IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 45, NO 1, FEBRUARY 1996 12 Measurement of Noise (Cross-) Power Spectra for Frequency-Domain Identification Purposes: Large-Sample Rik Pintelon, Member, IEEE, Patrick Guillaume, Member IEEE, and Johan Schoukens, Senior Member, IEEE Abstract-Several frequency-domain estimators of parametric transfer function models assume that the (cross-) power spectra of the disturbing noise sources are known [l].This paper presents a time-efficient method to measure these noise (cross-) power spectra and studies its influence on the asymptotic properties of the estimated model parameters. anti-alias filter I. INTRODUCTION HE basic experimental setup used in frequency-domain identification of parametric transfer function models is shown in Fig. 1. Since both the input and the output of the device under test (DUT) are measured, the appropriate stochastic description of the setup is an errors-in-variables model [2], [3] T Fig, 1. Experimental setup used in frequency-domain system identification. I disturbing noise sources N X (f ) , NY (f ) where X ( f ) is the true (unknown) input spectrum H ( f , P ) the parametric transfer function model, P the vector of the model parameters, and LPx (f),LPy ( f )the anti-alias filter characteristics. N y ( f ) and N x ( f ) are zero mean, jointly correlated, independent (over the frequency) complex random variables (see Appendix A for a definition of complex noise) which are related to the generator noise N,(f), the process and the measurement noise sources M x ( f ) , noise Np(f), M Y ( f ) as Several frequency-domain estimators of the model parameters P , such as the generalized total least squares [4], bootstrapped total least squares [5],the maximum likelihood [6] . . . (see are known at the frequencies of interest. Equation (3) is called a nonparametric noise model. Note that if LPx ( f )# LPy ( f ) then the measured output spectrum Y, (f ) , the variance & ( f ) and the covariance p ( f ) should be compensated by the ratio LPx( f ) / L P y( f ) in the following way: Ym(f) Ym(f)LPX(f)/LPY(f) + [l] for an overview), assume that the (co)variances of the Manuscript received October 26, 1994; revised May 9, 1995. This work was supported by the Belgian National Fund for Scientific Research, the Flemish cpmmunity (Concerted Action IMMI), and the Belgian government (IUAP 50). The authors are with Vrije Universiteit Brussel, Department ELEC, 1050 Brussels, Belgium. Publisher Item Identifier S 0018-9456(96)00847-9. P(f) + P(fWx(f)/LPy(f). (4) The ratio L P x ( f ) / L P y ( f )can easily be obtained via a relative calibration of the acquisition channels. Note that 0018-9456/96$05.00 0 1996 IEEE Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. 13 PINTELON et al.: MEASUREMENT OF NOISE (CROSS-) POWER SPECTRA systems identified in a feedback loop also fit within the stochastic framework (2) since it allows correlation between the input and output disturbances. In [2] and [ 3 ] ,the noise (cross-) power spectra are measured by removing the excitation signal ( X ( f ) = 0). During this noise measurement no information about the device under test (DUT) is gathered. The total measurement time equals the time necessary to measure the DUT ( X (f ) # 0) and the time required for the noise analysis ( X ( f ) = 0). In this paper, it will be shown that using periodic excitations ( X ( f ) # 0) it is possible to measure simultaneously the DUT and the noise (cross-) power spectra. Hence for a given measurement time the new procedure will acquire more information about the system (1) AND the noise model ( 3 ) than the former method. As it will be explained in the paper the new method consists of calculating the sample mean and the sample (co)variance of correlated (not stochastically independent) experiments. While the statistical properties of the sample mean and the sample (c0)variance of INDEPENDENT experiments are well known [ 7 ] , no results for the DEPENDENT (correlated) case are available. The contributions of the paper, hence, are 1) time-efficient measurement of noise (cross-) power spectra in a frequency-domain system identification context, 2) statistical analysis of the sample mean the sample (co)variance of DEPENDENT experiments, and 3) study of the influence of the measured noise model on the estimated model parameters. Note that throughout the paper it will be implicitly assumed that the disturbing noise is stationary. 11. MEASUREMENT OF THE NOISEPOWER SPECTRA B. New Approach The new approach consists of a statistical analysis of M periods of the input and output signals. The discrete Fourier transform (DFT) of eaclh period is calculated, resulting in M DEPENDENT samples y?)