Abstract:
In recent years, the spectrum demand for wireless communication services and applications
has been increasing drastically besides spectrum resource management and allocation have
become a hot issue. Cognitive Radio (CR) is designed and implemented to overcome this
existing problem by allocating a spectrum band to Primary and Secondary users
dynamically. One of the key features to decide over spectrum utilization for a CR is the
spectrum sensing (SS) unit which detects and identifies spectral data from the environment.
Conventional SS schemes such as Energy detection (ED), Cyclo-stationary and matched
filters were first developed and employed on CRs. Their drawbacks such as the inability to
exploit both spatial and temporal features of data, high false alarms and less detection
probability over noisy data lead to further studies to develop AI, particularly Machine
learning (ML) and Deep learning (DL) integrated models. This thesis work is mainly
focused on the performance of DL-based models to sense, predict and classify a spectral
dataset. Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and
Long-Short Term Memory (LSTM) models are adapted and their performance in spectrum
classification has been evaluated. One of the major contributions of this thesis work was to
adapt a hybrid Convolutional Recurrent neural network (CRNN) and to compare its
performance with the above existing Neural Network models. CNN is a good performing
model in extracting spatial features whereas RNN performs well in extracting temporal
features of spectral data. The performance of these DL models has been evaluated using
metrics such as classification accuracy, probability of detection (Pd), probability of false
alarm (Pfa), Sensing error (SE) and confusion matrix metric formulations. The signal samples
were generated with GNU for SNRs from -20 dB to 18 dB with step size of 2 dB over flat
fading channel and AWGN. This reliable synthetic dataset consists of 11 modulations with
varying SNR levels to train, validate and test our DL models. The simulation experiment
was carried out in Python Notebook and virtual Google- Colab environment. The results
show our proposed hybrid model outperforms the other DL models in terms of high
classification accuracy, high probability of detection and less SE. The LSTM model also
performed better than CNN and RNN models with its less probability of false alarm in
identifying a signal feature. Although all the DL models proved their better performances,
CNN was less accurate in identifying the signal feature particularly in low SNR ranges.