PSD Computation Using Modified Welch Algorithm
Uploaded by
ijsretPSD Computation Using Modified Welch Algorithm
Uploaded by
ijsretInternational Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 9, September 2015
PSD Computation using modified Welch algorithm
Akhilesh Krishna P.R, Jency Andrews
1
Electronics and Communication, M.G University, Kottayam
2
Electronics and Communication, M.G University, Kottayam
ABSTRACT
Power Spectral Density (PSD) shows strength of energy
as a function of frequency. PSD computation can be
done by the method called FFT. Welch algorithm is used
to calculate Power Spectral Density. High computational
complexity is the main drawback of Welch method. By
using 50% overlap we are reducing the computational
complexity of Welch method. Here N/2 point FFT is
computed instead of N-point FFT and we are using
Windowing in frequency domain for high performance.
In this approach even samples are computed exactly but
odd samples required half sample delay shift. So,
overlap method reduces the computational complexity as
half of the normal method. To reduce area, pipeline stage
is introduced instead of parallel structure.
Keywords - Power Spectral Density, Low complexity,
Low power, Welch method, Windowing, Fast Fourier
Transform.
1. INTRODUCTION
PSD is used for analysis and tracking the signal. PSD
shows input signal over range of frequencies. PSD is
divided into two groups Parametric and Non-parametric.
Parametric generate underlay signal and it consist of
different methods such as, autoregressive moving
average(ARMA), model identification, minimum
variance distortion less response method (MVDR) and
Eigen decomposition based methods. Non-parametric
method consist of two methods periodogram based
method and multiple-window method. The periodogram
methods include the computation of the Fourier
transform of the signals. Welch PSD method is to
estimate the power spectral density and is a nonparametric approach based on the periodogram method.
This low-complexity architecture suitable for low-power
embedded systems. Parameters such area, performance
and power consumption should be a consideration while
PSD computation into biomedical monitoring systems.
FFT is the main part of the Welch method. Here 50%
overlap is used and dividing the input signal into
multiple segments. Also 50% overlap is used and
dividing the input signal into multiple segments. Half of
the samples are same over two consecutive FFT
functions.
The PSD architecture, which reduce the number of
operations required to compute the PSD. The
architecture needs N/2-point FFT instead of N-point
FFT, where N is the length of the window. When change
time domain into frequency domain, reduced the
operations on the basis of cosine functions. Here the
design merges two N/2-point FFT into N-point FFT.A
bidirectional filter is proposed to create the half-sample
delay filter. The advantages of the this approach are
limited to applications where errors introduced due to
short filters are acceptable.In biomedical signal
processing and some of application where algorithm is
tolerant to errors (noise).
PSD estimation based on the Welch method which
involve dividing the signal into multiple segments, then
the modified periodograms of these segments and
averaging these modified periodograms.The normal
algorithm is described as follows,
The input signal X (n) is divided into N
overlapping segments.
The specified window is applied to each
segment.
FFT is applied to the windowed data.
Each periodograms of new windowed data
segment is computed which is modified
periodogram.
Taking the average of periodograms to obtain
spectral density.
2. LOW COMPLEXITY PSD
To calculate the FFT of the individual non overlap parts
and obtain the FFT of the overlapped segments by
combining those of the non-overlapped segments. The
window function will be different for these nonoverlapped parts. So, the windowing operation has to be
performed in the frequency domain which is more
reliable. It consists of mainly two steps, windowing
operation in frequency domain and merges two N/2point FFT.
2.1. Windowing in frequency domain
To take advantage of the overlap segments, window
function will be applied in frequency domain instead of
time domain.
www.ijsret.org
951
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 9, September 2015
N/2 FFT and Window filter is used. The FFT consists of
a butterfly structure and FIR filter. Even samples are
computed by butterfly structure while odd samples are
computed by FD filter.
