Detection of Array Signal Number With Multiple Sensors Based on Transfer Component Analysis

Jiaqi Zhen

doi:10.1109/access.2019.2939865

Abstract

The conventional algorithms for estimating number of array signals are only suitable for the background of Gaussian white noise, and need many snapshots, but their performance will reduce seriously in the circumstance of impulse noise and small samples. Therefore, a new method of detecting array signal number with multiple sensors based on transfer component analysis is proposed in this paper. First, the array signals in Gaussian white and impulse noise are respectively modeled. Then the received array data are transformed into a common hidden space by the mapping function, thus, data in the hidden space have the same distribution, and most initial characteristics are retained. Finally, a support vector machine or K-means clustering are used for classifying the mapped data into two categories, on this basis, the array signal number can be estimated.

Highlights

Signal number estimation is always one of the hottest topics in array signal processing, it is often the precondition of further processing [1]–[8], the research dates from the late 1950s
The information theory criteria, such as modeling by shortest description(MDL) [10] and Akaike information criterion(AIC) [11] are widely concerned with their good estimation performance, but they are only suitable for the background of Gaussian white noise (GWN), in order to solve this problem, gerschgorin disk method [12] was proposed in the colored noise, some modified techniques are presented successively
A support vector machine(SVM) is obtained through training with the mapped data, on this basis, the array signal number can be estimated by classifying the source and impulse noise

Summary

INTRODUCTION

Signal number estimation is always one of the hottest topics in array signal processing, it is often the precondition of further processing [1]–[8], the research dates from the late 1950s. The information theory criteria, such as modeling by shortest description(MDL) [10] and Akaike information criterion(AIC) [11] are widely concerned with their good estimation performance, but they are only suitable for the background of Gaussian white noise (GWN), in order to solve this problem, gerschgorin disk method [12] was proposed in the colored noise, some modified techniques are presented successively Another kind of method is based on Bootstrap, its essence is the resample process to the received signals, in 2000, Brcich et al [13] first put forward the method to estimate source number through constructing hypothesis test statistics by Bootstrap. We can use K-means clustering for the classification

SIGNAL MODEL

RESULTS

CONCLUSION