(f) (i = 1, 2 , . . . , M ) of the input and output spectra X m ( f ) , Y m ( f ) at the frequencies of interest. The sample (co)variances are used as estimates of the noise model (3) CONSECUTIVE ~-k)(f), . M a=1 . ( V $ ( f ) - m v ( f ) ) * (7) where V and W equal X and/or Y , and with m v ( f ) (V = X , Y ) the sample mean of respectively the input and the output spectrum DEPENDENT samples It is reasonable to assume here that the N $ ) ( f ) , N $ ) ( f ) (i = 1, 2 , . . . , M ) of the disturbing errors N X ( f ) ,N y ( f ) are jointly mixing of order 4 (see Appendix A, Definitions 3 and 4). Intuitively a mixing condition means that the span of dependence is 'limited', for example, the class of filtered white Gaussian noise sequencer; is mixing of infinite order. The mixing condition on the frequency-domain noise implies the same mixing condition on the time-domain noise [8]. The sample mean (8) of the M DEPENDENT samples is still an unbiased estimate of the true mean value V ( f ) E h " =V(f) (V = X,Y ) . (9) This is no longer true for the sample (co)variances (7) A. Classical Approach + E{cwv(f)} = a w v ( f ) O(1/M) (V, w = x,Y ) In the classical approach, the noise (cross-) power spectra are measured when no excitation is applied to the DUT ( X ( f )= 0). The measurement procedure consists of acquiring M time records of the input and output disturbances and calculating the sample (co)variances of the corresponding discrete Fourier spectra '(10) where O( 1/M) is a deterministic function that tends to zero as l/M for M -+ cc (proof: see Appendix B). Under the mixing condition of order 4 the sample (co)variances and the sample mean converge in the mean square sense to their expected value with a convergence rate of 1/J?i7 where V and W equal X and/or Y . Assuming that INDEPENDENT time records are gathered, the calculated noise Fourier spectra N $ ) ( f ) , N $ ) ( f ) (i = 1, 2 , . . . ,M ) are INDEPENDENT samples of the jointly correlated noise sources N X (f ), N y ( f) . If in addition the fourth-order moments of the noise exist, then it is well known that the sample (co)variances (5) are unbiased estimates which converge ( M cm)in the mean square sense (see Appendix A, Definition 5) to their expected value with a convergence rate of 1/d% [71 (proof for the sample mean: apply Theorem 1, Appendix C to (8); proof for the sample (co)variances: first note that a linear combination of mixing random variables, is mixing of the same order [8] and next apply Theorem 2, corollary to Theorem 2 and Theorem 1 to (7)). It is well known that the (co)variances of the sample means of M INDEPENDENT samples equal the (c0)Yariances of one sample divided by M [7]. This is not longer true if the samples (i = 1, 2, . . . , M ) are correlated. Terms in addition to cwv ( f ) appear which do not vanish for M large -+ ( O m s ( l / a ) is a random variable which converges in the mean square sense to zero as 1/A2 for M + 00). (V, w = x,Y ) Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. (12) IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 45, NO. 1, FEBRUARY 1996 14 with then the series (18) limited to the K first terms Here o g ; ) ( f ) is the (cross)-correlation between the noise samples IV$)(~), IV$)(~)(i = 1 , 2 , . . . , ~ ) is a consistent estimate of the series (13) limited to the K first terns K (k) k = 1, 2 , . . . , M - 1 o w v ( f )= E{N$+”(f)(N$)(f))*} ab$)(f) = E { N $ ) ( f ) ( N F + ” ( f ) ) * }i = 1, 2 , . . . ,M - k (14) (proof: see Appendix D). Note that o&$’ ( f ) = (o?; (f))’ . The sample (cross)-correlations are used to estimate the terms ( k = 1, 2 , . . . , M - 2) in the series (13) og;’(f) ~ (15) (cL-,f“)(f) = ( $ $ ( f ) ) * , and V, W = X, Y ) . Note that no &(M-1) useful value of cWv ( f )can be calculated (only one sample available). The calculation of the expected value and the mean square convergence proof of the sample (cross)-correlations (15) is done in exactly the same way as for the sample (co)variances (7). It follows that = &)(f) + 0(1/M) IC = 1, 2, ’ . . , M 2 (16) - and CF$)(f) = (V,W = X , Y ) .If (19) is not fulfilled then a trade off should be made between the bias and the variance of the truncated series b w v , ~ ( (1 f ) 5 K < M-1). The following estimates of the (co)variances of the sample means (12) are hence proposed 1 Cwv(f)= ~ ( C W V ( S ) b w v , ~ c ( f ) 1) I K < M - 1 (22) with + \ M-k (v$?k+z) ( f m v ( f ) )* E{cG,)(f)} k=l .G$)(f) + o m s ( l / m ) k=1,2,...,M-2 (17) (V,W = X , Y ) .Although