sine wave fit algorithm

Transcription

1 TECHNICAL REPORT IR-S3-SB-9 1 Properties of the IEEE-STD-57 four parameter sine wave fit algorithm Peter Händel, Senior Member, IEEE Abstract The IEEE Standard 57 (IEEE-STD-57) provides algorithms for fitting the parameters of a sine wave to noisy discrete time observations. The fit is obtained as an approximate minimizer of the sum of squared errors, i.e. the difference between observations and model output. The contributions of this paper include a comparison of the performance of the four-parameter algorithm in the standard with the Cramér-Rao lower bound on accuracy, and with the performance of a nonlinear least squares approach. It is shown that the algorithm of IEEE-STD-57 provides accurate estimates for Gaussian and quantization noise. In the Gaussian scenario it provides estimates with performance close to the derived lower bound. In severe conditions with noisy data covering only a fraction of a period, however, it is shown to have inferior performance compared with a one-dimensional search of a concentrated cost function. Keywords IEEE standards, Measurement standards, Signal analysis, Frequency estimation, Electronic equipment testing I. Introduction In testing digital waveform recorders, an important part is to fit a sinusoidal model to recorded data, and to calculate the parameters that result in the best fit (in leastsquares). Several algorithms were evaluated in [1], [2], and algorithms have been standardized in IEEE Standard 57 [3]. A survey of the standard can be found in [4]. The standard [3] was prepared by a working group that is a part of the Waveform Measurement and Analysis Technical Committee (TC-) of the IEEE Instrumentation and Measurement Society. TC- is currently working on a standard for analog-to-digital converters (IEEE-STD-1241) [5]. In IEEE-STD-1241, test methods for signal-to-noise and distortion ratio (SINAD) and effective number of bits (ENOB) rely on the sine wave fit [3]. In this paper, the 4-parameter sine wave fit algorithm of [3, ] is studied in some detail. The considered estimation problem is nonlinear with respect to the parameters, and thus the 4-parameter sine fit algorithm is an iterative method where in each iteration an updated frequency estimate is obtained based on estimated parameters. The algorithm does not exploit the fact that, although it is a nonlinear optimization problem, it is linear in three out of four parameters. Thus, it can be expected that the algorithm of [3] has inferior performance compared to algorithms that utilize such a fact. In particular, one can expect that it may suffer from ill-convergence when the digital waveform recorder utilizes course quantization. In the The author is with the Department of Signals, Sensors and Systems, Royal Institute of Technology, SE- 44 Stockholm, Sweden. This work was supported in part by the Junior Individual Grant Program of the Swedish Foundation for Strategic Research. ( ph@s3.kth.se). worst case scenario with 1-bit quantization the magnitude of the sine wave amplitude is not observable from data, and thus ill-convergence may occur. The recommended procedure to obtain initial estimates virtually eliminates convergence problems in most practical cases [4]. The main interest in this paper is to study the properties of the obtained frequency estimate. Further, we are interested in the small error properties, i.e. performance of the algorithm when it is properly initialized with initial values close to the correct ones. The contributions of this paper include an alternative derivation of the 4-parameter matrix algorithm [3, ] in Section III. In Section IV, an alternative nonlinear least-squares algorithm with improved convergence properties is derived. This algorithm utilizes the fact that the error criterion that is minimized can be concentrated with respect to three of the parameters. A theoretical performance assessment is performed in Section V, where the Cramér-Rao bound for this estimation problem is derived when the additive noise is Gaussian. In Section VI, some illustrative numerical examples are presented. The conclusions are drawn in Section VII. II. Measurements and data model Assume that the data record contains the sequence of samples y 1,..., y N (1) taken at time instants t 1,..., t n,..., t N. It is further assumed that data can be modeled by y n [A, B, C, ω] = A cos ωt n + B sin ωt n + C (2) where A, B are C are unknown constants. The angular frequency ω may be known, or not, leading to models with three or four parameters, respectively. Equivalently, we may write y n [θ] in (2) as y n [θ] = α sin(ωt n + φ) + C where A = α sin φ, B = α cos φ and θ is a vector of unknown parameters. Stressing the dependence of y n [θ] on the generic parameter vector θ turns out to be convenient for the following discussion. III. Algorithms of IEEE-STD-57 The Standard 57 provides algorithms both for 3- parameter (known frequency) and 4-parameter (unknown frequency) models. For easy reference, the 3-parameter algorithm is reviewed below [3, ]. It is exploited by the nonlinear least squares fit in Section IV. In Section III-B, a derivation of the 4-parameter algorithm in [3, ] is provided, a derivation that differs from the one in [3]. Thus, it is believed to be informative.

