This paper revisits the characteristics of windowing techniques with various window functions involved, and successively investigates spectral leakage mitigation utilizing the Welch method. The discrete Fourier transform (DFT) is ubiquitous in digital signal processing (DSP) for the spectrum analysis and can be efficiently realized by the fast Fourier transform (FFT). The sampling signal will result in distortion and thus may cause unpredictable spectral leakage in discrete spectrum when the DFT is employed. Windowing is implemented by multiplying the input signal with a window function and windowing amplitude modulates the input signal so that the spectral leakage is evened out. Therefore, windowing processing reduces the amplitude of the samples at the beginning and end of the window. In addition to selecting appropriate window functions, a pretreatment method, such as the Welch method, is effective to mitigate the spectral leakage. Due to the noise caused by imperfect, finite data, the noise reduction from Welch’s method is a desired treatment. The nonparametric Welch method is an improvement on the periodogram spectrum estimation method where the signal-to-noise ratio (SNR) is high and mitigates noise in the estimated power spectra in exchange for frequency resolution reduction. The periodogram technique based on Welch method is capable of providing good resolution if data length samples are appropriately selected. The design of finite impulse response (FIR) digital filter using the window technique is firstly addressed. The influence of various window functions on the Fourier transform spectrum of the signals is discussed. Comparison on spectral resolution based on the traditional power spectrum estimation and various window-function-based Welch power spectrum estimations is presented.
The method of information extraction depends on the type of signal and the nature of the information being carried by the signal. A fundamental aspect of digital signal processing (DSP) is filtering, which is the process of transformation of input sequence to obtain desired output sequence. A digital filter is a system that performs mathematical operations on a sampled, discrete-time signal to reduce or enhance certain aspects of that signal. Digital filters are pervasive and play an important role in DSP, which is an important field in the present era of communication systems, digital computer technology, and consumer devices.
The input and output signals in a digital filter are discrete time sequence. Digital filters are linear time invariant (LTI) systems which are characterized by unit sample response. They are portable and highly flexible with minimum or negligible interference noise and other effects. Various methods for filter designs and analysis have been developed. Digital filters are categorized in two types: Finite impulse response (FIR) and infinite impulse response (IIR). Digital filters with infinite duration impulse response are referred to as IIR filters. The present output of FIR type filters depends on present input and past inputs. In comparison to IIR filters, the FIR filters are non-recursive and have greater flexibility to control the shape of their magnitude response. FIR digital filters are widely used than IIR ones since FIR filters are always stable, and have an exactly linear phase, and arbitrary amplitude-frequency characteristic, etc. Bandpass filtering plays an important role in DSP applications. It can be used to pass the signals according to the specified frequency passband and reject the frequency other than the passband specification.
Spectrum analysis [
Simple in operation, window function provides greater flexibility in DSP applications. In frequency resolution problems, a narrow main-lobe width of window function in frequency domain is usually required. Direct truncation is equivalent to the rectangular window shaping. A rectangular time window has narrow main lobe and wide side lobe with small main lobe to side lobe ratio, resulting in serious side lobe leakage. Effort has been made to mitigate the influence of spectral leakage of the Fourier transform spectrum by selecting appropriate window functions [
The spectral leakage suppression algorithms [
In spread spectrum system, the individual data sets are commonly overlapped in time to mitigate the information loss. The frequency-domain interference suppression algorithms always use data windows to reduce the spectral leakage associated with truncation, and employ overlap to mitigate the signal-to-noise ratio (SNR) degradation due to windowing. Known as the method of averaged periodograms used for estimating power spectra, the Bartlett’s method provides a way to reduce the variance of the periodogram at the expense of a reduction in resolution. A final estimate of the spectrum at a given frequency is obtained by averaging the estimates from the periodograms derived from non-overlapping portions of the original series. Bartlett’s method provides a method to reduce the variance of the periodogram in exchange for a reduction of resolution in compared to standard ones. The variance of the periodogram is reduced by applying a window to the autocorrelation estimate to decrease the contribution of unreliable estimates to the periodogram. As a generalization of Barlett’s method using windowed overlapping segments, the nonparametric Welch’s method [
The remainder of this paper is organized as follows. In Section 2, preliminary background on the Fourier transform is reviewed, including the discrete Fourier transform, and the fast Fourier transform. The spectrum leakage is introduced in Section 3. In Section 4, window functions for spectrum leakage suppression are discussed. The Welch method is introduced in Section 5. In Section 6, simulation experiments are carried out to evaluate the performance for various designs. Conclusions are given in Section 7.
