed.
(The two traffic noise recordings were employed such
that one recording consistently was added to the white-
noise and the other recording consistently represemed
the excitation-alone, "actual"-magnitude measuremem.
[The separate recordings were used to preclude possible
false favorable results that may have arisen from com-
paring an excitation to itself.] The error associated v.ith
the sampling alone, i.e., aside from the presence of
background noise, can be determined by estimating the
magnitude of one recording from the other using the
correction procedure with infinitesimal background
noise. The results of this comparison are given in Ta-
ble 3. With periodic sampling, the error is relatively
large [> 1 dB] for n :s 10; with random sampling the
error is small 0.2 dB] for all n.)
The resultant tests had actual signal-lo-noise ratios
ranging from -9 to +13 dB, relatively high compared to
the nominal magnitudes. The differences between the
nominal and actual SIN ratios are the result of the man-
ual setting of amplifier gains.
Periodic Sampling. The results, summarized in Ta-
ble 4, are:
Accuracy-mean error magnitudes =:: 1.4 dB for
n 10 and about 2 dB for n =6;
Consiscency-standard deviation of errors <
3 dB for n 10 and 3-4 dB for n =6;
Assurance-<:onfidence interval contains error
:S 75% of lhe time for nearly all cases; and
Success-all tria15 were successful.
While the accuracy, consistency and success of these
results are quite good, the assurance performance is
unacceptable since at least 2.5% of the trials resulted in
errors exceeding the confidence interval. Funhermore,
the assurance did not improve significantly as the num-
bers of measurementS increased from 6 to 56. The poor
performance of the correCtion procedure may be the
result of twO factOrs-the periodicity exhibited by the
'\vhite noise" excitatio nand the possibility that the
highway noise samples taken at 2-s intervals may be at...
least panially correlated. In fact. comparison of
Tables 3 and 4 shows the periodically sampled results to
be strongly influenced by sampling error. Therefore, at
least for quasi-random and/or highway-noise-like proc-
esses such as those tested, a modified sampling scheme
is necessary. random sampling of the stored los-interval
Leq values. sampling with intervals significamly greater
than 2 s, and/or greatly increased numbers of measure-
ments.
Random Sampling. Random sampling of the stOred
L values was evaluated with the same numbers of
The results, summarized in Table 5, are:
Accuracy-mean error magnitudes :s 1.2 dB for
all cases (and :s 0.5 dB for n 30);
Cons iSlency-sta nda rd devia tio n 0 f er ro rs
:s 1.2 dB for nominal SiN 0 dB, and for SIN
:::: -10 dB with n 30, but about 3 dB for SIN =::
-10 dB with n :S 10;
7
TABLE 5. RANDOMLY SAMPLED MEASUREMENTS TEST RESULTS
n MEAN ---RROR--- ---e.I. for a --- -ERR C.l. for a OK
S!N Mean SO 95.0 99.0 99.5 95.0 99.0 99.5 TRlS
(dB) (dB) (dB) (dB) (dB) (dB) (%) (%) (%) (%)
5 12.2 0.7 1.4 1.5 2.3 2.6 75 88 100 100
10 13.0 -0.1 1.6 1.1 1.6 1.8 75 100 100 100
30 12.9 0.1 0.6 0.7 LO 1.1 88 88 88 100
56 12.7 -0.1 0.3 0.5 0.7 0.7 100 100 100 100
6 1.8 1.2 1.4 1.6 2.4 2.7 50 88 88 100
10 2.5 0.1 1.4 1.1 1.6 1.8 63 100 100 100
30 2.3 0.2 0.7 0.7 1.0 1.1 75 88 88 100
56 2.3 0.1 0.4 0.6 0.8 0.9 88 88 88 100
6 -8.0 0.7 3.5 3.3 4.6 5.1 63 63 75 100
10 -7.3 0.9 2.7 2.2 3.0 3.3 63 75 75 100
30 -7.6 0.5 1.5 1.7 2.3 2.5 75 75 75 100
56 -7.7 0.5 0.5 1.2 1.6 1.7 75 100 100 100
Assurance--amfidence interval contains error
88%of the timefor nominal SIN 2': 0dB with
I-a2': 99%, and 100% ofthe timefor SIN =:<
-10dB withn =56andl-a 2': 99%.
Success-all trialsweresuccessful.
Randomizedsamplinggreatlyimproved theassurance
oftheprOceduresuch thai the correction method again
producedusefulestimates,givensufficient numbersof
measurements. With I-a 2': 99%, as few as six meas-
urementsgavegoodresults for SIN > 0dB. However,
for very poorsignaltonoiseconditions,a substantial
numberofmeasurementsarerequired,n 2': 56.
Discussion.TheexperimentallestSwereperformed
usingmeasurements of1-s equivaleol sound levels.
Equivalentsound levelsofotherdurations couldhave
been used-or for that mauer other noise
descriptors-providedthe"total"and"noise"measure-
mentsareofacommondescriptOrand duration and the
measurement results can be input validly into
Equation {I}. Certainly meeting the sampling require
ments for somedescriptors, e.g" L (l hr), would be
laborious and other as stationarity
ofthevariables--difficulttomeet.
Thefavorable resultsoftheexperimentaltests may
have beenenhanced by the normallydistributed
ler ofthe constituent processes. However, as indicated
by therandomnumberleSts, good resullscanbe ob-
tainedwithnon-Gaussianparameters. Theusefulness
ofthis procedurefor stronglyskewed and/or multi
modaldistributions must bedemonstrated. However, it
is speculated that useful resulls will be obtainablegiven
sufficientsamplesizeandsuitablesamplingtechnique.
