Mobile QR Code QR CODE

  1. (Department of Electrical Insulation and Cable Engineering, National Technical University "Kharkiv Polytechnic Institute" / Kharkiv, Ukraine)



Spectral methods, Bias of estimation, Sidelobes, Orthogonal functions

1. Introduction

Similarly to the problem of estimation of power spectrum of preliminary sampled single frequency or multifrequency signals, the issue of accurate frequency estimation belongs to a classical group of problems which are usually considered in signal processing theory and have a long record of related studies. Typical applications for which these problems are of a significant interest include the assessment of quality of electrical energy [1] and the analysis of features of a human voice [2]. The majority of conventional approaches for the solution of the mentioned problem are based on applying parametric methods or rely on applying the discrete Fourier transform. Many parametric methods rely on the estimation of autoregressive coefficients. Among various approaches, this can be accomplished according to the classical approach by solving Yule-Walker equations [3] or by using the Burg method [4]. Some other methods that also belong to the parametric group of methods eventually employ the specific properties of the autocorrelation matrix calculated for the analyzed signal. The Pisarenko frequency estimator [5] belongs to this group of methods and implies estimation of an eigenvector associated with the minimum eigenvalue of an autocorrelation matrix.

The frequently used method MUSIC also employs eigenvalue decomposition of the autocorrelation matrix and is usually considered as a generalization of a Pisarenko frequency estimator. This method uses the orthogonality of signal and noise spaces and the projection of vector of harmonic frequencies on the noise subspace [2]. Applying this method implies that the exact number of tones in the analyzed signal is known. Many parametric methods display increased computational complexity due to the necessity of estimation of the appropriate order of the used autoregression model [6]. One study [7] did a comparative analysis of noise stability of an autocorrelation-based frequency estimator and non-parametric discrete Fourier Transform-based methods (DFT). The results showed a higher level of noise stability of non-parametric algorithms.

For the case of applying the DFT-based frequency estimators, the accuracy of estimation is affected by spectral leakage and is usually increased by applying interpolation algorithms [8,9], which are often combined with spectral windowing [10]. Although the additional computational procedures used in order to implement interpolation algorithms allow us to increase the accuracy of frequency estimation, they also complicate the processing of the analyzed signal. Another obstacle for this approach is possible ambiguity in the selection of the most appropriate type of spectral windowing and interpolation algorithm, especially when taking into account that the analyzed signal does not necessarily fit a mathematical model of finite sum of sine functions with some initial phases and frequencies. This circumstance encourages the development of slightly modified frequency estimators, which do not imply the necessity of applying spectral windowing and additional interpolation but are still based on calculation of the inner product of the analyzed signal and trigonometric functions. The substantiation of one such approach is carried out in this study.

Carried out in Section 2 studies discuss the affect of the initial phase of the analyzed signal on the bias of frequency estimation caused by the refusal to complement the sine Fourier transform by the corresponding cosine transform. This Section also provides the relations for the selection of frequencies of the auxiliary sine functions which allow us to eliminate this bias. The sum of specifically selected points of sidelobes is used as a criterion for making the conclusion about the proximity of a particular frequency to the exact frequency of the analyzed signal. It was shown that the best available frequency estimation is attained for the case if the sum of local minimum values of sidelobes is equal to zero. This Section also presents the discussion of frequency properties of the auxiliary function which can be used for the recognition of the frequency interval which contains the exact frequency of the analyzed signal. Section 3 presents the results of simulations which illustrate the proper selection of the duration of analyzed signal, which allows us to calculate the array of frequencies of sine functions with the exact frequency of the analyzed signal and also the improper selection of the duration, which does not allow us to eliminate the bias of frequency estimation. The conclusions were made in Section 4.

2. Substantiation of the Proposed Approach for Frequency Estimation

Analysis was carried out for continuous time dependence of an analyzed signal. It was done on the basis of the inner product determined according to (1) for an analyzed signal and auxiliary sine function with zero initial phase and arbitrarily selected frequency:

(1)
$ S=\int _{0}^{b}\sin \left(\omega _{1}t\right)A_{2}\sin \left(\omega _{2}t+\phi _{2}\right)dt $

where A$_{2}$ denotes the amplitude of the analyzed signal, ${\varphi}$$_{2}$ and ${\omega}$$_{2}$ correspondingly denote the values of the initial phase and angular frequency of the analyzed signal, b denotes the duration of the analyzed signal, and ${\omega}$$_{1}$ denotes the value of the angular frequency of the employed auxiliary sine function. All three parameters of the analyzed signal, A$_{2}$, ${\omega}$$_{2}$ and ${\varphi}$$_{2}$, are considered to be unknown. The value of the initial phase ${\varphi}$$_{2}$ is considered to stay within the range of values from 0$^{\circ}$ to 90$^{\circ}$. The inner product (1) represents the sine Fourier transform of the continuous analyzed signal determined over some time interval from 0 to b. According to one study [11], for the case of ${\omega}$$_{1}$ ${\neq}$ ${\omega}$$_{2}$, the results of the calculation of the inner product (1) can be expressed as:

