Classification of R.F Spectrum Measurements Principal Investigator : Dr. Gary J Minden ([e-mail]) Student Investigator : Dinesh Datla

1
Classification of R.F Spectrum Measurements
Principal Investigator Dr. Gary J Minden
(gminden_at_ittc.ku.edu)
Student Investigator
Dinesh Datla
Research Overview
Iterative one-sided hypothesis testing

• Problem Statement Given , the measurements
  collected from Spectrum over time t frequency f
  ,
• To classify M(fi,tj) as a signal or noise
  measurement
• Applications -
• 1) Quantify efficiency of spectrum usage
• 2) Assess feasibility for usage of cognitive
  radios
• 3) Develop models of spectrum usage patterns
• ( As Part of NRNRT , a NSF funded project
  )

• Assume that the band contains more noise than
  signal. View the spectrum as having a Gaussian
  distribution of power measurements.
• Classify measurements on far right of the
  Gaussian distribution ,p standard deviations away
  from mean , as signal and discard the signal
  portion.
• Iteratively apply above procedure to remaining
  unclassified measurements, till standard
  deviation of the Gaussian PDF doesnt reduce
  further.
• Adaptive version of this algorithm classifies
  measurements using a local threshold.

Results Performance Evaluation
Threshold based segmentation
P(Classification Error) P(Miss) P(False
Alarm) Where
Miss Signal
wrongly classified
as noise ,
False Alarm noise wrongly
classified as signal


Performance for P 1.6449
P(Miss) 2.7

P(False Alarm) 34.4

• Histogram Based thresholding is the simplest
  example The minimum point in the valley between
  bimodal peaks in the histogram is taken as
  threshold

Bimodal Histogram of spectrum image pixel gray
values
Spectrum Image - (a) segmented
(b) original
Gaussian PDF after each iteration