Transfer learning-based spectrum sensing over unknown scenarios

• Main Author: Zhang, Qishu
• Other Authors: Teh Kah Chan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Cognitive radio; OFDM; TDL channel; Spectrum sensing; Transfer learning; VGG-16
• Online Access: https://hdl.handle.net/10356/180691
• Institution: Nanyang Technological University

Efficient spectrum sensing is crucial for the optimal utilization of the increasingly crowded electromagnetic spectrum in cognitive radio networks (CRNs). Traditional spectrum sensing methods, including energy detection, matched filtering, and cyclostationary feature detection, often struggle in dynamic and unknown environments due to their dependence on predefined signal characteristics and extensive prior knowledge. To address these limitations, this dissertation explores the application of transfer learning-based deep learning techniques to enhance spectrum sensing performance in diverse and unpredictable scenarios. A simulation environment has been developed based on MATLAB. It emulates an orthogonal frequency-division multiplexing (OFDM) system combined with a tapped delay line (TDL) channel model variant. This setup represents multipath propagation, changing channel conditions, and real-world noise interference most accurately. It thus forms the strong basis for making and testing advanced spectrum sensing algorithms. It comes with the capability to easily generate big datasets using a simulation environment. These datasets are important for the training of deep learning models under various difficult conditions. Confronted with the various challenges brought about by spectrum sensing in unknown scenarios, we need to design and optimize a VGG-16 architecture suitable for this specific task. Major modifications include incorporating Batch Normalization layers, accelerating the training process and maintaining stability; reduction of fully connected layers to avoid overfitting; and incorporation of Dropout regularization to improve generalization. Besides, in this report, transfer learning methods were proposed to enable the convolutional neural network (CNN) model to adapt to new and unknown spectrum environments with very little extra training. The proposed architecture uses the learned features effectively by using pre-trained weights and fine-tuning the network for target domains. This will improve the detection accuracy and robustness for different spectral conditions. This way, it will significantly reduce the requirements for huge datasets in every new environment. It enables the quick deployment of spectrum sensing systems in diverse and dynamically changing scenarios.