Automatic Modulation Classification in Practice

Automatic Modulation Classificationin Practice

Guide: Prof. Saravanan Vijayakumaran

Sreeraj Rajendran,10307933,

IIT Bombay

June 21, 2013

Intro AMC Algorithms SDR Implementation Conclusion extra

Contents

I Motivation and Problem statement

I AMC in Literature and Challenges

I AMC Algorithms

I SDR architecture

I Implementation Details

I Future work

2 / 32


Motivation

I Numerous publications: No practical implementation inopen literature

I Importance in defence applications

I Cryptanalysis can be done only after demodulation anderror correction

I Modulation classification scenariosI Nearly all parameters of the single signal is knownI No parameters are known in advance

3 / 32


Decision Theoretic Approach

I Classifiers presented in literature assume that signalparameters are known

I Maximum Likelihood ClassifierI Signal classified from likelihood ratios from the matched filter

outputI The most likely candidate is given by the modulation scheme

which provides the maximum value for the likelihood ratio

I Practical implementations suffer from computationalcomplexity

4 / 32


Feature Based Detection

I Amplitude, phase and frequency features are used

I ExamplesI FM and AM : Can be distinguished by analysing amplitude

I Suboptimal but easily implementable

I Signal CharacteristicsI Analog: Sample Based, e.g FM, AMI Digital: Symbol Based, e.g MPSK, QAM

I Example Signal Features

Modulation Scheme Amplitude Phase Frequency

FM Constant Continuous Continuous

AM Variable Continuous Continuous

FSK Constant Continuous Discrete

QAM, MPSK - Discrete -

5 / 32


Why these algorithms fail?

I Not robust in the presence of frequency or phase offsets

I Assumptions made do not hold in practical scenarios.

I Constellation Demo

6 / 32


Problem Statement

I Implementation of a practical realtime AMC tool

I Practical algorithms to classifyI OFDM, AM, DSB, SSB, FM, ASK, PAM, PSK (MPSK),

QAM, FSK (MFSK), or CPM

I AssumptionsI No prior knowledge of carrier frequency, symbol rate, pulse

shaping function, frequency deviation, symbol constellation ormodulation index

I For CPM, modulation index is assumed to be an integer or0.5× integer

7 / 32


OFDM Classification

I Structure of OFDM signallingNS

NG ND

I For a received signal y[n]

E[y[n]y∗[n+ ∆]] =

σ2s + σ2

n ∆ = 0,

σ2sej2πξ∆tND ∆ = ND,

0 otherwise.

ND

8 / 32


ND Estimation

I Received signal assumed to have D samplesI Discrete correlation represented as

Ry(∆) =1

D −∆

D−∆∑n=1

y[n]y∗[n+ ∆]

I With proper averaging

Ry(∆) =

σ2s + σ2

n ∆ = 0,NG

ND+NGσ2sej2πξ∆tND ∆ = ND,

0 otherwise.

I ND estimation formulated as

ND = argmax∆|Ry(∆)|

9 / 32


ND Estimation

I Discrete correlation of a received OFDM signal with datalength 512.

10 / 32


Cyclic Prefix Length (NG) Estimation

I CP length chosen as a multiple of symbol duration ND

I Less correlation value for NG

ND

I High correlation vale for NG

ND

11 / 32


Likelihood function for OFDM classification

I The likelihood function evaluated at the estimated valuesis given by

γ = Λ(y;ND, NG, θ)|ND,NG,θ = Λ(y; ND, NG, θ)

γ =

⌊D

ND+NG

⌋NGσ

2s H1,

0 H0.

I For 0dB SNR the value of threshold

ζ =σ2s + σ2

n

2

⌊D

ND +NG

⌋NG

Reference: Tevfik Yucek and Huseyin Arslan “OFDM signal identification and transmission parameter estimation

for cognitive radio applications”, IEEE GLOBECOM 2007 proceedings

12 / 32


Cyclic Moments

Definition

For a discrete time input y(n) the time-varying kth order momentis defined as

mky(n; τ ) , Ey(n)y(n+ τ1) . . . y(n+ τk−1)

Definition

The process y(n) is said to be kth order cyclo-stationary ifmky(n; τ ) is periodic and has a Fourier Series expansion