c G $ ) ( f ) (IC fixed) is a consistent estimate of o$$) ( f ) ,the variance of the series (V,W = X , Y ) .In Section IV, it will be studied how many significant terms in each series (20) can be identified. Note that the results of this section also apply to the special case where one analyzes the disturbing noise during the dead time in between two consecutive bursts of the input and output signals. C. Alternatives To avoid the problems related to the correlation of neighboring signal periods, one could think of using every second or third period or of making repeated INDEPENDENT measurements. If the measurement time is not important then these are indeed the easiest solutions. However, if the goal is to gather as much information as possible concerning the device under test in a given measurement time, then these easy solutions will result in smaller signal-to-noise ratios. The total uncertainty on the estimated model parameters is a combination of the uncertainty on the sample means (8) and the uncertainty on the sample (co)variances (22). It turns out that the first effect (signal-to-noise ratio) is more important, and hence, the easy solutions will result in larger parameter uncertainties compared to the method proposed in Section.11. EI. c$F-~)(~) cF$)(f) = 0 + Q k >K (19) MEASURED NOISE MODEL ESTIMATED MODELPARAMETERS INFLUENCE OF THE ON THE does not decrease to zero as M tends to infinity. This can easily be verified by noticing that the variance of each of the Q last terms ( q = 2, 3 , . . . ,Q 1) in the series (18) is 0(1/ q ) , independent of M . Hence, bwv ( f ) (18) is an inconsistent (in the mean square sense) estimate of Pwv(f) (13). More restrictive noise assumptions are necessary to ensure consistency of the estimate (18). For example, if for IC sufficiently large the (cross-correlation) of the experiments is zero I Several frequency-domain estimators of the node1 parameters P , like the genralized total least squares (GTLS) [4], the bootstrapped total least squares (BTLS) [5],the iterative quadratic maximum likelihood (IQML) [l], and the maximum likelihood (ML) [2], [ 3 ] , require the knowledge of the noise model (3). Their asymptotic (number of spectral lines F t 00) properties (consistency, asymptotic efficiency, asymptotic normality) are studied assuming that the noise model is exactly known [l]. In this section, the influence of the measured noise model on the estimated model parameters is studied through the (equivalent) cost function L F ( . )of the identification method. Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. PINTELON et al.: MEASUREMENT OF NOISE (CROSS-) POWER SPECTRA 15 A. Classical Approach TABLE I In the classical approach, the measured noise model ( 5 ) is stochastic independent of the measured input and output spectra (1). Convergence in probability of the cost function with the measured noise model ( 5 ) to the cost function with the EXACT noise model (3) follows immediately: P1imM-w LF(P, V“, .wv(f)) = LF(P, Vm(f), .wv(f)) (24) where V , W equals X andlof Y (proof use the interchangeability of the probability limit and a continuous function, and the fact that convergence in the mean square sense (6) implies convergence in probability [9]). The convergence rate is O p ( l / m ) ( O p ( l / m ) is a random variable which converges in probability to zero as 1 / a for M -+ CO); the proof is analogous to that in Appendix E. Hence, the asymptotic ( F -+ CO) properties are still valid within an 0p(1/v‘Z) term. B. New Approach Here, the sample mean of the input and the output spectra (8), and the corrected sample (co)variances (22) are fed to the frequency-domain identification methods. The difficulty in the analysis of the (equivalent) cost functions is that the sample means are correlated with the sample (co)variances. Nevertheless, if (19) is true, convergence in probability of the cost function with the measured noise model (22) to the cost function with the (co)variance model (12) with a convergence rate of can be shown 1/m where V , W equals X and/or Y (proof see Appendix E). The extra scaling by M is necessary to prevent that the cost function blows up for increasing M . The Op(l/d%) term in the cost function (25) shows up in the asymptotic ( F + CO) properties of the estimators. If (19) is not true, then the (co)variance model (22) is a biased estimate of the (co)variance of the sample means (12). This bias will introduce a bias in the GTLS, BTLS and ML estimates of the model parameters P , which can be calculated explicitly if the true model belongs to the model set H ( f , P ) [lo]: it is a function of the model, the selected frequency band and the number of frequencies, and is proportional to a linear combination of the following noise-to-signal ratio terms .xf) + Pxx(f) .