Fig. 1 Segmentation of overlapped segments [1]
The convolution operation is complex, compared to
simple multiplication operation. The window function
used in the PSD computation which represented cosine
functions and this can be shown by 3 or 5 non-zero
coefficients in the frequency domain.
2.2. Merging of 2 N/2-point FFT
To compute N-point FFT, merging two N/2-point FFT
instead of N-point FFT. So we can reduce the
complexity. Overlapping and non overlapping segments
are combining to form N-FFT.
Fig. 3 Flow chart of Welch method [1]
Fig. 2 Combination of 2 N/2-point FFT [1]
To compute PSD using the original Welch method,
addition and multiplication are required and in the
overlap approach, where N is the size of the FFT and N
is the length of the fractional-delay filter. It can be see
that the overlapping approach requires less number of
calculations compared to the existing Welch method.
3. WELCH ALGORITHM
New method for PSD [1] computation is described as,
The input is divided into (L+1) non overlap
segment where L is the number of segments.
An N/2-point FFT is applied to segment.
N-point FFT is computed by combining two
N/2-point FFT.
Windowing is applied in frequency domain.
The periodogram of each window segment is
computed.
Taking the average of modified periodogram.
The main components are FFT and absolute-square
multiple accumulator (AMAC) circuit. This is actually
used to compute FFT.AMAC circuit is to compute the
periodogram and average them over segment. AMAC
block, which stores the values of L different
periodograms. To correctly accumulate the periodogram
outputs the control block which controls the address
decoder and multiplexer. In the case of overlap method
Fig .4 Filter circuit for windowing operation [1]
The power consumption will be less even number of
addition operations are more in the case of higher-order
FD filters. This is due to the reduction in the number of
multiplication operations that will case the reduction in
the size of the FFT that needs to be computed.
3.1. Review of STFT Computation
The short-time Fourier transform [1] with a 50% overlap
can be computed as follows,
The input signal X (n) is separating into nonoverlapping segments of length N/2.
www.ijsret.org
952
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 9, September 2015
N/2-point FFT is applied to each segment.
Two consecutive N/2-point FFT are merging to
form N-point FFT.
Window function is applied to new segment.
The short-time Fourier transform does not need to
calculate average over multiple segments.
4. PROPOSED METHOD
So from the literature study done, the existing PSD
computation could be improved but using a different
multiplier. Here the overlap architecture, method uses a
pipelined digital slicing multiplier is used instead of
parallel digital slicing multiplier. Using parallel
multiplier would able to process multiple inputs and fast
processing of inputs gives a better yield design than the
existing architecture. So it is more reliable. As the
design was implemented and synthesized it was proved
with better results. In the new method ROM allowed to
perform the multiplication process without any real
multiplier.Area and complexity which can be reduced by
using pipelined architecture.
complexity is reduced at the cost of slight performance
loss. The proposed approach computes the even samples
exactly.
METHODS
AREA(s)
DELAY(ns)
EXISTING
METHOD
6527
16.29
MODIFIED
METHOD
1215
18.33
Table.1Simulation results
Simulation Waveforms
Fig .6 PSD computations using existing Welch method.
Fig .5 Digital slicing [2]
5. IMPLEMENTATION AND RESULTS
Proposed architecture was implemented in VHDL-HDL
on Xilinx Spartan (xc7va100t-3csg324) FPGA as the
target using Xilinx 13.2 Synthesis and Implementation
Tool. Veried its functionality using different test
vectors using Xilinx ISim Simulator. Implementation
followed the step explained as above. Resource
utilization and maximum frequency of design were
observed with the existing architecture. In the case of
modified method used a pipelined architecture instead of
parallel architecture. The new approach will reduce area
and complexity. Slight performance loss is the
disadvantage of the system. This paper has proposed
area reduction technique to compute Power spectral
density using the Welch method. The computational
Fig.7 PSD computation using pipelined architecture.