(2)
$ S=\sum _{i=1}^{5}s_{i} $

where all terms s$_{1}$..s$_{5}$ can be determined according to:

(3)
$ s_{1}=\frac{A_{2}\sin (\phi _{2})\omega _{1}}{(\omega _{1}-\omega _{2})(\omega _{1}+\omega _{2})} \\ $
(4)
$ s_{2}=-\frac{A_{2}\omega _{1}\cos (b\omega _{1})\sin (b\omega _{2})\cos (\phi _{2})}{(\omega _{1}-\omega _{2})(\omega _{1}+\omega _{2})} \\ $
(5)
$ s_{3}=-\frac{A_{2}\omega _{1}\cos (b\omega _{1})\cos (b\omega _{2})\sin (\phi _{2})}{(\omega _{1}-\omega _{2})(\omega _{1}+\omega _{2})} \\ $
(6)
$ s_{4}=\frac{A_{2}\omega _{2}\sin (b\omega _{1})\cos (b\omega _{2})\cos (\phi _{2})}{(\omega _{1}-\omega _{2})(\omega _{1}+\omega _{2})} \\ $
(7)
$ s_{5}=-\frac{A_{2}\omega _{2}\sin (b\omega _{1})\sin (b\omega _{2})\sin (\phi _{2})}{(\omega _{1}-\omega _{2})(\omega _{1}+\omega _{2})} $

The inner product determined according to (1) of the analyzed signal and auxiliary sine function has not been complemented by the corresponding inner product of the analyzed signal and auxiliary cosine function. This is due to not applying the complete system of the orthogonal functions, which cannot be directly used for frequency estimation. Therefore, as the maximum value of (1) is attained for the case when ${\omega}$$_{1}$ ${\neq}$ ${\omega}$$_{2}$, the frequency estimator is biased. Increasing the duration of the analyzed signal b partially allows us to mitigate this error of the frequency estimation. But practically applying the inner product (1) for the purposes of frequency estimation is complicated. In order to attain an admissible level of accuracy, various values of the initial phases will require various durations of the analyzed signal. Besides, the duration of the analyzed signal for practical applications is limited for the vast majority of cases, and this limitation of the analyzed signals causes additional difficulties in the practice of frequency estimation. For analyzed signals with different duration, different values of the initial phase ${\varphi}$$_{2}$ obviously will lead to different levels of accuracy.

2.1 The Affect of the Selected Frequencies of the Auxiliary Sine Functions on Sidelobe Properties

Eq. (2) can be substantially simplified by means specific selection of the frequencies of the auxiliary sine functions. This simplified form of the inner product can be efficiently used for the purposes of frequency estimation. The simplification of (2) can be attained by the selection of the frequencies used in (1) for auxiliary sine functions according to the following relations:

(8)
$ f_{s1}=\frac{1}{4b}+\frac{k}{2b} \\ $
(9)
$ f_{s2}=\frac{k}{2b} $

where f$_{s}$$_{1}$ denotes the frequency used in (1) for the auxiliary sine function selected in order to simplify (2) by eliminating terms s$_{2}$ and s$_{3}$. f$_{s}$$_{2}$ denotes the frequency used in (1) in the auxiliary sine function to simplify (2) by eliminating terms s$_{4}$ and s$_{5}$, k = 1..n. It also denotes the number of a particular auxiliary sine function from the total amount of auxiliary sine functions n.

As it is easy to infer from the analysis of (2), for the case when the values of the frequencies used in auxiliary sine functions fit (8), the value determined according to (2) of the inner product S will contain only terms s$_{1}$, s$_{4}$, and s$_{5}$. Due to the presence of the dependence on the ${\omega}$$_{1}$ multiplier in cos(b${\omega}$$_{1}$), the dependence of the terms s$_{2}$ and s$_{3}$ in (2) on ${\omega}$$_{1}$ is represented by the cosine functions. The selection of the frequencies of the auxiliary sine functions according to (8) allows us to eliminate the impact of the terms s$_{2}$ and s$_{3}$on the frequency dependence of the results of the inner product calculation S.