$(f> + ’ MlX(f)12 PYY(f) P ( f ) MlY(f)12 + PYX(f) ’ MY(f)X*(f) BOUNDS ON THE CORRELATION COEFFICIENT d L, 0.5000 1 0.7071 2 0.8090 3 0.8660 4 0.9001 5 AS A FUNCTION OF K Iv. STUDY OF THE CORRECTION TERMS(20) Simulations with colored continuous-time noise indicate that in most practical situations one correction term ( K = 1) in each series (20) is sufficient, while only in some extreme cases three terms ( K = 3) are required. In this section it will be shown how to estimate the number K of relevant correction terms in the series (20) for the case V = W . Assuming that assumption (19) is true, the question arises whether the expected value of the corrected variance (23) is guaranteed to be positive. For M :sufficiently large (23) (V = W ) can be written as with @ ’ ( f )the correlation coefficient @ ’ ( f )= real ( & ! ~ ) ) / c ~ v v ( f ) . (28) Since 6$’ E [-1, 11, it is clear that (27) can become negative, which is highly undesirable. What is the interpretation of this phenomenon? Consider thereto the case K = 1. In Appendix F it is shown for linearly filtered white noise that if assumption (19) is true for K = 1, then IS$’I < 0.5, which guarantees the non-negativeness of (27). Consequently, if IS$)I > 0.5 then, the hypothesis (19) with K = 1 is false, (27) can be negative, and K should be increased. Similar conclusions hold for the case K > 1 (see Table I). As a result, the value of 6 ; ) gives a lower bound on K ; e.g., if I@’l E ((0.7071,0.80901, then (27) should contain at least three correction terms ( K = 3). Another important aspect in the estimation of the number of relevant correction terms is the uncertainty of the estimated variance of the sample mean (22) (V = W ) .For M sufficiently large it can be written as with d r ) (f) the sample correlation coefficient * (26) From (26) it can be seen that the bias on the estimated model parameters due to the bias on the noise model tends to zero as 1/M. Since the noise on the sample mean tends to zero as 1 / a ( l l ) , it can be concluded that the asymptotic ( M + CO) bias on the model parameters can be neglected. d$)(f) = real (c$$(f))/cvv(f). (30) Exact calculation of the uncertainty bounds on the estimates cvv and d$) is very difficult (the experiments are NOT independent) and is out of the scope of this paper. One can however get an idea of the uncertainty by looking to the results Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. 16 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 45, NO 1, FEBRUARY 1996 TABLE II 68% AND 95% CONFIDENCE BOUNoS ON THE SAMPLE VARIANCE CI/v AND THE SAMPLE CORRELATION COEWICENT d v ) E [-0.5. 0.31 FOR INDEPENDENT SAMPLES AS A FUNCTION OF THE SAMPLE SIZE M ~ 10 f03 k0.5 100 fO.l f0.2 M.14 M.3 250 f0.05 33.1 33.09 33.17 1E3 f0.03 k0.05 fo.045 M.09 1E4 fO.O1 f0.02 fo.014 M.03 1E5 f0.003 k0.005 +_00.0044 H.009 i0.45 for INDEPENDENT experiments. While the uncertainty on the sample variance cvv is proportional to C V V , it tums out that the uncertainty on the sample correlation coefficient d:) is almost independent of d:) in the range [-0.5, 0.51 [7]. This means that a large number M of experiments are necessary to obtain an acceptable uncertainty on d$), especially for small values of the correlation coefficient (see Table 11). The total uncertainty on the quantity between brackets in (29) (1 + 2cf=’_, d c ) ( f ) ) is bounded above by 2KAd$). It can be used in combination with Table I1 to calculate an upper bound on the number of relevant correction terms K by imposing that 2 K A d t ’ << 1. For example, if the sample size M equals 250, then no more than K = 5 correction terms should be identified. The following correction strategy is hence proposed. First calculate the correlation terms d$) ( f ) , and the lower (Klower) and upper (Kupper)bounds on K (to be deduced from Tables I and 11). Choose a value of K << M in the interval [Klower, Kupper] such that CVV(f) 20 (V = X , Y ) Fig. 2. Experimental setup. systems. Since memory becomes cheaper the tendency is that this number will still increase so that the correction discussed in Section II-B will gain importance in the future. The correlation between the M samples { X k )( f ), Y i ’ ( f ); ‘I = 1, 2 , . . . , M } is only important IF the generator noise N g ( f ) and/or process noise N , ( f ) sources (see Fig. 1) are dominant over the usually white measurement noise sources Mx(f) and M y ( f ) , and IF the