www.ijsret.org
953
International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 4, Issue 9, September 2015
6. CONCLUSION
In the presented paper designed a pipelined architecture
instead of parallel architecture in Welch PSD calculation
which gives better complexity reduction than parallel
design. The main advantage of this design is, it reduces
area. Similarly, it could increase maximum throughput
by designing in pipelined architecture. Overlapping
(50%) technique is used to reduce complexity of PSD
computation. Another benefit of this design is that it
technology independent. This design gives maximum
performance in FPGA.This paper discussed about
Pipeline Digit-Slicing Multiplier less circuit for FFT
structure. It concluded that FPGA Implementation of
Pipeline Digit-Slicing Multiplier-Less circuit for FFT
Structure is an enabler in solving problems that affect
communications capability in FFT and possesses huge
potentials for future related works and research areas.
REFERENCES
[1]Keshab K. Parhi, Fellow, IEEE, and Manohar
Ayinala, Member, IEEE, Low Complexity Welch
Power Spectral Density Computation. Vol. 61, No. 1,
January 2014.
[2]Yazan Samir Algnabi,RozitaTeymourzadeh, Masuri
Othman, MdShabiul Islam, FPGA Implementation of
Pipeline Digit-Slicing Multiplier-Less Radix 22 DIF
SDF Butterfly for Fast Fourier Transform Structure.
UniversitiKebangsaan Malaysia, 43600 Bangi, Selangor,
Malaysia.
[3] Y. Zhou, J. M. Noras, and S. J. Shephend, "Novel
design of multiplier-less FFT processors," Signal Proc.,
vol. 87, pp. 1402-1407, 2007.
[7] L. Yang, K. Zhang, H. Liu, J. Huang, and S. Huang,
"An Efficient Locally Pipelined FFT Processor," IEEE
Transactions on Circuits and Systems II: Express Briefs,
vol. 53, pp. 585 - 589, 2006.
[8] D. P. Kolba and T. W. Parks, "A prime factor FFT
algorithm using high-speed convolution," IEEE Trans
Acoust. Speech, Signal Process, vol. 25, pp. 281-294,
1977.
[9]M. Ayinala and K. K. Parhi, FFT architectures for
real-valued signals based on radix-23 and radix-24
algorithms, IEEE Trans. Circuits Syst.I Reg. Papers,
vol. 60, 2013.
[10] M. Garrido, K. K. Parhi, and J. Grajal, A pipelined
FFT architecture for real-valued signals, IEEE Trans.
Circuits Syst. I, Reg. Papers, vol. 56, no. 12, pp. 2634
2643, Dec. 2009.
[11] M. Ayinala, M. Brown, and K. K. Parhi, Pipelined
parallel FFT architectures via folding transformation,
IEEE Trans. VLSI Syst., vol. 20, no. 6, pp. 10681081,
June 2012.
[12] K. K. Parhi, VLSI Digital Signal Processing
Systems: Design and Implementation.Hoboken, NJ,
USA: Wiley, 1999.
[13] P. D. Welch, The use of fast Fourier transform for
the estimation of power spectra: A method based on time
averaging over short, modified periodograms, IEEE
Trans. Audio Electro acoustics, vol. AU-15, pp.7073,
Jun. 1967.
[4] B. Mahmud and M. Othman, "FPGA implementation
of a canonical signed digit multiplier-less based FFT
Processor for wireless communication applications," in
ICSE2006 Proc Kuala Lumpur, Malaysia, 2006, pp. 641645.
[5] S. He and M. Torkelson, "A new approach to
pipeline FFT processor," in Parallel Processing
Symposium, Proceedings of IPPS '96, The 10th
International Honolulu, HI, USA, 1996, pp. 766 770.
[6] S. He and M. Torkelson, "Designing pipeline FFT
processor for OFDM (de)modulation," in International
Symposium on Signals, Systems, and Electronics
(ISSSE'98), 1998, pp. 257 262.
www.ijsret.org
954