The elimination of the impact of the terms s$_{2}$ and s$_{3}$ also leads to a simultaneous increase of the impact of the terms s$_{4}$ and s$_{5}$on the frequency dependence of this inner product. The selection of the frequencies of the auxiliary frequencies according to (9) leads to an antithetical result and allows us to eliminate the impact of the terms s$_{4}$ and s$_{5}$ with a simultaneous increase of the impact of s$_{2}$ and s$_{3}$ on the dependence of the inner product S determined according to (1) on the frequencies of the auxiliary sine functions. Therefore, the selection of the frequencies of the auxiliary sine functions according to (8) and (9) allows us to segregate the input of the terms s$_{2}$ and s$_{3}$ from the input of the terms s$_{4}$ and s$_{5}$ to the ultimate value of the inner product S. Fig. 1 presents typical dependencies of the inner product (2) on frequencies calculated according to (8) used in auxiliary sine functions. All calculations have been carried out for the case of f$_{2}$ = 46 Hz, ${\varphi}$$_{2}$ = 60$^{\circ}$, A$_{2}$= 1, and b = 0.2.

Fig. 2 presents the same frequency dependence calculated for the case when the frequencies of the sine functions fit (9).

Fig. 1. Typical frequency dependence of inner product S calculated according to(1)determined for the case when the frequencies of auxiliary sine functions fit the values determined according to(8).
../../Resources/ieie/IEIESPC.2021.11.1.14/fig1.png

For the case where the frequencies of the auxiliary sine functions have been selected according to (9), frequency estimation can be carried out on the basis of the observation of the particular points of the sidelobes in the frequency dependence of the inner product S. The results of simulations have shown that for the case when some particular frequency coincides with the frequency of the analyzed signal, the sum of the specifically selected values of the frequency dependence of the inner product S is equal to zero, regardless of the value of the initial phase of the analyzed signal. This behavior of the sidelobes is illustrated in Fig. 3, which presents the fragment of the frequency dependence of the inner product (1) for the case of b = 3 and, f$_{2}$ = 40 Hz. A previously mentioned area of interest which contains characteristic elements of the sidelobes with a sum equal to zero is marked with an elliptical curve.

In the case of calculated frequencies according to (9) of the auxiliary sine functions for the selected duration of the analyzed signal with b = 3 and k = 240, the value of f$_{\mathrm{s2}}$ will be equal to 40 Hz. Therefore, this frequency will exactly coincide with the frequency of the analyzed signal. Fig. 4 presents the frequency dependence of the inner product (1) calculated for the case of f$_{2}$ = 40 Hz and b = 2 s.

For all results of the calculations presented in Figs. 1-4, the values of the inner product (1) have been calculated numerically by applying the trapezium formula. The results of the calculations presented in Fig. 4 show that the existence of some difference between the frequency of the analyzed signal and auxiliary sine function used in (1) leads to the gradual increase of the previously distinguished characteristic points of the sidelobes. Therefore, the sum of these elements is not equal to zero.

The described behavior of the sidelobes which allows us to remove the bias of the frequency estimation caused by the initial phase of the analyzed signal is determined by the previously mentioned removal of the terms s$_{4}$ and s$_{5}$ in (2). For the case when the initial phase of the analyzed signal coincides with the initial phases of the auxiliary sine functions (i.e., for the case of ${\varphi}$$_{2}$ = 0 in (1)), the frequency dependence of S will be determined only by terms s$_{2}$ and s$_{4}$. Since the selection of the frequencies according to (9) allows us to remove only the terms s$_{4}$ and s$_{5}$ and does not allow us to eliminate the term s$_{1}$, such a selection of the frequencies disrupts a possible approach for frequency estimation based on analysis of the interaction of the terms s$_{2}$ and s$_{4}$. However, it provides another option suitable for frequency estimation.

Provided s$_{4}$= s$_{5}$ = 0, which is attained by the selection of the frequencies of the auxiliary sine functions according to (9), and for the case of the duration of the analyzed signal fitting the condition sin(b${\omega}$$_{2}$) = 0, the value of s$_{2}$ = 0 and, consequently, the value of inner product S will be determined by a single alternating term s$_{3}$. The residual of the independent on b term s$_{1}$ and the remaining alternating term s$_{3}$determines the presented in Fig. 3 frequency dependence of marked with an elliptic curve characteristic points. The described interaction of the terms s$_{1}$, s$_{2}$, and s$_{3}$ and the resulting frequency dependence of the inner product S are illustrated in Fig. 5.

Further analysis will be carried out based on the following premises:

1. The exact value of frequency of the analyzed signal stays within the frequency interval limited by the maximum value of frequency dependence of the inner product (1) and the closest minimum value of this frequency dependence which follows the closest zero crossing of the main lobe. These particular points of the dependence of the inner product (1) on frequencies of the auxiliary sine functions allow us to distinguish the array of the frequencies, which is supposed to contain the exact frequency of the analyzed signal, and are correspondingly marked by the points A, B, and C in Figs. 1 and 2.

