Open Collections

Pan, Junyang

University of British Columbia

In the era of 6G wireless communications and cognitive radio (CR), the pivotal tasks of spectrum sensing and modulation classification are integral to optimizing spectrum utilization and enhancing communication system efficacy. Conventional methods frequently grapple with diminished accuracy in low signal-to-noise ratio (SNR) environments, as well as susceptibility to noise and interference. Notwithstanding their computational expediency, these methodologies prove inadequate in intricate and fluid communication settings. Deep learning (DL) approaches have emerged as a potential remedy for these deficiencies, as they possess the capability to autonomously extract features and adapt to evolving environments. Nonetheless, the majority of DL models entail substantial computational complexity and voluminous parameter sizes, hence rendering them less feasible for real-time or resource-constrained applications. The Kolmogorov-Arnold Networks (KANs) represent a novel architectural proposition, with limited practical application studies published to date. The authors emphasize the advantages of KANs over Multilayer Perceptrons (MLPs) in terms of both accuracy and interpretability, while also mandating significantly fewer parameters. In order to transcend the limitations of traditional methods and DL approaches, this thesis proceeds to employ KANs in two lightweight solutions: LWSS-KAN for spectrum sensing and SMC-KAN for modulation classification. LWSS-KAN conceptualizes spectrum sensing as a multi-output binary classification predicament, employing mean squared error (MSE) Loss function, while SMC-KAN delineates modulation classification as a multi-class classification undertaking, employing a cross-entropy loss function. By harnessing KAN’s adaptive activation functions, both LWSS-KAN and SMC-KAN attain elevated classification accuracy with diminished complexity, particularly in high-SNR conditions. The simulation results evince that both models deliver accuracy commensurate with traditional algorithms, while necessitating markedly fewer parameters. Furthermore, the research delves into the impact of hidden layers and input size in KAN. Additionally, LWSS-KAN manifests heightened interpretability, endowing superior insights into the decision-making process vis-à-vis conventional neural networks. These models proffer invaluable solutions for contemporary communication systems, especially in intricate environmental scenarios characterized by constrained computational resources.

Text

2024-11-07

Attribution-NonCommercial-ShareAlike 4.0 International

http://hdl.handle.net/2429/89626

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-sa/4.0/