Enhanced Spectrum Sensing with Self-Organizing Maps and Deep Belief Networks: A Time-Efficient Approach

Abstract

In today’s wireless communication era, Cognitive radio systems (CRS) report two pressing problems: the shortage or scarcity of usable frequencies and second is the efficient use of those frequencies by the legal (or "licensed") users. Spectrum sensing (or “Spectrum detection”) is the key process in the CRS that allows these systems to be "cognitive." Demand of wireless communications is not narrowing and in fact, it is accelerating day by day, particularly in emerging economies. An increasing number of methods, both classical and modern, are being developed for the robust and timely spectrum sensing. Classic methods for spectrum sensing, like MLP, CNN, and LSTM, routinely encounter difficulties in a balancing between two main design targets as: accuracy and computational efficiency. The real-time nature of typical application scenarios makes this operational challenge particularly noticeable. In this paper's introduction and the review of related work we underline the same. Thus, a very strong approach should use to explore the new deep learning techniques which preferably have relatively low in required computational overhead at same time secure high detection accuracy. The application of Deep Belief Networks (DBNs) and Self-Organizing Maps (SOMs) for enhanced spectrum sensing in CRNs is the subject of our study. The critical task is to evaluate these two models for accuracy and speed in detecting spectrum availability. Insights into the trade-offs between detection performance and computational efficiency are especially vital concerning SOM and DBN models because they are fairly new as compared to traditional spectrum sensing methods. Challenging conditions for spectrum sensing were utilized to assess the performance of DBN and SOM models, and the results were compared to benchmarks established by traditional methods. The obtained results shows that SOM gets to an amazing accuracy of 95.77% with an execution time of only 0.02 seconds, makes it extremely suitable for real-time applications. Meanwhile, DBN achieves a supreme accuracy of 99.85%, but with a moderate execution time of 147.08 seconds, demonstrating its ability to extract features hierarchically and in a very deep way. Both models outperform traditional methods, and they strike a superior balance between accuracy and computational requirements with modern hardware. These gained results spotlight SOM and DBN based methods are next-generation CRN system competitors. SOM’s architecture required low processing and power, and DBN’s near-perfect detection capabilities. This meant that CRN systems using deep learning could handle the RF spectrum under even the worst conditions also. Their potential gains offered insights into several key trade-offs.