2. For the case of a multifrequency signal, the duration of the analyzed signal and the distance between its frequency components are supposed to have values which allow us to consider that the main lobe and its closest sidelobes of a particular frequency component do not affect the location and size of the sidelobes, which are determined by other frequency components. This stipulation in fact implies that the case of a multifrequency signal can be reduced to a separate analysis of the contribution of each frequency component to the overall frequency dependence of the inner product S.

3. The exact value of frequency of the analyzed signal should not necessarily correspond to the maximum value of inner product (1), but belongs to the previously distinguished frequency interval limited by the maximum value of frequency dependence of the inner product (1) and the closest minimum value of this frequency dependence.

4. Frequency estimation is carried out on the basis of the consecutive calculation of the series of the inner products of the analyzed signal and auxiliary sine functions with various frequencies. For each inner product calculation, the analyzed signal is limited by its various durations, which allow us to determine various arrays of the frequencies calculated according to (9) for the auxiliary sine functions.

5. According to the results of the simulations presented in Fig. 3, the most accurate frequency estimation is attained for the case when the sum of the results marked by the elliptical curve in Fig. 3 of the inner product calculation is closest to zero. Thus, the criterion for the assessment of the proximity of the estimated value of frequency to the exact frequency of the analyzed signal should be based on calculation of the sum of these particular values of S. This criterion will be accomplished for the case when for a certain duration of the analyzed signal among the frequencies determined according to (9) for the auxiliary sine functions some particular frequency will have its value as close possible to the exact frequency of the analyzed signal.

6. For the case of conventional algorithms for frequency estimation, based on applying the discrete Fourier transform, the frequencies used in auxiliary sine and cosine functions usually are selected to be a multiple of the sampling frequency. For this case, their circular frequencies can be calculated according to the following relation [12]:

Fig. 2. Typical frequency dependence of the inner product S calculated according to(1)determined for the case when the frequencies of auxiliary sine functions fit the values determined according to(9).
../../Resources/ieie/IEIESPC.2021.11.1.14/fig2.png
Fig. 3. The affect of the calculated according to(9)frequencies of the auxiliary sine functions on sidelobes of the frequency dependence of the result of the inner product calculation for the case when a certain frequency of the auxiliary sine function coincides with the frequency of the analyzed signal.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig3.png
Fig. 4. Typical frequency dependence of inner product S calculated according to(1)determined for the case when the frequencies used in auxiliary sine functions fit the values determined according to(9)and do not coincide with the frequency of the analyzed signal.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig4.png
(10)
$ \omega _{k}=\frac{k\omega _{s}}{N} $

where N denotes the total amount of signal samples, and k denotes the number of a particular component of the Fourier transform. For the considered in frequency estimator instead of applying (10) for the circular frequencies of the auxiliary trigonometric functions, the values of frequencies will be selected according to (9). Also, it should be noted that calculation of the frequencies of the auxiliary sine functions according to (9) usually allows us to attain much lower frequency resolution in comparison with the conventional approach based on applying (10). In the case of applying (9) for the selection of the frequencies, a high level of frequency resolution can be attained only for analyzed signals with a sufficiently large duration. For analyzed signals with the same duration determined according to (9), frequency resolution is considerably lower than for the case where the frequencies of the auxiliary sine functions are selected according to (10).

Frequency estimation is carried out based on the assumption that the exact value of frequency of the analyzed signal belongs to the range of frequencies limited by the points A and B in Figs. 1 and 2. The existence of undesirable noise components in analyzed signals, however, might complicate visual recognition of this frequency interval. Therefore, for practical applications, it is desirable to develop complementary methods which would provide additional tools which could be used in order to identify this frequency interval. In the current study, such a tool was developed on the basis of applying the antiderivative function determined for the dependence of the result of the inner product calculation (2) on values of angular frequencies ${\omega}$$_{1}$of auxiliary sine functions.

2.2 Estimation of the Frequency Interval which Contains the Frequency of the Analyzed Signal

The necessity of employing the distinguished frequency interval is determined by the selected approach for frequency estimation, which implies that the exact frequency of the analyzed signal does not necessarily correspond to the maximum value of the determined according to (1) inner product. Since the value of frequency cannot be attributed to the maximum value of the inner product, it should be associated with some other specific properties of the dependence of the inner product on frequencies of the auxiliary sine functions.