transfer function H ( f ) of the device under test exhibits large amplitude dynamics in the frequency band of interest. This is typically the case for vibrating mechanical structures. The method described in Section 11-B assumes accurate synchronization between excitation and measurement. Otherwise, due to a slight slip between the generation and acquisition clock frequencies, the variance of the discrete Fourier spectra will virtually be increased. This assumption is fulfilled for all measurement devices where the generation and acquisition clock frequencies are derived from the same reference clock, like for example dynamic signal analyzers. Most network analyzers and VXI-based measurement systems also allow measurement according to that principle. VI. EXPERIMENTAL VERIFICATION If such a value of K cannot be found, then no useful correction can be made and M should be increased. Under normal circumstances the correlation between the samples { X k ) ( f ) Y , i ) ( f ) :z = 1, 2 , . . . , M } is not that large (lPwv(f)I << a w v ( f ) )so that (31) is practically always fulfilled. V. SOMEPRACTICALCONSIDERATIONS Dozens of simulations with continuous-time noise indicate that an autocorrelation length of the continuous-time noise of 1/5th and 4/5th of the signal period results in, respectively, a 10% and 30% correction (20) of the sample (co)variance (22). Hence, several hundreds of signal periods (see Table 11) are typically necessary to make sensible estimates of the correction terms. Modern measurement devices already have the required amount of memory to measure these large numbers of signal periods; for example, 1E6 samples per channel in VXI-based We have chosen to measure a lowly damped system and to add generator noise to the excitation signal such that var(Ng(f)) >> var ( M v ( f ) )(V = X , Y ) .The experimental setup is shown in Fig. 2: the device under test is an aluminium plate (52.5cm x 2.5 cm x 2 mm) which is excited in its center by a mini shaker (B&K 4810). The excitation voltage v(t) of the shaker is the sum of a multisine source (Wavetek 75) and a random noise source (HP 3562A). The multisine has a flat amplitude spectrum and consists of 58 components equally distributed in the band [lo2 Hz, 267 Hz]. It is calculated using the time-frequency-domain swapping algorithm [ 111, [ 121. The applied force f ( t ) and the acceleration a ( t ) of the Al-plate are measured at the excitation point by an impedance head mounted on the shaker. Both signals are amplified and lowpass filtered (cutoff frequency of 300 Hz) before being fed to the acquisition unit. M = 243 consecutive periods ( N = 256 samples per period, f s = 741.053 Hz) of the input and output signals have been measured with a transient recorder (Nicolet 490). Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. 17 PINTELON et al.: MEASUREMENT OF NOISE (CROSS-) POWER SPECTRA 50 3 = n -10 5 c 50 -30 B -50 310 1 , , , , 50 ~~~~ 100 150 200 250 300 -50 , , , , 50 100 150 200 250 300 frequency (Hz) frequency (Hz) (a) (b) Fig. 3. Measured acceleration-to-force transfer function: (a) amplitude, and (b) phase. Fig. 5. Pole (x), zero (0)plots in the complex plane with the 68% confidence ellipsoids (dotted lines): (a) estimates with uncorrected sample (co)variances ( K = O), and (b) estimates with corrected sample r(co)variances(22) ( K = 1). TABLE I11 VALUE OF THE COST FUNCTION WITH 95% CONFIDENCE BOUNDS FOR TWOMODELCOMPLEX~IES value cost function model -1 L n = d = 2 n = d = 3 - 50 100 150 200 250 frequency (Hz) (a) 300 1 50 L ' ' 1, I I 100 150 200 250 frequency (Hz) " " " " " uncorrected sample corrected sample theoretical value cov. (22), K = 0 cov. ( 2 3 , K =: 1 (no model errors) 11.4 f 6.8 83.7 21 + 55 - 68.2 & 18 54 I 300 (b) Fig. 4. Sample correlation coefficients (30) (solid lines) d!$)(f) and d$)!f) of (a) the input errors, and the (b) output errors and the 95% uncertainty bounds (dashed lines). The mean values (8) and the sample (co)variances (22) are estimated using the M = 243 measurements (samples) of the input and output spectra. The measured frequency response function m y ( f ) / m x ( f ) is shown in Fig. 2. Following the strategy explained in Section IV it is found that only one correction term ( K = 1) should be added in the calculation of the sample (co)variances (22). Fig. 4 shows the corresponding estimated sample correlation coefficients (30). Taking into account the confidence bounds it follows that the correction of the sample variance of the input and output errors is significant respectively around 200 Hz and 150 Hz, which corresponds exactly to the