13 / 32


Joint Classification of Analog and Digital modulation

I Used moments for classification

Mα1y , lim

N→∞

1

N

N−1∑n=0

E[y(n)]e−j2παn

Mα2,0y , lim

N→∞

1

N

N−1∑n=0

E[y2(n)]e−j2παn

Mα2,1y , lim

N→∞

1

N

N−1∑n=0

E[y(n)y∗(n)]e−j2παn

14 / 32


Modulation schemes and cyclic frequencies

Mα1y Mα

2,0y Mα2,1y

AM fc 2fc N/A

DSB N/A 2fc N/A

LSB N/A N/A N/A

USB N/A N/A N/A

NBFM f∗∗c N/A N/A

WBFM N/A N/A N/A

MASK fc + n×Rs 2fc + n×Rs ♦ n×Rs;n 6= 0=

MPAM, PSK2 N/A 2fc+ n×Rs ♦ n×Rs;n 6= 0=

MPSK(M≥4), QAM N/A N/A n×Rs;n 6= 0=

MFSK

fc+0.5× (2m− 1−M)× fd,m = 1 . . .M

2fc+0.5× (2m− 1−M)× fd,m = 1 . . .M

N/A

CPM: h = integerMultiplecycle frequencies

Multiplecycle frequencies

N/A

CPM: h = 0.5× odd N/AMultiplecycle frequencies

N/A

15 / 32


Cyclic moment estimator

I Cyclic moment Mαky(τ ) can be estimated by

Mαky(τ ) =

1

N

N−1∑n=0

fky(n; τ )e−j2παn

fky(n; τ ) , y(n)y(n+ τ1) . . . y(n+ τk−1)

I Tests developed Dandawate and Giannakis in [14] can beused to detect the presence of cycles without theknowledge of input data distribution

16 / 32


Tests for the presence of cyclostationarity

I Cyclic frequencies and test statistics for conjugatedsecond order moment

17 / 32


Classification Flowchart

Input: OFDM, AM, DSB,

SSB, FM, ASK, PAM,

PSK, QAM, FSK, or CPM

γ > ζClassify

as OFDM

yes

Mα21y has a

cycle

frequency

Classify in group

ASK, PAM, PSK,

QAM, FSK, CPM

Classify in group AM,

DSB, SSB, FM, FSK, CPM

no

yes

no

A

B

18 / 32




SSB, FM, ASK, PAM,


γ > ζClassify

as OFDM

yes

Mα21y has a

cycle

frequency

Classify in group

ASK, PAM, PSK,

QAM, FSK, CPM



no

yes

no

A

B

18 / 32




SSB, FM, ASK, PAM,


γ > ζClassify

as OFDM

yes

Mα21y has a

cycle

frequency

Classify in group

ASK, PAM, PSK,

QAM, FSK, CPM



no

yes

no

A

B

18 / 32


Classification Flowchart2

A

Number of

CF of Mα1y

Classify

as FSK

Classify

as ASKNumber of

CF of Mα20y

Classify

as CPM:

h = 0.5× odd

Classify as

PAM, BPSK

Classify

as QAM,

MPSK:

M ≥ 4

≥ 2

= 1

= 0

≥ 2= 1

= 0

19 / 32


Classification Flowchart3

B

Number of

CF of Mα1y

Classify

in group

AM, DSB,

SSB, FM

Classify as

MFSK,CPM:

h = integer

Classify

in group

DSB, SSB,

FM, CPM:

h = 0.5× odd

Is Max of

Mα20y a CF

Number of

CF of Mα20y

Classify as

AM,DSB

Classify

as LSB,

USB, FM

Classify

as CPM:

h = 0.5× odd

Classify

as DSB

= 0

≥ 2

= 1

yes

no

= 0

≥ 2

= 1

20 / 32


Radio in Software

I Communication modules implemented in software

I GNURadioI Applications written as signal flow graphsI Signal processing blocks written in C++ and converted to

PythonI Modular ArchitectureI GUI: GNU Radio Companion (GRC)

21 / 32


USRP (Universal Software Radio Peripheral)

FPGA

ADC

ADC

ADC

ADC

DAC

DAC

DAC

DAC

USBController

Host P C

ReceiveBoard

TransmitBoard

ReceiveBoard

TransmitBoard

22 / 32


First Step: Spectrum Sensing

I Scan the interested frequency range by tuning the receiverfront-end

I Usually averaged FFT is good enough for locating a signal

I Energy detection based on Power Spectral Density (PSD)

I Welch Estimate: For close estimate of PSD

23 / 32


Block Schematic

USRP Downconverter Filter AMC Module

GUI Frontend

24 / 32


AMC Module

ResamplerStream

to vectorDecimator

OFDM

Classifier

Stream

to vectorDecimator FFT block Averager Mα

1y Classifier

Squaring blockStream

to vectorDecimator FFT block Averager

Mα20y

Classifier

Complex to

magnitude

squared

Stream

to vectorDecimator FFT block Averager

Mα21y

Classifier

I Tool Demo

25 / 32


Conclusions and Challenges

I Real-time classification is achievableI 7000 lines of Python codeI Good understanding of large GNURadio codebaseI Understanding various papers to find ones suitable for

implementation.