2 2 TECHNICAL REPORT IR-S3-SB-9 A. Known frequency When the frequency ω is known, estimates of the unknown parameters in θ = x = (A, B, C) T (where T denotes transpose) are obtained by a least squares fit. Gather the data record in y, y = (y 1,..., y N ) T. (3) Then, y obeys the linear set of equations y = Dx (4) where D is the N 3 matrix cos ωt 1 sin ωt 1 1 cos ωt 2 sin ωt 2 1 D =.... (5) cos ωt N sin ωt N 1 Equation (4) is an overdetermined (i.e. N > 3) set of linear equations, with the least-squares solution ˆx (in general, ˆ denotes an estimate) given by [3] B. Unknown frequency ˆx = (D T D) 1 D T y. (6) Assuming that an estimate in iteration i, say ˆω i, of ω is available, a Taylor series expansion around the estimate ˆω i gives and cos ωt n cos ˆω i t n t n sin ˆω i t n ω i (7) sin ωt n sin ˆω i t n + t n cos ˆω i t n ω i (8) where ω i = ω ˆω i. Inserting (7)-(8) into (2) gives y n [θ] A cos ˆω i t n + B sin ˆω i t n + C At n ω i sin ˆω i t n +Bt n ω i cos ˆω i t n (9) where θ = x i is the parameter vector x i = (A, B, C, ω i ) T. () Equation (9) is still nonlinear in the parameters, but may be linearized using the observation that ω i. Putting available estimates of A and B from previous iteration, i.e. Â i 1 and ˆB i 1, in place of the unknown parameters in the two last terms in (9) results in an equation linear in the components of x i. Gathering the data record in y gives, similarly as in (4) y = ˆD i x i (11) TABLE I IEEE-STD-57 for four parameter least squares fit to sine wave data using matrix operations. a) ˆx = (A, B, C, ) T, ˆω b) next iteration i = i + 1 c) ˆω i = ˆω i 1 + ˆω i 1 d) create ˆD i, see (12) e) ˆx i = ( ˆD T i ˆD i ) 1 ˆDT i y f) optional [3] g) repeat b)-f) until convergence The basic idea behind the algorithm in [3, ] is repeatedly to solve the linear system (11), i.e. at iteration i use ˆD i in order to obtain a new set of estimates ˆx i. The algorithm is summarized in Table I. The initial estimates ˆx and ˆω are, for example, obtained by peak-picking the Discrete Fourier Transform (DFT) of data, followed by a prefit using the algorithm in Section III-A. Alternative methods for finding initial estimates are discussed in [1]. The algorithm in Table I is an iterative process to find the parameters that minimize the sum of squared differences, i.e. N (y n y n [θ]) 2. (13) In (13), y n [θ] is given by (2). In each iteration, an updated frequency estimate ˆω i is obtained based on ˆω i 1 and ˆω i 1, estimates that also depend on estimated values of A and B. The recommended procedure to obtain initial estimates virtually eliminates convergence problems in most practical cases [4]. However, convergence problems may occur, especially for short noisy data records or signals at low frequency (see, Section VI). An alternative to the 4- parameter algorithm in [3] is derived below. IV. Nonlinear least-squares by grid search Using (4)-(5) the criterion (13) can be rewritten as (y Dx) T (y Dx) (14) where x = (A, B, C) T, and D is defined in (5). The criterion (14) can be concentrated with respect to x, and ω can be found by a one-dimensional search for the maximum of g(ω), i.e. [6], [7] ˆω = arg max g(ω) (15) ω where ˆD i the N 4 augmented D matrix with an extra column ˆD i = (12) cos ˆω i t 1 sin ˆω i t 1 1 Âi 1t 1 sin ˆω i t 1 + ˆB i 1 t 1 cos ˆω i t 1 cos ˆω i t 2 sin ˆω i t 2 1 Âi 1t 2 sin ˆω i t 2 + ˆB i 1 t 2 cos ˆω i t cos ˆω i t N sin ˆω i t N 1 Âi 1t N sin ˆω i t N + ˆB i 1 t N cos ˆω i t N where g(ω) = y T D(D T D) 1 D T y. (16) The three parameters in x are then obtained from the leastsquares fit in Section III-A with ω there replaced by ˆω given by (15). The maximization of (16) may be implemented by iterative methods, or simply by a one-dimensional grid search, as presented in Table II. The frequency grid may