frequency bands of the device under test with large amplitude dynamics. From Fig. 4 it can also be seen that values of &'(f) and d $ ) ( f ) smaller than -0.5 are included within the 95% confidence bounds. Hence with some probability the estimated sample variances (29) ( K = 1) can become negative. This has been experimentally confirmed by repeating the same experiment dozens of times. Adding a second correction term does not help here: since for this experiment Id@'(f)l << I d g ' ( f ) l (V = X , Y ) its only effect is to double the uncertainty of the expression between brackets in (29). For the experiments with negative sample variances (29) ( K = 2 ) the only sound solution is to increase M to 1000 to more (see Table 11). The parameters P of the rational transfer function model H(s, P) = ~ ~ , , . i ~ s k / / C ~ = , , & ares kidentified using a maximum likelihood estimator [3], [6], [ 121. The identification is done two times on the data set { m x ( f k ) ,m y ( f k ) :k = 1, 2, . . . , F } : once with the uncorrected sample (co)variances {.xX'(f.k)/A4,.YY(f/c)/M, . Y x ( f k > / J M : k = 1, 2 , * * . , F } , and once with the corrected sample (c0)variances { C x x ( f k ) , C y y ( f k ) , Cyx(f/c): k = 1, 2 , . . . , F ; k = 1). The results for the model ( n = d = 2 ) are shown in Fig. 5. It follows that the estimates of the parameters (or poles and zeros) in both cases are equal w&hin their uncertainty. The difference between the corresponding uncertainties and the value of the cost functions (see Table 111) is, however, significant. Using the uncorrected sample (co)variances one would wrongly conclude from Table 111 that the model order n = d = 2 is too large (overmodeling' ), while the results with the corrected sample (co)variances indicate that the model order n = d = 2 is too low (undermodeling' ). No significant model errors can be detected when the model complexity is increased to n = d = 3 (see Table 111). VU. CONCLUSIONS A time-efficient method to estimate (cross-) power spectra for frequency-domain identification purposes has been presented. It consists of measuring M consecutive periods of the input and output signals and requires a statistical analysis of M DEPENDENT samples (experiments) of jointly correlated complex random variables. For large sample values (A4 >> 100) the classical formulae to calculate the sample (co)variances must be corrected with terms which account 'Overmodeling means that the model not only fits the dynamics of the system but also tries to follow the noise. A cost function value smaller than its theoretical expected value indicates overmode ling. *Undermodeling means that the model is unable to fit all the dynamics of the system. A cost function value larger than its theoretical expected value is a strong indication for undermodeling. Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 45, NO. 1, FEBRUARY 1996 18 for the correlation between the experiments. These correction terms are important if the device under test has large amplitude dynamics and if the generator noise and/or the process noise are dominant over the measurement noise. No sensible correction can be made for small sample values (A4< 100) unless the interexperiment correlation is very high (absolute value correlation coefficient close to 1). These conclusions can be extended to any experiment where variance information about the mean value is required and where the coloring of the noise is important within the measured frequency band. The influence of the noise model on the estimated model parameters has been discussed for several estimators. Although the correction of the sample (co)variances does not change much the value of the estimated parameters, it may change a lot the calculated uncertainty bounds, the value of the cost function and the selected order qf the model. where (*) denotes that the complex conjugate can be taken or not at any place. Intuitively a mixing condition means that the span of dependence is 'small'. Example: white Gaussian noise passed through a stable linear time-invariant filter satisfies (34) with Q = 00. For stationary noise cum(x,(*I , (*) , x k(*p1+ , ) is independent of U and (34) reduces to c o c o co k z = l ks=l kp=l P n-m + Dejinition 1: z = x j y is a zero-mean complex random variable if E { z } = 0, E { x 2 }= E{y2} and E{xy} = 0. Property 1: Using definition 1 it can easily be verified that the variance E{lx12} = 2E{z2} = 2E{y2} and that E ( z 2 ) = 0. Definition 2: V = [ ~ 1 2 2 is] ~a zero-mean complex noise vector if z1 = x1 j y 1 and z2 = x2 jy2 are zeromean complex random variables and