I Practical fine tuning is essential for various parameterslikeI Signal strength thresholdI Window length parameter and threshold for cyclic frequency

testsI Decimation factors to match processing speed of PC

I Automatic demodulation requires more parameterestimation

26 / 32


Future Work

I Test AMC tool on various real life wireless signals

I Implement automatic demodulation modules

I Detect packet structure of the transmitted digitalmodulation data

I Detect the presence of error correcting codes

27 / 32


References I

Wikipedia SIGINT, http://en.wikipedia.org/wiki/Signals_intelligence.

Tevfik Yucek and Huseyin Arslan “OFDM signal identification and transmission parameter estimation for

cognitive radio applications”, IEEE GLOBECOM 2007 proceedings.

Jeffrey H.Reed, “Software Radio: A modern approach to radio engineering”, Prentice Hall PTR, 2002.

Ettus webpage, http://www.ettus.com/resource.

GNURadio webpage, http://gnuradio.org/.

O.A. Dobre, A. Abdi, Y. Bar-Ness and W. Su, “Survey of automatic modulation classification techniques:

Classical approaches and new trends”, IET Commun, 2007. pp. 137-156.

J.A. Sills, “Maximum-Likelihood modulation classification for PSK/QAM”, Proc. IEEE MILCOM, 1999, pp.

57-61.

William A. Gardner, “Introduction to random process with applications to signals and systems”,

McGraw-Hill Inc, 1986.

William A. Gardner, “Signal Interception: A unifying theoretical framework for feature detection”, IEEE

Transactions on Communications, Vol. 36, No. 8, August 1988.

William A. Gardner, “Spectral correlation of modulated Signals: Part I-Analog Modulation”, IEEE

Transactions on Communications, Vol.Com-35. No. 6, June 1987.

William A. Gardner, “Spectral correlation of modulated Signals: Part II-Digital modulation”, IEEE

Transactions on Communications, Vol. Com-35. No.6, June 1987.

28 / 32

http://en.wikipedia.org/wiki/Signals_intelligence

http://www.ettus.com/resource

http://gnuradio.org/


References II

Anantha Krishna Karthik N, “Blind modulation classification and symbol rate estimation in Rayleigh fading

multipath environment with frequency and timing Offsets”, Master thesis IIIT, Hyderabad , 2005.

Zaihe Yu, Yun Q Shi and Wei Su, “Automatic classification of joint analog and digital communication

signals in blind environment”, MILCOM 2007, IEEE.

Amod V Dandawate and Georgios B Giannakis, “Statistical tests for presence of cyclostationarity”, Signal

Processing, IEEE Transactions, Volume:42, Issue:9, Sep 1994.

Zaihe Yu, “Automatic modulation classification of communication signals”, Dissertation, Department of

Electrical and Computer Engineering, New Jersey Institute of Technology, August 2006.

Fred Harris, “Let’s assume the system is synchronized”, Globalization of Mobile and Wireless

Communications, pp.311-325, 2011.

Petre Stoica, Randolph Moses, “Spectral analysis of signals”, Pearson Prentice-Hall.

29 / 32


THANK YOU

30 / 32


ML Classifier

I

rn = rI,n + jrQ,n

I Pr(rI,n, rQ,n/m = mi) can be expressed as

1

M(i)

M(i)∑k=1

1

2πσ2e

−(rI,n −√EsµI,k)

2 − (rQ,n −√EsµQ,k)

2

2σ2

I The governing PDF of a vector r of length N is

Pr(rI , rQ/m = mi) =N∏n=1

Pr(rI,n, rQ,n/m = mi)

I The most likely candidate is given by

arg max1≤i≤M

Pr(rI , rQ/m = mi)

31 / 32


NG, θ = arg maxNG,θ

Λ(y; ND, NG, θ)

= arg maxNG,θ

∣∣∣∣∣∣∣∣∣∣∣

⌊D

ND+NG

⌋∑m=0

NG∑p=1

y[m(ND +NG) + p + θ] y∗[m(ND +NS) + ND + p + θ]

∣∣∣∣∣∣∣∣∣∣∣

Definition

The process y(n) is said to be kth order cyclo-stationary ifmky(n; τ ) is periodic and has a Fourier Series expansion

mky(n; τ ) =∑

α εΩm,k

Mαky(τ )e

j2παn

Mαky(τ ) , lim

N→∞

1

N

N−1∑n=0

mky(n; τ )e−j2παn

32 / 32

Documents

Automatic Modulation Classification in Practice