3 DEPARTMENT OF SIGNALS, SENSORS AND SYSTEMS, ROYAL INSTITUTE OF TECHNOLOGY, JUNE 3 TABLE II A non-linear least-squares fit by grid search. a) frequency grid ω i, i = 1,..., M b) for i = 1 to M c) create D from ω i, see (5) d) g i = y T D(D T D) 1 D T y e) end f) ˆω = ω k where g k = max[g i i = 1,..., M] g) ˆx is obtained from (6) be obtained by peak-picking the DFT of data and with ω 1 corresponding to the frequency bin to the left of the maxima, and ω M corresponding to the bin on the opposite side. The number of grid points M is chosen depending on the desired resolution. V. Cramér-Rao bound The algorithms in Tables I-II are expected to produce consistent estimates of the unknown parameters, that is the estimation error is small for large N. In the next section, we assess the performance of the considered algorithms by comparing their frequency error variance with the Cramér- Rao bound (CRB) [7]. A lower bound on the variance of any unbiased estimator is given by the CRB. Assume that data are given by z n = y n [θ] + e n (17) where e n is zero-mean white Gaussian noise with variance σ 2. The CRB is given by the inverse of the information matrix I(θ). The elements of I(θ), i.e. [I(θ)] k,p are given by [7] [I(θ)] k,p = 1 σ 2 N. (18) θ k θ p With θ = (A, B, C, ω) T (that is, for k, p = 1,..., 4) the derivatives in (18) can be calculated from (2) as follows A B C ω Thus, I(θ) can be written as where I(θ) = 1 σ 2 N = cos ωt n (19) = sin ωt n () = 1 (21) = At n sin ωt n + Bt n cos ωt n. (22) I AA I AB I AC I Aω I AB I BB I BC I Bω I AC I BC I CC I Cω I Aω I Bω I Cω I ωω (23) I AA = cos 2 ωt n (24) I BB = sin 2 ωt n (25) I CC = 1 (26) I ωω = ( At n sin ωt n + Bt n cos ωt n ) 2 (27) I AB = cos ωt n sin ωt n (28) I AC = cos ωt n (29) I BC = sin ωt n (3) I Aω = cos ωt n ( At n sin ωt n + Bt n cos ωt n ) (31) I Bω = sin ωt n ( At n sin ωt n + Bt n cos ωt n ) (32) I Cω = At n sin ωt n + Bt n cos ωt n. (33) One can note that the information matrix is independent of the offset C. We are mainly interested in the frequency estimation error, viz. CRB(ˆω) = [I(θ) 1 ] 4,4. Decompose I(θ) as ( ) I1 I I(θ) = 12 I T (34) 12 I 22 where I 1 is the upper left 3 3 matrix and I 22 = I ωω /σ 2. Then [8] CRB(ˆω) = [I(θ) 1 ] 4,4 = 1 I 22 I T 12 I 1 1 I. (35) 12 For uniform sampling t n = n/f s with f s being the sampling frequency, an approximation of the CRB of the absolute frequency ˆf = f s ˆω/2π for frequencies f well inside (, f s /2) and large N is CRB( ˆf) = ( ) 2 fs 2σ π (A 2 + B 2 ) N(N 2 1) ( ) 2 fs 12 2π SNRN(N 2 1) (36) where SNR denotes the signal-to-noise ratio, i.e. SNR = (A 2 + B 2 )/2σ 2. The asymptotic result (36) only depends on the SNR. In particular, it is independent of absolute frequency and initial phase of the sine wave. A. Monte Carlo simulations VI. Numerical examples In order to verify the performance of the considered algorithms, the theoretical CRB is compared with the normalized sum of squared errors obtained from Monte Carlo simulations, based on independent realizations. The empirical mean square error (mse) is computed as mse( ˆf) = 1 ( ˆf l f) 2 (37) l=1 where ˆf l denotes the frequency estimate (f = f s ω/2π) in the l-th realization. B. Data records with random initial phase In Figure 1, the 4-parameter matrix algorithm in [3] (Table I) and the nonlinear least-squares (NLS) method in

