COMMITTEE CHAIR: Dr. Cajetan Akujuobi

TITLE: DESIGN AND PERFORMANCE ANALYSIS OF ML-ASSISTED COOPERATIVE WIDEBAND ENERGY DETECTION FOR DYNAMIC SPECTRUM ACCESS IN 5G NETWORKS

ABSTRACT: Dynamic spectrum access is a key strategy for enhancing the utilization of the increasingly congested wideband radio spectrum, particularly in 5G networks where traffic is bursty, users are heterogeneous, and spectral opportunities are fleeting. Although energy detection remains attractive for spectrum sensing due to its simplicity and low cost, its performance drops markedly at low SNR, under noise uncertainty, and in fading channels. This dissertation addresses these limitations by investigating the design and performance analysis of ML-assisted cooperative wideband energy detection for dynamic spectrum access in 5G networks. It employs an analytical and data-driven evaluation framework to model detector behavior under diverse noise, interference, and fading conditions, and then augments the detector with learning-based adaptation. The proposed architecture targets operation in both complex AWGN and Rayleigh fading environments over multiple subbands and multiple cooperating cognitive radios. It integrates three key components: Maximal Ratio Combining (MRC) before energy detection to exploit spatial diversity across cognitive radios; adaptive, noise-aware thresholds designed per scene and per subband to maintain a target PFA = 0.01 despite noise and channel variability; and an ML-based k-of-N fusion rule at the decision center, where the fusion level is inferred from sensing-context features (such as per-CR SNR statistics, number of cooperating CRs, estimated noise power, subband index, and MRC-combined energy) rather than fixed a priori. To realize the ML-based k-of-N fusion rule, several supervised classifiers were trained and benchmarked, including Support Vector Classifiers (SVC), Random Forests (RF), Extremely Randomized Trees (ET), and Gradient Boosted Decision Trees (GBDT). Among these, GBDT emerged as the most effective model, delivering the highest k-prediction accuracy and best-calibrated decisions on the cooperative wideband datasets, and generalizing well to an external RadioML_2016-based evaluation. Using MATLAB-based simulations under complex AWGN and Rayleigh fading, the model consistently improves the probability of detection and AUC relative to static-threshold, fixed-k cooperative baselines while preserving the desired false-alarm constraint. This confirms that combining MRC-driven diversity, adaptive thresholding, and GBDT-based ML-k fusion yields a robust, scene-adaptive framework for low-SNR spectrum sensing and spectrum sharing in 5G and beyond.

Keywords: Dynamic Spectrum Access, Cooperative Wideband Spectrum Sensing, Energy Detection, Maximal Ratio Combining, Adaptive Thresholding, ML-Based k-of-N Fusion, Gradient-Boosted Decision Trees, Complex AWGN, Rayleigh Fading.

Room Location: Electrical and Computer Engineering Department Conference, Room 315D