The Fourier transform is employed to transform signals in time domain to frequency domain and to do the reverse and has been very important in a broad spectrum of expertise, such as the engineering, mathematics, and the physical sciences. The discrete counterpart of the Fourier transform is the DFT for implementation on digital hardware. DFT is can be efficiently realized by fast Fourier transform (FFT). Equivalent to the continuous Fourier transform, the DFT converts a finite sequence of equally-spaced samples of a function into a same-length sequence of equally-spaced samples of the discrete-time Fourier transform (DTFT), which is a complex-valued function of frequency.
The DTFT transform pair is symbolically denoted by
The DTFT is a continuous and periodic function of frequency, whereas the DFT provides discrete samples of one cycle. Operation on a finite N point sequence,
The corresponding inverse DFT (IDFT) is
A more simplified formulation for DFT can be represented as
with the IDFT given by
where
As the sampled Fourier transform, the DFT does not contain all frequencies forming an image, while only a set of samples which is large enough to fully describe the spatial domain image. If the original sequence is one cycle of a periodic function, the DFT provides all the non-zero values of one DTFT cycle.
N | Standard DFT | FFT |
---|---|---|
2 | 4 | 1 |
4 | 16 | 4 |
8 | 64 | 12 |
16 | 256 | 32 |
32 | 1024 | 80 |
64 | 4096 | 192 |
128 | 16384 | 448 |
In 1965, Cooley and Tukey proposed an efficient calculation algorithm of DFT that reduces the computational complexity to
Signal truncated by time window results in spectral leakage, which causes the signal levels to be reduced and redistributed over a broad frequency range, and is important for implementing DSP properly. The flow chart of the DFT calculation with various processing techniques involved is shown as in
The sampling signal length is not integer multiple of the signal period length. Through the DFT, the sampling signal will result in larger distortion and cause unpredictable spurious components and spectral leakage. Assume that the signal period is
Windowing is implemented by multiplying the input signal with a window function. Windowing amplitude modulates the input signal so that the spectral leakage is evened out. Thus, windowing reduces the amplitude of the samples at the beginning and end of the window. An input signal can be of any number of dimensions and can be complex.
As data length is increased, the rate of fluctuation in this is also increased. The variance of the periodogram is reduced by applying a window to the autocorrelation estimate to decrease the contribution of unreliable estimates to the periodogram. Increasing the width of window function is equivalent to enlarging the length of segment with large main-lobe-to-side-lobe ratio, which results in less side-lobe leakage. The spectrum resolution is improved since the main-lobe width becomes narrower as the length (N) of the data is increased. However, decreasing the main-lobe width also reduces the suppression capability on spectrum leakage since the energy on the side lobes is also increased.
The method of window functions has been very popular in DSP for inhibition of spectral leakage by cutting the signal when the period is delay and two-side is continuous. Many different window functions have been developed for truncating and shaping a length-N signal segment for spectral analysis. Frequency domain function in representation of window function is calculated through
A rectangular window is a function that is constant inside the interval and zero elsewhere, which provides information about every component of frequency within the range. The simple direct truncation window has a periodic sinc DTFT, which has the narrowest main-lobe width,
The corresponding frequency domain can be expressed as
Rectangular window has a narrow main lobe and wide side lobes with small main lobe to side lobe ratio, resulting in serious side lobe leakage. The minimum stopband attenuation is 21 dB.
Triangular windows in time domain are given by either of the followings:
or
which can be represented by the unified formulation
where
The main-lobe width for the triangular window is
As a raised-cosine window, the Hanning window takes the form:
with the corresponding frequency domain defined as
It has a main-lobe width considerably larger than that of the rectangular window, but with much lower largest side-lobe peak, at about −31 dB. The side lobes also taper off much faster. For a given length, it is worse than the boxcar window at separating closely-spaced spectral components of similar magnitude, but better for identifying smaller-magnitude components at a greater distance from the larger components. It has a main-lobe width about
Like the Hanning window, the Hamming window also belongs to a kind of the raised cosine window, and thus exhibits similar characteristic to the Hanning window, but further suppresses the first side lobe. The Hamming window is defined as
which has the frequency domain
where
The Blackman method is used to reduce variance of the estimator, thus presents improvement in stopband attenuation. As compared to other windows, the Blackman window possesses good characteristics for audio processing, which has the time domain defined as
It has the frequency domain given as
The maximum side lobe for the Blackman window is 57 dB lower than the main lobe, which is three times as that of rectangular window, and the minimum stopband attenuation is 74 dB.