4 4 TECHNICAL REPORT IR-S3-SB-9 N=16, SNR= db N=16, SNR= db 5 6 LOG(MSE) (db) 4 3 IEEE 57 NLS AS CRB LOG(MSE) (db) IEEE 57 NLS Exact CRB Fig. 1. Mean square frequency estimation error versus frequency for noisy sinusoidal signal with random initial phase Fig. 2. Mean square frequency estimation error versus frequency for noisy sinusoidal signal with fixed initial phase. Table II are evaluated for short records of N = 16 noisy samples, with sampling frequency f s = 1. A sinusoidal signal with random (uniformly distributed in [, π]) initial phase in additive white Gaussian noise was generated, with a SNR of db. As estimator of the initial frequency, a DFT-based estimator was used with 4 times zero-padding, and peakfinding by triple parabolic interpolation. The grid search for the algorithm in Table II was performed (rather arbitrary) in M = 16 points in the symmetric interval (of length corresponding to twice the frequency resolution of the DFT) around the maxima of the DFT. The initial estimates of the nuisance parameters required for the algorithm in Table I were estimated using the 3-parameter fit with ˆω in place of ω. The 4-parameter matrix algorithm was aborted after 5 iterations. Figure 1 shows the mse (37) as a function of frequency f. As reference the asymptotic CRB (36) is shown. From the figure one may note an excellent performance of both algorithms for frequencies well inside (,.5). For frequencies near or.5, the proposed algorithm in Table II outperforms the one in [3]. C. Data records with fixed initial phase In order to compare the performance of algorithms with the exact CRB (35) the above experiment was repeated for low-frequency signals (f <.1). The conditions were set as in the experiment above, but now the initial phase φ was fixed and set (rather arbitrary) to φ = π/7. The comparison with the exact CRB in Figure 2 reveals that the empirical mse can be seen as an empirical variance for frequencies above f =.3 (for NLS) and f =.45 (for IEEE-57), respectively. For frequencies below f =.3, the empirical mse increases for both methods, indicating that for (very) low frequencies both methods suffer from bias errors. For N = 16, the frequency f =.3 corresponds to less than half a period of a uniformly sampled sine wave. LOG(MSE) (db) N=16, SNR= db IEEE 57 (2 BITS) NLS (2 BITS) IEEE 57 (1 BIT) NLS (1 BIT) Fig. 3. Mean square frequency estimation error versus frequency for quantized noisy data (1 and 2 bits, respectively). D. Quantized data records The sensitivity for course quantization is studied in this experiment. The noisy sine wave in Section VI-B, i.e. with random initial phase, is quantized with one and two bits, respectively. The results for this scenario are displayed in Figure 3. Similar results as in Figure 2 are obtained. Repeating the experiment with noise-free quantized data results in similar performance as in Figure 3, except for the IEEE 57 algorithm applied to 1-bit data where the performance coincides with the performance of NLS for f >.55 (an improvement compared with f >.65 in Figure 3). VII. Conclusions The four-parameter sine wave fit algorithm in IEEE Standard 57 [3, ] has been studied in some detail. In most practical applications it seems to be an excellent method for parameter estimation, and for reconstruction of a sine wave from noisy or quantized data. Its performance (in terms of frequency estimation error) has been compared

