Abstract:
Signal detection theory has long been studied in the contexts of mathematics and engineering. Among its many applications, radar is a notable example in which signal detection theory is applied for the interpretation of the radar echo. Despite its long history, the signal detection problem has received renewed attention in recent years, in the context of “spectrum sensing in cognitive radios”. The concept of cognitive radio is viewed as a potential solution to accommodate the growing demand for the radio spectrum, by efficiently utilizing the under-utilized spectrum. Spectrum sensing is an enabling technology that allows a cognitive radio to detect the presence of a licensed user’s signal. This thesis focuses on blind sensing of low powered wireless signals (corresponding to a signal to noise ratio < -15 dB), while using covariance-based detection (CBD) algorithms. A software defined radio (SDR) based multi-antenna receiver is used to receive multiple signals over its synchronized front-ends. All the existing algorithms considered in this thesis and the newly proposed algorithms are tested using over-the-air wireless signals acquired through this receiver system, and their sensing performance is compared. Such a performance comparison with real signals (rather than simulations), is a major contribution of this thesis. The existing CBD algorithms evaluated in this work are: covariance absolute value (CAV), maximum-minimum eigenvalue (MME), energy with minimum eigenvalue (EME), maximum eigenvalue detection (MED) and feature template matching (FTM). A novel blind signal detection algorithm is proposed, that uses principal component (PC) analysis. The experimental results show that the PC algorithm outperforms the MED and EME algorithms under all conditions and it performs better than the MME and CAV algorithms under certain conditions. The development of the PC algorithm is a significant contribution of this thesis. In addition to eigenvalue based detection (as in MME, EME and MED), leading eigenvectors can also be used to detect a signal. This concept is used in a recently proposed technique, called the FTM algorithm. However, FTM involves a feature learning stage that has several drawbacks for implementation in actual systems. This thesis proposes a multiple feature matching (MFM) algorithm, that finds similarity between multiple (current) features that are extracted from signals received at closely spaced antennas over several sensing instants. The experimental results show that without even relying on feature learning, the MFM algorithm still offers a comparable performance with the FTM algorithm. Modification of the FTM algorithm to form a more robust MFM algorithm is another contribution of this thesis. The final contribution of this thesis is to investigate the impact of some of the practical aspects of a multi-antenna signal detection system on the performance of CBD algorithms. In the presence of noise calibration error across the front-ends (which is inevitable in a practical system), the CAV algorithm is found to exhibit significant improvement in detection performance as the number of antennas is increased, whereas the MME, MED and EME algorithms show no improvement. The effect of physical antenna spacing on the received signal correlation and its subsequent impact on the detection performance is also investigated. Significant performance improvement is achieved for an antenna spacing of 0:1 c, however to avoid any adverse effect due to mutual coupling at such a close antenna spacing, proper impedance matching should also be ensured.