Sub- Nyquist Reconstruction Final Presentation Winter 2010/2011

Sub-NyquistReconstruction

Final PresentationWinter 2010/2011

By: Yousef Badran

Supervisors:Asaf ElronIna Rivkin

TechnionIsrael Institute of Technology

Project Overview

• Part of the Modulated Wideband Converter project cluster.– A system for sub-Nyquist sampling of multiband

signals.• Sub-module of the reconstruction block. • Implemented on a single NI FlexRio Virtex-5

SX95T FPGA device.– NI LabVIEW interface

Sub-Nyquist Xampling Project Block Diagram

Project Goals

Reconstruction Algorithm outline:1. Covert complex-valued sequences to real-

valued sequences.2. Estimate power spectral density.3. Find band edges.

COMPLEX FFT!!

Input and Output

Input:Sequences z[n] representing spectrum slices.At most N sequences, one per slice.

Spectral support slices s Bnt

lBtjl thetxnz

|)())((][ 2

Problem:Spectral support and carrier frequencies of bands are unknown.A single slice may contain more than one band (at most N/2).A single band might be divided between two slices.

Input and Output

Output:• A sequence of vectors, one per spectral slice.• Each vector describes energy distribution in the slice; areas in the slice that

hold energy.• Requires high resolution vectors.

Welch’s Method for Power Estimation

Algorithm— Divide the time signal into successive, overlapping blocks.— Form the periodogram for each block:

— The Welsh estimate of the power spectral density is given by:

(K= # of blocks)

Contradiction: Maximize M for spectral resolution vs. Maximize K for better averaging results and greater spectral stability.

Typical choice:

21( ) | ( ) |kx kP f FFT x

M

1

1( ) ( )k

k

X xi

S f P fK

_ _M K Length of vector

Spectrum Quantization

Final step: Define band location

1. Find PSD threshold.

2. Unite support regions that are closer than .

3. Prune isolated regions with widths smaller than .

bandsbetweenspacingMinimalwidthbandMinimalB

min

min

min

minB

Implementation Demands

Demands

– 18-bit signed fixed-point representation (input).

• 17 bit word length

• 1 bit fraction length

– Input rate of 20 MHz

– Complex FFT

– Implementation on a single FPGA

• Use 30% of FPGA resources.

Project Block Diagram

Low rate sequences

CTF module FIFO

20 MHZ

Power Estimation

Band detection

output

Testing Environment

Host.vi

Host.vi (Write)

Parse data(64bit

double to U32)

Display Input Signal

Host to FPGA FIFO

(WRITE)

2048 elements

1024 Samples

data

Generate Sequence

MU

X

Select

1 MHz rate

Complex data

Real data

Img data

Join numbers(U64bit)

Host.vi (Read)

Write to file

Display Output Signal

FPGA to Host FIFO

(READ)

2048 elements

1024 Samples(U64 bit) 1024 Samples

(double)

DBL

WELCHDisplay and

compare results

FPGA

Real

Split numbers

(I32)

Host to FPGA FIFO

(READ)

2048 element

s

FPGA FIFO

1024 element

s

FPGA to Host FIFO

(WRITE)

2048 elements

FPGA FIFO

1024 elements

FXP

FXP

Img.Welch Power

Spectral Estimation

Spectrum Quantization

Welch PSD – Block Diagram

Data out

en

CounterMemory

Read Addr.

Data outData in

Write Addr.

Hamming Window

Real data in

Img. data in

Real data

Img. dataComplex

FFT

Real data

Img. data

Out validShift reg

Quantization – Block Diagram

Welch’s Method for Power EstimationComparison between MATLAB’s PWELCH, and my own: (M=256, R=128)

ResultsComplex FFT size = 1024.

Welch’s Method for Power EstimationComparision between MATLAB’s PWELSH, and my own: (M=256, R=128)

ResultsComplex FFT size = 1024.

Mathematical Tools

Project WorkflowAlgorithm

• Resources & requirements• Understand theory and reference implementation• Propose alternative and equivalent calculation

MATLAB

• Floating-point representation• 18-bit fixed point representation• Simulink FXP FFT model

• Find an alternative fixed point FFT Solution• Find implementations of required mathematical operations in VHDL

NI-LabVIEW

• Creating a block diagram

• Implement algorithm on the host

LabVIEW FPGA

• Integrate system on LabVIEW FPGA

• NI-FlexRIO Virtex-5 SX95T model.

Debug

• Debug system

• Timing considerations

• Compare results

Resource Estimation

New Device Utilization ( NO HOST/FPGA FIFOS)

Total Slices 3292 14720 22.36

Slice Registers 10626 58880 18

Slice LUTs 13869 58880 23.5

Block RAMs 24 244 9.8

@ 40MHz, FFT size of 1024

Next Steps

• Detect bands and isolate them.

• Estimate carrier frequency per band.• 6 low rate sequences (20MHz) instead of 1.

Next Steps (Cont.)

Serial FFT!

Low rate sequences

(20MHz each) FIFO1024

elementsFIFO1024

elementsFIFO1024

elementsFIFO1024

elementsFIFO1024

elementsFIFO1024

elements

Complex FFT&

Hamming Window

PSD&

Band DetectionM

UX

High rate Calculation 120 MHz FIFO1024

elementsFIFO1024

elementsFIFO1024

elementsFIFO1024

elementsFIFO1024

elementsFIFO1024

elements

Thank you!