if E{xlxz} = E{yly2}, E{xlY2} = -E{y1z2}. Property 2: Using Dejinition 2, it can easily be verified that E { V V T } = 0 and that + 1, a , . . . , & . (35) Definition 5: Convergence in the mean square sense ([ 131, p. 56) is 1.i.m. x, = x APPENDIXA COMPLEXNOISE [61 = U lim E { ( x , - x)~}= 0. n-00 (36) APPENDIXB EXPECTEDVALUESAMPLE(CO)VARIANCE By definition of the mean value (33), and since the disturbances N $ ) ( f ) , N $ ) ( f ) have zero mean, it follows that + E { V V H }= [% $1 + with a? = E{/z1I2},ai = E{lz2I2}, p = E{zzzT} (complex covariance), superscript H the hermitian transposition operator, and * the complex conjugate. Definition 3 ([8],p. 19): The Qth-order cumulant cum(Ic1, z 2 , . . . ,Z Q ) is the coefficient of j Q t l t z . . . t~ in the Taylor series expansion of Zn(E{exp(j E:='=z,k t k ) } ) about the origin tl = t 2 = . . . = tQ = 0 (another equivalent definition can be found in [SI). The cumulants up to order 3 are identical to the central moments, for example, cum (xk) = E { x k } ~ kz ):, = cov (zk, xn). (V,W = X , Y). Since the noise samples N $ ) ( f ) , N $ ) ( f ) are jointly mixing of order 4, their joint second-order cumulants are absolutely summable (see Appendix A, Definition 4), and hence the second term in (38) tends to zero as l/M for M -+ Co. APPENDIXC THEOREMS cum ( Z k , xi) = var (xk) cum ( where V, W equal X and/or Y . Due to the multilinearity of the cum(.) operator ([XI, p. 19) and using the relationship V $ ) ( f )= V ( f ) N $ ) ( f ) ,(37) can be written as (33) Theorem 1: Take a complex random noise variable xk which is mixing of order 2, and define a sequence of complex numbers a k with ' d k E [l,001: l a k l < a < 00. Then the sum This is no longer true for higher order (>3) cumulants [7]. Definition 4: A random noise sequence xk is mixing of order Q if its cummulants up to order Q are absolutely summable: M (39) converges in the mean square sense to its expected value Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. 19 PINTELON et al.: MEASUREMENT OF NOISE (CROSS-) POWER SPECTRA with a convergence rate of 1/@ YM (41) = E { y M ) -k where O m s ( l / a ) denotes a random variable that tends to for M 4 CO. zero as Proof of Theorem 1: It is sufficient to show that the variance of YM tends to zero as 1/M for M ---f 00 (see Appendix A, Definition 5). By definition of the variance (33), and due to the multilinearity of the cum(.) operator ([8], p. 19), it follows that Each cumulant appearing in (44) 3s absolutely summable since z k and z k are jointly mixing of order 4. Hence, each term in (44) is absolutely summable since it consists of the multiplication of absolutely summable cumulants3. It can be concluded that l/m var ( Y M ) = cum ( Y M , = YR) --y2 u k u i cum M2 l M M (xk, zt). (42) k=l n=l The expression for the variance (42) can be bounded above by U2 5 -C M (43) with C a positive real number independent of M . The last inequality in (43) is true since the second-order cumulants of z k are absolutely summable (see Appendix A, Definition 4). Theorem 2: If two complex random noise variables x k and z k are jointly mixing of order 4, then their product yk = x k z k is mixing of order 2. Proof of Theorem 2: According to [15] (or 181, theorem 2.3.2, p. 21) the second-order cumulant of yk can be written as a sum of products of cumulants of z k and xk l lim M M+oo M F;Mlcum(yk, Y,)I < CO. (45) k = l n=l Corollary to Theorem 2: If Xk is mixing of order 4, then = x i and yk = 1x1~1~ are mixing of order 2 (proof replace Zk in theorem 2 by, respectively, xk and xi). yk APPENDIXD (CO)VARIANCE SAMPLEMEANS(9) Using the definition of the covariance (33) and the multilinearity property of the cum (.) operator ([8], p. 19) it follows that where V , W equal X and/or Y . Due to the stationarity of the noise the joint cumulant in (46) depends only on the difference (2 - n) P I Making the change of variables k = i - n and r = n, (46) can be rewritten as PROBABILITY APPENDIXE LIMITOF SOME COST FUNCTIONS Formula (25) will be proven for the maximum likelihood cost function. The proof for the generalized total least squares (GTLS), bootstrapped total least squares (BTLS) and iterative maximum likelihood (IQML) is analogous. Assuming that (19) is true, the maximum likelihood cost function based on the The series formed by the term-by-term multiplication of 2 absolutely convergent series is absolutely convergent [ 141. Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. 