According to the previously made assumption the exact frequency is associated with the frequency interval limited by the maximum value of S and the closest minimum value of this dependence (points A and B in Fig. 1). The contamination of the analyzed signal by noise components in analyzed signals, however, distorts the dependence of the inner product (1) on frequencies f$_{1}$ used in auxiliary sine functions and consequently complicates recognition of this frequency interval. Nevertheless, the mentioned problem of the frequency interval recognition partially can be mitigated on the basis of the analysis of the properties of the antiderivative function determined for the dependence of the inner product (1) on values of angular frequencies of the auxiliary sine functions S = f(${\omega}$$_{1}$). The selection of this particular function for the purpose of the identification of the necessary frequency interval was motivated by the intention to smooth the function S = f(${\omega}$$_{1}$) and conceal its random oscillations caused by the contamination with some undesirable noise component. In a simplified form, assuming zero integration constant, this antiderivative function can be expressed as:

(11)
$ B=0.5A_{1}\left(\sum _{i=0}^{4}B_{i}\right) $

where all terms can be determined according to:

(12)
$ B_{0}=\sin (\phi )\ln (\left(\omega _{1}-\omega _{2}\right)\left(\omega _{1}+\omega _{2}\right)) \\ $
(13)
$ B_{1}=\cos (\phi )\mathrm{Si}(b\omega _{1}-b\omega _{2}) \\ $
(14)
$ B_{2}=-\cos (\phi )\mathrm{Si}(b\omega _{1}+b\omega _{2}) \\ $
(15)
$ B_{3}=-\sin (\phi )\mathrm{Ci}(b\omega _{1}-b\omega _{2}) \\ $
(16)
$ B_{4}=-\sin (\phi )\mathrm{Ci}(b\omega _{1}-b\omega _{2}) $

The results of the calculations presented in Fig. 6 illustrate typical frequency dependencies of the antiderivative function according to (11).

The results of the calculations presented in Fig. 7 illustrate the impact of the duration of the analyzed signal on frequency dependence of the expressed antiderivative function according to (11).

The previously mentioned impact of the undesirable noise components on S(f$_{1}$) is illustrated in Figs. 8 and 9. Fig. 9 presents this frequency dependence for an analyzed signal without any undesirable noise component, while Fig. 9 presents the same frequency dependence calculated for a signal contaminated with noise. Fig. 10 presents the frequency dependence of the antiderivative function for the cases of the absence and presence of noise components in analyzed signals.

For all cases, noise contamination has been represented by normally distributed white noise. The results of the calculations illustrate the distortions in the calculated frequency dependence of the inner product determined according to (1) caused by the presence of the noise components in analyzed signals. In Fig. 8, the frequency of the auxiliary sine function approaching the exact frequency of the analyzed signal leads to gradual increase of the sidelobes. According to the results in Fig. 9, the presence of the noise components disrupts the gradual increase and can lead to a false conclusion about the existence of a certain frequency component in the analyzed signal. Also, it may cause difficulties in the identification of the previously distinguished frequency interval, which contains the exact frequency of the analyzed signal.

The results of the calculations presented in Fig. 10 indicate that the presence of the noise components also distorts the frequency dependence of the antiderivative function. However, in the vicinity of the frequency interval limited by the maximum value of B(f$_{1}$) and its closest local minimum value, such a disruptive impact becomes less evident. The maximum value of frequency dependence presented in Fig. 8 still corresponds to the main lobe frequency equal to 1002 Hz. However, from the prospective of the impact of the noise components on the antiderivative function according to the results presented in Fig. 10, it is possible to distinguish a certain frequency region in the vicinity of the exact frequency of the analyzed signal for which the impact of the noise components does not lead to such drastic distortions. This frequency interval is distinguished by the frequencies of 998.1 Hz and 1009 Hz in Fig. 10. Therefore, this impact becomes mitigated for the range of frequencies in the vicinity of the exact frequency of the analyzed signal.

Although the presence of the noise components can lead to some shifting of the previously distinguished frequency interval, it still covers the exact frequency of the analyzed signal. Therefore, for the case of the analyzed signals with noise components, the existence of a certain frequency component can be established by calculating the antiderivative function (11), at least in some range of signal-to-noise ratios. The selection of the frequencies of the auxiliary sine functions according to (9) changes the dependence of the antiderivative function on values of frequencies of the auxiliary sine functions. For the case of frequencies selected according to (9), this dependence can be expressed as:

(17)
$ D\left(\omega _{1}\right)=B_{0}+\sum _{n=1}^{8}D_{n} $

where D(${\omega}$$_{1}$) denotes the antiderivative function determined for inner product (1) and calculated for the case when the frequencies of the auxiliary sine functions fit the values determined according to (9), and the expressions for D$_{1}$..D$_{8}$ can be determined according to:

(18)
$ D_{1}=\frac{A_{2}}{2}\cos (\phi )\sin ^{2}\left(b\omega _{2}\right)\mathrm{Si}(b\omega _{1}-b\omega _{2}) \\ $
(19)
$ D_{2}=-\frac{A_{2}}{2}\cos (\phi )\sin \left(b\omega _{2}\right)\cos \left(b\omega _{2}\right)\mathrm{Ci}(b\omega _{1}-b\omega _{2}) \\ $
(20)
$ D_{3}=-\frac{A_{2}}{2}\cos (\phi )\sin ^{2}\left(b\omega _{2}\right)\mathrm{Si}(b\omega _{1}+b\omega _{2}) \\ $
(21)
$ D_{4}=-\frac{A_{2}}{2}\cos (\phi )\sin \left(b\omega _{2}\right)\cos \left(b\omega _{2}\right)\mathrm{Ci}(b\omega _{1}+b\omega _{2}) \\ $
(22)
$ D_{5}=\frac{A_{2}}{2}\sin (\phi )\sin \left(b\omega _{2}\right)\cos \left(b\omega _{2}\right)\mathrm{Si}(b\omega _{1}-b\omega _{2}) \\ $
(23)
$ D_{6}=-\frac{A_{2}}{2}\sin (\phi )\cos ^{2}\left(b\omega _{2}\right)\mathrm{Ci}(b\omega _{1}-b\omega _{2}) \\ $
(24)
$ D_{7}=\frac{A_{2}}{2}\sin (\phi )\sin \left(b\omega _{2}\right)\cos \left(b\omega _{2}\right)\mathrm{Si}(b\omega _{1}+b\omega _{2}) \\ $
(25)
$ D_{8}=-\frac{A_{2}}{2}\sin (\phi )\cos ^{2}\left(b\omega _{2}\right)\mathrm{Ci}(b\omega _{1}+b\omega _{2}) $

Fig. 11 illustrates the dependencies of the antiderivative function on frequency of the auxiliary sine functions.

The results of the calculations presented in Fig. 11 indicate that despite the presence of the noise contamination, the location of the minimum value of the antiderivative function that has not been affected by the noise component is close to the location of the minimum of the antiderivative function that has not been distorted by noise contamination. In the vicinity of this minimum value, two antiderivative functions also display higher similarity than in distant regions of the considered frequency dependence.

For the case of frequencies selected according to (9) for the auxiliary sine functions mentioned, the similarity of the antiderivative functions in the vicinity of their minimum value becomes more evident than for the case of results of the calculations in Fig. 10, which were carried out for arbitrarily selected frequencies of the auxiliary sine functions. Such similarity of the antiderivative functions of the contaminated and uncontaminated signals confirms that the existence of the certain frequency component in noise can be determined by identifying the minimum value of the antiderivative function calculated for the dependence of the inner product (2) on frequencies used for auxiliary sine functions.

For the purposes of simplicity, all antiderivative functions in Figs. 10 and 11 have been represented by cumulative sums of the corresponding elements of the frequency dependence S(f$_{1}$). Consequently, the proposed approach for frequency estimation for a single tone signal which does not contain any noise component consists of the following procedures:

1. Limitation of the whole array of values of the preliminary sampled signal by some intermediate value which will correspond to some intermediate signal duration b.

2. Calculation of the array of frequencies of the auxiliary sine functions with zero initial phases according to (9).

3. Calculation of the values used in auxiliary sine functions in particular moments of time which are determined by the value of sampling frequency used in order to obtain the array of values of the analyzed signal.

4. Calculation of the series of the inner products of the samples of the analyzed signal and auxiliary sine functions.

5. Calculation of the sum of the results marked with an ellipse in Figs. 3-5 for the inner product calculation. The first element of this sequence directly follows the maximum calculated value of the inner product (point A in Figs. 1 and 2). All other elements of this sequence alternate with non-zero results of the inner product calculation, as illustrated by the results of the calculations presented in Figs. 3-5.

6. The selection of the other intermediate value among the array of samples of the analyzed signals which will correspond to other intermediate signal duration b and reiteration of procedures 2-5.

7. The selection of the most accurately estimated value of frequency. The value of frequency which corresponds to the maximum value (point A in Figs. 1 and 2) of the dependence of the inner product on frequencies of the auxiliary sine functions indicates the upper boundary of the frequency interval which contains the exact frequency of the analyzed signal. Frequency associated with the minimum value of this frequency dependence (point B in Figs. 1 and 2) indicates the lower boundary of the frequency interval with the exact value of frequency. The value of frequency which corresponds to the sum closest to zero of the elements marked with ellipse in Figs. 3-5, which stays between the upper and lower boundaries and represents the best available frequency estimation.