5 DEPARTMENT OF SIGNALS, SENSORS AND SYSTEMS, ROYAL INSTITUTE OF TECHNOLOGY, JUNE 5 with theoretical bounds on accuracy, and with the performance of a nonlinear least squares (NLS) approach. The comparison reveals that its performance is close to optimal in a Gaussian scenario, but it also reveals that it is inferior compared with the performance of the NLS in severe experimental conditions, characterized by a small number of samples at a low-frequency sine wave. It is worth noticing that the performance of the 4- parameter algorithm is studied in a small error context, i.e. it relies on precise initial estimates. In this paper, the initial estimates were obtained following the recommended procedure in [3]. In general, its convergence properties depend on the initial estimates, whereas the NLS relies on a one-dimensional grid search, i.e. its convergence is ensured. References [1] J. Kuffel, T. McComb and Malewski, Comparative evaluation of computer methods for calculating the best fit sinusoid to the high purity sine wave, IEEE Transactions on Instrumentation and Measurement, Vol. IM-36, No. 2, June 1987, pp [2] T.R. Mccomb, J. Kuffel, and B.C. Le Roux, A comparative evaluation of some practical algorithms used in the effective bits test of waveform recorders, IEEE Transactions on Instrumentation and Measurement, Vol. 38, No. 1, February 1989, pp [3] IEEE Standard for digitizing waveform recorders, IEEE Standard 57, [4] T.E. Linnenbrink, Waveform recorder testing: IEEE standard 57 and You, Instrumentation and Measurement Technology Conference, IMTC/95, pp [5] S.J. Tilden, T.E. Linnenbrink and P.J. Green, Overview of IEEE-STD Standard for terminology and test methods for analog-to-digital converters, Proc. 16th IEEE Instrumentation and Measurement Technology Conference, IMTC/99, Vol. 3, 1999, pp, [6] P. Stoica and R. Moses, Introduction to Spectral Analysis, Prentice-Hall, Upper Saddle River, NJ, [7] S.M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory, Prentice Hall, Upper Sadle River, NJ, [8] T. Söderström and P. Stoica, System Identification, Prentice- Hall, Hemel Hempstead, UK, Peter Händel(S 88-M 94-SM 98) received the M.Sc. degree in Engineering Physics and the Lic.Eng. and Ph.D. degrees in Automatic Control, all from the Department of Technology, Uppsala University, Uppsala, Sweden, in 1987, 1991, and 1993, respectively. During , he was a Research Assistant at The Svedberg Laboratory, Uppsala University. Between 1988 and 1993, he was a Teaching and Research Assistant at the Systems and Control Group, Uppsala University. In 1996, he was appointed as Docent at Uppsala University. During , he was with the Research and Development Division, Ericsson Radio Systems AB, Kista, Sweden. During the academic year 96/97, Dr. Händel was a Visiting Scholar at the Signal Processing Laboratory, Tampere University of Technology, Tampere, Finland. In 1998, he was appointed as Docent at the same university. Since August 1997, he has been an Associate Professor with the Department of Signals, Sensors and Systems, Royal Institute of Technology, Stockholm, Sweden. Dr. Händel is a former President of the IEEE Finland joint Signal Processing and Circuits and Systems Chapter. He is a registered engineer (EUR ING).