20 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 45, NO. 1, FEBRUARY 1996 sample mean (8) of the input and output spectra is given by [6] =-E 1 with n k a zero mean white noise sequence with variance The correlation coefficient of the process ~k is 02. F I m x ( f k ) N ( f k , PI-- m Y ( f k ) N f k l F k=l P)I2 (49) 4 f k ) with N ( f , P ) and D ( f , P ) , respectively, the numerator and denominator of the transfer function model H ( f , P ) , and a + P X X , K ( f ) ) l N ( f , P)I2 + ( O Y Y ( f )+ P Y Y , K ( f ) . ) l D ( f , m2 - area1 ( ( . Y X ( f ) + P Y X , K ( f ) ) D ( f , P ) N * ( f ,PI). Maximizing S(l) with respect to the coefficients a1 gives the results shown in Table I. a2, . . . aK m =(CXX(f) (50) Subtracting and adding the cost function (49) with the exact noise model from the cost function with the measured noise model (22) gives where Since the sample (co)variances ( M C X X( f ) M C y y ( f ) , M C y x ( f ) ) converge in the mean square sense to their expected value with a Convergence rate of 1 / a this is also true for s & ( f ) REFERENCES R. Pintelon, P. Guillaume, Y. Rolain, J. Schoukens, and H. Van hamme, “Parametric identificationof transfer functions in the frequency-domain, a survey,” IEEE Trans. Automat. Contr., vol. 39, no. 11, pp. 2245-2260, 1994. R. Pintelon and J. Schonkens, “Robust identification of transfer functions in the s- and z-domains,” IEEE Trans. Instrum. Meas., ,vol. 39, no. 4, pp. 565-573, 1990. J. Schoukens and R. Pintelon, Identijkation of Linear Systems: A Practical Guideline To Accurate Modeling. Oxford: Pergamon Press, 1991. J. Swevers, B. De Moor, and H. Van Brussel, “Stepped sine system identification errors-in-variables and the quotient singular value decomposition,” Mechanical Systems and Signal Processing, vol. 6, pp. 121-134, 1992. H. Van hamme and R. Pintelon, “Application of the bootstrapped total least squares @TLS) estimator in linear system identification,” Signal Processing VI: Theories and Applications, J. Vandewalle et al., Ed. Amsterdam: Elsevier, 1992, pp. 73 1-734. R. Pintelon, P. Guillaume, Y. Rolain, and F. Verbeyst, “Identification of linear systems captured in a feedback loop,” IEEE Trans. Instrum. Meas., vol. 41, no. 6, pp. 747-754, 1992. A. Stuart and J. K. Ord, Kendall’s Advanced Theory of Statistics, vol. 1. London: Charles Griffin, 1987. D. R. Brillinger, Time Series: Data Analysis and Theory. New York: McGraw-Hill, 1981. E. Lukacs, Stochastic Convergence. New York Academic, 1975. P. Guillaume, R. Pintelon, and J. Schoukens, “Robust parametric transfer function estimation using complex logarithmic frequency response data,” IEEE Trans.Automat. Contr., vol. 40, no. 7, pp. 1180-1 190, 1995. E. Van der Ouderaa, J. Schoukens, and J. Renneboog, “Peak factor minimization using a time-frequency-domain swapping algorithm,” ZEEE Trans. Instricm. Meas., vol. 37, no. 2, pp. 207-212, 1988. I. KollAr, Frequency-Domain System Ident@cation Toolbox for Use with Matlab. Natick, M A The Mathworks, 1994. A. H. Jazwinski, Stochastic Processes and Filtering Theory. London: Academic, 1970. T. J. I’a Bromwich, Introduction to the Theoly of InJnite Series. Loudon: MacMdlan & Co. Ltd., 1965. V. P. Leonov and A. N. Shiryaev, “On a Method of Calculation of Semi-Invariants,” Theory of Probability and its Applications, vol. IV, no. 3, pp. 319-329, 1959. (proof: apply the linearity property of the 1.i.m. operator [13]). It follows that the second term in the right-hand side of (51) is O p ( l / f i ) in probability (proof: convergence in the mean square sense implies convergence in probability [9], p. 33, theorem 2.2.2; and the probability limit and a continuous function may be interchanged [9], p. 42, Theorem 2.3.3). Rik Pintelon, (M’90) for a photograph and biography, see this issue, p. 11. APPENDIXF EXTREMEVALUES OF THE CORRELATION COEFF’ICENT Linearly filtered white noise satisfying (19) must be of the form Ek =nk + K amnkpm m=l (54) Patrick Guillaume, (S’86-M’87) for a photograph and biography, see this issue, p. 11. Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. PINTELON et al.: MEASUREMENT OF NOISE (CROSS-) POWER SPECTRA Johan Schoukens(M’9CSM’92) was born in Belgium in 1957. He received the degree of engineer in 1980, the degree of doctor in applied sciences in 1985 from the Vrije Universiteit Bmssel (VUB) and Brussels, Belgium. He is presently a Research Director of the National Fund for Scientific Research (NFWO) Brussels, Belgium, and part-time Lecturer at the VUB. The prime factors of his research are in the field of system identification for linear and non-linear systeins. Authorized licensed use limited to: Rik Pintelon. Downloaded on December 1, 2008 at 09:02 from IEEE Xplore. Restrictions apply. 21