Discovered by James Kaiser, the Kaiser window is a simple approximation of the DPSS (digital prolate spheroidal sequence) window using Bessel functions. The Kaiser window function is given by
where
As a kind of adjustable window function, the Kaiser window provides independent control of the main-lobe width and ripple ratio. The variable parameter
The zero-phase version of Kaiser Window function is given by
Window used | Main-lobe width | Maximum side lobe relative to main lobe (dB) | Minimum stopband attenuation (dB) | |
---|---|---|---|---|
Exact | Approximate | |||
Rectangular | −13 | 21 | ||
Triangular | −25 | 25 | ||
Hanning | −31 | 44 | ||
Hamming | −41 | 53 | ||
Blackman | −57 | 74 | ||
Kaiser | Adjustable | Adjustable | Adjustable | Adjustable |
Shown in
where
where
In Welch’s method, the original data segment is split up into several data segments where there might be overlapping points. A time domain window is applied to the individual data segments after the data is split up into overlapping segments. It is an improvement on the standard periodogram spectrum estimating method and on the Bartlett’s method, for which the following procedures are involved: (1) Split up the original data segment into several non-overlapping segments; (2) Estimate the PSD/Compute the periodogram by computing the DFT for each segment, and then compute the squared magnitude of the result, followed by dividing this by the length; (3) Average over these local estimates for the non-overlapping data segments. The Welch method is carried out by dividing the time signal into successive overlapping segments, forming the periodogram for each block, and averaging, which reduces the variance of the individual power measurements.
The Welch method is to window the data segments prior to compute the periodogram.
where
The Welch PSE is the average of modified periodogram
with the mean value given by
The resolution of PSE is determined by the spectral resolution of each segment which is of length L, which is window independent.
The windowing processing of the segments is what makes the Welch method a modified periodogram.
To investigate the capability for inhibition of spectral leakage, the power spectra are illustrated using various window functions before and also after the Welch method is incorporated. Simulation experiments are conducted to evaluate the performance for various designs. Computer codes were developed using the Matlab® software for testing and verification of the PSE performance.
In the first example, the continuous signal
There are two requirements for the window functions: (1) The main-lobe width should be as narrow as possible for better resolution; (2) The side lobes should be as smaller as possible, so as to increase energy on the main lobe, and increase the stopband attenuation. Resolution represents the ability to discriminate spectral feature and is a key concept on the analysis of spectral estimator performance. Estimation with low spectrum resolution causes the signal easily disturbed by the noise, leading to the failure of detection.
It can be seen that, the rectangular window, with minimum stopband attenuation, has the clear peaks showing in the graph of the power spectrum estimation and has admirable resolution characteristics for sinusoids of comparable strength. Ripples in pass band are more in rectangular as compared to the other windows. Therefore, the rectangular window is generally not recommended due to high side lobes. Both the rectangular and Kaiser windows possess clear peaks showing in the graph of the PSE, and they have narrower main-lobe width in frequency domain. As can be seen, the spectral resolution based on the rectangular window and Kaiser windows is relatively higher, whereas the noise level near the fundamental frequency of the signal is higher. It is considered that both the rectangular and Kaiser windows are suitable for high precision spectrum estimation of signals with high SNR among others. As compared to the other ones the Kaiser window is known as the optimal window since it has maximum attenuation according to the given width of the main lobe.
The advantage of using Hanning window method is that it reduces the side lobes and possesses good frequency resolution over the rectangular window. The Hanning window possesses better suppression capability on spectrum leakage, while the frequency resolution is relatively lower, hence suitable for general frequency estimation of signals with low SNR. The Hamming window is a modification over Hanning window and both present similar characteristics. The difference between the two is that, at the edges the Hamming window does not get as close to zero as does Hanning window. The Hamming window is employed to minimize the maximum nearest side lobes, which possesses a form similar to the Hanning window but with slightly different constants. It has a similar main-lobe width, but with a much lower largest side-lobe peak. However, the side-lobes also taper off much more slowly than that with the Hanning window. For a given length, the Hamming window outperforms the Hanning at separating a small component relatively near to a large component, but worse than the Hanning for identifying very small components at considerable frequency separation.