Fig. 5. Interaction of the terms s$_{1}$, s$_{2}$, and s$_{3}$ which result in equality to zero of the points of the frequency dependence of the inner product S marked with elliptical curve particular for the case of A$_{1}$ = A$_{2}$ = 1, b = 3 s, and f$_{2}$ = 46 Hz.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig5.png
Fig. 6. The dependencies of the antiderivative function on frequencies of the auxiliary sine functions for the case of A$_{2}$ = 1 and b = 1.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig6.png
Fig. 7. The dependencies of the antiderivative function for the case of ${\varphi}$ = 45$^{\circ}$, A$_{2}$ = 1.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig7.png
Fig. 8. Typical frequency dependence of the inner product S calculated according to(1)determined for arbitrarily selected frequencies of the auxiliary functions.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig8.png
Fig. 9. Typical frequency dependence of the inner product S calculated according to(1)determined for arbitrarily selected frequencies of auxiliary sine functions and signal-to-noise ratio of 0.026.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig9.png
Fig. 10. B(f) for the case of arbitrarily selected frequencies of the auxiliary sine functions: 1 - for the case of the analyzed signal without any noise component, 2 - for signal-to-noise ratio equal to 0.026.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig10.png
Fig. 11. The dependence of the antiderivative function D(f$_{1}$) for the case of the frequencies determined according to(9)of the auxiliary sine functions: 1- for the case of the analyzed signal without any noise component, 2 - for the case of presence of the additive normally distributive white noise.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig11.png

3. The Results and Discussion of the Simulations

All simulations have been carried out for the case of a simple sine signal with a preliminary selected amplitude, frequency, and initial phase. Simulations also have been carried out without taking into consideration the possible presence of noise components in analyzed signals, which may arise from diverse sources and have their own specific spectral properties. The criterion for the most appropriate duration of the analyzed signal was based on the results of the calculation of the finite differences between the necessary for frequency estimation elements of the frequency dependence marked with an ellipse in Fig. 3. For the purposes of the simplification instead of applying the relation (26), which directly stems from the trapezium rule, all simulations have been carried out on the basis of the calculation of the auxiliary value P which was determined according to (28):

(26)
$ S\approx h\cdot \left(0.5\left(Q_{1}+Q_{n}\right)+\sum _{i=2}^{n-1}Q_{i}\right) $

where h denotes the time duration between the adjacent samples of a signal, and all Q$_{i}$ for i = 1..n can be calculated as:

(27)
$ Q_{i}=\sin (\omega _{1}t_{i})A_{2}\sin (\omega _{2}t_{i}+\phi _{2}) \\ $
(28)
$ P=(0.5(Q_{1}+Q_{n})+\sum _{i=2}^{n-1}Q_{1}) $

All simulations have been carried out for an analyzed sine signal with the values of A$_{2}$ = 1, f$_{2}$ = 50 Hz, and ${\varphi}$$_{2}$ = 60$^{\circ}$. The results of the simulations are presented in Fig. 12

Figs. 13 and 14 present frequency dependencies of the inner product (1) for different durations of the analyzed signal.

According to the results of the simulations presented in Fig. 12, it can be concluded that the sum closest to zero of the finite differences of the sequence determined according to (26) exhibits periodical alternations. The results presented in Fig. 13 illustrate proper selection of the duration of the analyzed signal, which corresponds to one of the local values closest to zero of P in Fig. 12 and allows us to use (9) for the array of the frequencies which contains the exact frequency of the analyzed signal. For this case, the maximum value of S corresponds to the upper boundary of the frequency interval which contains the exact frequency of the analyzed signal (f$_{1}$ = 50.17 Hz). The minimum value of S corresponds to the lower boundary of this frequency interval (f$_{1}$ = 49.83 Hz), whereas the intermediate value (f$_{1}$ = 50 Hz) represents the exact frequency of the analyzed signal, as it corresponds to the value of b which matches with one of the local values closest to zero of P in Fig. 12.

The results presented in Fig. 14 illustrate the unsuitable selection of the duration of the analyzed signal b, which corresponds to one of the local minimum values of P in Fig. 12. For this case, the intermediate value of S which stays between the estimated upper (50.08 Hz) and lower (47.75 Hz) boundaries of the frequency interval does not allow us to carry out accurate frequency estimation. In further analysis, special attention should be paid to the analysis of the multifrequency signals and the effect of the noise contamination, which might distort the frequency dependence of the inner product. Consequently, the criteria based on calculation of the finite differences or on calculation of the sum used for frequency estimation characteristic points of the sidelobes could be inapplicable.

