RFI Classification using Neural Networks

madbrainedmudlickAI and Robotics

Oct 20, 2013 (3 years and 10 months ago)

96 views

RFI Classification using Neural Networks

Mike McCarty

and

Mark Whitehead

(NRAO)

Introduction

Radio frequency interference with single dish telescopes observations is a well
-
known problem
10,11

which has been restricting the scientific output of telescopes such as the GBT. While some RFI can be
avoided through mitigation techniques such as “Quiet Skies” legislation, RFI will continue to persist due
to

satellite

and

other non
-
ground emissions.

Removal of R
FI from the data is, however, a

difficult
problem. Many types of RFI are predictable and can be

classified and removed from

data provided a
good description of the
signal
properties is known, allowing for
an

excision algorithm to find the pr
e
-
identified RFI lines and remove them from the data. This technique works extremely well but can be
extremely time and equipment consuming for the initial analysis of the signal
10,11
. The other common
technique for RFI removal is to look for anomalies
within the data and flag those anomalies as RFI.
Again this technique works well, and is in fact common
ly

use
d
, but it often leaves behind a small
signature from weak or improperly removed RFI.

One possible means for more
effectively

and efficiently re
moving RFI from radio astronomy data is
through the application of neural network classification.
Recent work in radio astronomy demonstrates
the efficacy of applying neural network classification to automate radio astronomy data analysis
2
-
9
.


Here, w
e
propose to produce a prototype

neural network classification system to identify and flag RFI in
archived GBT data
. We plan to use existing research and data to construct an RFI classification model,
implement the model via a back propagation neural networ
k and apply the network to existing data to
quantitatively
determine if this approach can detect RFI in radio wavelength observations
. If successful,
this technique could be used
as part of the standard
NRAO
data reduction techniques to open up
frequency s
pace otherwise not useful to astronomers
.

Activities and Goals

We will use
archival, high time resolution, GBT data combined with the techniques outlined in
Fisher

and
Singhal

10,11

to empirically characterize RFI

in the 700
-
900 MHz frequency range
.

(The f
requency range is
chosen such that a successful technique will be
immediately

usable for the HI Intensity Mapping
project.)
This will allow us to identify a set of representative parameters that describe RFI in
observational data. The parameters

will then

be used to i
dentify an input vector and document any
required pre
-
processing procedures
9
. The scope of this part our work will be limited to spectral line data
from the GBT.


Next we will evaluate available packages and l
ibraries useful for constructing

classifiers. Some of the
packages we have identified to date are Fast Artificial Neural Network Library (FANN)
12
, neuralnet
13
, and
annie
14
. Once we select a package, we will begin implementing the preprocessing procedures to
produce an input vector, from

observational data, in a format suitable for the chosen library and
programming language.


Finally, we will construct prototype classification models and evaluate them to determine a useful
classification method and neural network topology. Once the clas
sification model is constructed using
actual observational data for training,

we will test its accuracy using test data sets. Both training and
testing data set
s

will have been manually identified as ei
ther “clean spectrum” or “spectrum containing
RFI”
.
Standard accuracy measurements found in classification literature
9,15

will be calculated using the
test data and neural network output.

We will
publish

our process a
nd results in a technical memo. These results could be used to define a
strategy for a larg
e
-
scale neural network to analyze GBT data and potentially a
product which
, much like
the GUPPI Pulsar Backend, may be shared or sold as a turn
-
key package to other radio observatories.


References


[1] McCarty, M.; Artifical Neural Network Applications in

Astronomy;
http://www.gb.nrao.edu/~mmccarty/ann_astronomy.pdf

[
2
] Eatough, R.P.; Molkenthin, N.; Kramer, M.; Noutsos, A.; Keith, M. J.; Stappers, B. W.; Lyne, A. G.; Selection
of radio pulsar candidates using artificial neural networks, 10/2010,
http://ad
sabs.harvard.edu/abs/2010MNRAS.407.2443E

[
3
] David A. Neufeld; Rate Coefficients for the Collisional Excitation of Molecules: Estimates from an Artificial
Neural Network, 1/2010, http://iopscience.iop.org/0004
-
637X/708/1/635

[
4
] Misra, Amit; Bus, S. J.; Ar
tificial Neural Network Classification of Asteroids in the Sloan Digital Sky Survey,
9/2008, http://adsabs.harvard.edu/abs/2008DPS....40.6003M

[
5
] H. U. Nørgaard
-
Nielsen; Foreground removal from WMAP 5 yr temperature maps using an MLP neural
network, 10/20
10, http://arxiv.org/abs/1010.1634

[
6
] S.R. Folkes, O. Lahav, S.J. Maddox; An Artificial Neural Network Approach to Classification of Galaxy
Spectra, 8/1996, http://arxiv.org/abs/astro
-
ph/9608073

[
7
] Yanxia Zhang, Lili Li, and Yongheng Zhao; Morphology cla
ssification and photometric redshift measurement of
galaxies, 4/2008, http://onlinelibrary.wiley.com/doi/10.1111/j.1365
-
2966.2008.14022.x/pdf

[
8
] C. Almeida, C. M. Baugh, C. G. Lacey, C. S. Frenk, G. L. Granato, L. Silva, and A. Bressan; Modelling the
dusty

universe


I. Introducing the artificial neural network

and first applications to luminosity and colour
distributions, 6/2009, http://onlinelibrary.wiley.com/doi/10.1111/j.1365
-
2966.2009.15920.x/pdf

[
9
] Bishop C. M., 1995, Neural Networks for Pattern Recogn
ition. Oxford Univ.

Press, Oxford

[10] Fisher, J. R.; Singhal, A.; A Study of Radar Signals Received by the Green Bank Telescope; 2006;
http://www.gb.nrao.edu/electronics/edir/edir316.pdf

[11] Fisher, J. R.; Signal Analysis and Blanking Experiments on DME Interference; 2004;
http://www.cv.nrao.edu/~rfisher/DME/dme_analysis.pdf

[12]
Günther
,
Frauke
;

Fritsch

Stefan
;
neuralnet: Training of Neural Networks
; 1/2010; R Journal;
http://journal.r
-
p
roject.org/archive/2010
-
1/RJournal_2010
-
1_Guenther+Fritsch.pdf

[13]
Fast Artificial Neural Network Library

website
http://leenissen.dk/fann/wp/

[14] Annie website
http://annie.sourceforge.net/

[15
] Han J., Kamber

M., Pei J., 2012, Data Mining Concepts and Techniques, Morgan Kaufmann