Hamming and Blackman, and possibly Kaiser (with appropriate selection of adjustable parameters) have larger stopband attenuation. The PSE based on rectangular window has narrowest main-lobe width, meaning that it has best resolution. However, its side lobes are higher than those of others, with larger leakage induced. Additionally, it has larger deviation between side lobes and thus has largest variance. On the other hand, the other windows (such as the Hamming and Blackman, and Kaiser windows, etc.) have wider main-lobe widths, thus possess decreased resolution. However, they have smaller side lobes, with improved suppression capability on spectrum leakage. Comparing other methods, the Blackman has the smallest side lobes but the main-lobe width is the largest. The spectrum resolution of the estimate is improved as the length N of the data increases in exchange for a reduction of suppression capability. Influence of data length to the PSE resolution is shown in
In this subsection, results on PSE utilizing the Welch method with various windows are presented. The simulation employed an input signal composed of three sine waves, with frequencies 250 Hz, 600 Hz, and 1000 Hz, respectively. A Gaussian-white sequence is added into the signal, which has the SNR 10 dB. The sampling frequency is 3000 Hz with the number of sampling points 3000. The data length of FFT is 512. The characteristics of window function are analyzed from spectrum resolution and noise level. Simulation based on the classical periodogram method was carried out and thereafter incorporation of the Welch method into various windows, all with length 512, was performed for PSE. Power spectrum estimation based on the classical periodogram approach is shown as in
Window used | Average noise power (dB) | Spectral leakage mean power (dB) | Improvement based on Welch’s method (dB) |
---|---|---|---|
Rectangular | −26.8947 | −17.8459 | 3.9867 |
Triangular | −28.6105 | −24.2959 | 3.0495 |
Hanning | −28.2051 | −24.3721 | 3.1629 |
Hamming | −28.5355 | −24.9049 | 3.2261 |
Blackman | −27.6137 | −23.1110 | 3.1939 |
Kaiser | −27.0262 | −18.1267 | 3.9376 |
This paper deals with the design of FIR digital filter using various window techniques to extract the unwanted noise from the signal. Profound insight has been provided in the statistical properties of the Welch power spectrum estimator. The effect of data length on power spectral density by the assistant of Welch method in various window functions is shown. Different parameters can be adjusted to meet the specific engineering requirements in the FIR filter design process on the basis of the desired filter characteristics. According to the kind of signal being analyzed a specific window function should be selected. How the spectral leakage is reduced using various window techniques is also involved. The effect of leakage and the role of windowing techniques to show the frequency domain measurements are interpreted. The various windows involved aim at satisfying two conflicting requirements: simplicity of implementation, comparable to that of window-based design methods and accuracy in the fulfillment of design requirements. The ideal window is expected to minimize the maximum nearest side lobes. The quality of the estimate is increased as the length N of the data is increased. A good resolution in PSE may be achieved by using appropriate size of data sample. The rectangular window has the narrower main-lobe width in frequency domain. Various window techniques are employed in this paper, including the rectangular, triangular, Hanning, Hamming, Blackman and Kaiser windows. Each window has its own unique advantages and disadvantages.
Investigation of the spectrum leakage suppression had also been carried out. The performance of DFT method inherently suffers from the spectral leakage, which is caused by the non-coherent sampling in the practical acquisition system. Estimation of the power spectrum was performed using various window functions with the Welch method involved. The superiority on spectrum leakage suppression algorithm was verified by comparing with the traditional spectrum analysis based on the DFT with windowing processing. The Welch algorithm splits the random signal into segments before DFT to reduce the spectral leakage, which uses a modified version of Bartlett method in which the portion of the series contributing to each periodogram are allowed to overlap. The PSE has been performed for variable data length using various windows with the Welch method, which enhances the performance in terms of both resolution and variance. The Welch method is used to find the PSD of a signal with reducing the effect of noise where it can be seen that the clear peak shows in the graph of the PSE. The Welch method enables the performance enhancement on PSE of the power spectral density for signals with less distortion caused by noise.