Fig. 12. The dependence of P calculated according to(28)on the duration of the analyzed signal b.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig12.png
Fig. 13. S(f) for the case of b = 2.97 s.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig13.png
Fig. 14. S(f) for the case of b = 2.955 s.
../../Resources/ieie/IEIESPC.2021.11.1.14/fig14.png

4. Conclusion

This study presented an approach for frequency estimation based on applying the sine Fourier transform. The elimination of the bias caused by the refusal to the apply the complete system of orthogonal functions (i.e., the conventional approach based on the applying methods which employ algorithms with both sine and cosine transforms) is attained by the specific selection of the frequencies of the auxiliary sine functions and by means of distinguishing the particular characteristic points of the sidelobes. The equality to zero of the sum of these elements is used as a criterion for a properly selected duration of the analyzed signal, which allows us to compute the array of frequencies of the auxiliary sine functions used for inner product calculation, which contains the exact frequency of the analyzed signal. For an analyzed signal with noise components, the distortions of the antiderivative function calculated for the dependence of the result of the inner product on frequencies of the auxiliary sine functions are mitigated in the vicinity of the frequency interval with the exact frequency of the analyzed signal. This property can be used for the purposes of identification of the existence of a particular frequency in an analyzed signal, which can be complicated for the case with the presence of undesirable noise components.

REFERENCES

1 
Chen Kaiming, Huang Lvchao, Liu Zhiu, 2020, A Frequency Measurement Algorithm for Power Quality Monitoring Terminal, Journal of Physics: Conference Series, Vol. 1626, No. 4, pp. 012071DOI
2 
Murakami T., Ishida Y., 2001, Fundamental Frequency Estimation of Speech Signals Using MUSIC Algorithm, Acoustical Science & Technology, Vol. 22, No. 4, pp. 293-297DOI
3 
Zielinski T. P., Duda K., 2011, Frequency and Damping Estimation Methods - an Overview, Metrology and Measurement Systems, Vol. 18, No. 4, pp. 283-302DOI
4 
Woodcock R., Muhamedsalih H., Martin H., Jiang X., 2016, Burg Algorithm for Enhancing Measurement Performance in Wavelength Scanning Interferometry, Surface Topography: Metrology and Properties, Vol. 4, No. 2, pp. 024003DOI
5 
Chan J. W., So H. C., March. 2003, An Exact Analysis of Pisarenko’s single-tone frequency estimation algorithm, Signal Processing, Vol. 83, No. 3, pp. 685-690DOI
6 
Borkowski J., Kania D., Mroczka J., 2018, Comparison of sine-wave Frequency Estimation Methods in Respect of Speed and Accuracy for a Few Observed Cycles Distorted by Noise and Harmonics, Metrology and Measurement Systems, Vol. 25, No. 2, pp. 283-302DOI
7 
Al-Khafaji S.D., Hussain Z.M., Neville K., 2017, Frequency estimation of Fm Signals Under Non-Gaussian and Colored Noise, International Journal of Applied Engineering Research, Vol. 12, No. 22, pp. 12342-12356URL
8 
Almayyali H. R., Hussain Z. M., 2021., Deep Learning versus Spectral Techniques for Frequency Estimation of Single Tones: Reduced Complexity for Software-Defined Radio and IoT Sensor Communications, Sensors, Vol. 21, No. 8, pp. 2729DOI
9 
Quinn B. G., May. 1994, Estimating Frequency by Interpolation Using Fourier Coefficients, IEEE Transactions on Signal Processing, Vol. 42, No. 5, pp. 1264-1268DOI
10 
Borkowski J., Kania D., Mroczka J., December. 2014, Interpolated-DFT-Based Fast and Accurate Frequency Estimation for the Control of Power, IEEE Transactions on Industrial Electronics, Vol. 61, No. 12, pp. 7026-7034DOI
11 
Kostiukov I., Oct. 2020, Estimation of Dissipation Factor by Applying Cross-Correlation Method, in Proc. of 2020 IEEE KhPIWeek, pp. 333-338DOI
12 
Sedlacek M., Krumpholc M., 2005, Digital Measurement of Phase Difference - A Comparative Study of DSP Algorithms, Metrology and Measurement Systems, Vol. 12, No. 4, pp. 427-448URL

Author

Ivan Kostiukov
../../Resources/ieie/IEIESPC.2021.11.1.14/au1.png

Ivan Kostiukov is an Associate Professor at National Technical University "Kharkiv Polytechnic Institute" (NTU “KhPI”), Ukraine. He received his B.S. and M.S. degrees in Electrical Engineering from NTU “KhPI”, in 2009 and 2011, respectively and his candidate of technical sciences degree in Control Engineering from NTU “KhPI” in 2015. His research interests include the development of methods for the testing of electrical insulation and digital signal processing techniques.