CONCLUSIONS
Corrections for High-Noise Measurements. The
correctionprocedurewas found to give useful results
for poor (SIN =:< 0 dB) and very poor (SIN "':: -10 dB)
signal-tO-noise ratios provided thenumbers ofmeas-
urements areadequate. In Table 6, the minimum re-
quired numbersofmeasurementsobservedin tbefOUf
testS ofthe procedurearesummarized. Forall evaluat
edparameterdistributions,variabilities, and other tirne-
historyCharacteristics-acceptableresults(I.e.)with
assuranceandsuccess equalingor exceeding 90%) were
obtained with some measurement scheme. The re-
quired numberofmeasurementsdecreased with de-
creasingparametervariabillty-toas few as 6 [orpoor
signal-la-noise and as few as 30 for very poorsignaltO
noise-even for non-Gaussian parameter distributions.
Usingwiderconfidence intervals (i.e., higher confidence
estimates),reduces the minimum Dumberofmeasure
ments required; however, a more preciseintervalmay
beworth theeXira measurements(e.g.,forGaussian
numbers and SIN :::: 0 dB-as I-a goes from 95.0 ......
99.5%, n . =30 ->0 6but theconfidenceintervalgoes
IDln .
from 1.5 - 5.1 dB). Forsomerequirements,acceptable
results may beobtained onlywjth a suitablesampling
scheme,suchasrandomsampling.
Confidence Intervals for LowNoise Measure-
ments. For low-noisecondilions (i.e., SIN :::: +10 dB),
theseresultsareprimarily useful inshowingtheneed
for samplingand confidenceintervals sinceexisting
standardsdo nOt address thevariability ofthe measured
parameters and the sampling requirements for
them-acceptingasinglemeasurementofeachas ade
8
-10
TABLE 6. OBSERVED MINIMUM REQUIRED NUMBER OF MEASUREMENTS
for acceptable accuracy, consistency, assurance, and success
TEST CONFIDENCE -REQUIRED n. for SiN (dB)- -e.!. (dB) for n, and SiN (dB)-
mIn mln
(l-a)% +10 0 -10 +10 0
Gaussian Random Numbers 95.0 100 30 0.5 1.5
(LI0 - L90 = 8 dB) 99.0 6 30 >100 2.9 2.1 =5.6
99.5 6 6 >100 3.3 5.1 =5.9
.,
Rayleigh-Like Random Numbers 95.0
30+
100 0.5 0.5
(LIO - L90 = 4 dB) 99.0 6 6 100 1.7 2.5 4.2
99.5 6 6 100 2.0 2.9 4.5
., ., ., ., .,
*
Periodically Sampled Measmts. 95.0
., .,
(LID - L90 = 2 and 6 dB) 99.0
.,
99.5
Randomly Sampled Measmts. 95.0 56
.,
0.5
(LID - L90 = 2 and 6 dB) 99.0 10" 10" 56 1.6 1.6 1.6
99.5 6" 10" 56 2.6 1.8 1-7
no acceptable number '" poorer performance obtained with greater n
TABLE 7. CONFIDENCE OF SIGNAL ESTIMATES WITH LOW BACKGROUND NOISE
SIN::::: +10 dB, n = 6 measurements
--------------TST------------ ----ERROR (dB)- CONFIDENCE INTERVAL (dB) for a
Mean S.D. 95.0 99_0 99.5
Gaussian Random Numbers 0.5 2.0 2.9 3.3
Rayleigh-Like Random Numbers 0.3 0.9 1.7 2.0
Periodically Sampled Measmts. 1.8 3.2
Randomly Sampled Measmts. 0.7 1.4 2.6
., assurance inadequate
quate. Appropriate confidence intervals for the low
background noise tests with six measurements aIe sum-
marized in Table 7. For six measurements and an ade-
quate sampling technique (e.g., random sampling if nec-
essary), the prediction error was about 0.5 dB with up-
per-bound confidence intervals about 3 dB. A lO-dB
SIN ratio is well within that permitted by all measure-
ment standards. Oearly, even in these favorable condi-
tions, multiple measurements are desirable.
RECOMMENDAnONS
The correction procedure described above appears to
give useful results for a number of realistic measure-
ment conditions. The procedure should be used in all
measurements with background noise to gain more ex-
perience with its use so that sampling requirementS can
be specified for measurement conditions not tested in-
this paper. Ultimately, the proposed seven-step correc-
tion procedure should be incorporated into field meas-
urement standards.
More tests are necessary with different parameter
variations especially those which are highly variable
and/or otherwise ill-behaved, e.g., bi-modal distr..ibu-
tions. In addition, sampling guidelines for Ihe imple-
mentation of this procedure need to be defined furlher:
the required numbers of measurements, appropriate
sampling intervals or schemes (i.e., random \IS. periodic
sampling), and the required confidence values appro-
priate for various measurements.
REFERENCES
American National Standards Institute, American Na-
tional Standard Methods for the Measurement of Sound
Pressure Levels, ANSI S 1.13-1971 (R 1986).
Hassall, J.R, "Noise Measurement Techniques:
Chapter 9, Handbook of Acoustical Measurement in
Noise Control-Third Edition, (Ed. eM. Harris),
McGraw Hill, Inc., 1991.
Bendat, J.S. and AG. Piersol, Random Data: Analvsis
and Measurement Procedure, WiJey-Imerscience, 1971.
Dixon, W.J. and F.J. Massey, Jr., Introduction to S13-
tistical Analvsis (Fourth Ed.), McGraw-Hili, Inc., 1983.
9